This book introduces the finite element method applied to the resolution of industrial heat transfer problems. Starting

*495*
*103*
*4MB*

*English*
*Pages 282*
*Year 2008*

- Author / Uploaded
- Jean-Michel Bergheau
- Roland Fortunier

- Categories
- Mathematics
- Computational Mathematics

*Table of contents : Finite Element Simulation of Heat Transfer......Page 6Table of Contents......Page 8Introduction......Page 14PART 1. Steady State Conduction......Page 201.1.1. Thermal equilibrium equation......Page 241.1.2. Fourier law......Page 251.1.3. Boundary conditions......Page 261.2. Mathematical analysis......Page 271.2.1. Weighted residual method......Page 281.2.2.Weak integral formulation......Page 301.3.1. Physical modeling......Page 331.3.2.1. Analytical integration......Page 351.3.2.2. The finite difference method......Page 361.3.3. Collocation methods......Page 371.3.3.1. Point collocation......Page 381.3.3.2. Sub-domain collocation......Page 391.3.4.1. Polynomial functions......Page 401.3.4.2. Piecewise linear functions......Page 422.1.1.Mesh......Page 462.1.2. Nodal approximation......Page 492.2.Discrete problem formulation......Page 512.2.1. Element quantities......Page 522.2.2. Assembly......Page 542.3.1. Application of temperature boundary conditions......Page 562.3.2. Linear system solution......Page 592.3.2.1. Direct methods......Page 612.3.2.2. Iterative methods......Page 632.3.3. Storing the linear system matrix......Page 652.3.4. Analysis of results......Page 662.3.4.1. Smoothing the heat flux density......Page 672.3.4.2. Result accuracy......Page 692.4. Working example......Page 712.4.1.1.Mesh......Page 732.4.1.2. Nodal approximation......Page 742.4.2.1. Element quantities......Page 752.4.2.2. Assembly......Page 772.4.3.1. Application of boundary conditions......Page 782.4.3.2. Solution......Page 803.1.1. Reference element......Page 823.1.1.1. Triangular element with linear transformation functions......Page 843.1.1.2. Quadrangle element with linear transformation functions......Page 853.1.1.3. Quadrangle element with quadratic transformation functions......Page 873.1.2. Isoparametric elements......Page 883.1.3. Interpolation function properties......Page 923.2. Calculation of element quantities......Page 933.2.1. Expression in the reference frame......Page 943.2.2. Gaussian quadrature......Page 963.2.2.1. 1D numerical integration......Page 973.2.2.2. 2D and 3D numerical integration......Page 1003.3. Some finite elements......Page 102PART 2. Transient State, Non-linearities, Transport Phenomena......Page 1044.1.1. The continuous problem......Page 1084.1.2. Finite element approximation......Page 1104.1.3. Linear case......Page 1124.2.1. Modal method......Page 1144.2.1.1. Determining the modal basis......Page 1154.2.1.2. Projection on the modal basis......Page 1174.2.2.Direct time integration......Page 1184.2.3. Accuracy and stability of a direct integration algorithm......Page 1224.2.3.1. Accuracy......Page 1234.2.3.2. Stability......Page 1244.2.3.3. Simplified analysis of the stability condition......Page 1254.2.4.1. Space oscillations during thermal shock simulation......Page 1274.2.4.2. Discrete maximum principle......Page 1314.2.4.3. Initial temperatures during thermal contact simulation......Page 1334.3.1. Physical modeling and approximation......Page 1384.3.2. Numerical applications......Page 1425.1.1. Formulation......Page 1465.1.2. Non-linear equation system solution methods......Page 1475.1.2.1. Newton-Raphson method......Page 1505.1.2.2. Substitution method......Page 1525.1.2.3. Quasi-Newton methods......Page 1535.1.3.Line search method......Page 1555.2.1. Physical properties......Page 1565.2.2. Flux or volumetric heat source boundary conditions......Page 1585.2.3. Modeling state changes......Page 1605.2.3.1. Equivalent specific heat method......Page 1615.2.3.2.Enthalpy solution method......Page 1635.3. A temperature-enthalpy formulation......Page 1655.3.1. Mathematical formulation......Page 1665.3.2. Example......Page 1696.1.1. Thermal balance......Page 1726.1.2.Treating a simple case......Page 1746.2. Resolution techniques......Page 1776.2.1. Upwind technique......Page 1786.2.2. SUPG method......Page 1806.2.3. 2Dand 3DPetrov-Galerkin formulation......Page 183PART 3. Coupled Phenomena......Page 1867.1. Modeling radiative heat exchanges in a cavity......Page 1927.1.1. Posing the problem......Page 1937.1.2.Calculation of view factors......Page 1977.1.3. Diffusion-radiation coupling......Page 2007.1.3.1. Tangent matrix......Page 2017.1.3.2. Substitution matrix......Page 2027.2.1. Radiation between two walls......Page 2037.2.2. Cylinder quenching......Page 2068.1.1. Physical model and mathematical formulation......Page 2108.1.2. Modeling the coupling......Page 2138.2.1. Physical and geometric modeling......Page 2158.2.2.Results......Page 2169.1.1.1.Avrami kinetics......Page 2189.1.2. Numerical integration......Page 2209.1.3. The case of several phase changes......Page 2239.1.4. Modeling the coupling......Page 2249.2. Examples......Page 2259.2.1. Phase transformation diagrams......Page 2269.2.2. Steel quenching......Page 23010.1. Finite element simulation of simultaneous diffusion and precipitation......Page 23410.1.1. Governing equations......Page 23510.1.2. Finite element formulation......Page 23710.2.1. Mathematical formulation......Page 23910.2.2. Numerical scheme......Page 24110.3.1. Calculation of a phase diagram......Page 24210.3.2.Carbon diffusion in a titanium steel......Page 24311.1.1.Weak formulation......Page 24611.1.2. Modeling the coupling......Page 24711.1.3. Solving the coupled problem......Page 24911.2. Resistance welding......Page 25111.2.1. Implementing the model......Page 25211.2.2.Results......Page 25412.1. Introduction......Page 25612.2. Magnetic vector potential formulation for magnetodynamics......Page 25712.3. Coupled finite element-boundary element method......Page 26012.3.1. Finite element formulation......Page 26212.3.2. Boundary element formulation......Page 26312.4. A harmonic balance method for the magnetodynamic problem......Page 26412.5.1. Iterative coupling......Page 26612.5.2. A direct method for magnetothermal coupling......Page 26812.6. Application: induction hardening of a steel cylinder......Page 269Bibliography......Page 272Index......Page 280*

Finite Element Simulation of Heat Transfer

Jean-Michel Bergheau Roland Fortunier

This page intentionally left blank

Finite Element Simulation of Heat Transfer

This page intentionally left blank

Finite Element Simulation of Heat Transfer

Jean-Michel Bergheau Roland Fortunier

First published in France in 2004 by Hermes Science/Lavoisier entitled “Simulation numérique des transferts thermiques par éléments finis” First published in Great Britain and the United States in 2008 by ISTE Ltd and John Wiley & Sons, Inc. Translated from the French by Robert Meillier Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licenses issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned address: ISTE Ltd 27-37 St George’s Road London SW19 4EU UK

John Wiley & Sons, Inc. 111 River Street Hoboken, NJ 07030 USA

www.iste.co.uk

www.wiley.com

© ISTE Ltd, 2008 © LAVOISIER, 2004 The rights of Jean-Michel Bergheau and Roland Fortunier to be identified as the authors of this work have been asserted by them in accordance with the Copyright, Designs and Patents Act 1988. Library of Congress Cataloging-in-Publication Data Bergheau, Jean-Michel. [Simulation numérique des transferts thermiques par éléments finis. English] Finite element simulation of heat transfer / Jean-Michel Bergheau, Roland Fortunier. p. cm. Includes bibliographical references and index. ISBN 978-1-84821-053-0 1. Heat--Transmission--Mathematical models. 2. Finite element method. I. Fortunier, Roland. II. Title. TJ260.B45413 2008 621.402'2015118--dc22 2008025105 British Library Cataloguing-in-Publication Data A CIP record for this book is available from the British Library ISBN: 978-1-84821-053-0 Printed and bound in Great Britain by CPI/Antony Rowe Ltd, Chippenham, Wiltshire.

Table of Contents

Introduction

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

PART 1. Steady State Conduction . . . . . . . . . . . . . . . . . . . . . . . . .

17

Chapter 1. Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . .

21

1.1. Physical modeling . . . . . . . . . . . . 1.1.1. Thermal equilibrium equation . . . 1.1.2. Fourier law . . . . . . . . . . . . . 1.1.3. Boundary conditions . . . . . . . . 1.2. Mathematical analysis . . . . . . . . . . 1.2.1. Weighted residual method . . . . . 1.2.2. Weak integral formulation . . . . . 1.3. Working example . . . . . . . . . . . . . 1.3.1. Physical modeling . . . . . . . . . 1.3.2. Direct methods . . . . . . . . . . . 1.3.2.1. Analytical integration . . . . 1.3.2.2. The ﬁnite difference method 1.3.3. Collocation methods . . . . . . . . 1.3.3.1. Point collocation . . . . . . . 1.3.3.2. Sub-domain collocation . . . 1.3.4. Galerkin method . . . . . . . . . . 1.3.4.1. Polynomial functions . . . . 1.3.4.2. Piecewise linear functions . .

. . . . . . . . . . . . . . . . . .

21 21 22 23 24 25 27 30 30 32 32 33 34 35 36 37 37 39

Chapter 2. The Finite Element Method . . . . . . . . . . . . . . . . . . . . .

43

2.1. Finite element approximation 2.1.1. Mesh . . . . . . . . . . . 2.1.2. Nodal approximation . . 2.2. Discrete problem formulation

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . . . . . . . . . . . . . . . .

. . . .

. . . .

43 43 46 48

6

Finite Element Simulation of Heat Transfer

2.2.1. Element quantities . . . . . . . . . . . . . . . . . 2.2.2. Assembly . . . . . . . . . . . . . . . . . . . . . . 2.3. Solution . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1. Application of temperature boundary conditions 2.3.2. Linear system solution . . . . . . . . . . . . . . . 2.3.2.1. Direct methods . . . . . . . . . . . . . . . . 2.3.2.2. Iterative methods . . . . . . . . . . . . . . . 2.3.3. Storing the linear system matrix . . . . . . . . . 2.3.4. Analysis of results . . . . . . . . . . . . . . . . . 2.3.4.1. Smoothing the heat ﬂux density . . . . . . 2.3.4.2. Result accuracy . . . . . . . . . . . . . . . . 2.4. Working example . . . . . . . . . . . . . . . . . . . . . 2.4.1. Finite element approximation . . . . . . . . . . . 2.4.1.1. Mesh . . . . . . . . . . . . . . . . . . . . . 2.4.1.2. Nodal approximation . . . . . . . . . . . . 2.4.2. Discrete problem formulation . . . . . . . . . . . 2.4.2.1. Element quantities . . . . . . . . . . . . . . 2.4.2.2. Assembly . . . . . . . . . . . . . . . . . . . 2.4.3. Solution . . . . . . . . . . . . . . . . . . . . . . . 2.4.3.1. Application of boundary conditions . . . . 2.4.3.2. Solution . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

49 51 53 53 56 58 60 62 63 64 66 68 70 70 71 72 72 74 75 75 77

Chapter 3. Isoparametric Finite Elements . . . . . . . . . . . . . . . . . . . .

79

3.1. Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1. Reference element . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1.1. Triangular element with linear transformation functions . . . 3.1.1.2. Quadrangle element with linear transformation functions . . 3.1.1.3. Quadrangle element with quadratic transformation functions 3.1.2. Isoparametric elements . . . . . . . . . . . . . . . . . . . . . . . . 3.1.3. Interpolation function properties . . . . . . . . . . . . . . . . . . . 3.2. Calculation of element quantities . . . . . . . . . . . . . . . . . . . . . . 3.2.1. Expression in the reference frame . . . . . . . . . . . . . . . . . . 3.2.2. Gaussian quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2.1. 1D numerical integration . . . . . . . . . . . . . . . . . . . . 3.2.2.2. 2D and 3D numerical integration . . . . . . . . . . . . . . . . 3.3. Some ﬁnite elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

79 79 81 82 84 85 89 90 91 93 94 97 99

PART 2. Transient State, Non-linearities, Transport Phenomena . . . . . . 101 Chapter 4. Transient Heat Conduction 4.1. Problem formulation . . . . . . . . 4.1.1. The continuous problem . . . 4.1.2. Finite element approximation 4.1.3. Linear case . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . 105 . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

105 105 107 109

Table of Contents

4.2. Time integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1. Modal method . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1.1. Determining the modal basis . . . . . . . . . . . . . . . 4.2.1.2. Projection on the modal basis . . . . . . . . . . . . . . . 4.2.2. Direct time integration . . . . . . . . . . . . . . . . . . . . . . 4.2.3. Accuracy and stability of a direct integration algorithm . . . 4.2.3.1. Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.3.2. Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.3.3. Simpliﬁed analysis of the stability condition . . . . . . 4.2.4. Practical complementary rules . . . . . . . . . . . . . . . . . 4.2.4.1. Space oscillations during thermal shock simulation . . 4.2.4.2. Discrete maximum principle . . . . . . . . . . . . . . . 4.2.4.3. Initial temperatures during thermal contact simulation . 4.3. Working example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1. Physical modeling and approximation . . . . . . . . . . . . . 4.3.2. Numerical applications . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

7

111 111 112 114 115 119 120 121 122 124 124 128 130 135 135 139

Chapter 5. Non-linearities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 5.1. Formulation and solution techniques . . . . . . . . . . . . 5.1.1. Formulation . . . . . . . . . . . . . . . . . . . . . . . 5.1.2. Non-linear equation system solution methods . . . . 5.1.2.1. Newton-Raphson method . . . . . . . . . . . . 5.1.2.2. Substitution method . . . . . . . . . . . . . . . 5.1.2.3. Quasi-Newton methods . . . . . . . . . . . . . 5.1.3. Line search method . . . . . . . . . . . . . . . . . . 5.2. Traditional non-linearities . . . . . . . . . . . . . . . . . . 5.2.1. Physical properties . . . . . . . . . . . . . . . . . . . 5.2.2. Flux or volumetric heat source boundary conditions 5.2.3. Modeling state changes . . . . . . . . . . . . . . . . 5.2.3.1. Equivalent speciﬁc heat method . . . . . . . . 5.2.3.2. Enthalpy solution method . . . . . . . . . . . . 5.3. A temperature-enthalpy formulation . . . . . . . . . . . . 5.3.1. Mathematical formulation . . . . . . . . . . . . . . . 5.3.2. Example . . . . . . . . . . . . . . . . . . . . . . . . . Chapter 6. Transport Phenomena

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

143 143 144 147 149 150 152 153 153 155 157 158 160 162 163 166

. . . . . . . . . . . . . . . . . . . . . . . . 169

6.1. Highlighting instabilities . . . . . . . . . . . . 6.1.1. Thermal balance . . . . . . . . . . . . . 6.1.2. Treating a simple case . . . . . . . . . . 6.2. Resolution techniques . . . . . . . . . . . . . 6.2.1. Upwind technique . . . . . . . . . . . . 6.2.2. SUPG method . . . . . . . . . . . . . . 6.2.3. 2D and 3D Petrov-Galerkin formulation

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

169 169 171 174 175 177 180

8

Finite Element Simulation of Heat Transfer

PART 3. Coupled Phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 Chapter 7. Radiation Exchanges in a Chamber

. . . . . . . . . . . . . . . . 189

7.1. Modeling radiative heat exchanges in a cavity . 7.1.1. Posing the problem . . . . . . . . . . . . . 7.1.2. Calculation of view factors . . . . . . . . 7.1.3. Diffusion-radiation coupling . . . . . . . 7.1.3.1. Tangent matrix . . . . . . . . . . . . 7.1.3.2. Substitution matrix . . . . . . . . . . 7.2. Examples . . . . . . . . . . . . . . . . . . . . . 7.2.1. Radiation between two walls . . . . . . . 7.2.2. Cylinder quenching . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

189 190 194 197 198 199 200 200 203

Chapter 8. Fluid-Structure Coupling in a Pipe . . . . . . . . . . . . . . . . . 207 8.1. Modeling the ﬂuid . . . . . . . . . . . . . . . . . . . 8.1.1. Physical model and mathematical formulation 8.1.2. Modeling the coupling . . . . . . . . . . . . . . 8.2. Example . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1. Physical and geometric modeling . . . . . . . . 8.2.2. Results . . . . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

207 207 210 212 212 213

Chapter 9. Thermometallurgical Coupling . . . . . . . . . . . . . . . . . . . 215 9.1. Modeling phase changes . . . . . . . . . 9.1.1. Rate of phase changes . . . . . . . 9.1.1.1. Avrami kinetics . . . . . . . . 9.1.1.2. Martensitic kinetics . . . . . 9.1.2. Numerical integration . . . . . . . 9.1.3. The case of several phase changes 9.1.4. Modeling the coupling . . . . . . . 9.2. Examples . . . . . . . . . . . . . . . . . 9.2.1. Phase transformation diagrams . . 9.2.2. Steel quenching . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

215 215 215 217 217 220 221 222 223 227

Chapter 10. Thermochemical Coupling . . . . . . . . . . . . . . . . . . . . . 231 10.1. Finite element simulation of simultaneous diffusion and precipitation 10.1.1. Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . 10.1.2. Finite element formulation . . . . . . . . . . . . . . . . . . . . . . 10.2. Calculation of precipitation . . . . . . . . . . . . . . . . . . . . . . . . 10.2.1. Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . 10.2.2. Numerical scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.1. Calculation of a phase diagram . . . . . . . . . . . . . . . . . . . 10.3.2. Carbon diffusion in a titanium steel . . . . . . . . . . . . . . . . .

231 232 234 236 236 238 239 239 240

Table of Contents

9

Chapter 11. Electrothermal Coupling . . . . . . . . . . . . . . . . . . . . . . 243 11.1. Electrokinetic modeling . . . . . . 11.1.1. Weak formulation . . . . . . . 11.1.2. Modeling the coupling . . . . 11.1.3. Solving the coupled problem 11.2. Resistance welding . . . . . . . . . 11.2.1. Implementing the model . . . 11.2.2. Results . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

243 243 244 246 248 249 251

Chapter 12. Magnetothermal Coupling . . . . . . . . . . . . . . . . . . . . . 253 12.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2. Magnetic vector potential formulation for magnetodynamics . 12.3. Coupled ﬁnite element-boundary element method . . . . . . . . 12.3.1. Finite element formulation . . . . . . . . . . . . . . . . . . 12.3.2. Boundary element formulation . . . . . . . . . . . . . . . 12.3.3. FEM-BEM coupling . . . . . . . . . . . . . . . . . . . . . 12.4. A harmonic balance method for the magnetodynamic problem 12.5. Coupling magnetodynamics with heat transfer . . . . . . . . . . 12.5.1. Iterative coupling . . . . . . . . . . . . . . . . . . . . . . . 12.5.2. A direct method for magnetothermal coupling . . . . . . . 12.6. Application: induction hardening of a steel cylinder . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

253 254 257 259 260 261 261 263 263 265 266

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

This page intentionally left blank

Introduction

In their daily practice, professional engineers are often confronted with problems involving complex physical phenomena. Whether they have to analyze the behavior of a product or optimize an implementation process, modeling these phenomena often makes it possible to solve these problems: – by gaining access to the physical magnitudes characterizing the product and the consequences of their modiﬁcation, – by highlighting the main parameters of the process and thus its optimal conditions of use. Physical modeling seems to be the key point when an engineer is thinking of how to solve a scientiﬁc problem. Generally speaking, modeling follows these three steps: 1) identifying the essential physical phenomena related to the behavior of a product or process within a given context, 2) deﬁning the mathematical equations (domain of study, partial differential equation(s), boundary conditions and stresses, initial conditions, etc.) governing these phenomena, 3) validating the equations with respect to the necessary data which must be available or accessible by means of measurements as well as the results provided, which must be reliable and reproducible. The ﬁrst two points need commenting upon. Firstly, physical modeling has to meet a set objective. It is then obvious that several modeling methods are possible to meet a similar objective. The most relevant model is the simplest one making it possible to meet the set objective. Modeling is ﬁrst and foremost making (simplifying) assumptions. Some are obvious. For instance, it is unnecessary to take into account heat radiation phenomena in low temperature applications; a mean

12

Finite Element Simulation of Heat Transfer

exchange coefﬁcient, independent of the temperature, will be sufﬁcient. In other cases, it is far more difﬁcult. For example, how can we deﬁne the analyzed ﬁeld and include the behavior of the parts not addressed by means of carefully selected boundary conditions? This is exactly where the professional engineer’s contribution comes in. We will come back to this later on. The last point of this deﬁnition is about the model input data and results. The selection of input data (availability, validity, etc.) often guides the users in the choice of their modeling methods. It is unnecessary, for instance, to use very sophisticated metallurgical transformation models if sufﬁciently accurate data on the processed material are not available. Result reliability and reproducibility guide the users in the choice and use of the model solution method. The aim of this method (numerical or analytical) is to determine a solution of the problem with minimum approximation. It is obvious that an accurate analytical solution will always be given priority; however, the growing complexity of the model often leads today’s users towards approximate numerical solutions. Moreover, commercial numerical simulation software often includes the functions and models required for industrial applications, thus making it possible to take full advantage of the progress of information technology (rapid calculation, memory resources, graphic visual aids, etc.). Consider for instance the boundary value problem related to steady state heat exchanges in a solid occupying a limited ﬁeld (equation [1.8]), which reveals: – a volumetric heat source Q most often resulting from other physical phenomena such as the Joule effect in conduction heating or induction heating applications. Is it necessary to model these phenomena by adding the corresponding partial differential equations to the heat equation? This is not certain but then we require simple analytical models to evaluate Q; – a surface density of the heat ﬂux q applied to a portion of the solid boundary, reﬂecting the heat exchange between this solid and the outside medium. This can be for instance heat radiation towards an inﬁnite medium or a liquid ﬂow into a quenching bath. Here again, is it necessary to include those phenomena in the modeling process with the addition of complementary equations? It depends on the application considered and the feasibility of such an approach. However, if, whatever the reason, ﬁne modeling of these phenomena was not envisaged, other approaches would then be necessary to determine the boundary conditions. These could be, for instance, simpliﬁed analytical models; – a temperature prescribed on the other portion of the solid boundary. This condition is a limit case of the previous condition. As a matter of fact, either the solid is steeped at this spot in a ﬂuid with a very high exchange coefﬁcient so as to prescribe the ﬂuid temperature on the corresponding boundary, or the user has a measurement.

Introduction

13

Experimental measurements will frequently be used to determine the missing input data, but these will be obtained by alternative routes as the data required for modeling are often inaccessible to direct measurements. The methods called inverse methods have developed signiﬁcantly over the last few years. They are able to couple physical modeling with accessible magnitude experimental measurements by adjusting iteratively an input data until the whole set of calculated results are as close as possible to the measurements. It is tempting to say after these comments that the easiest model to elaborate is one in which all inﬂuential physical phenomena are ﬁnely modeled. This is partly true; however: – Are we really in a position to deﬁne the mathematical equations governing these phenomena with a sufﬁcient degree of accuracy to make it worthwhile? – Are we certain that we are able to have the data required to feed these models all the more so that these models will frequently require unusual data on a different scale, which is thus hardly accessible? Using inverse methods may solve this type of difﬁculty. – Will it not be often more difﬁcult to interpret the results than with a simpler model given the amount of information to process? Indeed, will it not be necessary to conduct a posteriori the analysis which will not have been carried out a priori to eliminate insigniﬁcant results? In fact, everything depends on the problem to solve. Moreover at which point of the domain of study considered should we stop? First, we will take into consideration all the symmetries presented by the problem for which the corresponding boundary conditions are written clearly from a mathematical point of view: – revolution symmetry: the most productive one as it allows us to carry out 3D analysis on a plane model representing a meridian section of the studied structure, – symmetry to a plane: very commonly used, – anti-symmetry to a plane: less natural and often forgotten, – periodicity conditions on a repetitive structure. The relevance of a 3D model will then be questioned. If it is true that calculation software and computers make it easy to carry out this type of analysis, a 2D (and even 1D) model is often sufﬁcient (in the case of a repetitive structure in one direction for example) and therefore preferable as the analysis and interpretation of results will always be easier with a 2D than with a 3D model. In all other cases, restricting the domain of study will lead to the deﬁnition of appropriate boundary conditions.

14

Finite Element Simulation of Heat Transfer

It is therefore in the interest of the engineers in charge of physical modeling to think in detail about the relevance of their choices and assumptions. It is the price to pay to take maximum advantage of their models. In any case, before reaching the intensive exploitation stage, it is imperative to ensure the model prediction quality by comparing it with the results of one experiment (at least!). Most often, analytical methods cannot be applied to the solution of the mathematical equations governing a set of physical phenomena, except if major assumptions reducing the modeling validity are made. The analog method takes advantage of the fact that conduction heat exchanges and electrical conduction phenomena are governed by the same equations. It is thus possible to study conduction heat exchanges by means of more easily accessible measurements carried out on a similar electrical device. However, the ﬁeld of application of this method is far too restrictive. Therefore, numerical techniques whose use is made possible by the performance of today’s computers are used to determine an approximate solution of the set of mathematical equations governing the problem. The ﬁnite difference method which replaces partial derivatives with ﬁnite differences at different points of a grid is highly regarded by mechanical engineers. The discrete equation system obtained can also be interpreted as resulting from a complex electrical diagram combining resistances and capacities. This technique is hardly possible with complex geometries. The ﬁnite volume method is also highly regarded by mechanical and thermal engineers. It is based upon a previous division of the geometric domain of study into element volumes. On each element volume, the thermal balance equations are solved. This method is particularly efﬁcient in the case of structured geometries. It is widely used in thermal science and ﬂuid mechanics, but more rarely in other disciplines. Among the various numerical techniques available today, the ﬁnite element method [TOU 81, ZIE 91] is the most widespread owing to: – its general ﬁelds of application (thermal, electromagnetic sciences, solid mechanics, ﬂuid mechanics, etc.), – its capacity to treat problems with complex geometries, – its easy implementation. This can be done as follows: 1) mesh-geometry: the geometric domain to be analyzed, most often resulting from CAD geometric modeling, is divided into a set of element sub-domains (ﬁnite elements) interconnected by nodes;

Introduction

15

2) solution: the continuous functions sought (the temperature in a heat conduction problem) are replaced with a set of values estimated at the mesh nodes. This approximation, applied to an integral formulation of the problem, leads to a system of equations (linear or non-linear) whose number is equal to the number of values to be estimated; 3) analysis-interpretation of results: at this point, other results can be calculated (for instance, the heat ﬂux density) from those obtained by the direct solution of the equation system. They are then analyzed and interpreted by the user, by means of very efﬁcient graphic visualization resources. The success of the ﬁnite element method is largely due to the signiﬁcant progress of information technologies, both from a numerical point of view (rapid calculation, memory size available) and from a graphic point of view (3D visual resources). Today a large amount of software makes use of this method. They offer a growing number of functions and are increasingly user-friendly. It should be noted that calculation software displays clear and user-friendly interfaces nowadays and can be used by non-specialist engineers. However, this apparent facility should not conceal the fact that, whatever the numerical method adopted, the discretization phase impairs the properties of the initial continuous model. Some phenomena present in the continuous model could be erased by the numerical model if we do not pay attention. The meshing phase in the ﬁnite element method is therefore very important and it is the user’s know-how that will produce a quality approximate solution. Therefore, the aim of this book is to present the basis and application of the ﬁnite element method to the solution of industrial thermal problems. It consists of three parts which the reader may possibly complete by reading a number of books related to this ﬁeld [COM 94, RED 94, LEW 96, MIN 06]. Part 1, dedicated to the solution of steady state heat conduction problems, introduces the ﬁnite element method. Starting with the partial derivative problem and the related boundary conditions, various formulations upon which various discretization methods are based, are presented. Part 2 extends the ﬁeld of application of the method to transient state conduction problems, the most common non-linearities and transport phenomena (diffusion convection problems). The last part, Part 3, deals with coupled problems: – coupled by boundary conditions: radiation problems, ﬂuid and structure coupling in a piping system,

16

Finite Element Simulation of Heat Transfer

– including additional state variables: thermometallurgical coupling, – coupled by partial differential equations: electrothermal magnetothermal coupling and thermochemical coupling.

coupling,

This book is a survey of the various thermal problems which professional engineers may have to simulate. The methods presented will allow readers to use in the best way possible a calculation software and design new calculation modules so as to complete their work.

PART 1

Steady State Conduction

This page intentionally left blank

Introduction

In this part we introduce the ﬁnite element method in the simplest ﬁeld of thermal science: steady state conduction. We therefore consider a solid Ω undergoing thermal → loads and try to determine the temperature ﬁeld T (− x ) in this solid, when it is in equilibrium with the external environment. In this part, it is assumed that no quantity involved depends on time. Thermal loads are therefore constant and the solid is motionless. The equilibrium state sought corresponds to a steady state. Besides, we will restrict our study to the linear case in which the mathematical ﬁeld is clearly identiﬁed. The ﬁnite element method is introduced according to the following three steps. Firstly, the physical problem to solve is analyzed from a mathematical point of view. This leads to a series of three problems (equations [1.8], [1.10] and [1.13]). Problem [1.8] is a partial differential equation resulting directly from physical modeling with its boundary conditions. Problem [1.10] turns this partial differential equation into a variational equation by means of the weighted residual method. Finally, problem [1.13], often called weak form, is the basis of the ﬁnite element method. It is possible to ﬁnd an approximate solution to each of the three problems [1.8], [1.10] and [1.13]. For this purpose, several methods are used. Each of them will be illustrated by a work example: an induction-heated plate. We will use a physical model to describe this example, then seek approximate solutions (the temperature ﬁeld) of problems [1.8], [1.10] and [1.13] related to this physical model. Chapter 2 deals with the ﬁnite element method. This method is based upon the weak form described in Chapter 1. The approximation used is called ﬁnite element approximation. It consists of discretizing the geometry (the mesh) and approximating the temperatures sought (nodal approximation). This step makes it possible to write

20

Finite Element Simulation of Heat Transfer

the problem to solve in a discretized form, which is very appropriate for a computer numerical solution. In this chapter, the various steps of the ﬁnite element method will be illustrated by a new work example: thermal conduction in a plate with holes. In Chapter 3, we introduce isoparametric ﬁnite elements starting from the notion of the reference element. This makes it possible to develop simple and rapid methods for the calculation of the element quantities involved in the discrete problem formulation. The major types of isoparametric elements are also described in detail in this chapter.

Chapter 1

Problem Formulation

1.1. Physical modeling In this chapter we describe the different steps of physical modeling leading in the next section to a boundary value problem. These steps are: 1) writing the equation expressing the solid thermal equilibrium, 2) introducing the Fourier law connecting the heat ﬂux to the temperature gradient, 3) formulating boundary conditions. 1.1.1. Thermal equilibrium equation Figure 1.1 illustrates a homogenous solid Ω. In order to write that this solid is in thermal equilibrium, consider any portion ΩA related to this solid and write that the heat produced on that portion is equal to the heat ﬂux coming out of it, i.e.: → − − ∀ΩA , Q dv = φ ·→ n ds [1.1] ΩA

∂ΩA

→ − In this equation, φ is a vector characterizing the heat ﬂux surface density (in → n is the outward unit normal W/m2 ) coming out of ΩA through its boundary ∂ΩA , − to this surface, and Q is a scalar representing an internal heat volumetric source (in W/m3 ) in ΩA . Among the physical phenomena represented by this volumetric term, we can include Joule effect heating (conduction or induction), heat dissipation by plastic deformation, etc.

22

Finite Element Simulation of Heat Transfer

n

Ω ΩA φ

Figure 1.1. Solid thermal equilibrium

If the divergence theorem is now applied (integration by parts) to the surface integral of equation [1.1], i.e.: − − → − → div φ dv φ ·→ n ds = ΩA

∂ΩA

it is possible to write the thermal equilibrium of any portion ΩA of the solid, Ω, in Figure 1.1 as follows: − → ∀ΩA , Q − div φ dv = 0 ΩA

If this equation is veriﬁed in any domain ΩA of the solid Ω we obtain the partial differential equation governing the thermal equilibrium: − → Q − div φ = 0 at any point of Ω

[1.2]

1.1.2. Fourier law The law connecting the heat ﬂux density to the temperature ﬁeld must satisfy the second principle of thermodynamics. In this book, we will not deal with the details of continuum thermodynamics. For that purpose, the reader should consult books related to this discipline. Here we will deal with the simplest law, the Fourier law, connecting linearly the heat ﬂux density to the local temperature gradient. This is

Problem Formulation

23

→ − −−→ written φ = −λgrad(T ), where λ is a positive scalar called thermal conductivity. −−→ A temperature variation according to a space direction (i.e. a component of grad(T )) will generate a heat ﬂux in the same direction but inverted (i.e. from hot to cold). The thermal conductivity of a material may be a function of the local temperature and the space direction in which thermal transfer occurs. In this book we will therefore use a law which is more general than the one described above: − → −−→ φ = −λ.grad(T )

[1.3]

In this equation, λ is a symmetric second-order tensor in which each component λij , possibly as a function of temperature, produces the inﬂuence of a thermal gradient in the direction j on the heat ﬂux density in the direction i. These components make up the thermal conductivity matrix. In the case of isotropic materials (similar thermal conductivity in all space directions), this matrix comes down to λ = λI, where I is the identity matrix, and we go back to our initial expression. If we now insert Fourier law [1.3] into equilibrium equation [1.2], this becomes: −−→ div λ · grad(T ) + Q = 0 at any point of Ω

[1.4]

The equation obtained is a partial differential equation in T . Indeed, the → → temperature ﬁeld sought is a function T (− x ), where − x characterizes the position of a point in the solid Ω. For example, in Cartesian coordinates (x, y, z), and in the case of an isotropic material whose thermal conductivity λ is independent of temperature, this equation can be written as follows: 2 ∂ T ∂2T ∂2T λ + Q = 0 at any point of Ω + + ∂x2 ∂y 2 ∂z 2 1.1.3. Boundary conditions In order to solve partial differential equation [1.4] expressing the thermal equilibrium of the solid Ω, it is necessary to add boundary conditions. Those conditions express the connection of Ω with the external environment. This connection is generally represented on Ω boundary either by a known ﬂux density → − − φ ·→ n , or by a known temperature T . A known ﬂux density condition q on the portion ∂Ωq of Ω boundary is generally written as follows: → − − → −−→ φ ·→ n = − λ · grad(T ) · − n = −q(T )

at any point of ∂Ωq

[1.5]

24

Finite Element Simulation of Heat Transfer

→ where − n is the outward unit normal at the point considered and q is the ﬂux density entering Ω at this point, this density possibly being a function of the temperature T at the same point. This type of condition is used to represent, for instance: – convection exchanges of the form q(T ) = h(Text − T ), where h is a local heat exchange coefﬁcient (in W/m2 /K) and Text is the external environment temperature (Fourier condition), 4 – radiation exchanges in an inﬁnite environment q(T ) = σ(T∞ − T 4 ) where σ is −8 2 4 Stefan constant (σ = 5, 67·10 W/m /K ), is the emissivity of the surface assumed to be “gray”, T is the Kelvin temperature at the point considered, and T∞ is the Kelvin temperature of the inﬁnite environment.

When the surface ﬂux density q is not a function of the temperature, this boundary condition is called the Neuman condition. It is used for instance to represent: – an adiabadicity condition (q = 0), – a symmetry condition (q = 0 in the symmetry plane). A temperature condition Td known on a portion ∂ΩT of Ω boundary is written as follows: T = Td at any point of ∂ΩT [1.6] This type of condition, known as the Dirichlet condition, is in fact a convection exchange limit case, with a very high heat exchange coefﬁcient h and Td equal to the temperature of the external environment. It is generally used when local experimental temperature measurements are available. It can also be used to represent an anti-symmetry condition, but this is not very frequent in practice. 1.2. Mathematical analysis For a thermal problem to accept a unique solution in a steady state (problem correctly posed), it is necessary: – to prescribe at any point of the boundary ∂Ω of Ω either a known ﬂux density condition (equation [1.5]), or a known temperature condition (equation [1.6]), – to prescribe at one point at least of ∂Ω either a known temperature condition or a Fourier condition. The boundary ∂Ω can then be broken down into a portion ∂Ωq on which a ﬂux condition is applied, and a portion ∂ΩT on which the temperature is known. These two portions must satisfy the following conditions: ∂ΩT ∩ ∂Ωq = ∅ and ∂ΩT ∪ ∂Ωq = ∂Ω

[1.7]

Problem Formulation

25

The boundary value problem expressing the equilibrium of the solid Ω can then be written as: → → Find T (− x ) at any point − x ∈ Ω such that: −−→ R(T ) = div λ · grad(T ) + Q = 0 in Ω [1.8] − −−→ → λ · grad(T ) · n = q on ∂Ωq T = Td on ∂ΩT The term R(T ) involved in this problem is said to be residual. It is a quantity to cancel in the solid Ω while satisfying the boundary conditions representing the connections of the solid with the external environment. It is possible to work directly on problem [1.8] to determine the temperature → ﬁeld for T (− x ) satisfying the thermal equilibrium of the solid. This can be done in a few simple cases in an analytical way, but in most cases, only an approximate solution can be determined. The ﬁnite difference method is the ﬁrst possibility to obtain an approximation of problem [1.8]. In this case, the ﬁrst- and second-order partial derivatives appearing are evaluated from the ﬁrst terms of a Taylor sequence development of the temperature ﬁeld. 1.2.1. Weighted residual method → In problem [1.8] we have to cancel a scalar residual R(T ) at any point − x of the → − solid Ω. First, note that if a temperature ﬁeld T ( x ) cancels this residual, then for any → function ψ(− x ): ψR(T )dv = 0 [1.9] Ω

→ The functions ψ(− x ) introduced by this method are called weighting functions. The only condition required for these functions is to give a sense to equation [1.9]. These functions will then have to be integrable on Ω. The set of weighting functions → ψ(− x ) will be written Eψ . → With regard to the temperature ﬁeld T (− x ) sought, it also has to give a sense to equation [1.9]. Due to the expression of the residual R(T ) these functions, called admissible, are to have second-order partial derivatives integrable over Ω. The set of → admissible functions T (− x ) will be written ET .

26

Finite Element Simulation of Heat Transfer

The weighted residual method therefore produces the following problem formulation: Find T ∈ ET such that for any ψ ∈ Eψ : −−→ ψ div λ · grad(T ) + Q dv = 0 Ω [1.10] −−→ λ · grad(T ) · − → n = q on ∂Ωq T = Td on ∂ΩT It is obvious that, if a temperature ﬁeld is the solution of [1.8], then it is also the solution of [1.10], whatever Eψ is. On the contrary, a temperature ﬁeld which is the solution of problem [1.10] depends on the choice of the sets of functions ET and Eψ . Problem [1.10] is ideal for seeking an approximation of the temperature ﬁeld → T (− x ). For this purpose, it is sufﬁcient to choose particular spaces Eψ and ET . Partial differential equation [1.4] is then no longer satisﬁed at any point of Ω; it is only satisﬁed in the sense of equation [1.10]. Point collocation and sub-domain collocation methods are based upon formulation [1.10]. Point collocation consists of deﬁning a ﬁnite number n of points in the solid Ω with → → coordinates − x 1, . . . , − x n , and choosing for Eψ the set of Dirac distributions related to them: − Eψ = δ→ x i , i = 1, . . . , n → The resolution of problem [1.10] then comes down to seeking a function T (− x) which can be differentiated twice, satisfying the boundary conditions, and canceling → the residual at each point − x i of Ω: − → − δ→ = 0 for i = 1, . . . , n x i R(T )dv = R(T ) → x =− x Ω

i

→ For example, if T (− x ) is chosen in a polynomial form, the resolution of these equations allows us to determine the polynomial coefﬁcients. The quality of the solution obtained with this method depends directly on the form selected for the temperature ﬁeld (set ET ) and the choice of collocation points (number and positions).

Problem Formulation

27

Sub-domain collocation is the basis of the ﬁnite volume method. It consists of dividing the solid Ω into element volumes Ωi with no intersection, and in such a way that their union forms Ω. Then, the following weighting functions are selected:

→ 1 if − x ∈ Ωi → − Eψ = ψi /ψi ( x ) = 0 otherwise → The resolution of problem [1.10] then comes down to seeking a solution T (− x) satisfying the thermal equilibrium on each element volume or control volume Ωi . As a matter of fact, using this type of weighting function makes it possible to write on each control volume: ψi R(T )dv = R(T )dv = 0 Ω

Ωi

With the divergence theorem and Fourier law [1.3], this equation can be expressed as follows: → − − ∀i, Q dv = φ ·→ n ds Ωi

∂Ωi

Then we go back to equation [1.1], which must be satisﬁed for a particular set of sub-domains Ωi . This equilibrium equation is the basis of the ﬁnite volume method. The temperature is then often assumed to be homogenous in an element volume. The heat ﬂuxes coming out of the volume are then calculated, in each direction, from local laws using the temperatures of adjacent volumes. 1.2.2. Weak integral formulation The weak integral formulation of problem [1.10] is introduced when the volume integral equation involved has a speciﬁc form. If the divergence theorem (integration by parts) is applied to this integral equation, we can write: −−→ ψ div λ · grad(T ) dv Ω

=

−−→ div ψ λ · grad(T ) dv −

Ω

=

∂Ω

−−→ → ψ λ · grad(T ) · − n ds −

−−→ −−→T grad (ψ) · λ · grad(T )dv

Ω

−−→ −−→T grad (ψ) · λ · grad(T )dv Ω

28

Finite Element Simulation of Heat Transfer

The temperature ﬁeld sought must now be such that, for any weighting function ψ of Eψ , the following equations must be satisﬁed: −−→ − −−→ −−→T → ψQ dv + ψ λ · grad(T ) · n ds − grad (ψ) · λ · grad(T )dv = 0 Ω ∂Ω Ω −−→ − → λ · grad(T ) · n = q on ∂Ω q T = Td on ∂ΩT We see that no more second derivatives of the temperature ﬁeld appear in the problem formulation. The regularity required on T to give a sense to the problem has therefore been lowered; that is the reason why it is called a weak integral formulation. On the contrary, the regularity required for functions ψ has increased since ﬁrst derivatives of these functions have appeared. The same regularity is required for functions of ET and Eψ . A detailed mathematical analysis, which will not be performed here, makes it possible to demonstrate that these regularity conditions correspond to Sobolev’s ﬁrst space H 1 (Ω), which gathers the functions square integrable on Ω and whose ﬁrst derivatives are also square integrable on Ω. An important step toward the weak integral formulation of the problem consists of inserting the boundary conditions on ∂ΩT into the space of admissible functions ET . This is done by deﬁning: ET = T ∈ H 1 (Ω)/T = Td on ∂ΩT

[1.11]

Dirichlet boundary conditions (prescribed temperature) are therefore directly inserted into the solution space. These conditions are said to be essential for the problem formulation. Besides, we select weighting functions canceling one another on ∂ΩT . Thus, these functions ψ can be considered to be admissible variations of functions of ET . Indeed, if T ∈ ET , then for any ψ of Eψ , T +ψ ∈ ET . The following space Eψ is then deﬁned: Eψ = ψ ∈ H 1 (Ω)/ψ = 0 on ∂ΩT

[1.12]

Taking into consideration the ﬂux boundary conditions (on ∂Ωq ) and the previous deﬁnition of Eψ , the surface integral term appearing when integration by parts is applied becomes: −−→ → −−→ → ψ λ · grad(T ) · − n ds = ψ λ · grad(T ) · − n ds ∂Ω

∂ΩT

+ ∂Ωq

−−→ → ψ λ · grad(T ) · − n ds =

ψq ds ∂Ωq

Problem Formulation

29

Now problem [1.10] to be solved can be written as follows: Find T ∈ ET (deﬁned by equation [1.11]) such that for any ψ ∈ Eψ (deﬁned by equation [1.12]): −−→ −−→T ψQ dv + ψq ds − grad (ψ) · λ · grad(T )dv = 0 Ω

∂Ωq

[1.13]

Ω

The ﬂux boundary conditions are included in the function to cancel in problem [1.13]. They are said to be natural boundary conditions for the formulation. Problem [1.13] deﬁned in this way now offers a continuous mathematical framework for seeking a solution. It can be demonstrated that it is equivalent to boundary value problem [1.8] but is far more appropriate to seek approximate → solutions. Indeed the only obligation required for the solution T (− x ) is to belong to the space ET , i.e. to belong to H 1 (Ω) (square integrable function whose ﬁrst derivatives are square integrable), and to satisfy the essential boundary conditions T = Td on ∂ΩT . As weighting functions ψ of Eψ can be assimilated to admissible variations of functions ET , it is easy to demonstrate that, in linear cases (λ independent of T , Q and q linear according to T ), problem [1.13] is equivalent to the minimization problem of a functional Π(T ). In particular, when Q and q are independent of T , this functional equation is written as follows: Find T ∈ ET minimizing the functional equation: −−→ −−→T Π(T ) = 1 (T ) · λ · grad(T )dv − T Q dv − grad 2 Ω

Ω

T q ds

[1.14]

∂Ωq

Minimizing Π then comes down to ﬁnding T of ET such that, for any admissible variation δT in Eψ , the corresponding variation δΠ of Π vanishes. This is expressed as follows: For any δT ∈ Eψ : −−→ −−→T δΠ = (δT ) · λ · grad(T )dv − δT Q dv − δT q ds = 0 grad Ω

Ω

∂Ωq

We then go back to problem [1.13]. Approximation methods consist of constructing an approximate solution of problem [1.13] by working in the sub-spaces ETn and Eψn of ET and Eψ of ﬁnite

30

Finite Element Simulation of Heat Transfer

dimension n. By constructing Eψ , a particular element T ∗ of ET makes it possible to express all the elements of this space in the form T = T ∗ + ψ, where ψ is an element of Eψ . The Galerkin method [GAL 15] makes use of this property when it considers a space ETn deﬁned from the space Eψn as follows: ETn = T /T = T ∗ + ψ with ψ ∈ Eψn The Galerkin method then consists of replacing continuous problem [1.13] with the following discrete problem (of dimension n): Find T ∈ ETn such that for any ψ ∈ Eψn : −−→ −−→T ψQ dv + ψq(T )ds − grad (ψ) · λ · grad(T )dv = 0 Ω

∂Ωq

[1.15]

Ω

It is obvious that if we increase the dimension n of spaces ETn and Eψn , it is possible to construct a series of approximate solutions. It is demonstrated that this series converges to the solution of continuous problem [1.13] when n tends to inﬁnity. 1.3. Working example In this section, we will deal with the example of a metal plate heated by induction. The aim is to illustrate the application of physical modeling and mathematical analysis processes used in the previous sections to a real case. Physical modeling will allow us to specify the assumptions made and the equations used. Then we will examine the various solution methods: – direct methods based upon partial differential equations and boundary conditions resulting from physical modeling (problem [1.8]). Here we will restrict our study to the analytical integration model and its approximation by the ﬁnite element method; – collocation methods based upon problem [1.10] obtained by the weighted residual method. We will restrict our study to point and sub-domain collocation methods; – the Galerkin method resulting from the weak integral form of problem [1.13]. We will study a polynomial approximation of the temperature ﬁeld then a piecewise linear approximation of this ﬁeld (ﬁnite element method). 1.3.1. Physical modeling The work example considered in this chapter is that of an induction heated plate. It is illustrated in Figure 1.2. Its physical modeling can be performed at different levels. In this chapter, we will restrict our study to a 1D thermal analysis of the plate thickness. A line of length E (the plate thickness) represents the solid.

Problem Formulation

31

Heat conduction is only considered in the direction x. The corresponding thermal conductivity is written λ and assumed to be constant. inductor

modeling plate

x

T

E

convection

E-P

0

Q(x) Td

Figure 1.2. Work example: induction heated plate

On the bottom part of the plate (x = 0), the temperature is assumed to be known and equal to a constant value Td . On the top part of the plate, a constant heat exchange coefﬁcient H is used to represent the plate convection exchanges with the external environment, whose temperature is equal to Text = Td . Induction heating is represented by the following heat volumetric density: Q = Q(x) = Q0 e

x−E P

where Q0 corresponds to the heating volumetric power (in W/m3 ) at x = E, and in which P is a characteristic heating depth (depending in particular on the inductor frequency). Low frequency heating for instance will be represented by a high value of P , and will therefore tend to heat the solid “in its volume”. On the contrary, high frequency heating will concentrate the power on the plate surface (skin effect). This effect is used by the induction hardening process. A graphic representation of the temperature ﬁeld obtained by various methods is given in this section. To do so, we have used the following values for the parameters: – plate thickness: E = 0.1 m; – material thermal conductivity: λ = 30 W/m/K; – heat exchange coefﬁcient at x = E: H = 2,000 W/m2 /K; – temperature at x = 0 and of the external environment: Td = 20°C; – heating parameters: Q0 = 5.107 W/m3 and P = 0.02 m.

32

Finite Element Simulation of Heat Transfer

1.3.2. Direct methods The thermal equilibrium in the plate is governed by problem [1.8] which, in this case, is written as follows: 2 d T for 0 ≤ x ≤ E λ 2 + Q(x) = 0 dx dT = H Td − T (x) at x = E λ dx at x = 0 T (x) = Td 1.3.2.1. Analytical integration It is possible to integrate this differential equation analytically. We then obtain the exact solution of the problem, i.e. the temperature ﬁeld T (x) which satisﬁes the thermal equilibrium and the boundary conditions: E x PH Q0 P 2 − E Q0 P T (x) = Td + 1+ 1 − e− P x + e P 1 − eP λ + EH λ λ The temperature ﬁeld obtained is illustrated in Figure 1.3. It is obvious that, since an analytical solution exists in this working example, it is preferable to any other. However, in order to illustrate approximation methods, we are going to apply them to this example, then compare the solutions obtained with this analytical solution.

Figure 1.3. Working example: analytical solution

Problem Formulation

33

1.3.2.2. The ﬁnite difference method To obtain an approximate solution of the initial differential equation, we discretize axis Ox between x = 0 and x = E by deﬁning n + 1 equidistant points with x-axis xi = ih (with h = E/n and i = 0, 1, . . . , n). Then, with Ti being the temperature at the x-axis point xi , an approximation of T second derivatives at the solid internal points i = 1, . . . , n − 1 can be obtained as follows: Ti+1 − 2Ti + Ti−1 d2 T = dx2 x=xi h2 Finally, the temperatures at the boundary points with x-axes x = x0 = 0 and x = xn = E are obtained by applying the boundary conditions. At x = 0, the temperature T0 is known, its value is Td . At x = E, the incoming ﬂux can be expressed as λ(Tn − Tn−1 )/e, which makes it possible to write: hH hH Tn − Tn−1 = Td 1+ λ λ The problem to solve is thus expressed in the form of a linear system whose unknowns are the temperatures Ti : Td h2 0 0 Q x1 T0 λ 0 0 2 T1 h Q x 2 0 0 λ T2 .. .. · · · = . . · · · h2 −1 0 T Q x n−2 n−2 λ 2 −1 Tn−1 2 h hH Q xn−1 Tn −1 1 + λ λ hH Td λ

1 0 −1 2 0 −1 . .. .. . 0 0 0 0 0 0

0 ··· −1 · · · 2 ··· .. . 0 0

··· ···

0

···

The solution of this linear system produces an approximate solution of problem [1.8] in our working example. This approximate solution is given by a discrete set of values Ti . Each Ti corresponds to an approximate value of the temperature at the point with x-axis xi . It should be noted that the ﬁnite difference method gives a solution only for points with x-axis xi . It gives no a priori indication of the form of the solution beyond these points. In practice, it is advised to construct an approximate solution on the whole domain with a linear interpolation of the solution on each interval.

34

Finite Element Simulation of Heat Transfer

Figure 1.4. Working example: the ﬁnite difference method

Figure 1.4 gives the approximate solutions obtained by ﬁnite differences with a number n of discretization segments equal to 10, 20 and 50. It is established that, when n increases, the approximate solution tends slowly to the analytical solution. The ﬁnite difference method is relatively easy to implement in cases with simple geometries (1D, 2D or 3D). For instance, in the 3D case, the calculation points are distributed on a Cartesian grid whose main directions are axes Ox, Oy and Oz. However, the problem is far more complicated when the grid does not make it possible to give a good representation of the geometry or write the boundary conditions correctly. In that case, curvilinear ﬁnite difference methods can be used. 1.3.3. Collocation methods Collocation methods are based upon problem [1.10] resulting from problem [1.8] by applying the weighted residual method. In this working example, problem [1.10] is expressed as follows: Find T ∈ ET such that for any ψ ∈ Eψ : E d2 T ψ λ + Q(x) dx = 0 dx2 0 dT = H Td − T at x = E λ dx T = Td at x = 0

Problem Formulation

35

1.3.3.1. Point collocation In order to obtain an approximate solution of the initial differential problem, we discretize axis Ox between x = 0 and x = E by deﬁning n + 1 equidistant points with x-axis xi = ih (with h = E/n and i = 0, . . . , n), and for Eψ , we select the set of Dirac distributions corresponding to those points. The solution to the problem then comes down to seeking a function T (x) which can be differentiated twice, satisfying the boundary conditions and canceling the residual at each point xi . This produces the n + 3 equations: 2 d T λ 2 + Q(x) = 0 at x = x0 , . . . , xn dx dT = H Td − T λ at x = xn dx T = Td at x = x0 For instance, choose for ET the space of n + 2 degree polynomials. The function T will then be expressed as T (x) = a0 + a1 x + a2 x2 + · · · + an+2 xn+2 , with n + 3 coefﬁcients to determine. We obtain a linear system in ai , of n + 3 equations with n + 3 unknowns:

0 0 .. . 0 H λ 1

1+

0

···

(n + 2)(n + 1)xn0

0 .. . 0

···

(n + 2)(n + 1)xn1 .. . (n + 2)(n + 1)xnn H n + 2 + xn xn+1 n λ

H xn λ 0

··· ··· 0

···

Q(x ) 0 − λ Q(x ) a0 1 − λ a1 . .. . . . = an Q(xn ) − λ an+1 H an+2 T d 0 λ Td

The solution of this linear system yields coefﬁcients ai of the polynomial used as the approximation of the function T (x) sought. Figure 1.5 gives the approximations obtained with a number of collocation points equal to 3 (n = 2), 4 (n = 3) and 6 (n = 5). We can see that with only 6 points, the solution obtained practically coincides with the analytical solution in this case. However, in this case a 7-degree polynomial is necessary. In 2D or 3D, the number of coefﬁcients to determine would soon become very high, making this situation ill-suited to the linear system solution.

36

Finite Element Simulation of Heat Transfer

Figure 1.5. Working example: point collocation

1.3.3.2. Sub-domain collocation Axis Ox is divided into n control volumes. These are segments of length h = E/n. They are limited by points xi−1 = (i − 1)h and xi = ih (with i = 1, . . . , n). If we make Ti the temperature (assumed to be constant) in the segment [xi−1 , xi ] it is possible to calculate the ﬁrst and second derivatives with a ﬁnite difference scheme between the segment centers. We then obtain a linear system on the temperatures Ti of each of the n element segments: Td x2 h Q(x)dx 1 0 ··· 0 0 λ x1 T1 −1 2 · · · 0 0 T2 h x3 0 −1 · · · 0 0 Q(x)dx T3 · · · · · · · · · · · · = λ x2 ··· · · · ··· 0 ··· 2 −1 T 0 xn−1 n−1 h hH Q(x)dx Tn 0 0 · · · −1 1 + λ xn−2 λ hH Td λ Figure 1.6 illustrates the sub-domain collocation method applied to our working example, with respectively 20, 50 and 100 segments. It should be noted that the approximate solution tends to the analytical solution when the number of segments

Problem Formulation

37

Figure 1.6. Working example: sub-domain collocation

increases. This method is the basis of the ﬁnite volume method, which is widely used in thermal science. It combines a simple formulation, like ﬁnite differences, with the very physical notion of thermal equilibrium per element volume. 1.3.4. Galerkin method The Galerkin method is based upon the weak integral form [1.13] of the problem to solve. In this work example, problem [1.13] is expressed as follows: Find T ∈ ET such that for any ψ ∈ Eψ : E ψ(x)Q(x)dx + Hψ(E) Td − T (E) − λ 0

E

0

dψ dT dx = 0 dx dx

Spaces ET and Eψ are deﬁned as follows: Eψ = ψ ∈ H 1 [0, E] /ψ(0) = 0 ET = T ∈ H 1 [0, E] /T (0) = Td The Galerkin method consists of selecting ﬁnite spaces Eψn and ETn , of size n, instead of spaces Eψ and ET . 1.3.4.1. Polynomial functions Here we will consider space Eψn generated by the n monomials xi (i = 1, . . . , n). In order to satisfy the condition ψ = 0 on the boundary ∂ΩT (here x = 0), the

38

Finite Element Simulation of Heat Transfer

elements ψ of Eψn will be expressed with n parameters as follows: ψ(x) =

i=n

bi xi

i=1

The quantity to cancel in the discrete problem then becomes: i=n bi i=1

E

xi Q(x)dx + HE i Td − T (E) − λ

0

E

ixi−1 0

dT dx dx

=0

As this quantity must be equal to zero for any function ψ ofEψn , i.e. for any set of components bi , the n quantities between brackets, functions of T , must be equal to zero. This produces n equations in T . To generate the space ETn of admissible functions, the temperature ﬁeld T (x) is expressed as the sum of a particular admissible function T0 (x) = Td and an element Eψn . Therefore, the temperature ﬁeld is sought in the following form, where ai represents the n parameters to determine: T (x) = Td +

n

aj xj

j=1

The n equations ﬁnally become the linear functions of the n unknowns ai in the form: E E n HE i+j + λ ijxi+j−2 dx aj = xi Q(x)dx j=1

0

0

Figure 1.7 illustrates the temperature ﬁeld obtained with this method for 2-, 3- and 4-degree polynomials. We can see that the approximate solution converges toward the analytical solution when the polynomial degree increases. The major advantage of this method is that the matrix and the load vector of the system to solve have general expressions, added to which, the matrix is symmetric. Conversely, these quantities include integral equations whose analytical estimation may be complex. Their numerical estimation is certainly possible, but generates new approximations. In our example, the integral equations have been estimated from their analytical forms. Moreover, the matrix is full. Applying this method to 2D or 3D cases presupposes that one is capable of constructing a polynomial approximation satisfying the essential boundary conditions, which does not happen very frequently in practice.

Problem Formulation

39

Figure 1.7. Working example: polynomial approximation

1.3.4.2. Piecewise linear functions Starting with the discretization of segment [0, E] performed from n + 1 points, with x-axis xi = ih with h = E/n and i = 0, . . . , n, we can consider for Eψn the space of continuous functions varying linearly on each interval [xi , xi+1 ]. It is easy to demonstrate that this space is generated by the n + 1 following functions Ni (x) (this is expressed as x−1 = −h and xn+1 = E + h): x − x i−1 h Ni (x) = xi+1 − x h 0

if xi−1 ≤ x ≤ xi if xi ≤ x ≤ xi+1 if not

In order to satisfy the condition ψ = 0, at x = 0, the functions of space Eψn will then be expressed as follows: ψ(x) =

n

Ni (x)ψi

i=1

Space Eψn is therefore of the dimension n. A function ψ of this space will be characterized by its n parameters ψ1 , ψ2 , . . . , ψn , corresponding to the values taken by the function respectively at the points x1 , x2 , . . . , xn . Between these points, the function ψ is obtained by linear interpolation. The quantity to cancel in the discrete

40

Finite Element Simulation of Heat Transfer 1L[

[L

K

[L

K

[L

[

Figure 1.8. Generating piecewise linear functions

problem now becomes: i=n i=1

E

ψi

Ni (x)Q(x)dx + HNi (E) Td − T (E) − λ

0

0

E

dNi dT dx dx dx

=0

As this quantity must be equal to zero for any function ψ, and hence for any set of components ψi , the term between brackets in this equation must be equal to zero for any index i. We now introduce a particular admissible solution T ∗ (x) = N0 (x)Td to express any function of space ETn as follows: T (x) = N0 (x)Td +

n

Ni (x)Ti

i=1

The approximate solution selected is then a piecewise linear function. Its value is Td at x = x0 = 0, Ti at x = xi (for i = 1, . . . , n), and is obtained by linear interpolation between two consecutive points. When such a function is inserted into the quantity to cancel, we ﬁnally obtain a linear system of the following type: j=n

Kij Tj = Fi

j=1

whose unknowns are temperatures T1 , . . . , Tn . The terms Kij of the matrix and Fi of the load vector of this linear system are expressed analytically as follows: = λ K ij

E

dNi dNj dx + HNi (E)Nj (E) dx dx 0 E dNi dN0 Fi = Td dx + HNi (x)N0 (x)Td Ni (x)Q(x) − λ dx dx 0

Problem Formulation

41

Figure 1.9. Working example: piecewise linear approximation

Figure 1.9 illustrates the approximate solution in the case of 5, 10 and 20 segments distributed on axis Ox (between x = 0 and x = E). We can see that this solution tends rapidly toward the analytical solution when the number of discretization segments increases. The approximation used here is a ﬁnite element approximation. It has the same advantages as a polynomial approximation (symmetric matrix, general and speciﬁc expression of the matrix and load vector), but has none of the drawbacks mentioned before. Indeed the calculation of integral equations included in the linear system is much easier. Speciﬁcally, the integrations involved in the expression of the terms Kij and Fi can be restricted to the only ﬁelds in which the function to integrate is not equal to zero. For instance, for the diagonal terms of the matrix [Kij ] we obtain:

xi+1

Kii = λ xi−1

dNi 2 dx + HNi2 (E) dx

Moreover, it should be noted here that matrix [Kij ] is sparse. In this example, the integral equations have been calculated in an analytical way. We will see later on in this book that effective numerical methods make it possible to estimate these integral equations accurately and rapidly.

This page intentionally left blank

Chapter 2

The Finite Element Method

2.1. Finite element approximation Finite element approximation consists of dividing (or meshing) the geometric domain of study into element domains or ﬁnite elements. The approximation is built upon each element domain from the function values at certain speciﬁc points, called nodes, while ensuring certain regularity conditions of the global approximation on the whole domain. This type of approximation is also called sub-domain nodal approximation. If the division, or meshing, performed includes a total of nt nodes, then the problem includes nt unknowns representing the values taken by the function → T (− x ) at each node. 2.1.1. Mesh Divide (discretize) a solid Ω into m element domains, termed Ωe , called ﬁnite elements and satisfying the following two conditions: – the union of the m element domains Ωe makes up the solid Ω; – the intersection of two different element domains is void. The major advantage of this type of discretization is that each element domain Ωe can be expressed in a simple geometric form (2D triangle or quadrangle, 3D tetrahedron, hexahedron, etc.). These domains are interconnected by nodes. The set of these element domains is called a mesh of the domain Ω. In this book, the total number of mesh nodes will be written nt .

44

Finite Element Simulation of Heat Transfer node

element

boundary

Figure 2.1. Mesh example of a domain Ω

Figure 2.1 illustrates the example of a mesh constructed on a 2D domain Ω. In that ﬁgure we can see that the triangular elements used give a polyhedric approximation of the domain boundary. To improve that approximation, we can, for example, increase the number of elements. In the next chapter, we will see that there are also elements with curved boundaries, making it possible to obtain a very good approximation of the boundary of the domain studied. In a mesh, the nodes and elements are numbered. The nt nodes included in the mesh are numbered sequentially from 1 to nt , following their order of creation, during discretization for instance. This numbering process is called global numbering. Then we will designate: – {T } the vector including the nt temperatures at the mesh nodes in a column; – ψ the vector including the nt weighting function values at the mesh nodes on a line. For each element e we now write ne the number of nodes corresponding to it. These nodes are numbered locally from 1 to ne . Thus: – {T e } will be the vector including in a column the ne node temperatures of the element e, – ψ e will be the vector including on a line the ne weighting function values at the nodes of the element e. The global and local numbering methods of the mesh nodes are illustrated in Figure 2.2. They must now be interconnected. This is done by means of a matrix, [Ae ], which makes it possible to move from the vectors {T } or ψ (of dimension nt ) to the vectors {T e } or ψ e (of dimension ne ): {T e } = [Ae ] · {T }

and

T

ψ e = ψ · [Ae ]

[2.1]

The Finite Element Method 7

4

1

e=1

4

e=3

e=2

e 1

e=4

2

9

6

3

3

8

5

2

45

global numbering

local numbering

Figure 2.2. Numbering example in a mesh

The matrix [Ae ] has ne lines and nt columns. Each term Aeij equals 1 if the node numbered (globally) j on the whole structure coincides with the node numbered (locally) i on element e and 0 otherwise. In Figure 2.2, the previous equation is written as follows for the node temperatures of element 1: T1 T 2 1 T3 T 1 010000000 T 4 T 1 000010000 2 · T5 = 1 000100000 T3 T6 1 100000000 T4 T7 ! " ! " 1] [A 1 {T } T 8 T9 ! " {T }

Likewise, the element node weighting functions are expressed as follows: 0001 1000 0000 0010 $ ψ9 · 0100 " 0000 0000 0000 0000 ! "

# 1 ψ1

ψ21

!

ψ31

ψ 1

$ # ψ41 = ψ1 "

ψ2

··· !

ψ

ψ8

[A1 ]T

46

Finite Element Simulation of Heat Transfer

2.1.2. Nodal approximation The quantities involved in the problem to be solved are now estimated in each element e from the values they take at the nodes corresponding to that element. If we deﬁne Ωe = Ωe ∪ ∂Ωe (inside of the element plus its boundary), we now write for → → each element e, T (− x ) and ψ(− x ) in the form: e T1 → − → → → e − e e − e − T( x ) = Ni ( x )Ti = N1 ( x ), . . . , Nne ( x ) · · · · e i=1 Tn e e

→ ∀− x ∈ Ωe ,

n

[2.2]

→ x ) · {T e } = N e (− e − → N1 ( x ) − → ··· ψ(→ x)= ψie Nie (− x ) = ψ1e , . . . , ψne e · e − i=1 x) Nne (→ e

→ ∀− x ∈ Ωe ,

n

[2.3]

→ = ψ e · {N e (− x )} In equations [2.2] and [2.3], Tie and ψie respectively represent the temperature and the value of the weighting function at node number i (local numbering) of element e. → The functions Nie (− x ) introduced are known functions, making it possible to interpolate the temperature in element e from the node values. Nie is the interpolation function, or basis function, or shape function corresponding to the node i (local numbering) of element e, {N e } is the vector of dimension ne including in a column all the functions corresponding to element e, and N e is the vector including them on a line. → → → Note that it is equally possible to write ψ(− x ) = N e (− x ) · {ψ e } or ψ(− x) = → − e e ψ · {N ( x )}. The choice of the second form is prompted by the fact that the parameters corresponding to the weighting functions will then be written as a line vector ψ . For instance, in the case of element 1 in Figure 2.2, it will be possible to use the following expressions to approximate the temperature and weighting functions in element 1: 1 T1 1 # 1 $ T 2 T (x, y) = N1 (x, y) N21 (x, y) N31 (x, y) N41 (x, y) · 1 ! " T3 N 1 T41 ! " {T 1 }

The Finite Element Method

# ψ(x, y) = ψ11

ψ21

!

ψ31

ψ 1

47

1 N1 (x, y) N 1 (x, y) $ 2 ψ41 · " N31 (x, y) 1 N4 (x, y) ! " {N 1 }

It is obvious that the nodal approximations deﬁned by equations [2.2] and [2.3] are continuously differentiable inside each element e (as long as the interpolation functions are). In order to generate continuous approximations in the whole domain Ω we make use of shape functions satisfying the following two continuity conditions: → → 1) if − x is the position of the node j of element e, then N e (− x ) = 1 if i = j and j

i

j

0 otherwise; → 2) for any point indexed − x located on a face or an edge of element e which does → e − not include node i, then Ni ( x ) = 0. An approximation with shape functions satisfying these conditions is said to be a conform approximation. It makes it possible to generate continuous functions in the whole domain Ω. We demonstrate that, if the functions Nie are differentiable in Ωe , these conditions are sufﬁcient to allow the approximation to belong to space H 1 (Ω). More regular approximations can be constructed, for example with conditions on the derivative connections between the elements. These will not be dealt with in this book. It is important to note at this point that nodal approximation makes it possible not only to estimate a function in an element from its node values, but also to calculate its gradient, thus: − ∀→ x ∈ Ωe ,

ne

−−→ −−→ → → grad(T (− x )) = x ))Tie grad(Nie (− i=1

e T1 %−−→ & −−→ e .. e = grad(N1 ), . . . , grad(Nne ) · . e Tn e & %−−→ → x )) · {T e } = grad(N e (− − ∀→ x ∈ Ωe ,

ne

−−→ −−→T → → grad (ψ(− x )) = ψie gradT (Nie (− x )) i=1

[2.4]

48

Finite Element Simulation of Heat Transfer

−−→ T e (N ) grad 1 .. e e = ψ1 , . . . , ψne · . −−→T e grad (Nne ) '−−→ ( → = ψ e · gradT (N e (− x ))

[2.5]

$ #−−→ In equation [2.4], grad(N e ) is a vector including ne terms on a line. These −−→ terms are gradients grad(Nie ), with a number of terms equal to the dimension of the −−→ problem dealt with. In equation [2.5], gradT (N e ) is a vector including ne terms in −−→ a column. These terms are transposed gradients gradT (Nie ), with a number of terms equal to the dimension of the problem dealt with. For example, with a 2D element in Figure 2.2, if we consider a Cartesian coordinate system (x, y), it will be possible to estimate the temperature gradient and the weighting function gradient in element 1 as follows: 1 1 1 1 1 T1 ∂T ∂N ∂N ∂N ∂N 1 2 3 4 ) * 1 ∂x ∂x ∂x ∂x ∂x T 2 = · ∂T ∂N 1 ∂N 1 ∂N 1 ∂N 1 T 1 3 1 2 3 4 1 ∂y ∂y ∂y ∂y ∂y T ! " ! " 4! " & % −−→ −−→ grad(N 1 )

grad(T )

∂ψ ∂x

!

∂ψ ∂y

−−→T grad (ψ)

"

# = ψ11

ψ21

!

ψ31

ψ 1

{T 1 }

∂N11 ∂N11 ∂x ∂y 1 1 ∂N2 ∂N2 $ ∂x ∂y 1 ψ4 · ∂N31 ∂N31 " ∂x ∂y 1 1 ∂N ∂N 4 4 ∂x ∂y ! " ' ( −−→T grad (N 1 )

Finally, note that if the temperature ﬁeld approximation is performed by continuous construction on the whole domain Ω, that of its gradient, and hence the heat ﬂux density, is not so a priori. 2.2. Discrete problem formulation In the previous chapter, we saw that it was possible to approximate continuous problem [1.13] with a discrete problem (Galerkin method, equation [1.15]). To this

The Finite Element Method

49

end, we deﬁne two spaces Eψn and ETn of dimension n, to replace respectively spaces Eψ (weighting functions) and ET (admissible functions). The ﬁnite element method is based upon the sub-domain nodal approximation of the functions ψ and T . First, we will decompose the integral equations for the entire solid Ω appearing in [1.15] as a sum of integral functions on the element domains Ωe , then introduce the nodal approximation of the functions ψ and T inside each element. This will lead us to deﬁne element quantities corresponding to each element of the mesh. Then we will perform the assembly operation, the goal of which is to form a global system to solve. This global system will be the same dimension nt as the number of nodes used for the mesh. The insertion of the essential boundary conditions (on T ) into this global system will then be studied in detail. 2.2.1. Element quantities → If we apply nodal approximation [2.3] to the weighting functions ψ(− x ) in discrete problem [1.15], the quantities involved can be written as follows: ψQ dv = Ω

ψ e ·

e=1

ψq ds = ∂Ωq

m

m

∂Ωe ∩∂Ωq

e=1

−−→ −−→T ψ e · grad (ψ) · λ · grad(T )dv =

Ω

e=1

{N e } Q dv

ψ e ·

m

Ωe

Ωe

{N e } q ds

'−−→ ( −−→ gradT (N e ) · λ · grad(T )dv

If these expressions are transferred into equation [1.15], the quantity to cancel then becomes: m −−→ −−→T ψQ dv + ψq ds − ψ e · {Re (T )} grad (ψ) · λ · grad(T )dv = Ω

Ω

∂Ωq

e=1

where the vector {Re (T )}, speciﬁc to each element e, is called element residual. If we apply nodal approximation [2.2] to the temperature ﬁeld, this element residual is expressed according to the nodal temperatures {T e }, as follows: {Re (T )} = {N e } Q dv + {N e } q ds Ωe

−

Ωe

∂Ωe ∩∂Ωq

'−−→ ( %−−→ & gradT (N e ) · λ · grad(N e ) {T e } dv

[2.6]

50

Finite Element Simulation of Heat Transfer

The element residual is a fundamental quantity resulting from the ﬁnite element method. In this section, we have restricted our study to the so-called linear cases. A linear case is characterized by the fact that the element residual can be expressed linearly according to the node temperatures of the corresponding element, i.e. according to the vector {T e }. Equation [2.6] shows that the element residual is a linear function of the node temperatures of the corresponding element if: – the thermal conductivity of the material (matrix λ) is not temperature-dependent; – the term Q, the volumetric heat source, is a temperature-independent function (as a matter of fact, it could vary linearly with the temperature, but as far as we know, this type of variation does not correspond to any physical example); – the heat ﬂux q applied to the boundary ∂Ωq of the solid is a linear function of the temperature (for instance q = h(Text − T ) + q0 , where h and q0 are temperature-independent). Then it is possible to write: {Re (T )} = {F e } − [K e ] · {T e } In this equation, [K e ] is a square matrix of dimension ne × ne . This is called an element matrix. It is also called an element conductance matrix, because we will see that the main contribution comes from thermal conductivity. It is also called an element stiffness matrix, an irritating term for thermal scientists but revealing the origins of the ﬁnite element method in mechanical engineering. The vector {F e } in this equation includes ne terms in a column. It is called an element load vector. The element matrix [K e ] and the element load vector {F e } are written as follows: ' %−−→ & ( −−→T e e e [Λ ] = (N ) · λ · ) dv grad grad(N

[K e ] = [Λe ]+[H e ] with: {F e } =

Ωe

Ωe

e [H ] =

{N e } Q dv +

∂Ωe ∩∂Ωq

∂Ωe ∩∂Ωq

[2.7] {N e } h N e ds

{N e } hText + q0 ds

[2.8]

In these general expressions, note that: – the matrix and element load vector are obtained by integration on the corresponding domain Ωe . In Chapter 3, we will introduce an efﬁcient numerical method (Gauss method) which makes it possible to evaluate these integral equations accurately;

The Finite Element Method

51

e – the element matrix [K e ] is symmetric. As a matter of fact, the general term Kij of this matrix is expressed as follows: + , −−→ −−→T e e Kij = Nie hNje ds grad (Ni ) · λ · grad(Nje ) dv + Ωe

∂Ωe ∩∂Ωq

2.2.2. Assembly One of the major advantages of the ﬁnite element method is that it highlights element quantities, which are therefore calculated in the corresponding element only. However, the temperature ﬁeld sought is related to the whole environment studied. Thus, the assembly operation consists of turning these element quantities into global quantities so as to construct the equation system to solve. We saw that the quantity to cancel in the discrete problem resulting from [1.13] was determined according to an element residual, {Re }, whose general expression is given by equation [2.6]. If we now use equation [2.1], applied to the weighting functions, this quantity to cancel can be expressed as follows: m

ψ · {R (T )} = ψ · e

e

e=1

m

T

[Ae ] · {Re (T )} = ψ · {R(T )} = 0

[2.9]

e=1

In this equation, ψ is a vector of dimension nt , including on a line the weighting function values ψ at the mesh nodes. With regard to the vector {R(T )} it is also of dimension nt and includes in a column nt functions of T resulting from the element residual assembly at each node. This vector is said to be residual. Since the above expression must be equal to zero whatever the weighting functions ψ, the nt components of the residual vector must be equal to zero. This is written as follows: {R(T )} =

m

T

[Ae ] · {Re (T )} = {0}

e=1

In fact, we will see later that the insertion of the essential boundary conditions may lead to either not canceling certain components of that residual vector or modifying its expression. In the linear case, the element residual is written linearly according to the temperatures. Indeed if we use equation [2.1] once more, but this time applied to the nodal temperatures, we obtain: {R(T )} =

m e=1

T

[Ae ] · {Re (T )} = {F } − [K] · {T }

[2.10]

52

Finite Element Simulation of Heat Transfer

In this equation, [K] is a square matrix of dimension nt × nt , which is obtained by assembling the element matrices [K e ]. It is called a global matrix or a conductance matrix or a stiffness matrix (after mechanical engineering once again): [K] =

m

T

[Ae ] · [K e ] · [Ae ]

[2.11]

e=1

Likewise, {F } is a vector of dimension nt , which is obtained by assembling the element quantities {F e }. It is called a load vector or stress global vector: {F } =

m

T

[Ae ] · {F e }

[2.12]

e=1

Canceling the global residual components leads us to solve the following linear system (of dimension nt ): [K] · {T } = {F } In Chapter 3, we will see that inserting essential boundary conditions modiﬁes either the linear system dimension or the matrix and load vector of that system. Figure 2.3 illustrates the construction of [K] and {F } by assembling the matrices and element load vectors as in the case of the mesh in Figure 2.2: – the component Kij of [K] corresponds to nodes i and j. It alone incorporates the contribution of elements including nodes i and j. For instance, the component K55 receives contributions from the element matrices of elements 1, 2, 3 and 4 because all these elements include node 5. On the contrary, only element 3 which includes nodes 5 and 7 brings a contribution to the component K75 . Finally many components Kij are equal to zero because no element incorporates the corresponding nodes (for example K37 = 0); – the component Fi of {F } corresponds to node i. It includes the contributions of all the elements including node i. For instance, F6 receives the contributions of the element load vectors of elements 2 and 4 which contain node 6. It can be noted that the solution of the previous linear system minimizes the following quadratic form: Π∗ =

1 T · [K] · {T } − T · {F } 2

This quadratic form corresponds to the discretization of the functional Π introduced in formulation [1.14].

The Finite Element Method 1

4

7

1

3

8

5

2 2

4

[K] =

1

1

1 2

2

2

2

9

6

3 1

1

1

1

1 2

53

1 1 2

2

2

2

2

1

1

1

3 1

3

3

3

1

3

1

1 2

1

3 4 2

3

2

3 1 2

4

3 4

4

1 2

3 4

2

2

2

4 2

4

4

4

2

4

{F} =

3

3

3

3

3

3 4

3 4

3 4

3

4

3 4

4

4

4

4

4

Figure 2.3. Example of an assembly on a simple mesh (contributions of the elements to the global matrix and load vector)

2.3. Solution In the previous section, we obtained a discrete formulation of the quantity to cancel in continuous problem [1.13]. In the linear case, this formulation leads to the solution of a linear system related to the temperatures at the mesh nodes. In this section, we list the main methods used to insert the essential (temperature) boundary conditions into the system to solve, then detail a few techniques for storing and solving the linear system obtained and end with an analysis of results. 2.3.1. Application of temperature boundary conditions Reconsider equation [2.9], with, in the linear case, expression [2.10] for the residual {R(T )}. We obtain: ψ · {F } − [K] · {T } = 0

[2.13]

In this equation, the nodal temperature vector {T } is of dimension nt . Likewise, ψ is a vector of dimension nt , including all the nodal parameters of the weighting functions.

54

Finite Element Simulation of Heat Transfer

It is important to note at this point that, among the nt nodes of the mesh used, some are located on the boundary ∂ΩT of the solid, in which the temperature T is known and in which the weighting functions ψ are therefore equal to zero. If we designate n∗ the number of mesh nodes located on that boundary, it is convenient to partition the nodal vectors ψ and {T } as follows:

# $ {Tc } ψ = ψc ψl , {T } = {Tl } We have gathered from the ﬁrst components of the vectors ψ and {T } the n nodal quantities known a priori (vectors ψc and {Tc }). In order to cancel the weighting functions ψ on the domain boundary ∂ΩT , it is sufﬁcient to cancel their values at the nodes of the boundary ∂ΩT , values gathered in vector ψc . Likewise, for a function T to be admissible, it is sufﬁcient to give it the temperature values known at the nodes of the boundary ∂ΩT which can be expressed as {Tc } = {Td }. In fact this natural course of action is not the one generally used in ﬁnite element software, while linear vectors and systems are never partitioned in this way (with the n∗ components numbered ﬁrst). This expression is introduced to simplify the presentation of the different methods used to apply the temperature boundary conditions. ∗

If the same partitioning method is used for the matrix [K] and the load vector {F } of the problem to solve, equation [2.13] can eventually be rewritten as follows: / -

.

0 [Kcc ] [Kcl ] {Tc } {Fc } · ψc ψl · − =0 [2.14] T {Fl } {Tl } [Kcl ] [Kll ] As this equation must be satisﬁed whatever the vector ψl and with ψc = 0 , this leads to the following linear system, of nt − n∗ equations, for nt − n∗ unknowns (nodal vector {Tl }): T

[Kll ] · {Tl } = {Fl } − [Kcl ] · {Td }

[2.15]

The resolution of this system yields the unknown nodal vector {Tl } and hence the temperature at any point of Ω by making use of the shape functions. This speciﬁc method, called the elimination method, is rarely used in computer codes. We might think that, as we know n∗ nodal temperatures (at the nodes of ∂ΩT ), the physical problem contains only nt − n∗ unknowns (vector {Tl }). As a matter of fact, this is not the case. We introduce concurrently to n∗ known nodal temperatures n∗ additional unknowns corresponding to the heat ﬂux prescribed by the external

The Finite Element Method

55

environment in order to satisfy the condition {Tc } = {Td }. Those n∗ additional unknowns correspond to the reactions of the external environment at the nodes where the temperature is prescribed; they are gathered in the vector {rc }. The introduction of the vector {rc } now allows us to constitute the linear system to solve, in which the number of unknowns is nt (the components of the vectors {Tl } and {rc }) as follows:

. [Kcc ] {Fc } + {rc } − T {Fl } [Kcl ]

/

[Kcl ] {Tc } · =0 {Tl } [Kll ]

[2.16]

Whatever the method used to take into consideration the (temperature) essential boundary conditions, the resolution of the problem yields the unknown nodal vector {Tl } by satisfying the condition {Tc } = {Td }. This makes it possible to obtain the − temperature ﬁeld T (→ x ) by nodal approximation. The penalty method is the most widely used in computer codes to determine the essential conditions {Tc } = {Td }. Applying those conditions to certain nodes of a boundary comes down to assuming that at this point the solid is plunged in a ﬂuid at the corresponding temperature, with a very high heat transfer coefﬁcient which will be written P here. This comes down to replacing the (temperature) essential boundary conditions with (node) ﬂux conditions, by using the following expression for the external environment reactions at these points: {rc } = P {Td } − {Tc } With this expression, use equation [2.16], with the vectors {Tl } and {Tc } (i.e. all the components of the unknown nodal vector {T }) as unknowns: / .

[Kcc ] + P [I] [Kcl ] {Fc } + P {Td } {Tc } · [2.17] = T {Tl } {Fl } [Kcl ] [Kll ] The penalty method therefore consists of adding P on the diagonal of the matrix [Kcc ] and P {Tc } to the vector {Fc }. In equation [2.17], [I] represents the identity matrix and, in practice, a value P Max(Kij ) is selected. This leads to a linear system with predominant terms on the diagonal. In addition to its good numerical stability, this linear system has the advantage of solving a system of similar dimension, whatever the temperature boundary conditions used. The dimension of this system is always equal to the number of mesh nodes. Another method is commonly used in computer codes. It is called Lagrange multipliers and consists of introducing an unknown vector {λc }, of dimension

56

Finite Element Simulation of Heat Transfer

n∗ , into the function to cancel. This unknown vector will be compensated by n∗ additional equations {Tc } = {Td }. The system to solve therefore includes nt + n∗ equations, for nt + n∗ unknowns (vectors {Tc }, {Tl } and {λc }): {Tc } [Kcc ] −k [I] [Kcl ] {Fc } · 1 {λc } = −k {Td } −k [I] 0 0 [2.18] k T {Fl } 0 [Kll ] {Tl } [Kcl ] A factor k is introduced into equation [2.18]. This factor is homogenous with a thermal conductivity and allows us to condition the system properly (for its numerical solution). Each node value included in the vector {λc } plays the role of a Lagrange multiplier corresponding to a prescribed temperature condition. System [2.18] can then be obtained directly (with k = 1) by solving the following saddle point problem: 1 Max Min T · [K] · {T } − T · {F } + T − Td · {λc } T λ≥0 2 It can be noted that the system to solve in equation [2.18] is always symmetric (as it is with other methods). One difﬁculty of this type of method, combined with pivot methods used to solve the linear system, is that it reveals diagonal terms equal to zero in the ﬁrst member matrix. This difﬁculty can be overcome by numbering the equations to solve. This method also has the drawback of increasing the size of the linear system to solve. In the case of Lagrange multipliers, the external environment reactions required to satisfy the condition {Tc } = {Td } i.e. the vector {rc } components, coincide with Lagrange multipliers {λc }. 2.3.2. Linear system solution The matrices involved in linear systems [2.15], [2.17] and [2.18] are generally large. Indeed, it is common today to process models including several tens and even hundreds of thousands of nodes. In order to solve the linear system, it is necessary to use the method best adapted to the problem being dealt with. Most of the time, the size of the problem and the accuracy desired in its resolution will be the key factors in the choice of the solution method to use. The problem size is now a major factor, all the more so as a matrix storage method is related to each solution method (see next section) and the problem size a computer is capable of processing is directly limited by its memory space.

The Finite Element Method

57

A comprehensive description of all existing methods would go beyond the scope of this book. We will therefore restrict our study to the most common methods used in academic and industrial computer codes. Interested readers are invited to consult specialized books such as [CIA 82]. Consider therefore a linear equation system similar to that obtained by the ﬁnite element technique. This system is as follows: [A] · {x} = {b}

[2.19]

where {x} represents the unknowns vector, [A] the matrix and {b} the load vector. The system size (number of unknowns and number of equations) will be written n. This size, as well as the expression of {x}, [A] and {b} according to the quantities deﬁned in this book, depends on the method used to apply the temperature boundary conditions: – with an elimination method, we have a system of dimension n = nt −n∗ , that is, total number of nodes minus the number of nodes at which the temperature is known, where the components are expressed according to equation [2.15]: [A] = [Kll ] {x} = {Tl } {b} = {F } − [K ]T · {T } l cl d – with a penalty method, the system is of dimension n = nt , the total number of nodes, and equation [2.17] allows us to express the system components as follows: / . [Kcc ] + P [I] [Kcl ] [A] = T [Kll ] [Kcl ]

{Tc } {x} = {Tl }

{Fc } + P {Td } {b} = {Fl } – with a Lagrange multiplier method, the linear system dimension becomes n = nt + n∗ , total number of nodes plus the number of nodes at which the temperature is known, and the components are written according to equation [2.18]:

58

Finite Element Simulation of Heat Transfer

[Kcc ] −k [I] [Kcl ] −k [I] 0 0 [A] = T 0 [Kll ] [Kcl ] {Tc } 1 {x} = {λc } k {Tl } {Fc } {b} = −k {T } d {F } l

−1

The solution of system [2.19] is simply written {x} = [A] · {b}. However, contrary to what this expression might suggest, computing {x} never requires us to calculate the inverted matrix [A]. As a matter of fact, computing an inverted matrix of dimension n × n requires the solution of n linear systems. Globally, linear system solution methods are arranged into 2 classes: – direct methods (Gauss, Cholesky, etc.), – iterative methods (Jacobi, Gauss-Seidel, conjugate gradients, etc.). 2.3.2.1. Direct methods The general principle of direct methods is to transform the initial equation system into a system with the same solution but whose resolution is simpler. Among the whole set of direct methods, the Gauss method is probably the most popular one. It is a general method which is applied to symmetric or non-symmetric matrices, providing that these matrices can be inverted, of course. It consists of two steps: – an elimination step which triangulates system [2.19], – a second step to solve the triangular system obtained. The triangulation of system [2.19], of dimension n, is performed in n − 1 steps. At each step, an equivalent linear system is constructed by eliminating the terms under the diagonal of an additional column of the matrix. Let [A](1) = [A], {b}(1) = {b}, (i) (i) and let [A] and {b} , respectively be the matrix and load vector of the equivalent (i) linear system obtained after i − 1 steps; [A] is written as follows:

The Finite Element Method

(i)

A11

= 0

(i)

[A]

··· .. .

(i)

···

(i)

···

(i)

···

A1i .. . Aii .. .

Ani Matrix [A]

(i+1)

(i+1)

and load vector {b}

59

(i) A1n .. . (i) Ain .. . (i)

Ann

in the following step are then obtained

by:

1

[A]

(i+1)

{b}

(i+1)

= [E]

(i)

= [E]

(i)

· [A]

(i)

· {b}

(i)

with

[E]

(i)

=

..

.

0 1 (i)

−

Ai+1,i

0

(i)

Aii .. .

1 0

..

.

(i)

−

Ani

(i)

Aii

1

(i)

It is clear that this operation is possible only if the terms Aii (called pivots) are not equal to zero. If matrix [K] is singular, we will necessarily have a pivot equal to zero even if [K] can be inverted. As a demonstration, examine the following system of 2 equations with 2 unknowns: 1

0 2

1 0

2 3 4 3 4 T1 100 . = T2 100

It is obvious that the ﬁrst pivot encountered with the Gauss method is equal to zero, whereas the system admits the unique solution T1 = 50, T2 = 100. To perform the triangulation, it is then necessary to replace the equation considered with one of the following. (Generally speaking, to obtain the highest possible solution accuracy, it will be sensible to process at each step the equation with the pivot whose absolute value is highest.) As a matter of fact, the Gauss method comes down to factorizing the matrix of system [2.19] as follows: [A] = [L] · [U ]

60

Finite Element Simulation of Heat Transfer

where [L] is a lower triangular matrix and [U ] an upper triangular matrix. The matrix [U ] coincides with the triangular matrix obtained by the elimination process described 6−1 5 (2) 6−1 6−1 5 5 previously and [L] = E (1) · E · · · E (n) . The solution of a linear 3 system with the Gauss method requires about 2n3 arithmetic operations. When the matrix of system [2.19] is symmetric positive deﬁnite as is the case in heat conduction linear problems, it will be wise to choose the Cholesky method which requires only 3 about n3 arithmetic operations. This method consists of factorizing [A] as follows: [A] = [L] · [L]

T

Matrix [L] is a lower triangular matrix whose elements are computed gradually. √ Ai1 We ﬁrst calculate L11 = A11 then, for i = 2, . . . , n,Li1 = L . The Lip being then 11 assumed to be calculated for p = 1, . . . , j − 1 and i = p + 1, . . . , n, we calculate consecutively: 7 8 j−1 8 Ljj = 9Ajj − L2jp p=1

∀i = j + 1, . . . , n,

: Lij =

Aij −

;j−1 p=1

Lip Ljp

Ljj

2.3.2.2. Iterative methods Iterative methods give the solution of system [2.19] as the limit of a series of (i) vectors {x} . They are very efﬁcient at processing problems of a large size insofar as they require less data than direct methods, as will be seen in the next section. The general principle of these methods is to factorize the matrix [A] of system [2.19] as follows: [A] = [M ] − [N ] then to approximate the solution iteratively by solving, at each iteration, the system: (i+1)

[M ] · {x}

= [N ] · {x}

(i)

+ {b}

Matrix [M ] will be chosen so that the solution is simple. The Jacobi method, the Gauss-Seidel method and the over-relaxation method operate on this principle and correspond to speciﬁc choices of matrices [M ] and [N ].

The Finite Element Method

61

Now assume [A] to be symmetric positive deﬁnite and spend some time on the conjugate gradient method. In this method, we consider the function f which relates the following scalar to any vector {x}: f ({x}) =

1 x · [A] · {x} − x · {b} 2

It is easy to note that the solution sought for linear system [2.19] coincides with the minimizing vector f . The conjugate gradient method consists of approximating step by step the solution by minimizing the function f at each iteration in a particular (i) direction, called descent direction. Consider {u} , an arbitrary vector for the (i+1) moment, as the direction of the iteration i. The iterated {x} is chosen as (i+1) (i) (i) (i+1) (i) (i) {x} = {x} + α {u} , and α is determined in such a way that {x} (i) minimizes f in the direction {u} . We ﬁnd: (i)

α(i) = (i)

where {r}

(i)

u

· {r}

(i)

· [A] · {u}

u

(i)

= {b} − [A] · x(i) is the residual vector of iteration i. (0)

(0)

(0)

= {r} , the Starting with a test vector {x} and a descent direction {u} (i) (i) (i−1) vector {u} is constructed at each iteration so that u · [A] · {u} = 0. For this (i) (i) (i−1) purpose, we write {u} = {r} + β (i) {u} . We then demonstrate by recursion that we have: (i) (k) r · {u} = 0 (i) (k) ∀k = 1, . . . , i − 1, r · {r} = 0 (i) (k) u · [A] · {u} = 0 According to the latter relationship, we say that the consecutive descent directions are [A]-combined. We eventually obtain: (i)

{u}

(i)

= {r}

(i)

= {r}

(i)

−

+

r

(i−1)

· [A] · {u}

(i−1)

u

· [A] · {u}

(i)

· {r}

(i−1)

· {r}

r r

(i−1)

(i−1)

{u}

(i) (i−1)

(i−1)

{u}

The conjugate gradient method converges to the solution in, at most, n iterations. In this sense it can be considered to be a direct method. Various pre-conditioning

62

Finite Element Simulation of Heat Transfer

methods make it possible to obtain an accurate solution with a relatively small number of iterations. This method is extended to the case of matrices which are not positive deﬁnite symmetric. 2.3.3. Storing the linear system matrix In practice, the matrix of the linear system to solve is stored by means of a method using its properties. For example, if it is symmetric, it will be sufﬁcient to store only the upper triangular part. If, in addition, it is “sparse” (i.e. it includes a great number of terms equal to zero), which is the case with the ﬁnite element method, the following methods are generally used: – the skyline storage of a matrix is the most widely used method in computer codes. The matrix is stored in columns (or lines) variable in length. It is particularly well-adapted to linear system direct solution methods (Gauss, Choleski, etc.), which consist of triangulating the matrix. Figure 2.4 illustrates the storage of the matrix [K] resulting from the assembly of Figure 2.3. It should be noted that only the terms of the upper part of the matrix have been stored. The maximum number of terms stored per line on a symmetric matrix is half the bandwidth. In Figure 2.4, the half bandwidth equals 5, whereas the number of nodes is 9. We can show that generally speaking the half bandwidth is equal to the maximum difference between the node numbers of a same element, increased by 1. In Figure 2.4, we obtain for instance (5 − 1) + 1 = 5 considering element 1. If we renumber the mesh nodes in a sensible way, it is possible to replace certain lines and columns in the matrix so as to reduce the bandwidth of the matrix stored. This renumbering operation is very important from a practical point of view. Indeed it allows us to substantially reduce the number of terms to store and to accelerate the solution process. If b is the system half bandwidth, then this method involves the storage of a number of values equal to about b times the matrix dimension to which it is necessary to add a pointer per line (or column) giving its length. As b is globally proportional to the matrix dimension, the number of terms to store is proportional to the square of the problem dimension; – the sparse or compact storage of a matrix is used mainly when an iterative solution method (Jacobi, Gauss-Seidel, conjugate gradients, etc.) of the linear system is employed. This method stores only the matrix terms different from zero. To illustrate this storage method, consider the mesh in Figure 2.4. We see that the maximum number of non-vanishing terms on a line of the matrix is equal to 9. This number corresponds to the maximum number of nodes connected to one node, which is reached here at node 5. If we reﬁne the discretization, by dividing our rectangle into 10 × 10 elements, we see that the maximum number of nodes connected to a same node, and hence of terms different from zero on a matrix line, remains equal to 9. Therefore, the number of terms different from zero to store is proportional to

The Finite Element Method 1

4

7

1

3

5

2

8

2

4

9

6

3

half bandwidth

[K] =

1

1

1

1 2 2

1

1

1

1 2 2

63

Skyline 1

1

1 2

1 2

2

2

2 2

1

3 1

1 2

1 2

3 4 2

2

2

2

3

3

3 3

3

3

3 4

3 4

4

4

4

4

3

3 3 4

4 4

Figure 2.4. Example of skyline storage

the number of nodes (equal to knt ). The proportionality factor k is given by the mean number of nodes connected to a given node (in the order of 9 in our example). According to the type of elements, the value of this factor ranges from 8 to 20 on a 2D mesh and 20 to 80 in 3D cases. However, for each term different from zero, it is necessary to store its value and position in the matrix (line and column numbers). Finally the number of terms to store is proportional to nt and equal to 3knt , which makes this storage method particularly efﬁcient for large equation systems. 2.3.4. Analysis of results Solving a steady state heat conduction problem by means of the ﬁnite element method leads to a vector {T } including the temperatures calculated at the mesh nodes. This vector makes it possible to calculate a temperature ﬁeld at any point of the domain of study by using the nodal approximation and therefore the shape functions. It also enables us to calculate the temperature gradient ﬁeld with the shape function gradients. In order to obtain that result, the method involves the calculation of element quantities obtained by numerical integration with the Gauss method (see Chapter 3). As a consequence, the variables involved in element quantities are calculated in each element of the mesh, at the integration points (Gauss points). The results obtained during analysis can then be arranged into two categories and they are not used in the same way:

64

Finite Element Simulation of Heat Transfer

– the values calculated at the nodes (nodal quantities), which in our case are the temperatures. If the temperature ﬁeld nodal approximation (equation [2.2]) is used, the temperature at each point of the solid in thermal equilibrium can be calculated. This can be visualized by drawing the temperature variation along a curve crossing the solid or temperature contour lines or surfaces. Figure 2.12 illustrates the drawing of temperature contour lines; – the values calculated at the element integration points (see section 3.2) are those involved in element quantities. In our case, it is essentially the temperature gradient → − and heat ﬂux density φ which can be derived from Fourier’s law [1.3]: %−−→ & −−→ grad(T ) = grad(N e ) · {T e } − → −−→ φ = −λgrad(T ) As seen in section 2.1, the heat ﬂux density approximation, which is proportional to the thermal gradient according to Fourier’s law [1.3], is not continuous on all the domain Ω. From a physical point of view, the heat ﬂux density is continuous in a homogenous environment. Therefore, it may be interesting to construct an approximation of the heat ﬂux density on the whole domain from the values calculated at the element integration points. To do so, it is traditional to use the ﬁnite element approximation from nodal values of that density. These nodal values can be obtained in several different ways, for instance: – by extrapolating at the nodes of each element the values calculated at the integration points, then computing the mean node values obtained; – by smoothing according to the method described hereafter. In addition to obtaining a more accurate approximation, this method allows us to use the same isovalue drawing tools as those used for the temperature ﬁeld. 2.3.4.1. Smoothing the heat ﬂux density Among the quantities estimated at the mesh integration points, we have selected → − the heat ﬂux density φ to illustrate a smoothing method. As mentioned above, a fairly natural way of smoothing these quantities consists of using the ﬁnite element approximation used for the analysis. We then have to seek a →→ − ﬁeld Φ (− x ) in the form of a nodal approximation:

→ ∀− x ∈ Ωe ,

ne '− → − → ( →− − → → Φ( x ) = Nie (− x ) Φ ei = N e · Φ e '→ ( i=1 '− →( − Φ e = [Ae ] · Φ

The Finite Element Method

65

In this equation, Ωe is an element domain (a mesh element), ne is the number of nodes of this element, N e is a vector including its ne shape functions on a line, and → − { Φ e } is a vector including in a column the ne node values of Ωe of the heat ﬂux density vector (the number of components of that vector depends on the geometric size of the problem dealt with). Besides, [Ae ] is the matrix connecting the vector of → − → − local unknowns, { Φ e }, to the vector { Φ } including in a column the nodal values of all the heat ﬂux density components. Therefore, the objective is to determine the nodal values included in the vector → − → − { Φ } from the ﬁeld φ calculated from the nodal temperatures. These values are calculated in order to minimize the following error function: − → − → − → − → F = φ − Φ · φ − Φ dv Ω

It should be noted at this point that the error functional F does not necessarily result from an integration on the whole domain studied. It is possible to restrict the integration to a portion of the domain here. This is necessary when, for instance, the environment is composed of the discrete assembly of several materials (for example, coated metal sheet). In this case, the thermal gradient is no longer continuous at the interface between the two materials. → − − → If we simply state that the variation δF of F implied by any variation δ Φ of Φ , − → at φ ﬁxed, is equal to zero, we obtain: → − → − − → δF = −2 δ Φ · φ − Φ dv = 0 Ω

→ − − → If we use the nodal approximation to express Φ and δ Φ , we ﬁnally obtain the following linear system to solve: '− →( '− →( [L] · Φ = S [2.20] → − In that system only the load vector { S } involves the heat ﬂux density estimated at the integration points. That load vector is expressed in the form of the assembly of → − element quantities { S e }: m '− →( e T '→ − ( S = [A ] · S e e=1

with

'− →e ( S = Ωe

→ − {N e } φ dv

66

Finite Element Simulation of Heat Transfer

− → It will then be possible to calculate the element quantities S e by means of the Gauss integration scheme, which directly uses the heat ﬂux densities known at the integration points. Matrix [L] of the system to solve for smoothing is in fact independent of the quantities to smooth. It is also expressed as the assembly of element quantities [Le ]: [L] =

m

T

[Ae ] · [Le ] · [Ae ]

with

[Le ] = Ωe

e=1

{N e } · N e dv

Note in equation [2.20] that smoothing heat ﬂux densities requires, in a 3D analysis the solution of a linear system including 3nt equations for 3nt unknowns (nt is the number of mesh nodes). This is a priori more costly than the original analysis. However, it can be noted that matrix [L] has a very special form, similar to the thermal capacitance matrix encountered in transient analyses (see Part 2). In the case of transient analyses, it is common to use approximate capacitance matrices in a lumped diagonal form because it dramatically simpliﬁes the calculations while producing acceptable results. By operating here in the same way, matrix [L] is replaced with a lumped diagonal 5 6 matrix Ld obtained in this instance by transferring on the diagonal the sum of the terms on each line (δij is the Kronecker symbol): Ldij = δij

nt

Lik

k=1

It is obvious that the solution of system [2.20] is then immediate. This method is applicable to any value calculated at the integration points (thermal gradient, ﬂux, phase proportions, etc.). It may lead to node values greater than those actually calculated at the integration points. This may be a problem when the physical values involved are limited (phase proportions for instance are to range from 0 to 1). It is then necessary to introduce physical boundaries into the minimization problem. 2.3.4.2. Result accuracy It is very important to underline at this point that the ﬁnite element method is a technique for the approximate solution of the problem encountered. Among the many sources of error appearing during the numerical simulation of a physical problem, we will focus in this book on those always present during ﬁnite element analysis: – errors in physical data input (material properties, boundary conditions and stresses);

The Finite Element Method

67

– geometric errors due to mesh representation; – numerical errors (element quantity integration and linear system solution); – discretization errors (nodal approximation). The most important errors are probably related to input data. Indeed, the material’s physical properties or boundary conditions are very often known to an approximation of 10%. It is then futile to try and reﬁne a mesh to gain 5% on the numerical result accuracy. Geometric errors result from the inability of the mesh to accurately represent the geometry of the object considered. It is therefore necessary for the user of ﬁnite element calculation software to create the mesh best adapted to the physical problem and geometry studied. Only geometric details superﬂuous to the physical model considered can be eliminated. Numerical errors related to the ﬁnite element method are mainly due to the integration diagram used to calculate the element quantities and to the method used to solve the linear system obtained. Numerical integration errors occurring when calculating element quantities are generally negligible when one does not depart from the traditional integration scheme (see Chapter 3). Numerical errors in the solution can easily be evaluated by calculating, after resolution, the residual: {Rl } = {Fl } − [Kll ] · {Tl } − [Kcl ]T · {Td } Strictly speaking, the norm of {Rl } must be equal to zero. In practice it must be very small compared with the values of the loads applied or the norm of the vector {rc } giving the heat ﬂuxes calculated at the nodes where the temperature was prescribed. Before beginning a discretization error calculation, it is possible to check a few element items, in particular by ensuring that the ﬂux density calculated at the domain → → − boundary satisﬁes accurately enough the ﬂux boundary conditions (i.e. φ · − n = −q on ∂Ωq ). It is difﬁcult to evaluate with accuracy the discretization error made. There are methods to do so, however. The ﬁnite elements presented in this book guarantee the temperature ﬁeld continuity on the whole domain of study Ω but not that of its derivatives (in particular the temperature gradient and heat ﬂux density). The methods to evaluate the approximation error are generally based upon the quantiﬁcation of these discontinuities (for example, the heat ﬂux density) on the element boundaries. Applied to thermal science, the a posteriori most widespread error evaluation method in computer codes [ZIE 87] consists of:

68

Finite Element Simulation of Heat Transfer

1) constructing a continuous heat ﬂux density ﬁeld on the whole domain from the densities calculated directly at the integration points; 2) quantifying the deviation between these continuous ﬁelds and the ﬁelds actually calculated. For example, it is possible to use the above method to construct a continuous heat ﬂux density ﬁeld. Then it is possible to deﬁne an error E e on each element as follows (V e is the volume occupied by Ωe and V that occupied by Ω): 7 8 1 < =− →=2 − 8 e e =→ φ − Φ = dv e V Ω 9 [2.21] E = =2 < =− 1 =→ Φ = dv V

Ω

Generally speaking, in order to reduce the discretization error, it will be necessary to ﬁnely mesh an area with high temperature gradient or heat ﬂux density variations. There are two main methods to increase accuracy in the solution obtained: – method h (the most common), which consists of increasing the number of nodes by reducing the mesh size in high gradient variation areas; – method p, which consists of increasing gradually the approximation degree while adding degrees of freedom to the problem. 2.4. Working example In this section, the objective is to illustrate the notions reviewed in this chapter on a 2D example. To do so, we consider the solid in Figure 2.5, inside which runs a heat convection exchanges

y

z

x

Figure 2.5. Working example

convection exchanges

The Finite Element Method

69

y a

Γ4 -1/ 2 n4 1/ 2

Γ2

Γ1 Γ3

n3 0 -1

a

x

Figure 2.6. Working example: 2D diagram

carrying ﬂuid at temperature Tint . The exchange coefﬁcient between the heat carrying ﬂuid and the solid is assumed to be high enough to consider the temperature to be constant and equal to Tint inside the solid. On its external wall, an exchange with the environment at temperature Text is represented by a convection ﬂux. The exchange coefﬁcient is written h and assumed to be constant. Now it is necessary to determine the temperature ﬁeld in the solid according to Tint , Text , h, and the heat conductivity λ of the material assumed to be isotropic and constant. The problem is represented by a 2D section (with unit thickness) of the solid along axis Oz. The problem symmetries make it possible to restrict the study to one-eighth of that section. Figure 2.6 illustrates this representation. The physical process to model is therefore the thermal equilibrium section. By taking the symmetric conditions into consideration, we obtain: 2 ∂2T ∂ T =0 in Ω + λ ∂x2 ∂y 2 on Γ1 T = Tint ∂T = h Text − T λ on Γ2 ∂x −−→ ∂T → =0 on Γ3 grad(T ) · − n3 = − ∂y √ −−→ ∂T 2 ∂T → − grad(T ) · n 4 = − − = 0 on Γ4 2 ∂x ∂y

70

Finite Element Simulation of Heat Transfer

y node 19 (global numbering)

3 2 19 1

local numbering of element 19 (‘‘lower’’ triangle) 3

1 6

2

local numbering of element 6 (‘‘upper’’ triangle)

x Figure 2.7. Working example: mesh

This problem will be addressed with the ﬁnite element method. After constructing the mesh, we will calculate the element quantities, make the assembly, apply the temperature boundary conditions (on Γ1 ) and solve the problem. 2.4.1. Finite element approximation 2.4.1.1. Mesh The mesh is constructed by means of ﬁrst order triangular elements (with three nodes). To do so, we use in this instance: – an angle division in p angle sector π/4p; – a division of the radiuses obtained into p segments of equal length. Thus, we obtain n = (p + 1)2 nodes and m = 2p2 triangular ﬁnite elements. This is described in Figure 2.7 in the case p = 4, i.e. 25 nodes and 32 elements. The local numbering of the elements is also described in this ﬁgure. If we now number globally the nodes “radius after radius”, then “from left to right”, a node will be characterized by an index I referring to the angular sector, varying from 1 to p + 1, and an index J referring to its position compared with the origin, varying from 1 to p + 1. It will always be possible to determine the global number i of a node (varying from 1 to

The Finite Element Method

71

n = (p + 1)2 ) by means of an angular index I and a distance index J in the form i = (I − 1)(p + 1) + J. The coordinates of node i will then be: J − 1 xi = R cos αI + a − R cos αI p J − 1 y = R sin α + a − R sin αI I i p

with

αI =

π (I − 1) 4p

If we number the elements in the same way, an element will be characterized by an index K referring to the angular sector, varying from 1 to p, and an index L referring to its position compared with the origin, varying from 1 to p. If we differentiate the “lower” triangles (with the odd index e) from the “upper” triangles (with the even index e), the connectivities are written as follows: – for the elements e = 2p(K − 1) + 2L − 1 (“lower” triangles), the connected nodes have numbers (K − 1)(p + 1) + L, K(p + 1) + L + 1 and K(p + 1) + L; – for the elements e = 2p(K − 1) + 2L (“upper” triangles), the connected nodes have numbers (K − 1)(p + 1) + L, (K − 1)(p + 1) + L + 1 and K(p + 1) + L + 1. This numbering method makes it possible to index the nodes in the mesh and construct for each element e the matrix [Ae ] locating it. For example, for element 19, in Figure 2.7, that matrix has 3 lines and 25 columns. Its only terms different from 19 19 zero are A19 1,12 = 1, A2,13 = 1 and A3,17 = 1. 2.4.1.2. Nodal approximation The nodal approximation will now be performed with the shape functions N e (x, y) satisfying the conditions mentioned in section 2.1. If i, j and k are the local numbers of the nodes making element e, then the triplet (i, j, k) is an even permutation of (1, 2, 3), and, for instance, Nie (x, y) must equal 1 at nodes i and 0 on the segment connecting nodes j and k. This is illustrated by Figure 2.8. From three points with coordinates (xi ,yi ), (xj ,yj ) and (xk ,yk ), it is possible to deﬁne a linear function equal to 1 at the ﬁrst node and 0 on the segment connecting the other two. That linear function is the shape function Nie of node i (local numbering) of element e. It is expressed as Nie (x, y) = aei x + bei y + cei , where the coefﬁcients aei , bei and cei , depend on the coordinates of the nodes forming element e in the following form: yj − y k aei = y x − y − x i j k j − xi − xj yk − yj e xk − xj bi = yi − yj xk − xj − xi − xj yk − yj xj yk − yj − yj xk − xj e ci = yi − yj xk − xj − xi − xj yk − yj

72

Finite Element Simulation of Heat Transfer

Nie(x,y) 1

y

node k

0 node i x

node j

Figure 2.8. Working example: shape function Nie (x, y)

2.4.2. Discrete problem formulation 2.4.2.1. Element quantities In order to estimate the element quantities, we select an element e. Its surface is termed S e . It is obtained from the coordinates of nodes i, j and k, belonging to the element by the following expression: Se =

1 xj − xi yk − yi − yj − yi xk − xi 2

The components of the element matrix [K e ] and of the load vector {F e } of element e are given by the general formulae previously obtained. In this working example, the zero ﬂux conditions (on Γ3 and Γ4 ) only yield zero terms in the load vector. Besides, as we considered a unit thickness for the representation in Figure 2.6, the volume and surface integrals become respectively surface and contour integrals. We ﬁnally obtain: ∂Nie ∂Nje ∂Nie ∂Nje e + ds + h Nie Nje dl Kij = λ S e ∂x ∂x ∂y ∂y ∂S e ∩Γ2 Nie dl Fie = hText ∂S e ∩Γ2

In this equation, note that the ﬁrst term of [K e ], related to the thermal conductivity, can be expressed directly for any element e in the form λS e (ai aj + bi bj ). On the contrary, the second term and the vector {F e }, which are the contributions of convection exchanges (coefﬁcient h), are more difﬁcult to determine.

The Finite Element Method

73

2

1

2

(-1)

(1)

1

ξ

change in variable

Figure 2.9. Working example: skin elements on Γ2

An intelligent means of estimating the terms due to convection exchanges (terms of e Kij and Fie involving an integration on S e ∩ Γ2 ) is to introduce an additional mesh on the boundary Γ2 on which these conditions are applied. That additional mesh is often referred to as a skin mesh. In this 2D example, it would be made of two-noded linear elements, with coordinates (x1 , y1 ) and (x2 , y2 ). Then by changing the variables, as described in Figure 2.9, it is possible to characterize the position of a point on this element with a parameter ξ varying from −1 to 1 as follows:

x(ξ) = N1 (ξ)x1 + N2 (ξ)x2 y(ξ) = N1 (ξ)y1 + N2 (ξ)y2

with

1−ξ N1 (ξ) = 2 N (ξ) = 1 + ξ 2 2

Now if we term Le the length of the skin element corresponding to the surface e2 e element e, we note that dl2 = dx2 + dy 2 = L4 dξ 2 , i.e. dl = L2 dξ, and it is possible to write the contribution of each skin element as follows: Le hLe 1 h Nie Nje dl = Ni (ξ)Nj (ξ)dξ 2 −1 0 Le hLe Text 1 Nie dl = Ni (ξ)dξ hText 2 −1 0 This makes it possible to fully deﬁne the element matrices and load vectors for each element e as follows: – if element e has no common edge with Γ2 , then: 2 a1 a2 + b1 b2 a1 + b21 e e [K ] = λS a1 a2 + b1 b2 a22 + b22 a1 a3 + b1 b3 a2 a3 + b2 b3 0 {F e } = 0 0

a1 a3 + b1 b3 a2 a3 + b2 b3 a23 + b23

74

Finite Element Simulation of Heat Transfer

– if element e has a common edge with Γ2 , then this edge is carried by the locally numbered nodes 2 and 3, and we obtain: 2 a1 a2 + b1 b2 a1 a3 + b1 b3 a1 + b21 hLe hLe 2 2 a + b + a a + b b + e e a1 a2 + b1 b2 2 3 2 3 2 2 [K ] = λS 3λS e 6λS e e hLe hL a1 a3 + b1 b3 a2 a3 + b2 b3 + a23 + b23 + e e 6λS 3λS 0 1 {F e } = hLe Text 2 1 2

2.4.2.2. Assembly Figure 2.10 illustrates the assembly operation for the calculation of the components K13,13 and K3,8 of the global matrix. The local node numbering method described in Figure 2.7 was used.

node 13

y

20

21

19 32

30

12

31 28 29

25 17

10 2

20

12 13 6

16 14

12

11 5

4

15

5

8

6

4 3

19

14

11

23

13

9 1

node 6

21

18 19

13

21

K13,13=K33+K33+K22+K22+K11+K11

22

20

26 27

24

14 12 13

3 node 3

7

x

4

5

K3,8=K23+K13

Figure 2.10. Working example: assembling the terms, K13,13 and K3,8 , of the global matrix

The Finite Element Method

75

× vanishing term eventually non-vanishing term × × ×

× × ×

×

× ×

×

×

×

× × × ×

× × ×

×

× × × × ×

×

× ×

[K] = ×

{F }=

×

×

×

× ×

×

× × ×

×

× ×

×

× × × half bandwidth

Figure 2.11. Working example: assembling the global matrix and load vector

By performing the assembly with p = 4 (as in Figure 2.7), we ﬁnally obtain a global matrix and load vector whose terms different from zero are given in Figure 2.11. In this ﬁgure, we can note that the half bandwidth is equal to 6. 2.4.3. Solution 2.4.3.1. Application of boundary conditions In order to apply the boundary conditions, we will only use the penalty method here. According to this method, a high constant P is added to the diagonal components of the matrix [K] corresponding to the nodes located on the boundary Γ1 . Likewise, a constant P Tint is added to the components of the load vector {F } referring to the same nodes. In the case p = 4 these nodes have numbers 1, 6, 11, 16 and 21. The application of the essential boundary conditions then involves the following modiﬁcations: K1,1 −→ K1,1 + P

F1 −→ F1 + P Tint

K6,6 −→ K6,6 + P

F6 −→ F6 + P Tint

K11,11 −→ K11,11 + P

F11 −→ F11 + P Tint

K16,16 −→ K16,16 + P

F16 −→ F16 + P Tint

K21,21 −→ K21,21 + P

F21 −→ F21 + P Tint

It may be noted that the global matrix [K] described in Figure 2.11 cannot be inverted. In fact, without any temperature boundary conditions, there is an inﬁnite

76

Finite Element Simulation of Heat Transfer Stationary state

Isovalue Temp Charge 1 Min = 30.011 Max = 100 20 30 40 50 60 70 80 90 100

SYSTUS + V 2.0 - SunOS-5.5.1 - Tue Jul 13 13:32:42 1999

y z

x

0

Figure 2.12. Drawing isotemperatures

Stationary state

Isovalue QX Charge 1 Min = 0.0717817 Max = 0.587643 0 0.1 0.2 0.3 0.4 0.5

SYSTUS + V 2.0 - SunOS-5.5.1 - Tue Jul 13 13:34:07 1999

y z

x

0

Figure 2.13. Drawing isoﬂux values in the direction x

number of solutions, all of them being separate from each other with a uniform temperature ﬁeld (T constant in Ω). In this example, the temperature boundary conditions are therefore necessary to obtain a single solution.

The Finite Element Method

77

2.4.3.2. Solution This section requires the creation of a short computer program incorporating the methods described in this chapter. This program is easy to write. Therefore, we invite the reader to do this, as an exercise, in order to check the temperature ﬁeld obtained. We have used existing software to perform this thermal analysis. Figures 2.12 and 2.13 have been made with the following data: – geometry: R = 2 mm, a = 6 mm; – thermal conductivity: λ = 0.03 W/mm/K; – boundary conditions: Tint = 100°C, Text = 20°C, H = 0.01 W/m2 /K. Figure 2.12 gives the temperature ﬁeld, while Figure 2.13 gives the thermal ﬂux in direction x.

This page intentionally left blank

Chapter 3

Isoparametric Finite Elements

3.1. DeÞ nitions In Chapter 2, we introduced the major principles of ﬁnite element approximation leading to an approximate solution of continuous problem [1.13] by solving an equation system. This approximate solution is deﬁned by a set of nodal values gathered in the vector {T }. The temperature at any point of the mesh is then obtained by interpolation inside each element (equation [2.2]). This section concerns the description of isoparametric elements which are by far the most widely used elements in most computer codes. In order to introduce them, the concept of a reference element is presented. This notion makes it possible to deﬁne any element by means of a simple geometric transformation. The isoparametric elements use a geometric transformation deﬁned by the element node location and the reference element’s shape functions. Finally there will be a presentation of the properties required by the interpolation functions so that the approximate solution obtained by the ﬁnite element method leads to the continuous solution when the mesh is reﬁned. 3.1.1. Reference element The construction of the shape functions of complex geometric elements is not always straightforward. For convenience, the notion of a reference element is used. A reference element is a very simple element, indexed in a reference space, which can be transformed into an actual element by means of a simple geometric transformation (Figure 3.1).

80

Finite Element Simulation of Heat Transfer

η

(x,y)

1 (ξ,η) -1

0

ξ

1

y

x

-1 reference element

actual element

Figure 3.1. Reference element and actual element

The geometric transformation of a reference element is deﬁned by means of the actual element’s geometric nodes. It is a priori appropriate to differentiate the geometric nodes from the interpolation nodes of the element. Indeed, the geometric nodes of an element characterize its shape and location in the mesh while the interpolation nodes carry the problem unknowns (the temperatures in the present instance). In practice, it is common to use a linear geometric transformation in relation to the coordinates of the geometric nodes. Thus, if element e includes pe geometric → → nodes with coordinates − x ei (with i = 1, . . . , pe ), then the position − x of any point → − of the element will be obtained from the position ξ of the corresponding point in the reference element as follows: pe

→ − → →e − − → x(ξ )= Gei ( ξ )− xi

[3.1]

i=1

The functions Gei are called geometric transformation functions. They take a scalar → − value and act on the coordinates ξ of a point on the reference element. The geometric transformation serves the following purposes: – to be bijective; which means that the Jacobian of the transformation must keep a constant sign and never be equal to zero inside an element; – match the geometric nodes of the reference and actual elements. This is true → − → − with the condition Gei ( ξ j ) = 1 if i = j and 0 if i = j, where ξ j represents the coordinates of the geometric node j in the reference element; – match the boundaries of the reference and actual elements.

Isoparametric Finite Elements

81

node 3

η 1 node 1 y node 2 0

1

x

ξ

reference element

actual element

Figure 3.2. “Linear triangle” element

The similarity of relationship [3.1] with relationship [2.2] is obvious. Relationship [3.1] can thus be considered to be a nodal approximation of the geometry. Now examine in detail three examples of elements constructed from a reference element. For simpliﬁcation purposes, we will examine the 2D case only. The reader will ﬁnd the characteristics of other elements at the end of this chapter. 3.1.1.1. Triangular element with linear transformation functions Figure 3.2 represents the case of a triangle element. Three geometric nodes are → − used. In this case, the coordinates of a point are written ξ = (ξ, η) in the reference → space and − x = (x, y) in the actual space. Consider the geometric transformation functions deﬁned by the three top nodes: Ge1 (ξ, η) = 1 − ξ − η Ge2 (ξ, η) = ξ Ge3 (ξ, η) = η Then the coordinates (x, y) of any point in the actual element can be obtained from the nodes 1 (x1 , y1 ), 2 (x2 , y2 ) and 3 (x3 , y3 ) in the form: x(ξ, η) = (1 − ξ − η)x1 + ξx2 + ηx3 y(ξ, η) = (1 − ξ − η)y1 + ξy2 + ηy3 In these expressions, ξ and η represent the coordinates of the corresponding point in the reference element (Figure 3.1). It is easy to see that this transformation

82

Finite Element Simulation of Heat Transfer

matches the nodes and boundaries of the reference and actual elements. Furthermore, the Jacobian J is written: ∂x ∂x ∂ξ ∂η x2 − x1 x3 − x1 = J = ∂y ∂y y2 − y1 y3 − y1 ∂ξ ∂η = x2 − x1 y3 − y1 − x3 − x1 y2 − y1 Note that the Jacobian transformation is constant. It is the component according to → → → → Oz of the cross product between − x2 −− x 1 and − x3 −− x 1 . Therefore, this Jacobian will not be equal to zero as long as the three geometric nodes of the triangle are not aligned. 3.1.1.2. Quadrangle element with linear transformation functions Figure 3.3 represents the case of a quadrangle using four geometric nodes. The geometric transformation functions selected are: 1 (1 − ξ)(1 − η) 4 1 Ge2 (ξ, η) = (1 + ξ)(1 − η) 4 1 Ge3 (ξ, η) = (1 + ξ)(1 + η) 4 1 Ge4 (ξ, η) = (1 − ξ)(1 + η) 4 Ge1 (ξ, η) =

Then the coordinates (x, y) of any point in the actual element are obtained from the nodes 1 (x1 , y1 ), 2 (x2 , y2 ), 3 (x3 , y3 ) and 4 (x4 , y4 ) in the form: x(ξ, η) =

y(ξ, η) =

1 1 (1 − ξ)(1 − η)x1 + (1 + ξ)(1 − η)x2 4 4 1 1 + (1 + ξ)(1 + η)x3 + (1 − ξ)(1 + η)x4 4 4 1 1 (1 − ξ)(1 − η)y1 + (1 + ξ)(1 − η)y2 4 4 1 1 + (1 + ξ)(1 + η)y3 + (1 − ξ)(1 + η)y4 4 4

Isoparametric Finite Elements

η

node 4

1

-1

0

83

node 3 y

ξ

1

node 2

node 1 x

-1 reference element

actual element

Figure 3.3. “Linear quadrangle” element

In these expressions, ξ and η represent the coordinates of the corresponding point in the reference element (Figure 3.1). Here again, note that this transformation matches the geometric nodes and sides of the reference and actual elements. Following on from the results obtained for the linear triangle, it is easy to demonstrate that, if Vi is the component according to Oz of the cross product → → → → between − xj −− x i and − xk −− x i , where nodes j and k are next to node i obtained by anticlockwise and clockwise rotation respectively in the reference element, then the Jacobian transformation is written as follows: J(ξ, η) =

4

Gei (ξ, η)Vi

i=1

Thus, given the deﬁnition of the geometric transformation functions, the Jacobian will keep a constant sign in the whole actual element if the Vi all have the same sign. In Figure 3.3, the geometric nodes of the reference and actual elements are numbered in the trigonometric sense. All the Vi are therefore positive since all the actual quadrangle angles are strictly between 0◦ and 180◦ . The corresponding Jacobian is therefore positive in all of the actual element. It is never equal to zero and the transformation is therefore bijective. Conversely, if one of the quadrangle angle values had been greater than 180◦ , then the corresponding term Vi would have been negative, and the Jacobian sign would have changed inside the actual element. This is illustrated by Figure 3.4. In this case the transformation is no longer bijective and the element cannot be used. The angles of a linear quadrangle must therefore range strictly from 0◦ to 180◦ .

84

Finite Element Simulation of Heat Transfer

correct case

incorrect case

Figure 3.4. Condition of non-singularity of the quadrangle element

3.1.1.3. Quadrangle element with quadratic transformation functions Figure 3.5 represents the case of a quadrangle using eight geometric nodes. The geometric transformation functions selected are: 1 Ge1 (ξ, η) = − (1 − ξ)(1 − η)(1 + ξ + η) 4 1 Ge2 (ξ, η) = (1 − ξ 2 )(1 − η) 2 1 Ge3 (ξ, η) = − (1 + ξ)(1 − η)(1 − ξ + η) 4 1 Ge4 (ξ, η) = (1 + ξ)(1 − η 2 ) 2 1 Ge5 (ξ, η) = − (1 + ξ)(1 + η)(1 − ξ − η) 4 1 Ge6 (ξ, η) = (1 − ξ 2 )(1 + η) 2 1 Ge7 (ξ, η) = − (1 − ξ)(1 + η)(1 + ξ − η) 4 1 e G8 (ξ, η) = (1 − ξ)(1 − η 2 ) 2 This geometric transformation does match the nodes and sides of the reference and actual elements. Indeed, consider for instance in Figure 3.5 the straight segment of the reference element formed by nodes 1, 2 and 3. Any point on this segment is deﬁned by −1 ≤ ξ ≤ 1 and η = −1. The only functions different from zero are Ge1 , Ge2 and Ge3 . Restricted to this segment the geometric transformation becomes: 1 1 x(ξ) = − ξ(1 − ξ)x1 + 1 − ξ 2 x2 + ξ(1 + ξ)x3 2 2 1 1 y(ξ) = − ξ(1 − ξ)y1 + 1 − ξ 2 y2 + ξ(1 + ξ)y3 2 2

Isoparametric Finite Elements

node7

η

node6 node5

1

node4

node8 -1

0

85

node1

ξ

1

y node3

node2 x

-1 reference element

actual element

Figure 3.5. Quadrangle element with quadratic transformation functions

In the equations, (x1 , y1 ), (x2 , y2 ) and (x3 , y3 ) are the coordinates of node numbers 1, 2 and 3 of the actual element. The straight segment of the reference element is transformed into a curve parameterized by ξ giving the coordinates (x, y) of the points of this curve in the actual element. 3.1.2. Isoparametric elements We have just examined the geometric construction of a ﬁnite element from a reference element. This construction is based upon the deﬁnition of a bijective transformation between the reference element and the actual element from the location of geometric nodes. This transformation is deﬁned by means of geometric transformation functions. Through this geometric transformation, it is now possible to substitute the approximation on the actual element for the approximation on the → − reference element. By calling Nie ( ξ ) the shape functions of the reference element (i varying from 1 to the number of interpolation nodes), the approximation inside the actual element is completely deﬁned by: → – the bijective geometric transformation connecting any point − x of the actual → − element to a single point ξ of the reference element as follows: pe

→ − → →e − − → x(ξ )= Gei ( ξ )− xi i=1

– the nodal approximation of the temperature in the reference element, which is written as follows: ne

− → → − T( ξ ) = Nie ( ξ )Tie i=1

86

Finite Element Simulation of Heat Transfer

This approach is obviously interesting. As the reference elements are geometrically very simple (triangle, square, cube unit, etc.), it is easy to determine the → − → − shape functions Nie ( ξ ) as well as the geometric transformation functions Gei ( ξ ). The elements most commonly used in computer codes use geometric nodes and geometric transformation functions respectively similar to the reference element interpolation nodes and shape functions. They are termed isoparametric elements. Other types of elements are sometimes used. Subparametric elements have fewer geometric nodes than interpolation nodes. On the contrary, superparametric elements have more geometric nodes than interpolation nodes. Later on in this chapter as well as in the whole book, we will deal with isoparametric elements only. In this case, the number of geometric and interpolation nodes in an element e are equal, pe = ne , and the geometric transformation functions → − → − of the reference element are such that Gei ( ξ ) = Nie ( ξ ). The nodal temperature approximation (equation [2.2]) can then be written for any mesh element Ωe as follows: e

− ∀→ x ∈ Ωe ,

− T (→ x)=

n

− − → x ( ξ ) Tie Nie →

i=1

− − → & x ( ξ ) · {T e } = Nie → %

→ − → − → x ( ξ )) can be formulated Nie ( ξ ). However, it is appropriate to Actually, the Nie (− bear in mind the fact that the interpolation functions on the reference element and on the actual element are different. They can be deduced from one another by changing → − → the variables between the coordinates − x and ξ . For instance, in the case of an eight-noded quadrangle, the interpolation functions Nie will be written for the reference element as follows: 1 N1e (ξ, η) = − (1 − ξ)(1 − η)(1 + ξ + η) 4 1 N2e (ξ, η) = 1 − ξ 2 (1 − η) 2 1 N3e (ξ, η) = − (1 + ξ)(1 − η)(1 − ξ + η) 4 1 N4e (ξ, η) = (1 + ξ) 1 − η 2 2 1 N5e (ξ, η) = − (1 + ξ)(1 + η)(1 − ξ − η) 4

Isoparametric Finite Elements N1(ξ,η)

N2(ξ,η)

1

1

η

η

1 0

-1

87

1 1

ξ

-1

0

-1

1

ξ

-1

midside node

corner node

Figure 3.6. Interpolation functions of a reference eight-noded quadrangle

1 1 − ξ 2 (1 + η) 2 1 N7e (ξ, η) = − (1 − ξ)(1 + η)(1 + ξ − η) 4 1 N8e (ξ, η) = (1 − ξ) 1 − η 2 2 N6e (ξ, η) =

These functions are represented in Figure 3.6 for both types of node present in the element. Thus, a corner node (node 1 for instance) is distinguished from an midside node located in the middle of a segment (node 2 for instance). Note that interpolation functions vanish for nodes different from the one they are attached to, and more generally, on all the sides of the element which does not include this node. It is important to note at this point that the location of midside nodes in a second order (or upper) isoparametric element may substantially modify the temperature approximation on the actual element. For instance, in the case of a 1D second-order element, the three nodes are deﬁned in the reference element by ξ equal to −1, 0 and 1. Given the shape functions of this element (see Table 3.3), the temperature approximation in the reference element is written in the form of a second degree polynomial in ξ: ∀ξ ∈ [−1, 1],

T (ξ) = T2 + ξ

T 3 − T1 T1 − 2T2 + T3 + ξ2 2 2

By choosing an index centered on the actual element and writing x1 = −a and x3 = a, the geometric transformation from the reference element to the actual element is written (Figure 3.7): ∀ξ ∈ [−1, 1],

x(ξ) = x2 (1 − ξ 2 ) + aξ

88

Finite Element Simulation of Heat Transfer 1

2

3

-a

0

a

1 x

(a) midside node in the middle of the segment [-a, a ]

2

-a -a/2 0

3 a

x

(b) midside node on the quarter of the segment [-a, a ]

Figure 3.7. Locating the midside node of a 1D three-noded element in the middle (a) or on the quarter (b) of the segment

Now examine the following situations: – the midside node (number 2) is located in the middle of the segment (Figure 3.7a), i.e. x2 = 0, – the midside node (number 2) is located on the quarter of the segment (Figure 3.7b), i.e. x2 = −a/2. In the ﬁrst instance (x2 = 0), the geometric transformation from the reference element to the actual element is deﬁned by x = ξa. Proportionality between x and ξ then leads to approximating the temperature on the actual element in the form of a second degree polynomial in x: ∀x ∈ [−a, a],

T (x) = T2 + x

T3 − T1 T1 − 2T2 + T3 + x2 2a 2a2

In the second instance, the geometric transformation becomes x = (ξ 2 + 2ξ − 1) a2 . The inversion of this relationship gives two possible solutions for ξ according > to x. We retain the formula with values of ξ going from −1 to 1, i.e. ξ = −1 + 2 + 2x/a. This expression leads to an approximation of the temperature on the actual element as follows: ? x+a ∀x ∈ [−a, a], T (x) = T1 + − 3T1 + 4T2 − T3 2a x + a + 2T1 − 4T2 + 2T3 2a > This expression now involves such terms as (x + a)/2a. This temperature approximation is now very different from that obtained when the midside node is located in the middle of the segment. In this case, the temperature gradient obtained by derivation with respect to x, and hence the heat ﬂux density, tends to inﬁnity as one gets near node 1 (x tending to −a).

Isoparametric Finite Elements

89

Therefore, it is necessary to be very careful when locating midside nodes on the sides of elements. Especially with second-order elements, we will try to locate the middle node on the perpendicular bisector of the segment connecting the corner nodes. Another location for the midside node must be restricted to very speciﬁc cases. For example, in linear elastic fracture mechanics, it is possible to reproduce the crack tip singular stress ﬁeld by using a mesh radiating from this tip and moving the midside nodes to the quarter of the segments connecting the corner nodes. 3.1.3. Interpolation function properties The ﬁnite element method makes it possible to obtain an approximate solution of problem [1.13] in the form of a vector {T } including the temperatures on the mesh nodes. Then the temperature at any point is obtained by nodal approximation. A few convergence conditions of the approximate solution to the continuous solution will be given hereafter. For the approximate solution to converge toward the solution of continuous problem [1.13] when the number of nodes increases, it is necessary for the nodal temperature approximation deﬁned by equation [2.2] to belong to the set ET of admissible solutions of the continuous problem, which means that: – inside each element, the temperature ﬁeld nodal approximation, hence the interpolation functions, have to be sufﬁciently regular. In our case, this approximation must be class at least C 1 , that is to say such that its ﬁrst derivatives are continuous; – on the boundary between two elements, the nodal temperature ﬁeld approximation must be continuous. The ﬁnite elements satisfying this condition are said to conform or compatible. Some elements, said to be incompatible, contravene this condition. The integral form of the problem then includes additional terms showing the function jumps on the element boundaries. The convergences of the approximate solution toward the continuous solution can only be ensured if these additional terms are bound and tend to zero when the element size itself tends to zero. This type of element is fairly often found in mechanics (some shell elements in particular), but very rarely in thermal science. Finally note that the continuous condition for compatible elements is valid for the temperature ﬁeld only, not for its gradient. As a general rule, the temperature gradient is discontinuous on the boundary between two elements; – as the mesh is reﬁned, the approximate temperature solution and its ﬁrst derivatives must tend toward constants (the values taken by the continuous solution). As in practice polynomial interpolation functions are used, this condition means that the approximate polynomial basis inside each element has to be complete at order 1,

90

Finite Element Simulation of Heat Transfer

that is to say it includes terms in x in a 1D problem, in x and y in 2D, and in x, y and z in 3D. In fact, using a complete polynomial basis of the order p + 1 ensures the convergence of the temperature ﬁeld derivatives up to the order p. The ﬁrst condition is easily satisﬁed by choosing polynomial interpolation functions while the second condition will be satisﬁed by choosing interpolation functions Nie equal to zero on any side or face of the reference element not including node i. The third condition forces the approximate polynomial basis to be complete at the order 1, which means that inside each element Ωe the nodal approximation → → → can be written at the ﬁrst order next to a point in the form T (− x ) = T0 + − a ·− x. → − e If this development is applied to the node with coordinates x i , we obtain → → Tie = T0 + − a ·− x ei . If in the nodal approximation the node temperatures Tie are substituted for the expression above, it is possible to write simultaneously: − → − → − → → a ·− x(ξ ) T → x ( ξ ) = T0 + − ne − → e − − → → T x(ξ ) = Ni ( ξ )Tie i=1 ne ne → − → →e − → − e = T0 Ni ( ξ ) + a · Nie ( ξ )− xi i=1 i=1 ne → − → − → → = T0 Nie ( ξ ) + − a ·− x(ξ ) i=1

Therefore, the third condition will be satisﬁed as soon as: e

n

→ − Nie ( ξ ) = 1

i=1

It is possible to mix several types of ﬁnite elements in the same analysis. However, these elements must be compatible to satisfy the temperature continuity condition. For instance, it is not advised to combine isoparametric elements whose faces or segments have different node numbers, as illustrated in Figure 3.8. 3.2. Calculation of element quantities The matrices and element vectors obtained by equations [2.6], [2.7] and [2.8] are written in the form of integrals over the volume Ωe or surface ∂Ωe of an actual

Isoparametric Finite Elements

correct association of linear elements

91

incorrect association of linear and quadratic elements

Figure 3.8. Mixing isoparametric elements

element e. In order to calculate these integrals, it is convenient to work on the reference element with the geometric transformation making it possible to move from the actual → − → coordinates − x to the reference coordinates ξ . For this purpose the Jacobian matrix of this transformation J is used; it is deﬁned from its tangent linear application to a given point: → − → − − → → → x =− x ( ξ ) =⇒ d− x = J · d ξ with

→ J = gradξ (− x)

In this equation, the operator gradξ () designates the gradient in the coordinate → − → system ξ . For example, in a Cartesian 3D case, written as follows: − x = (x, y, z) and → − ξ = (ξ, η, ζ), we obtain: ∂x ∂x ∂x ∂ξ ∂η ∂ζ ∂y ∂y ∂y J = ∂ξ ∂η ∂ζ ∂z ∂z ∂z ∂ξ ∂η ∂ζ Firstly, we will express the element quantities in the reference frame related to the element e under consideration. Then we will note that these quantities can be numerically evaluated in a very accurate way by means of the Gauss method. Finally we will illustrate this method on 1D, 2D and 3D reference elements. 3.2.1. Expression in the reference frame The fundamental quantity to calculate during a ﬁnite element analysis is the element residual {Re } (equation [2.6]). This quantity is the result of the integration on element Ωe or on its boundary ∂Ωe of functions involving: – the thermal conductivity matrix λ; – the heat production volumetric term Q;

92

Finite Element Simulation of Heat Transfer

– the ﬂux q entering through the boundary ∂Ωq ; – the geometric characteristics of the element through its interpolation functions −−→ Nie and their gradients grad(Nie ). In order to turn an integral on Ωe into an integral on the reference element, it is appropriate to transport these quantities by means of the (bijective) geometric → − → transformation connecting the coordinates − x of a point on Ωe to the coordinates ξ of the corresponding point in the reference element. → The transport of those quantities which are explicit functions of the location − x of → − e the point on Ω is easy to perform. It is sufﬁcient to replace x with its expression → − according to ξ (equation [3.1] applied to an isoparametric element). It is, however, more difﬁcult to transport the interpolation function gradients. It is necessary to use the Jacobian transformation matrix. In fact the derivation rule of composed functions → allows us to write, in the case of Cartesian coordinates, by designating − x = (x, y, z) → − and ξ = (ξ, η, ζ): ∂x

∂Nie ∂ξ

∂Nie ∂η !

−−→T gradξ (Nie )

∂Nie ∂ζ

= "

∂Nie ∂x

∂Nie ∂y !

−−→T grad (Nie )

∂Nie ∂z

∂ξ ∂y · ∂ξ " ∂z ∂ξ

∂x ∂η ∂y ∂η ∂z ∂η !

∂x ∂ζ ∂y ∂ζ ∂z ∂ζ

"

J

In general, the terms involved in the expression of element quantities and including the gradient operator can be written as follows: −−→ −→ gradT (Nie ) = − gradTξ (Nie ) · J −1 [3.2] −−→ −→ − grad(Nie ) = J −T · gradξ (Nie ) In this equation, J −1 is the inverse Jacobian matrix, and J −T is the transposed inverse Jacobian matrix. The last step for transporting element quantities toward the reference element is the transformation of the volume element dv. If dv0 designates the volume element in the reference element coordinate system, then we obtain: dv = J dv0 where J is the Jacobian matrix determinant, also called the Jacobian.

Isoparametric Finite Elements

93

With these results, the element residual {Re } can be expressed for each element e, in the form of integrals on the volume or surfaces of the reference element Ω0 corresponding to it. For instance, the linear case in the previous chapter, in which the element residual was expressed linearly according to the element node temperatures (equation [2.10]), corresponded to a heat conductance matrix λ and a volumetric term Q independent of the temperature, and a ﬂux q appears as follows: q = q0 + h(Text − T ). This residual becomes: e e {N } QJ dv0 + {N } q0 + hText Js ds0 Ω0 ∂Ω0q ! " {F e } ' %−−→ & −−→T e ( −1 e {R } = gradξ (N ) · J · λ · J −T · gradξ (N e ) J dv0 Ω0 − · {T e } e e + {N } h N Js ds0 0 ∂Ωq ! " [K e ]=[Λe ]+[H e ] In this expression, Ω0 represents the reference element corresponding to the actual element e, ∂Ω0q is the face of the reference element corresponding to that of the actual element on which the ﬂux q is applied, ds0 is the surface element corresponding to that face, and Js is the determinant of that face transport Jacobian matrix. Note therefore that the calculation of element quantities comes down to the < → − calculation of integrals of type Ω0 f ( ξ )dv0 . Generally speaking, these integrals cannot be calculated analytically. Numerical integration techniques are then used. 3.2.2. Gaussian quadrature Numerical integration consists of replacing an integral with a sum of contributions → − on a ﬁnite number ng of points with coordinates ξ k (with k = 1, . . . , ng ): Ω0

→ − → − f ( ξ )dv0 ≡ wk f ( ξ k ) ng

k=1

→ − In this equation, ξ k locates the position of the integration point number k, and wk is the weight of this point contribution to the integral value. The Gauss method (Gaussian quadrature) is different from other methods essentially because both the integration point coordinates and weights are determined

94

Finite Element Simulation of Heat Transfer

to obtain maximum accuracy. They are calculated in order to minimize the → − approximation error in the integration of a polynomial function p( ξ ) interpolating → − the function f ( ξ ) at those points. 3.2.2.1. 1D numerical integration In the 1D case, the reference element is always characterized by ξ varying from → − −1 to 1. The integration of a function f ( ξ ) on the reference volume Ω0 can then be written as: 1 → − I= f ( ξ ) dv0 = f (ξ) dξ Ω0

−1

The evaluation of the integral I depends on the calculation of an approximating integral I:

1

I=

p(ξ) dξ −1

where p(ξ) is a polynomial interpolating function f (ξ) on ng points. It is obvious that the approximation I depends directly on the number and location of interpolation points of f (ξ). Assume ﬁrst that the positions of the integration points are known, and write them ξ1 , . . . , ξng . The polynomial p(ξ) interpolating the function f (ξ) on these points can then be expressed in the form of a linear combination of Lagrange polynomials Lk (ξ), satisfying the condition Lk (ξj ) = δkj : p(ξ) =

ng

Lk (ξ)f (ξk )

k=1

Lk (ξ) =

(ξ − ξ1 ) · · · (ξ − ξk−1 )(ξ − ξk+1 ) · · · (ξ − ξng ) (ξk − ξ1 ) · · · (ξk − ξk−1 )(ξk − ξk+1 ) · · · (ξk − ξng )

If we now insert this polynomial expression into the approximating integral I, we obtain a formula in which the weights wk of the contribution of each integration point ξk to the integral I can be calculated analytically according to the locations of these points: I=

ng k=1

Lk (ξ)dξ f ξk −1 ! " 1

wk

Isoparametric Finite Elements

95

The a priori choice of the integration point locations ξk can be made in various ways. For instance, in the Newton-Cotes integration method, the points ξk are evenly distributed in the form ξk = −1 + 2 nk−1 on the segment [−1, 1]. It is demonstrated g −1 that this method, like all those a priori setting values ξk , allows us to integrate accurately any polynomial function f (ξ) of a degree less than or equal to ng − 1. In the Gauss method, the position ξk of the integration points is also optimized to ensure a more accurate estimation. In this case, we therefore obtain 2ng parameters to optimize (ng positions ξk and ng weights wk ), which makes it possible to integrate accurately with the same number of points any polynomial function f (ξ) of a degree less than or equal to 2ng − 1. For instance, a 3-degree polynomial will be accurately integrated with 2 points by the Gauss method, whereas 4 will be necessary with the Newton-Cotes method. This method makes it possible to obtain the same accuracy by substantially reducing the number of points where the function to integrate has to be estimated, which reduces the calculation time. In order to calculate the optimal positions of the integration points ξk , the expression of the interpolation polynomial p(ξ) is enhanced with ng additional terms in the form: ng ng Lk (ξ)f (ξk ) + αk ξ k−1 q(ξ) p(ξ) = k=1

k=1

ng @ ξ − ξ q(ξ) = l l=1

Note in this equation that p(ξ) is now a polynomial of degree 2ng − 1 with 2ng parameters (the ξk and the αk ). Moreover, this polynomial always coincides with function f (ξ) at the points ξk , since q(ξk ) = 0. Now the integral I approximating I is written as follows: I=

ng k=1

1

−1

ng Lk (ξ)dξ f ξk + k=1

1

−1

ξ k−1 q(ξ)dξ αk

Since the integration is to be exact for any polynomial of a degree less than or equal to 2ng − 1, the integral I must be independent of the coefﬁcients αk , which leads us to choose the ξk so that: ∀l = 1, . . . , ng ,

1

ξ l−1 q(ξ)dξ = 0 −1

96

Finite Element Simulation of Heat Transfer

For instance, if we decide to choose two integration points ξ1 and ξ2 , then the location of these points will be determined with q(ξ) = (ξ − ξ1 )(ξ − ξ2 ): 1 ξ − ξ1 ξ − ξ2 dξ = 0 −1 1 ξ ξ − ξ1 ξ − ξ2 dξ = 0 −1

√ 3 ξ1 = − 3 =⇒ √ ξ2 = 3 3

2 3 + 2ξ1 ξ2 = 0 =⇒ 2 ξ1 + ξ2 = 0 3

The calculation of weights w1 and w2 corresponding to the integration points then generates: 1 1 ξ − ξ2 L1 (ξ)dξ = dξ = 1 w1 = ξ −1 −1 1 − ξ2 1 1 ξ − ξ1 w2 = L2 (ξ)dξ = dξ = 1 ξ −1 −1 2 − ξ1 More generally speaking, Table 3.1 gives the integration point locations and corresponding weights for the Gaussian quadratures mostly used in computer codes. ng

positions

ξk

wk

1

ξ

0.0

2.0

2

ξ

±0.57735 02691 89626

1.0

3

ξ

-0.77459 66692 41483 0.0 +0.77459 66692 41483

0.55555 55555 55556 0.88888 88888 88889 0.55555 55555 55556

4

0.34785 48451 37454

ξ

±0.86113 63115 94053

±0.33998 10435 84856

0.65214 51548 62546

Table 3.1. 1D Gauss integration points and weights

The number of integration points in an element depends on the complexity of the functions to integrate on that element for the calculation of element quantities. For instance, for a 1D two-noded element, the interpolation functions are linear (see Table 3.3) and a single integration point will be sufﬁcient to accurately integrate the element quantities (as well as their gradients, of course). However, the element quantities

97

temperature (°C)

Isoparametric Finite Elements

x (mm)

Figure 3.9. Gauss method induced error

require other functions such as the volumetric heat source Q. If these functions do not vary linearly according to the space coordinates, then this method will introduce an error in the estimation of these element quantities. Figure 3.9 illustrates this error in the working example of Chapter 1, with a discretization on ﬁve ﬁnite elements. 3.2.2.2. 2D and 3D numerical integration Extending the Gauss integration method to 2D or 3D elements is relatively easy with square or cubic geometries. In such a case, the integration directions are independent and it is possible to write:

1 −1

1

−1

1

−1

nξ

1

−1

f (ξ, η) dξ dη ≈

−1

f (ξ, η, ζ) dξ dη dζ ≈

wkξ wlη f ξk , ηl

k=1 l=1 nξ

1

nη

g g

nη

nζ

g g g

ζ wkξ wlη wm f ξk , ηl , ζm

k=1 l=1 m=1

The point numbers (nξg , nηg , nζg ), their coordinates (ξk , ηl , ζm ) and their weights ζ ) are obtained from the results of section 3.2.2.1 applied in the (wkξ , wlη , wm corresponding direction.

98

Finite Element Simulation of Heat Transfer ng

ξk

positions

1 η

ξ

η

7

ξ

w1 =0.5

3

wk

ξ1 =0.33333 33333 33333 η1 = ξ1

ξ

ηk

η

ξ1 =0.16666 66666 667 ξ2 =0.66666 66666 667 ξ3 = ξ 1

η1 = ξ1 η2 = ξ1 η3 = ξ2

w1 =0.16666 66666 667 w2 = w1 w3 = w 1

ξ1 =0.10128 65073 235 ξ2 =0.79742 69853 531 ξ3 = ξ 1 ξ4 =0.47014 20641 051 ξ5 = ξ 4 ξ6 =0.05971 58717 898 ξ7 =0.33333 33333 333

η1 η2 η3 η4 η5 η6 η7

w1 =0.06296 95902 724 w2 = w1 w3 = w 1 w4 =0.066197 07639 425 w5 = w 4 w6 = w 4 w7 =0.1125

= ξ1 = ξ1 = ξ2 = ξ6 = ξ4 = ξ4 = ξ7

Table 3.2. Gauss integration schemes for triangles Type

Geometry

Linear element with two nodes

Quadratic element with three nodes

Interpolation functions 1−ξ 2 1+ξ N2 (ξ) = 2 1 N1 (ξ) = ξ(ξ − 1) 2 N2 (ξ) = 1 − ξ 2 N1 (ξ) =

ξ

ξ

1 ξ(ξ + 1) 2 1 9 (ξ − 1) − ξ2 16 9 27 2 1 (ξ − 1) ξ − 16 3 27 1 2 (1 − ξ ) ξ + 16 3 9 1 2 (ξ + 1) ξ − 16 9

N3 (ξ) = N1 (ξ) =

Cubic element with four nodes

N2 (ξ) =

ξ

N3 (ξ) = N4 (ξ) =

Table 3.3. Most common 1D ﬁnite elements

For other types of elements (triangle, tetrahedron, wedge, etc.), it is impossible to act in the same way. A great deal of research work has enabled us to determine several numerical schemes, the most common of which appear in Table 3.2 for a triangle. These schemes can then be combined to construct new schemes adapted to tetrahedra and wedges (Table 3.5).

Isoparametric Finite Elements Type

Geometry

Interpolation functions

η

N1 (ξ, η) = 1 − ξ − η

Linear element with three nodes (triangle)

99

N2 (ξ, η) = ξ

N3 (ξ, η) = η

ξ

N1 (ξ, η) = −α(1 − 2α) N2 (ξ, η) = 4ξα

η

Quadratic element with six nodes (triangle)

N3 (ξ, η) = −ξ(1 − 2ξ)

N4 (ξ, η) = 4ξη

N5 (ξ, η) = −η(1 − 2η)

ξ

N6 (ξ, η) = 4ηα

η

Linear element with four nodes (quadrangle)

ξ

η

Quadratic element with eight nodes (quadrangle)

ξ

(with α = 1 − ξ − η) 1 N1 (ξ, η) = (1 − ξ)(1 − η) 4 1 N2 (ξ, η) = (1 + ξ)(1 − η) 4 1 N3 (ξ, η) = (1 + ξ)(1 + η) 4 1 N4 (ξ, η) = (1 − ξ)(1 + η) 4 1 N1 (ξ, η) = − (1 − ξ)(1 − η)(1 + ξ + η) 4 1 N2 (ξ, η) = (1 − ξ 2 )(1 − η) 2 1 N3 (ξ, η) = − (1 + ξ)(1 − η)(1 − ξ + η) 4 1 N4 (ξ, η) = (1 + ξ)(1 − η 2 ) 2 1 N5 (ξ, η) = − (1 + ξ)(1 + η)(1 − ξ − η) 4 1 N6 (ξ, η) = (1 − ξ 2 )(1 + η) 2 1 N7 (ξ, η) = − (1 − ξ)(1 + η)(1 + ξ − η) 4 1 N8 (ξ, η) = (1 − ξ)(1 − η 2 ) 2

Table 3.4. Most common 2D ﬁnite elements

3.3. Some Þ nite elements In this section, we will deal only with the most commonly used reference elements, for which the geometry, local node numbering and interpolation functions are speciﬁed. The 1D elements are in Table 3.3. The four-noded cubic element is hardly used. The 2D elements are in Table 3.4. Some of them are widely used in this chapter (the three-noded triangle for instance). Finally, Table 3.5 includes a few 3D elements. There are others, such as the 15-noded wedge or the 20-noded cube (quadratic

100

Finite Element Simulation of Heat Transfer

interpolation). There are also elements mixing a quadratic interpolation in one or two directions with a linear interpolation in the remaining directions. The interpolation functions are then obtained by applying the relationships presented in this chapter. Type

Geometry ζ

Linear element with four nodes (tetrahedron)

η

ξ

ζ

Quadratic element with ten nodes (tetrahedron)

ζ

Linear element with six nodes (wedge)

η

ξ

η

ξ

ζ

Linear element with eight nodes (hexahedron)

η

ξ

Interpolation functions N1 (ξ, η, ζ) = 1 − ξ − η − ζ N2 (ξ, η, ζ) = ξ N3 (ξ, η, ζ) = η N3 (ξ, η, ζ) = ζ N1 (ξ, η, ζ) = −α(1 − 2α) N2 (ξ, η, ζ) = 4ξα N3 (ξ, η, ζ) = −ξ(1 − 2ξ) N4 (ξ, η, ζ) = 4ξη N5 (ξ, η, ζ) = −η(1 − 2η) N6 (ξ, η, ζ) = 4ηα N7 (ξ, η, ζ) = 4ζα N8 (ξ, η, ζ) = 4ξζ N9 (ξ, η, ζ) = 4ηζ N10 (ξ, η, ζ) = −ζ(1 − 2ζ) (with α = 1 − ξ − η − ζ) 1 N1 (ξ, η, ζ) = (1 − ξ − η)(1 − ζ) 2 1 N2 (ξ, η, ζ) = ξ(1 − ζ) 2 1 N3 (ξ, η, ζ) = η(1 − ζ) 2 1 N4 (ξ, η, ζ) = (1 − ξ − η)(1 + ζ) 2 1 N5 (ξ, η, ζ) = ξ(1 + ζ) 2 1 N6 (ξ, η, ζ) = η(1 + ζ) 2 1 N1 (ξ, η, ζ) = (1 − ξ)(1 − η)(1 − ζ) 8 1 N2 (ξ, η, ζ) = (1 + ξ)(1 − η)(1 − ζ) 8 1 N3 (ξ, η, ζ) = (1 + ξ)(1 + η)(1 − ζ) 8 1 N4 (ξ, η, ζ) = (1 − ξ)(1 + η)(1 − ζ) 8 1 N5 (ξ, η, ζ) = (1 − ξ)(1 − η)(1 + ζ) 8 1 N6 (ξ, η, ζ) = (1 + ξ)(1 − η)(1 + ζ) 8 1 N7 (ξ, η, ζ) = (1 + ξ)(1 + η)(1 + ζ) 8 1 N8 (ξ, η, ζ) = (1 − ξ)(1 + η)(1 + ζ) 8

Table 3.5. Most common 3D ﬁnite elements

PART 2

Transient State, Non-linearities, Transport Phenomena

This page intentionally left blank

Introduction

Part 1 of this book dealt with the detailed description of the ﬁnite element method using, to this end, the simple case of steady state conduction. This allowed us to introduce the basic physical models used in thermal science (thermal balance, Fourier law, etc.), then show how the ﬁnite element method can be used to solve a particular problem. In Part 2, complementary notions will be introduced and their consequences for the ﬁnite element method will be shown. These concepts are: – the transient state; – non-linearities; – transport phenomena. The study of transient heat conduction requires consideration of the thermal capacitance of materials. Indeed, a solid undergoing a thermal load does not instantly reach a steady state. The temperature in the solid varies over time. In order to obtain this variation, the ﬁnite element method must be coupled with a time integration scheme. The behavior of a material in the steady or transient state is rarely linear. As a matter of fact, its thermal conductivity is often temperature-dependent. Moreover, the boundary conditions and the loads applied to a solid are often non-linear. For example, the heat exchange coefﬁcient between a solid and an external medium often depends on the local temperature of the solid surface. The traditional methods of solving non-linear thermal problems in the steady or transient state will be described. Transport phenomena are present when the solid is no longer assumed to be motionless as a result of the thermal loads it is subjected to. Such is the case when

104

Finite Element Simulation of Heat Transfer

a welding operation is simulated, the heat source moving along the weld seal. A solution in a reference frame related to that source will have to take into consideration the relative movement of the solid. Transport phenomena also introduce numerical instabilities and the remedial methods for this will be described.

Chapter 4

Transient Heat Conduction

4.1. Problem formulation Transient heat conduction involves time. In order to apply the ﬁnite element method to the numerical simulation of this phenomenon, we use a course of action similar to that introduced in Chapters 1 and 2. This method leads us to a time differential system for the temperatures at the mesh nodes. 4.1.1. The continuous problem If we perform the energy balance as in [RAP 98] for instance, the left-hand term of −−→ the steady state equilibrium equation [1.4], i.e. div(λ· grad(T ))+Q, appears to be the heat brought to a volume element of the solid Ω. The ﬁrst principle of thermodynamics speciﬁes that this heat produces an internal energy change on the domain over time. That internal energy change, equal to zero in the steady state, can be expressed here ˙ where ρ is the material volumetric mass and H˙ the variation over time of its as ρH, speciﬁc (i.e. per mass unit) enthalpy H. Equilibrium equation [1.4] thus becomes: −−→ div λ · grad(T ) + Q = ρH˙

at any point of Ω

[4.1]

In this case, the boundary conditions applied to the solid are similar to those mentioned in the steady state, i.e: – an input ﬂux q known on a portion ∂Ωq of Ω boundary; – a known temperature Td on the other portion ∂ΩT of that boundary.

106

Finite Element Simulation of Heat Transfer

However, in transient states, these boundary conditions may change over time, as well as the source term Q. It follows from this that the boundary value problem of transient heat conduction is expressed as: → → Find T (− x , t) at any point − x ∈ Ω verifying: at t = 0: − → − T (→ x , 0) = T0 ( x ) at any instant t ∈ ]0, tf [: −−→ ˙ R(T, t) = div λ · grad(T ) + Q(T, t) − ρH(T, t) = 0 in Ω −−→

→ λ · grad(T ) · − n = q on ∂Ωq with ∂Ωq ∪ ∂ΩT = ∂Ω T = Td on ∂ΩT

[4.2]

The term R(T, t) present in this problem is called residual, as in the steady state. It is a quantity to cancel at any point of the solid Ω, and at any instant of the study, i.e. between instants t = 0 and t = tf , while satisfying at these instants the boundary conditions illustrating the exchanges of the solid with the external environment. These boundary conditions are applied on the boundary ∂Ω of the solid Ω, and we must check that ∂Ωq ∪ ∂ΩT = ∂Ω. Conversely, in this boundary value problem an initial condition appears, the initial → temperature ﬁeld T0 (− x ) in the solid, as well as the enthalpy derivative with respect to time. We could therefore think of applying a time ﬁnite element approximation by using the initial conditions as boundary conditions. However, time variations are treated differently from space variations in most computer codes. The ﬁnite element approximation is only used for space variations, time integration requiring a ﬁnite difference technique like the one described in section 4.2. If a method similar to that of Chapter 1 is applied to space variations, the following weak integral formulation is obtained, where the spaces ET and Eψ are those deﬁned by equations [1.11] and [1.12]: → Find T : t ∈ [0, tf [ −→ T (− x , t) ∈ ET such that: at t = 0: − → − T (→ x , 0) = T0 ( x ) for t ∈ ]0, tf [ and for any ψ ∈ Eψ : −−→ −−→T ˙ ψ(Q − ρH)dv + ψq ds − grad (ψ) · λ · grad(T ) dv = 0 Ω

∂Ωq

Ω

[4.3]

Transient Heat Conduction

107

The Galerkin method constructs at any instant a temperature approximation by reasoning on n-dimensional spaces ETn and Eψn . Continuous problem [4.3] is then approximated by the following discrete problem (of dimension n): → Find T : t ∈ [0, tf [ −→ T (− x , t) ∈ ETn such that: at t = 0: T (− → − → x , 0) = T0 ( x ) for t ∈ ]0, tf [ and for any ψ ∈ Eψn : −−→ −−→T ˙ ψ(Q − ρ H)dv + ψq ds − grad (ψ) · λ · grad(T ) dv = 0 Ω

[4.4]

Ω

∂Ωq

4.1.2. Finite element approximation → In discrete problem [4.4], the temperature ﬁeld T (− x , t) is estimated inside each element e of a mesh, and at any instant t, from the values it takes at the nodes corresponding to that element, at that same instant, while the weighting functions → ψ(− x ) are estimated as in the steady state: e

→ ∀− x ∈ Ωe , ∀t ∈ [0, tf [,

− T (→ x , t) =

n

→ Nie (− x )Tie (t)

i=1

e T1 (t) → → ··· = N1e (− x ), . . . , Nnee (− x ) · e Tne (t) → x ) · {T e (t)} = N e (−

[4.5]

e

− ∀→ x ∈ Ωe , ∀t ∈ [0, tf [,

− ψ(→ x)=

n

− ψ e Nie (→ x)

i=1

e − → N1 ( x ) .. = ψ1e , . . . , ψne e · . e − Nne (→ x) → x )} = ψ e · {N e (−

[4.6]

In equations [4.5] and [4.6], Tie (t) and ψie respectively represent the temperature at instant t, and the weighting function value at the node number i (local numbering) of → element e. The shape functions introduced, Nie (− x ), are estimated similarly to those described in the previous chapters.

108

Finite Element Simulation of Heat Transfer

→ If nodal approximation [4.6] is applied to the weighting functions ψ(− x ) in the discrete problem [4.4], the quantities involved in that problem can be written as follows: m ˙ ˙ ψ(Q − ρH)dv = ψ e · {N e } (Q − ρH)dv Ω

ψq ds = ∂Ωq

Ωe

e=1 m

ψ e ·

e=1

−−→ −−→T ψ e · grad (ψ) · λ · grad(T ) dv = m

Ω

e=1

∂Ωe ∩∂Ωq

Ωe

{N e } q ds

'−−→ ( −−→ gradT (N e ) · λ · grad(T ) dv

If these expressions are transferred into equation [4.4], the quantity to cancel becomes: ˙ ψ(Q − ρH)dv + ψq ds m Ω ∂Ωq ψ e · {Re (T, t)} = − − → − − → e=1 − gradT (ψ) · λ · grad(T ) dv Ω

where the vector {Re (T, t)}, speciﬁc to each element e, is called the element residual, − and is written according to the temperature ﬁeld T (→ x , t) and time t as: ˙ {N e } (Q − ρH)dv + {N e } q ds e e Ω ∂Ω ∩∂Ωq {Re (T, t)} = [4.7] '−−→ ( −−→ T e − grad (N ) · λ · grad(T ) dv Ωe

In order to estimate the element residual, it is necessary to express the variation rate ˙ according to temperature and time. In many instances, the of the material enthalpy, H, material enthalpy depends on temperature only. The material speciﬁc heat C = dH dT can then be introduced to write ρH˙ = ρC T˙ . Term T˙ can then be approximated by a nodal approximation (equation [4.5]) as: e

− ∀→ x ∈ Ωe , ∀t ∈ [0, tf [,

→ T˙ (− x , t) =

n i=1

→ Nie (− x )T˙ie

' ( − = N e (→ x ) · T˙e

[4.8]

Transient Heat Conduction

109

In this equation, the vector {T˙ e } includes in a column the temperature change rates at the nodes of element e. If these expressions are inserted into the element residual (equation [4.7]), we obtain: {N e } Q dv + {N e } q ds Ωe e ∂Ω ∩∂Ωq ' ( {Re (T, t)} = − ρC {N e } N e dv · T˙e Ωe ' −−→ −−→T e ( − grad (N ) · λ · grad(T ) dv Ωe

Note that this expression differs from the steady state (equation [2.6] in Chapter 2) only by adding a term representing the quantity ρH˙ into the thermal balance (equation [4.1]). That term produces the deﬁnition of a new element matrix, called the capacitance matrix, heat capacity matrix or element mass matrix. It is a symmetric matrix which is written as: [C e ] = ρC {N e } N e dv Ωe

Further on in this chapter, as in Part 1, we will restrict our study to the linear case to fully express the element residual and its assembly. Non-linearities will be addressed in the next chapter. 4.1.3. Linear case If the results of Part 1 are used, we see that the element residual will be expressed linearly according to the problem unknowns if: – the material heat conductivity λ does not depend on the temperature; – the term Q, the heat volumetric source, is independent of the temperature (in fact, as in the steady state, it could depend on it linearly, but as far as we know, this corresponds to no physical instance); – the heat ﬂux q applied to the solid boundary ∂Ωq is a linear function of the temperature (i.e. q = h(Text − T ) + q0 , where h and q0 are independent of the temperature); – the material volumetric mass ρ and speciﬁc heat C are independent of the temperature. This element residual will then be expressed as: ' ( {Re (T )} = {F e } − [C e ] · T˙e − [K e ] · {T e }

110

Finite Element Simulation of Heat Transfer

The element terms: load vector {F e }, capacitance matrix [C e ] and conductance matrix [K e ] are written as follows: e {N e } Qdv + {N e } (q0 + hText )ds [4.9] {F e } = Ω

[C e ] =

Ωe

∂Ωe ∩∂Ωq

ρC {N e } N e dv

[4.10]

' %−−→ & −−→T e ( e e ] = (N ) · λ · ) dv [Λ grad grad(N Ωe e e e [4.11] [K ] = [Λ ]+[H ] with: [H e ] = {N e } h N e ds ∂Ωe ∩∂Ωq

Finally it is possible to assemble the element quantities as in the steady case, which allows the global problem of the following nodal residual to cancel: ' ( {R} = {F } − [C] · T˙ − [K] · {T } [4.12] In this equation, the capacitance and conductance global matrices [C] and [K], as well as the load vector {F }, are obtained by traditional assembly operations: m {F } = [Ae ]T · {F e } e=1 m [C] = [Ae ]T · [C e ] · [Ae ] [4.13] e=1 m T [Ae ] · [K e ] · [Ae ] [K] = e=1 Solving a transient heat conduction problem leads us to solve a ﬁrst order differential system with respect to time: ' ( [C] · T˙ + [K] · {T } = {F } [4.14] It is appropriate to note at this point that this differential equation does not take account of the essential boundary conditions (known temperature), which are to be inserted as the solution is in progress. Inserting these boundary conditions can be done in the same way as in the steady case. We will not go into the details here.

Transient Heat Conduction

111

4.2. Time integration The integration of differential system [4.14] gives the time evolution of the temperatures at the mesh nodes and, by means of a nodal approximation, the temperature and its gradient at any point of the solid. To do so, it is simple to use a time-stepping solution method for [4.14] by means, for example, of an implicit ﬁnite difference scheme. The temperature change rate at the mesh nodes is written as: ' ( {T (t + ∆t)} − {T (t)} T˙ (t + ∆t) = ∆t The column vector {F } and the matrices [C] and [K] being known at instant t + ∆t, as well as the nodal temperatures at instant t (i.e. vector {T (t)}), the nodal temperatures at instant t + ∆t are obtained by solving differential system [4.14] at t + ∆t. This yields:

[C] + ∆t [K] · {T (t + ∆t)} = ∆t {F } + [C] · {T (t)}

Finally, when using such a method, we start with an initial condition at t = 0 (i.e. temperature {T (0)}), then estimate the solution at each instant by consecutive time steps ∆t. This is a space approximate solution (by the ﬁnite element method) and a time approximate solution (by the ﬁnite difference method). It should be noted that the time step ∆t may change during the analysis process. We will now describe other numerical methods to integrate this system over time. 4.2.1. Modal method The modal method seeks a solution to the transient heat conduction problem as follows: − T (→ x , t) =

n

→ βk (t)τk (− x)

[4.15]

k=1

→ where the n functions τk (− x ) form a basis of the space of admissible solutions (ETn ). Applying the ﬁnite element method then leads us to look for a solution to the differential system of the form: β1 (t) .. {T (t)} = βk (t) {τk } = {τ1 } , . . . , {τn } · . k=1 βn (t) n

[4.16]

112

Finite Element Simulation of Heat Transfer

where the n vectors {τk } are a basis of Rn , called the modal basis. The modal method determines a modal basis then expresses the system to solve on that basis to obtain the solution in that basis. 4.2.1.1. Determining the modal basis To determine the modal basis, differential system [4.14] with no load vector is considered: ' ( [C] · T˙ + [K] · {T } = 0 [4.17] We are then tempted to seek solutions of the form {T (t)} = {τ } e−αt , so that this system becomes:

[K] − α [C] · {τ } = 0

[4.18]

Matrix [C], being symmetric positive deﬁnite, can be decomposed as [C] = [L] · [L] , where [L] is a lower triangular matrix. It would have been possible to choose to decompose matrix [K] as well. Choosing [C] here is preferable since [K] is not T necessarily deﬁned. This decomposition produces the system ([K] − α [L] · [L] ) · −1 {τ } = 0 which, multiplied on the left by [L] , yields: T

[L]

−1

· [K] · [L]

−T

T − α [I] [L] · {τ } = 0

−T

In this equation, [L] represents the transpose of the inverse of matrix [L]. We therefore go back to the solution of an eigenvalue problem whose general form is written as: −1 −T [A] = [L] · [K] · [L] [A] − α [I] · {Ψ} = 0 with [I] : identity matrix [4.19] T {Ψ} = [L] · {τ } Matrix [A] being symmetric, this problem admits n real eigenvalues α1 , . . . , αk , . . . , αn . To each eigenvalue αk corresponds an eigenvector {Ψk } which can be T written as {Ψk } = [L] · {τk }. If we reverse the process which led us from system [4.18] to system [4.19], it is easy to demonstrate that the vectors {τk } satisfy the following condition: [K] · {τk } = αk [C] · {τk }

[4.20]

Transient Heat Conduction

113

Matrix [C] being symmetric positive deﬁnite and matrix [K] symmetric but only positive semi-deﬁnite, we can write: τk · [K] · {τk } = αk τk · [C] · {τk } ! " ! " ≥0

>0

We deduce that the n real eigenvalues αk are positive or equal to zero: 0 ≤ α1 ≤ α2 ≤ · · · ≤ αn To deﬁne the eigenvectors in a single way, it is necessary to norm them. Several norms are able to do so; the most traditional are: – The inﬁnite norm: Maxi=1···n (τki ). – The [C]-norm: τk · [C] · {τk }. Now consider 2 vectors {τj } and {τk }, corresponding to 2 different eigenvalues αj = αk . It is then possible to determine equation [4.20] for each value and derive: τk · [K] · {τj } = αj τk · [C] · {τj } τj · [K] · {τk } = αk τj · [C] · {τk } Given the symmetry of the matrices [K] and [C], we can write: τk · [K] · {τj } = τj · [K] · {τk } τk · [C] · {τj } = τj · [C] · {τk } From the above equations, we can derive: αj − αk τk · [C] · {τj } = 0 then: τk · [K] · {τj } = τk · [C] · {τj } = 0

[4.21]

In the multiple mode case, i.e. if the eigenspace corresponding to an eigenvalue αk is of a dimension strictly greater than 1, we can construct an orthogonal basis of that space in the sense of the scalar product L({τj } , {τk }) = τj · [C] · {τk }. It is obvious that equation [4.21] remains valid.

114

Finite Element Simulation of Heat Transfer

Equation [4.21] thus shows that the {τk } for k = 1 . . . n form an orthonormal basis of Rn called a modal basis. We write: cj = τj · [C] · {τj } kj = τj · [K] · {τj }

[4.22]

As in mechanics, we can call cj and kj respectively generalized capacitance and k conductivity. Note that kj and cj verify αj = cjj . The description of the calculation methods of eigenvalues and vectors is not dealt with in this book. It is possible for interested readers to consult [ZIE 91] or [BAT 96]. 4.2.1.2. Projection on the modal basis We therefore seek a solution of the differential system in the form given by [4.16], where ({τk })k=1...n is the modal basis. By projecting the differential system onto each vector {τj }, we obtain: τj · [C] ·

n

β˙ k (t) {τk } + τj · [K]

k=1

·

n

βk (t) {τk } = τj · {F (t)}

k=1

Given [4.21] and [4.22], we then obtain n ﬁrst order differential equations, each now including only one unknown βj : cj β˙ j (t) + kj βj (t) = fj (t)

with fj (t) = τj · {F (t)}

[4.23]

To integrate equation [4.23] over time, we can use one of the direct integration algorithms presented hereafter. It is equally possible to use the analytical solution of equation [4.23] which is written: −αj (t−t0 )

βj (t) = e

1 t αj (s−t0 ) βj t0 + e fj (s)ds cj t0

where βj (t0 ) is the initial solution obtained from the initial temperature by: βj (t0 ) = τj {T (t0 )}

[4.24]

Transient Heat Conduction

115

Most of the time, the right-hand side of [4.24] cannot be calculated analytically. However, it is possible to calculate a time-stepping solution by assuming that, on each time step, fj (t) of [4.23] varies linearly. For each time step, we obtain: e−αj ∆t − 1 fj (t + ∆t) 1+ + βj (t + ∆t) = βj (t)e kj αj ∆t −αj ∆t −1 fj (t) −αj ∆t e e − + kj αj ∆t −αj ∆t

In practice, it is unnecessary to consider a modal basis including all the problem eigenmodes. Only the n ﬁrst modes corresponding to the lowest eigenvalues and allowing us to give an accurate representation of the temperature are necessary. The number of differential equations [4.23] is consequently reduced. The use of this method, also designated modal superposition method, requires on the one hand the calculation of the n ﬁrst eigenmodes and eigenvalues and on the other hand the integration of n uncoupled differential equations [4.23]. This method is widely used in mechanics, but fairly rarely in thermal science. 4.2.2. Direct time integration We consider the following ﬁrst order differential equation which we will try to integrate over time: u(t) ˙ = g(u, t)

[4.25]

In equation [4.25], if we refer to equation [4.23], we see that u represents the projection of the nodal temperature vector onto a vector {τj } of the modal basis, and g(u, t) = c1j (fj (t) − kj u(t)). In the general case of differential system [4.12], u represents the nodal temperature vector {T } and g the vector [C]

−1

({F }−[K]·{T }).

Direct integration algorithms make it possible to integrate an equation like [4.25] step by step over time. Thus, to calculate the quantities at instant t + ∆t, we assume all the quantities at the previous instants (t, t − ∆t, etc.) to be known and apply a numerical scheme to replace the differential equation with an algebraic equation whose unknowns are the velocities at t + ∆t. There is a large number of integration schemes of differential equation [4.25]. Among the most widely used are the generalized central difference algorithm and the generalized trapezoidal rule.

116

Finite Element Simulation of Heat Transfer velocity

velocity

.

.

u(t+∆t)

u(t+∆t)

. u(τ)

.

u(τ)

.

.

u(t)

u(t)

t

τ

t+∆t

time

(a) generalized central differences

t

τ

t+∆t

time

(b) generalized trapezoidal rule

Figure 4.1. Time integration schemes: (a) generalized central difference method and (b) generalized trapezoidal rule

The generalized central difference algorithm (Figure 4.1a) assumes that the velocity u˙ is constant over each time interval [t, t + ∆t] and equal to the value at an intermediate instant t + ν∆t (with 0 ≤ ν ≤ 1): u(t + ∆t) − u(t) = u(t ˙ + ν∆t) = g u(t + ν∆t), t + ν∆t ∆t

[4.26]

In the generalized trapezoidal algorithm (Figure 4.1b), the velocity u˙ is also assumed to be constant, but is obtained as a weighted mean velocity of the velocities at times t and t + ∆t. With 0 ≤ ν ≤ 1 we obtain: u(t + ∆t) − u(t) = (1 − ν)u(t) ˙ + ν u(t ˙ + ∆t) ∆t = (1 − ν)g(u(t), t) + νg u(t + ∆t), t + ∆t

[4.27]

Actually, these two families of numerical schemes are very close. They even coincide when the function g is a linear function of u, as is the case in linear conduction. In non-linear cases, the slight differences can be seen in Figure 4.1. In the above numerical schemes, u(t + ∆t) is the unknown. It is obvious that if the right-hand side of [4.26] or [4.27] is independent of this unknown, then we directly (or explicitly) obtain u(t + ∆t) = u(t) + g(u(t), t)∆t. Then the numerical scheme is called explicit. For the numerical scheme families considered here, only the case ν = 0 corresponds to an explicit scheme, designated as an Euler explicit scheme. In the case ν = 1, we obtain an Euler implicit scheme. The intermediate algorithms corresponding to 0 < ν < 1 are said to be semi-implicit. The central difference

Transient Heat Conduction

117

method (also called the Crank-Nicolson algorithm) corresponds to equation [4.26] with ν = 1/2 and the trapezoidal rule to equation [4.27] with ν = 1/2. To apply the generalized central difference scheme to the transient heat conduction problem, we start again with equation [4.14] which is considered at instant t + ν∆t: ' ( [Cν ] · T˙ν + [Kν ] · {Tν } = {Fν } In the above equation, index ν is used to locate the quantities evaluated at instant t+ν∆t. Unless otherwise indicated, this index will be used later on. In addition, index 0 will be used for quantities evaluated at t and index 1, for those evaluated at t + ∆t. The variable change rate {T } being assumed to be constant over the time interval considered, we can write: {Tν } = {T (t + ν∆t)} = (1 − ν) {T0 } + ν {T1 } In that expression, vector {T0 } is known (it is the one at instant t), whereas vector {T1 } is the one we are trying to determine. Inserting this expression into the differential equation to solve then gives the following equation linear system: 1 1 [Cν ] + ν [Kν ] · {T1 } = [Cν ] − (1 − ν) [Kν ] · {T0 } + {Fν } [4.28] ∆t ∆t ! " Gν

The solution of this linear system allows us to obtain {T1 }, i.e. the temperatures at the mesh nodes at instant t + ∆t, from vector {T0 }, i.e. the temperatures at the mesh nodes at instant t. In the linear case being dealt with here, the other vectors and matrices involved are not time-dependent. Now apply the generalized trapezoidal scheme to equation [4.14] considered at instant t + ∆t: ' ( [C1 ] · T˙1 + [K1 ] · {T1 } = {F1 } We obtain: 1 [C ] · {T0 } ∆t 1 ' ( 1 [C1 ] + ν [K1 ] · {T1 } = + (1 − ν) [C ] · T˙ 1 0 ∆t + ν {F1 }

[4.29]

118

Finite Element Simulation of Heat Transfer

The generalized trapezoidal scheme requires at each instant vector {T0 } and vector {T˙0 } which is calculated gradually by means of formula [4.27]. At the initial instant, {T˙0 } is calculated by solving equation [4.14] at that instant, the initial temperature vector {T0 } being assumed to be known. Equations [4.28] and [4.29] do highlight the differences between both algorithm families, resulting mainly, in non-linear cases, from the fact that the quantities involved are not evaluated at the same instant (t + ν∆t for the generalized central difference algorithm and t + ∆t for the generalized trapezoidal algorithm). Consider again equation [4.28] in the case of the Euler explicit scheme corresponding to ν = 0. We then obtain the following system: 1 1 [C0 ] · {T1 } = {G0 } with {G0 } = [C0 ] − [K0 ] · {T0 } + {F0 } [4.30] ∆t ∆t The matrix to inverse is simply [C0 ], the capacitance matrix, resulting from the assembly of the element capacitance matrices [C e ] calculated at instant t. The conductance matrix [K0 ] is not involved in this matrix to inverse, but only in the load vector. The strict application of the ﬁnite element method produces a capacitance matrix [C0 ], symmetric and sparse, with non-diagonal terms different from zero. [C0 ] is called a consistent capacitance matrix (with the ﬁnite element formulation). However, in order to facilitate the inversion of the linear system obtained, matrix [C0 ] is made diagonal. This operation keeps the model’s total capacitance while eliminating the coupling terms between two nodes. All the thermal capacitance is then transferred to the diagonal terms of the matrix. In fact, with this technique, the speciﬁc heat is no longer distributed evenly in the material volume but is lumped at the mesh nodes. The most common lumping technique transfers on the diagonal the sum of the terms of each line: C11 C12 · · · [C0 ] = C21 C22 · · · ··· ··· ··· C11 + C12 + · · · 0 0 =⇒ [C0 ] ≈ 0 C21 + C22 + · · · 0 0 0 ···

Transient Heat Conduction

119

This approximation considerably simpliﬁes the solution of linear system [4.30] as it is reduced to n scalar equations yielding the components of {T1 } directly. The problem unknowns at instant t + ∆t are thus obtained explicitly without solving a linear system. Furthermore, storing the matrix only requires storing the n diagonal terms. However, the explicit scheme has a major drawback. The algorithm is conditionally stable, i.e. the time step must never exceed a limit value. How to obtain this limit value is described in the next section. That is the reason why it is restricted to strongly non-linear problems, involving rapid phenomena already requiring, for physical reasons, very small time steps. When an implicit or semi-implicit scheme (0 < ν ≤ 1) is used, the temperatures at the mesh nodes at instant t + ∆t (i.e. vector {T1 }) are obtained by solving equation system [4.28]. When the problem is linear, matrices [Cν ] and [Kν ] involved in the system are independent of the temperature. Therefore, they are constant over time, and the matrix to invert in equation system [4.28] is determined by the time step ∆t. The assembly and triangulation of this matrix will only be necessary when the time step ∆t is modiﬁed during the calculation process. Implicit schemes are more complicated to process than explicit schemes. However, their stability conditions are less severe. This explains why they are frequently used in thermal science. 4.2.3. Accuracy and stability of a direct integration algorithm In order to analyze the stability and accuracy of the integration schemes presented in section 4.2.2, we will cover the linear case only, and assume that the time step used, ∆t, is constant. Thus, we can designate with index p the quantities estimated at instant p∆t. {T0 } thus gives the temperatures at the mesh nodes at t = 0, {Tp } the temperatures estimated at t = p∆t, and {Tp+ν } those estimated at t = (p + ν)∆t, i.e. during the time interval, [p∆t, (p + 1)∆t], with 0 ≤ ν ≤ 1. Besides, restricting the study to the linear case allows us to write matrices [Cν ] and [Kν ] as [C] and [K], since they are constant. Now it is important to differentiate the accuracy and stability conditions of an algorithm: – the algorithm accuracy is given by the truncation error made in formula [4.26] or [4.27]; – the stability is the capacity of an algorithm to reduce or amplify the errors made during the consecutive time steps.

120

Finite Element Simulation of Heat Transfer

4.2.3.1. Accuracy The variation of a quantity u during a time step is approximated as: up+1 − up = ∆t

u˙ p+ν (1 − ν)u˙ p + ν u˙ p+1

in the central difference method in the trapezoidal rule

In order to evaluate the truncation error made with the generalized central difference method, the following Taylor series expansions are written: 2 2 ∆t ¨p+ν + O ∆t3 up+1 = up+ν + (1 − ν)∆tu˙ p+ν + (1 − ν) 2 u ∆t2 u = u u ¨p+ν + O ∆t3 ˙ p+ν + ν 2 p p+ν − ν∆tu 2 By subtracting these two expressions then dividing the result by ∆t, we obtain: ∆t up+1 − up − u˙ p+ν = (1 − 2ν) u ¨p+ν + O ∆t2 ∆t 2 Likewise, the truncation error made with the generalized trapezoidal rule can be obtained from the following Taylor series expansions: 2 up+1 = up + ∆tu˙ p + ∆t u ¨p + O ∆t3 2 up + O ∆t2 u˙ p+1 = u˙ p + ∆t¨ From these expressions we immediately ﬁnd that: 1 up+1 − up − (1 − ν)u˙ p + ν u˙ p+1 = − ν ∆t¨ up + O ∆t2 ∆t 2 We can see that both families produce ﬁrst order algorithms when ν = 12 . The numerical schemes are said to be coherent at order 1 for differential equation [4.25]. This means that the errors in the velocity u˙ are in the order of ∆t. When ν = 12 , the algorithm is coherent at order 2 and the errors made are in the order of ∆t2 . The conclusion of this might be that the Crank-Nicolson algorithm (ν = 12 ) is, in any case, the best possible choice. In fact, it is not so obvious, for, as will be seen later on, this scheme produces oscillating solutions when the time step is too great. With strong non-linearities, like those resulting from state changes for example, the Euler implicit algorithm is generally preferred.

Transient Heat Conduction

121

4.2.3.2. Stability We have just seen that at each time step, the solution obtained is erroneous. Imagine these errors aggregated in the case of Euler implicit algorithm (ν = 1). We saw previously that this algorithm is coherent at order 1. Therefore, there is a constant such that: up+1 − up − u˙ p+1 ≤ k∆t ∆t The error would then be less than the product N k∆t where N represents the number of time steps performed. Unfortunately, these are not the only errors that add up because numerical scheme [4.26] or [4.27], uses numerical values up which are also erroneous. The numerical scheme may therefore amplify the errors to the point of producing an unacceptable solution. We say that a numerical scheme is stable if the error ampliﬁcation is bounded, i.e. if the very solution of the discretized problem is bounded. We can demonstrate that if a numerical scheme is coherent and stable, then the numerical solution converges toward the continuous solution when the time step tends to 0. In order to analyze the algorithm stability condition, it is useful to consider the eigenvector basis. We therefore consider the linear conduction case and go back to equation [4.23], which is re-written as: 1 β˙ j (t) + αj βj (t) = fj (t) cj

[4.31]

Applying the generalized central difference algorithm to the equation above over the interval [p∆t, (p + 1)∆t]; we obtain: 1 1 1 + ναj βj p+1 = − (1 − ν)αj βj p + fj p+ν ∆t ∆t cj which, once rearranged, becomes: 1 − (1 − ν)αj ∆t ∆t 1 βj p + fj p+ν βj p+1 = 1 + ναj ∆t 1 + ναj ∆t cj

[4.32]

By recurrence this relationship yields (βj )p+1 as a function of (βj )0 . The algorithm will be stable if all the (βj )p+1 remain bounded, regardless of j and p. The algorithm stability condition is therefore written as: 1 − (1 − ν)αj ∆t ≤ 1, i.e. ∀j, (1 − 2ν)αj ∆t ≤ 2 ∀j, 1 + ναj ∆t

122

Finite Element Simulation of Heat Transfer

From this relationship it can be deduced that if 1/2 ≤ ν ≤ 1, then the scheme used is unconditionally stable. On the contrary, if 0 ≤ ν < 1/2, then the scheme is conditionally stable, and the time step used, ∆t, must be less than or equal to the critical stability step, ∆tsta : ∆t ≤ ∆tsta

with

∆tsta =

2 (1 − 2ν)αmax

[4.33]

where αmax represents the greatest eigenvalue. When going back to relationship [4.32], it is clear that the solution oscillates when < 0. A non-oscillation condition of the solution over time can be derived from this. This condition is written as: 1−(1−ν)αj ∆t 1+ναj ∆t

∀j,

1 − (1 − ν)αi ∆t ≥ 0, 1 + ναj ∆t with ∆tosc

i.e.

∆t ≤ ∆tosc

1 = (1 − ν)αmax

[4.34]

We can see that relationship [4.34] is always satisﬁed for the Euler implicit algorithm. Therefore, this algorithm never produces an oscillating solution, whatever the time step value. 4.2.3.3. Simpliﬁed analysis of the stability condition Relationship [4.33] gives the stability condition in a stringent manner. However, to be applied practically, this relationship requires a previous calculation of the problem eigenvalues and modes, which increases the global cost of the analysis. Thus, it is useful to have a stability condition generated by a simpliﬁed analysis. To do so, we consider a 1D case with no source term and a regular mesh composed of n 2-noded elements of length e. The complete mesh thus consists of n + 1 nodes. Besides, we consider a capacitance thermal matrix lumped by means of the technique described in section 4.2.3.2. It is then easy to obtain: 1 −1 0 · · · · · · · · · −1 2 −1 0 · · · · · · 1 2 λ λ 1 −1 0 −1 2 −1 0 · · · e [K ] = =⇒ [K] = e −1 1 e · · · · · · · · · · · · · · · · · · · · · · · · 0 −1 2 −1 · · · · · · · · · 0 −1 1

Transient Heat Conduction

1 e

[C ] = ρCe

1/2 0 2 0 0 =⇒ [C] = ρCe ··· 1/2 ··· ···

1/2 0

0 1 0 ··· ··· ···

123

0 ··· ··· ··· 0 0 ··· ··· 1 0 0 ··· ··· ··· ··· ··· 0 0 1 0 · · · · · · 0 1/2

The generic equation corresponding to node 1 < j < n + 1 is written as: λ − Tj−1 + 2Tj − Tj+1 = 0 ρCeT˙j + e Applying the generalized central difference algorithm then yields, after a few mathematical manipulations: + , + 2 T − T (1 − ν) − T j−1 p j p j+1 p Tj p+1 − Tj p + F o + , = 0 + ν − Tj−1 p+1 + 2 Tj p+1 − Tj+1 p+1 [4.35] where F o is the Fourier number: Fo =

λ∆t ρCe2

[4.36]

To study the stability, the Fourier method is used here. This method analyzes the amplitude variation of each term of the Fourier series development of a disturbance ; of the initial solution. We therefore consider a disturbance in the form l V l (t)eilx , i being the complex number such that i2 = −1, and we study the variation of the lth harmonic V l when equation [4.35] is applied to it. If we term xj the x-coordinate of node j and given that xj−1 = xj − e and xj+1 = xj + e, we obtain: +

V

l

p+1

− V

p

(1 − ν) V l p − e−ile + 2 − eile ilxj + Fo =0 e l −ile ile + ν V p+1 − e +2−e

, l

eilxj

from which we derive:

V l p+1 1 + 4F o(ν − 1) sin2 (le/2) = Vl p 1 + 4F oν sin2 (le/2)

124

Finite Element Simulation of Heat Transfer

It is obvious that the stability condition, also called the von Neumann condition, is written as: 1 + 4F o(ν − 1) sin2 (le/2) 1 + 4F oν sin2 (le/2) ≤ 1 For this condition to be satisﬁed whatever the harmonic l considered, it is sufﬁcient that: 1 (1 − 2ν)F o ≤ [4.37] 2 Of course, for 1/2 ≤ ν ≤ 1, condition [4.37] is always satisﬁed. In the other cases, the stability time step derived from [4.37] is given by: ∆tsta =

ρCe2 2λ(1 − 2ν)

[4.38]

This stability condition is only approximated. It is valid only in the conditions considered to establish it. In 2D and 3D cases, and for elements of the same size in the 2 or 3 directions, it is possible to demonstrate that the stability condition related to the Euler explicit scheme is written respectively as F o ≤ 14 and F o ≤ 16 . When all the elements are not the same size (which is the most common case), the stability criterion must be calculated with the smallest element. 4.2.4. Practical complementary rules We have just seen that, for reasons of stability or non-oscillation of the solution over time, the time step ∆t used when simulating heat exchanges with ﬁnite elements, was always to remain less than two limit values (∆tsta deﬁned by equation [4.33] and ∆tosc deﬁned by equation [4.34]). In this section, a few complementary rules to choose this time step ∆t are given. These complementary rules illustrate the fact that the choice of the time step depends not only on the time integration algorithm selected (value of ν in particular) and the material properties (eigenvalues of the −1 matrix [C] [K]) but also on the mesh chosen to represent the structure on which the simulation is performed. 4.2.4.1. Space oscillations during thermal shock simulation In the case of a thermal shock, if the terms of the capacitance matrix [C] are not lumped on the diagonal, the simulation time step must remain greater than a value ∆ts , so as to avoid the appearance of space oscillations of the solution during the initial moments of the calculation process. This phenomenon will be illustrated on a 1D problem.

Transient Heat Conduction 1

i-1

i

i+1

n

125

n+1

1 x

0 e=1/n

Figure 4.2. Mesh of the segment [0, 1] with n elements

Consider the case of a body initially at homogenous temperature, put in sudden contact with another body at a different temperature (thermal shock problem). To represent this thermal shock without losing the general characteristic of the results obtained, the following 1D problem is used: Find T (x, t) satisfying at each instant t: ∀x ∈ ]0, 1[,

2

∂ T ρC T˙ = λ 2 ∂x

∂T (0, t) = 0 ∂x T (1, t) = 1 with the initial condition: ∀x ∈ [0, 1[,

T (x, 0) = 0

To solve this ﬁnite element problem, the segment [0, 1] is divided into n ﬁrst order elements (with linear interpolation) of length e = 1/n (Figure 4.2). The mesh obtained includes n + 1 nodes. The resolution of the problem leads us to seek the vector {T } including the nodal values of the temperature ﬁeld at each instant (Ti is the temperature value at node i at that instant), and satisfying: ' ( Tn+1 = 1 for any t > 0 [C] · T˙ + [K] · {T } = {0} with Ti = 0 at t = 0 for any 1 ≤ i ≤ n Conductance matrix [K] and capacitance matrix [C] can easily be calculated from the element matrices: 1 −1 0 · · · · · · · · · −1 2 −1 0 · · · · · · 1 2 λ λ 1 −1 0 −1 2 −1 0 · · · e =⇒ [K] = [K ] = e −1 1 e · · · · · · · · · · · · · · · · · · · · · · · · 0 −1 2 −1 · · · · · · · · · 0 −1 1

126

Finite Element Simulation of Heat Transfer

[C e ] =

1 ρCe 2 6 1

2 1 2 ρCe 1 0 =⇒ [C] = 2 6 · · · · · · ···

1 4 1 ··· ··· ···

0 ··· ··· ··· 1 0 · · · · · · 4 1 0 · · · · · · · · · · · · · · · 0 1 4 1 ··· ··· 1 2

The generic equation at one node 2 ≤ i ≤ n is then written as: 6F o − Ti−1 + 2Ti − Ti+1 = 0 T˙i−1 + 4T˙i + T˙i+1 + ∆t Now we will apply the generalized trapezoidal algorithm to estimate the temperatures Ti at the beginning of the thermal shock at t = ∆t. By writing the previous equation at time ∆t, and considering that the terms Ti and T˙i are equal to zero at t = 0, we get T˙i = Ti /ν∆t for each index i and obtain: with Ti = A Ti−1 + Ti+1

A=

6F oν − 1 4 + 12F oν

If we write Di = Ti − Ti−1 , we see that the above equation is satisﬁed in the same manner by these quantities, i.e: Di = A Di−1 + Di+1 In order to avoid space oscillations of the solution, Figure 4.3 suggests that all the Di must have the same sign. We derive from this a non-oscillation condition of the solution which is written A ≥ 0, i.e.: 6νF o ≥ 1 =⇒ ∆t ≥ ∆ts

with

∆ts =

ρCe2 6λν

[4.39]

If condition [4.39] is not satisﬁed, then the solution obtained will produce space oscillations contrary to physical laws. In fact, this condition must be generally considered as a constraint on the size e of the ﬁrst element affected by the thermal shock. This constraint gives the element maximum size according to the time step used: A 6λν∆t e ≤ emax with emax = [4.40] ρC

Transient Heat Conduction

127

T 1

Di+1 Di

xi-1 xi xi+1

0

x

Figure 4.3. General aspect of the problem solution

Consider a capacitance matrix obtained by lumping the thermal capacitance on the nodes (see section above). In our simple example, this comes down to using the following matrix [C]:

3 0 ρCe 0 [C] ≈ 6 · · · · · · ···

0 6 0 ··· ··· ···

0 ··· ··· ··· 0 0 · · · · · · 6 0 0 · · · · · · · · · · · · · · · 0 0 6 0 ··· ··· 0 3

If the same method as before is used, the following relationship is then obtained: with Ti = A Ti−1 + Ti+1

A =

6F oν 6 + 12F oν

The constant A thus deﬁned remains always positive or equal to zero, so that the solution obtained does not oscillate, whatever the time step ∆t used. In fact, this lumped approximation of the capacitance matrix consists of introducing an additional

128

Finite Element Simulation of Heat Transfer

conduction term. Indeed, the constant A can easily be expressed in the same form as the constant A previously deﬁned, but by means of a modiﬁed thermal conductivity: A =

6F o ν − 1 4 + 12F o ν

with F o =

λ ∆t ρCe2

and

λ = λ +

ρCe2 6ν∆t

4.2.4.2. Discrete maximum principle We saw that in thermal shock problems, space oscillations of the solution could appear if the time step chosen was less than a value ∆ts . In fact for time steps ∆t

λ1 ρ1 C1

and

b2 =

>

λ2 ρ2 C2

In this equation, b1 and b2 represent the thermal effusivity of media 1 and 2 respectively. Choosing Tc as the initial value of the contact temperature is not always the best choice. In order to calculate the temperature to use, which will be written Tc , we start with the initial distribution (at t = 0) of the temperatures in both media, resulting from the ﬁnite element discretization. This distribution is illustrated in Figure 4.5. The only exchanges possible, when both media are in contact, are related to the two media only. If the temperature Tc at the contact node is to be physically acceptable, it must reproduce the fact that the quantity of heat gained by one medium is equal to that lost by the other medium. By writing Qi the quantity of heat gained by medium i

132

Finite Element Simulation of Heat Transfer

T T2 Tc

medium 1

medium 2

λ1 ρ1 C1

λ2 ρ2 C2

T1

e1

x

e2

Figure 4.5. Initial temperature distribution

between instants − and 0 ( > 0, so that instant − represents the instant immediately before contact), we obtain: Qi =

0

−

ρi Ci T˙ dx dt =

Ωi

ρi Ci T (x, 0) − T (x, −) dx

Ωi

Given that the temperatures at instant − are T1 in medium 1 and T2 in medium 2, that the function T (x, 0) is linear in x due to the elements used, and assuming that the characteristics of the media are homogenous and are not temperature-dependent, the quantity of heat Qi can be written as follows: Qi =

1 ρi Ci ei Tc − Ti 2

The thermal balance between the two media prescribes Q1 + Q2 = 0, which produces the following expression for the contact temperature: Tc =

ρ1 C1 e1 T1 + ρ2 C2 e2 T2 ρ1 C1 e1 + ρ2 C2 e2

This physically coherent value of the contact node temperature is signiﬁcantly different from the theoretical value Tc obtained for inﬁnite media. In fact, it is possible to choose the sizes of the elements close to the contact, e1 and e2 , such that Tc and Tc coincide. To do so, it is sufﬁcient that the ratios of the coefﬁcients bi (producing Tc ) to ρi Ci ei (producing Tc ) of each medium are equal. The following mesh criterion is obtained: 7 8 λ2 8ρ C e2 = 9 2λ1 2 e1 ρ C 1

1

Transient Heat Conduction

133

This criterion is also written F o1 = F o2 , where F oi is a Fourier number, deﬁned by equation [4.36,] corresponding to the medium i. According to this criterion, the square ratio of the sizes of the ﬁrst elements in media 1 and 2 must be equal to the ratio of the thermal diffusivity of these media. This criterion can also be obtained directly, and in a more general framework, by writing T˙ = 0 at t = 0 at the interface level, from the discretized equations. By choosing Tc as the temperature at the contact node, and satisfying that criterion when meshing these two media in contact, a physically coherent value is used, because it satisﬁes the thermal balance between both media. In the case of the solidiﬁcation of a cast iron piece in a sand mold, hence a sand-cast iron interface (a traditional casting problem) the criterion established produces very different element sizes on each side of the interface [BER 86]. Indeed, consider the following mean characteristics of the two media: – λ1 = 0.4 W/m/K, ρ1 = 1,500 kg/m3 and C1 = 1,150 J/kg/K for sand (medium 1); – λ2 = 40 W/m/K, ρ2 = 7,500 kg/m3 and C2 = 820 J/kg/K for cast iron (medium 2). Applying the criterion produces e2 /e1 = 5.3. If it is not satisﬁed, we can show that: – if e2 /e1 5.3, then the cast iron cooling process accelerates, and the contact temperature becomes less than the analytical value Tc . We have established this mesh criterion in a simple 1D case. It is easily extended to 2D and 3D cases. On each side of the interface, the mesh must necessarily be in the form of regular layers whose thicknesses are in a ratio satisfying the criterion. For instance, in the case of the cast iron-sand interface, the ﬁrst layer regularity on each side of the interface is a strict obligation, otherwise it will signiﬁcantly affect the positions of the hot points. Indeed, the local cooling kinetics will depend on the satisfaction or the non-satisfaction of this criterion. Some computer codes include meshing procedures, enabling them to make successive layers of elements with controlled thicknesses on each side of the interface. It is then possible to satisfy the criterion previously established. Figure 4.6 illustrates a 2D mesh example performed from the layer mesher of the software SY ST U S T M .

134

Finite Element Simulation of Heat Transfer

Figure 4.6. Mesh of a casting and its mold

One way of obtaining, with certainty, and independently of the mesh used, physically coherent initial conditions at the interface of a perfect thermal contact consists of: – assuming ﬁrst of all that both media are at the same initial temperature, that of medium T1 , for instance; – calculating the energy required to bring medium 2 to temperature T2 ; – prescribing that energy as an internal heat source distributed in the volume of the elements representing medium 2 for a very short time.

Transient Heat Conduction

135

Very short time means the time allowing no signiﬁcant gradient to settle down beyond the ﬁrst layer of elements on each side of the interface. The interface temperature then adjusts itself to the local value Tc which will coincide with the value Tc if the element thicknesses satisfy the criterion. Another way of dealing with thermal interfaces is to perform two separate meshes, and introduce on the interface a thermal contact resistance. The interface heat ﬂux will then be given locally according to the temperatures of each medium obtained by nodal approximation. If T1 is the local temperature of medium 1, and T2 that of medium 2, then the heat ﬂux Q1 gained by medium 1 is equal to Q1 = (T2 −T1 )/Rc , whereas that gained by medium 2 is equal to Q2 = −Q1 = (T1 − T2 )/Rc . In these expressions, Rc is a characteristic introduced for the interface, often called thermal contact resistance. If a very low value is used for Rc , the situation is close to that of a perfect contact. 4.3. Working example The aim of this section is to illustrate with a simple example the notions introduced in this chapter to simulate transient heat conduction in a solid. To do so, the induction heating example used to illustrate the formulation of the steady state conduction problem is considered again. Firstly, the physical modeling process used in this example and its ﬁnite element approximation are studied in depth, then a few numerical applications are performed to illustrate the notions seen in this chapter. 4.3.1. Physical modeling and approximation The induction heating of a plate is illustrated in Figure 4.7. In Part 1 of this book, this example was dealt with using a heating volumetric term Q and an exchange coefﬁcient h independent of time and the steady state corresponding to those conditions was sought. Here, a thermal capacitance of the material is introduced and inductor

modeling plate

x

T

E

convection

E-P

0

Q(x) Td

Figure 4.7. Working example: plate induction quenching

136

Finite Element Simulation of Heat Transfer

the study of transient states allows us to use the heating term Q and cooling term h as x−E functions of time. The volumetric heat source is expressed as Q(x, t) = Q0 (t)e P , where only the term Q0 (t) is a function of time. The exchange coefﬁcient h will simply be written as h(t) for the moment. In this example, for convenience, we choose physical properties λ, ρ and C independent of temperature. The solution of the thermal conduction problem in the plate consists, in the transient state, of seeking a temperature ﬁeld T (x, t) which veriﬁes at any point x ∈ [0, E] and at any instant t: 2 ∂ T λ 2 + Q(x, t) = ρC T˙ (x, t) ∂x ∂T = h(t) Td − T (x, t) λ in x = E ∂x at x = 0 T (x, t) = Td at t = 0 T (x, t) = T0 We added to this equation the term T0 which represents the initial plate temperature (at t = 0). This temperature is assumed to be constant over time at point x = 0. On the contrary, the external temperature at point x = E was set at Td . To solve the problem, the ﬁnite element method is used. To do so, we discretize axis 0x in Figure 4.7 between 0 and E by means of n+1 nodes at uniform distances of e = E/n. The node coordinates are xi = ie, with i = 0, 1, . . . , n. It is now possible to choose the type of element. We select n two-noded isoparametric elements with linear interpolation. We term Ωi = [xi−1 , xi ], with i = 1, 2, . . . , n, the domain occupied by element i. With this discretization, the following nodal approximation is used for the temperatures: ∀x ∈ Ωi , ∀t ≥ 0,

T (x, t) = N1i (x)T1i (t)

xi − x N1i (x) = e + N2i (x)T2i (t) with x − xi−1 N i (x) = 2 e

In that case, the assembly of thermal conductance and capacitance matrices, and that of the element load vector, produces the following expressions for matrices [K] and [C] and for vector {F }:

Transient Heat Conduction

1 −1 0 λ [K] = · · · e · · · · · · 0

−1 0 · · · · · · · · · 0 2 −1 0 · · · · · · · · · 0 −1 2 −1 0 · · · · · · · · · · · · · · · · · · · · · · · · · · · + h(t) · · · · · · 0 −1 2 −1 0 0 · · · · · · 0 −1 2 −1 · · · · · · · · · 0 −1 1

2 1 0 ρCe [C] = · · · 6 · · · · · · 0

··· ··· ··· ···

··· ··· 0 0

137

0 · · · 0 1

1 0 ··· ··· ··· 0 4 1 0 · · · · · · · · · 1 4 1 0 · · · · · · · · · · · · · · · · · · · · · · · · ··· 0 1 4 1 0 ··· ··· 0 1 4 1 ··· ··· ··· 0 1 2

, x1 −E x1 −E P 2 + x0 −E e P −e P + Pe P − e , 2+ x1 −E x0 −E x2 −E P 2 P P P 2e − −e −e 0 e 0 , 2+ x2 −E x1 −E x3 −E P 2 P P P 2e − −e −e 0 e ··· {F } = Q0 (t) + h(t) · · · , xn−2 −E xn −E P 2 + 2 xn−1 −E 0 P P P 2e −e −e − e 0 + , 2 x −E x −E P xn −E n−1 n+1 Td 2 P P P 2e − − e − e e , 2+ x −E x −E xn −E P n+1 n+1 − e P + Pe P −e P e The integrals involved here were calculated analytically. We have deliberately emphasized in these expressions the presence of the two time functions Q0 (t) and h(t), which are used to model induction heating. Only the capacitance matrix is not time-dependent in this example. Now it is necessary to seek the temperature change at the n nodes over time. Indeed the temperature of node 0 is known: it is constant and equal to Td . The problem to solve is written:

138

Finite Element Simulation of Heat Transfer

Find at each instant t the nodal temperature vector {T (t)} such that: ' ( [K(t)] · {T (t)} + [C] · T˙ (t) = {F (t)} {T (0)} = {T0 } To obtain the variation of the temperature vector at the nodes {T (t)} it is now necessary to choose a time integration scheme. The generalized trapezoidal rule ' will ( be used. Thus, over time, if we know the temperature vector at time t, {T (t)}, T˙ (τ ) at any instant τ ∈ [t, t + ∆t] is expressed as (see equation [4.27]): ∀τ ∈ [t, t + ∆t],

'

( {T (t + ∆t)} − {T (t)} T˙ (τ ) = ∆t ' ( ' ( = (1 − ν) T˙ (t) + ν T˙ (t + ∆t) with 0 ≤ ν ≤ 1

In practice, the previous expression is used to express the quantity {T˙ (t + ∆t)} as follows: ' ( 1−ν ' ˙ ( 1 T˙ (t + ∆t) = − T (t) + ({T (t + ∆t)} − {T (t)}) ν ν∆t The problem to solve at instant t + ∆t is therefore written: ' ( Knowing {T (t)} and T˙ (t) , ﬁnd the vector {T (t + ∆t)} such that: [C] . {T (t)} + ν∆t {F (t + ∆t)} ([C] + ν∆t [K(t + ∆t)]). {T (t + ∆t)} = ' ( + (1 − ν)∆t [C] . T˙ (t) T0 (t + ∆t) = Td Note here that the matrix involved in this linear system can be written with the addition of ∆ts deﬁned by equation [4.39]:

∆t ∆ts ∆t 1 − ∆ts ρCe 0 [C] + ν∆t [K] = 6 ··· 0 2+

∆t ∆ts ∆t 4+2 ∆ts ··· 1−

0 ···

0 ∆t ∆ts ··· ∆t 1− ∆ts

1−

0

··· 0 ··· ∆t 4+2 ∆ts ∆t 1− ∆ts

0

··· 0 ∆t 1− ∆ts ∆t 2+ ∆ts

Transient Heat Conduction

139

The non-diagonal terms of this matrix will change signs according to the ratio of the time step used ∆t to the minimum time step to use ∆ts . In practice, the capacitance matrix [C] is approximated by a lumped matrix, obtained by transferring on the diagonal the sum of the terms on each line. This produces the following expression for the matrix involved in the linear system to solve:

∆t ∆ts ∆t − ∆ts 0

∆t ∆ts ∆t 6+2 ∆ts ···

···

0

0

···

3+

ρCe [C] + ν∆t [K] ≈ 6

−

0 ∆t ∆ts ··· ∆t − ∆ts

−

0

··· 0 ··· ∆t 6+2 ∆ts ∆t − ∆ts

0

··· 0 ∆t − ∆ts ∆t 3+ ∆ts

In that expression, whatever the time step ∆t chosen, no term of the matrix changes signs. It follows that there will be no signiﬁcant change in the result (no numerical instability) when the time step ∆t is modiﬁed. 4.3.2. Numerical applications The numerical applications shown in this book are used to illustrate the notions introduced in this chapter. To do so, we have chosen some ﬁxed values, and others which will change. The ﬁxed values are the following: – plate thickness: E = 0.1 m; – number of ﬁnite elements: n = 10; – material thermal conductivity: λ = 30 W/m/K; – external medium temperature: Td = 20°C; – induction heating depth: P = 0.02 m; – volumetric mass: ρ = 7,800 Kg/m3 ; – speciﬁc heat: C = 500 J/Kg/K. The aim of the ﬁrst simulation is to show that the simple introduction of the capacitance term without modifying the other data with respect to the steady state calculation, only delays the implementation of the steady state. Indeed, Figure 4.8 shows a calculation performed with:

140

Finite Element Simulation of Heat Transfer

Figure 4.8. Variation of the temperature proﬁle over time

– a constant exchange coefﬁcient: H = 2,000 W/m2 /K; – a constant heating power: Q0 = 5.107 W/m3 ; – a plate initial temperature: T0 = 20°C; – a time integration scheme (trapezoidal rule): ν = 0.5; – a constant time step: ∆t = 1 s. In this ﬁgure, we note that the temperature proﬁle changes gradually from a constant value T0 at t = 0 toward the steady state proﬁle (calculated in the ﬁrst part of this book) for an “inﬁnite” time. Indeed, after a fairly long time, the temperature does not change over time, so that {T˙ } = {0} in the differential equation to solve. The solution is then obtained in the steady state. The second simulation illustrates the limit time step ∆ts highlighted in the previous equations. To do so, we have restricted our study to the cooling phase, starting with a high initial temperature: – a constant exchange coefﬁcient: H = 10,000 W/m2 /K; – a heating power equal to zero: Q0 = 0; – a plate initial temperature: T0 = 700°C; – a time integration scheme: ν = 1.

Transient Heat Conduction

141

Figure 4.9. Plate quenching – temperature proﬁle at t = ∆t for ∆t = 1.3 s and ∆t = 13 s

We can see in Figure 4.9 that the solution obtained after the ﬁrst time increment depends on the time step chosen. With ∆t = 13 s, the solution is relatively conform to physics, whereas with ∆t = 1.3 s there are oscillations contrary to physics. In fact the 2 threshold time step in this example being ∆ts = ρCe 6λ ≈ 2.166 s, using a time step lower than this value produces space oscillations. The last simulation shows that our simple model makes it possible to represent an induction quenching sequence. Indeed, induction quenching consists of two phases. First, the plate is heated rapidly by induction, so that the convection exchanges are negligible. Then the heating system is turned off and the plate is cooled abruptly, for instance with an atomized jet of water projected onto the face x = E. This can be illustrated as follows: – between instants t = 0 and t = t0 , the heating period is given by Q0 (t) = Q0 and h(t) = 0. An exchange coefﬁcient equal to zero comes down to assume that there is no exchange in x = E, between the piece and the external medium. This is a simpliﬁcation of the problem of course, in which the exchange is considered to be small enough to be neglected; – from instant t = t0 , the quenching period is given by Q0 (t) = 0 and h(t) = H. The value of H will depend on the quenching medium considered (air, water, atomized water, etc.).

142

Finite Element Simulation of Heat Transfer

Figure 4.10. Induction quenching – temperature proﬁle over time

The numerical values used are the following: – constant exchange coefﬁcient (when it is different from zero): H 6,000 W/m2 /K;

=

– heating power (when it is different from zero): Q0 = 2.108 W/m3 ; – plate initial temperature: T0 = 20°C; – time integration scheme: ν = 1; – heating time: t0 = 15 s. Figure 4.10 is a good illustration of the temperature proﬁles obtained during induction quenching. The heating phase produces a gradual increase of the temperature at the plate surface (in x = E) and a positive thermal gradient. This thermal gradient strongly depends on the depth P introduced to model induction heating. This depth depends on the induction current frequency. A high frequency will produce surface heating and a high thermal gradient (low value of P ), whereas a lower frequency will heat the plate in its thickness (high value of P ) with a lower thermal gradient as well. Then the quenching phase produces rapid cooling of the surface, and a local sign change in the thermal gradient. The cooling rapidity depends directly on the thermal exchange coefﬁcient H.

Chapter 5

Non-linearities

5.1. Formulation and solution techniques In the previous chapters, we introduced the ﬁnite element method applied to the solution of permanent or transient heat conduction problems. Generally speaking we assumed that the thermal and physical properties of the materials concerned did not depend on the temperature and restricted our analysis to boundary conditions expressed linearly according to the temperature by means of a constant exchange coefﬁcient. Applying the ﬁnite element method then leads to the solution, possibly at different moments, of a linear equation system. We then speak of a linear problem solution. Practically speaking, the various parameters (thermal and physical properties, boundary conditions, etc.) involved in mathematical modeling are rarely expressed so simply. Such is the case when the application to simulate covers an extensive temperature range or involves phase changes. The problem formulation no longer produces a linear equation system and the solution techniques have to be appropriate. 5.1.1. Formulation The ﬁnite element method applied to a thermal conduction problem leads to the cancelation of a residual vector {R(T )} obtained by assembling element residuals {Re (T )} (equation [2.6] in Part 1 of this book for the steady state, equation [4.7] in Part 2 for the transient state): {R(T )} =

m e=1

T

[Ae ] · {Re (T )} = {0}

[5.1]

144

Finite Element Simulation of Heat Transfer

In a linear problem, we saw that this residual can be expressed as follows: {R(T )} = {F } − [K] · {T } (steady state) ' ( {R(T )} = {F } − [K] · {T } − [C] · T˙ (transient state) In these expressions, the thermal conductance matrix [K], the capacitance matrix [C] and the load vector {F } are independent of the temperature. It follows that in this case the ﬁnite element method leads to the solution of linear equation systems. In the transient state case, this system is obtained after applying a time integration scheme such as that of generalized central differences. When, for instance, the thermal conductivity λ or the speciﬁc heat C depend on temperature, matrices [K] and [C] also depend on temperature and the residual vector becomes a non-linear function of the nodal temperature vector {T } which can be written as follows: {R(T )} = {F } − [K(T )] · {T } (steady state) ' ( {R(T )} = {F } − [K(T )] · {T } − [C(T )] · T˙ (transient state) We then obtain a non-linear equation system whose unknowns are the temperatures at the mesh nodes. In fact, strictly speaking, the size of the non-linear system to solve [5.1] is not always exactly equal to the number of mesh nodes. Indeed, as in the linear case, the temperatures of certain nodes can be prescribed. This is even necessary in the case of a steady state problem to obtain a single solution. In the non-linear case, the application techniques for this type of condition are the same as those already introduced in Part 1 of this book. Therefore, we will not mention this aspect hereafter and assume that the non-linear system [5.1] is expressed in the form of n equations with n unknowns, where n is the number of mesh nodes, possibly reduced with the elimination method by the number of nodes at which the temperature is known, or increased by this number of nodes with the Lagrange multipliers method. 5.1.2. Non-linear equation system solution methods In order to solve equation [5.1], most computer codes follow two steps. In the ﬁrst, they try to predict the solution and in the second, they solve the non-linear equation system iteratively as indicated in Figure 5.1 using the prediction of step one as an initial estimation.

Non-linearities

145

estimation (prediction): {T }(0)

calculation of residual vector: {R} (it )

convergence test ||{R}(it)|| < ε

(iteration it)

solution

correction of the temperature vector {T}(it+1) = {T}(it) + {∆T }(it)

Figure 5.1. Non-linear equation system solution principle

(0)

The estimation {T } plays an important part in the iterative process convergence. In a steady state problem, it is difﬁcult to obtain a reliable estimation in all possible cases. Conversely, an estimation can be obtained at any moment in a transient problem from the solutions at the previous moments. For instance, it is possible to choose a linear extrapolation of the values calculated at the two previous time steps {T (t − ∆t)} and {T (t)}: (0)

{T }

(0)

= {T (t + ∆t)}

= 2 {T (t)} − {T (t − ∆t)}

or a quadratic extrapolation of the values calculated at the three previous steps {T (t − 2∆t)}, {T (t − ∆t)} and {T (t)}: (0)

{T }

(0)

= {T (t + ∆t)}

= 3({T (t)} − {T (t − ∆t)}) + {T (t − 2∆t)}

At each iteration it of the process described in Figure 5.1, we have an estimation of (it) (it) the solution {T } . Then the residual vector {R} corresponding to it is calculated. A convergence test is then carried out by comparing the norm of the residual (it) {R} with a threshold value R given by the user. The norm considered can be

146

Finite Element Simulation of Heat Transfer

the Euclidian norm (root-sum square of each component of the vector), but some computer codes give other measurements like the vector’s largest component in absolute value (called the “inﬁnite norm”). If the residual vector norm is less than the limit value prescribed by the user, the algorithm is then considered to have converged. (it)

If not, a correction {∆T } of the nodal temperature vector is calculated in order to approximate the optimal solution. This correction is obtained by solving a linear system of the following type: (it)

[M ]

· {∆T }

(it)

(it)

= {R}

[5.2]

(it)

In this equation, [M ] is a matrix of dimension n × n depending on the solution method used, as will be seen in the next section. The nodal temperature vector is then updated as follows before being used during the iteration it + 1: {T }

(it+1)

(it)

= {T }

+ {∆T }

(it)

[5.3]

(it)

If, during these iterations, the norm of the vector {∆T } is less than a threshold T deﬁned by the user, most computer codes consider that the iterative process has converged. However, we must bear in mind that the only reliable convergence indicator is the residual norm value. As a matter of fact, from a mathematical point of view, (it) (it) the condition {∆T } ≤ T shows that the sequence made up with {T } is a Cauchy sequence. Moreover, we know that in Rn , any Cauchy sequence converges. (it) However, the criterion {∆T } ≤ T does not indicate that the sequence of (it) {T } is very close to its limit. In all cases, it would be wise to check that the value of the residual vector norm is small enough to be acceptable. (it)

It is important to note that the choice of the matrix [M ] in equation [5.2] does not interfere with the solution quality but only with the speed at which it can be obtained. In fact there is no method guaranteeing the convergence of the iterative process in all cases. Therefore, choosing the best method to solve [5.1] as effectively as possible is not simple. This is all the more difﬁcult as the number of iterations is (it) not the only criterion. The time required for assembling [M ] and solving linear system [5.2] has a part to play as well. The user will always choose a strategy for solving a given problem, according to his/her experience and the nature of this problem (non-linearities of thermal and physical properties, phase change problem, thermal radiation, etc.). The main methods are illustrated in Figure 5.2 for a problem with one degree of freedom. The reader may go into the detail of this point with the help of specialized books [DEN 76].

Non-linearities

R

Newton-Raphson method

modified Newton-Raphson method: tangent matrix kept constant during two iterations

R

T

R

147

T

Substitution method

BFGS method

R

T

T

Figure 5.2. Most common methods for the solution of a non-linear equation

5.1.2.1. Newton-Raphson method This method consists of determining the temperature correction from a ﬁrst order Taylor expansion of the residual in the vicinity of the current solution, which means that the residual can be differentiated close to the solution sought: (it)

Ri = R i

+

n ∂Ri j=1

∂Tj

(it)

Tj − Tj

In that expansion, the components of the matrix involved are the partial derivatives of the components of the residual vector, with respect to the components of the vector’s (it) nodal temperatures calculated for {T } = {T } . A linear approximation of the (it) residual close to the current solution is therefore obtained. The correction {∆T } is calculated in order to cancel this approximation of the residual: 1 (it)

{R}

+

∂R ∂T

2(it) · {∆T }

(it)

= {0}

148

Finite Element Simulation of Heat Transfer (it)

The method therefore consists of selecting as matrix [M ] the tangent matrix of the non-linear system to solve. This tangent matrix is obtained at each iteration by assembling element tangent matrices: [M ] =

m

T

[Ae ] · [M e ] · [Ae ]

with

e Mij =−

e=1

∂Rie ∂Tje

The Newton-Raphson method converges in a quadratic way. To demonstrate this, consider the non-linear scalar equation r(u) = 0, where r is a derivable function near the solution. Now apply Newton-Raphson’s iterative process. In this case, the recurrence relationship between two consecutive estimations gives: u(it+1) = u(it) −

dr (it) u with r(it) = r u(it) and r (it) = du

r(it) r (it)

Denoting by e(it) = u(it) − u the error obtained during the iteration (it), u being the exact solution, we obtain: e(it+1) = e(it) −

r(it) r (it)

Now assume the function r to be derivable twice near the solution. A development of r near u(it) then gives: 1 r(u) = r(it) + r (it) (u − u(it) ) + r (it) (u − u(it) )2 + · · · 2 2 1 = r(it) − r (it) e(it) + r (it) e(it) + · · · 2 =0 hence e(it) −

r(it) 1 (it) 2 r (it) e = + ··· 2 r (it) r (it)

i.e., if we restrict our operation to the second-order terms: e(it+1) =

1 (it) 2 r (it) e 2 r (it)

with r (it) =

d2 r (it) (u ) du2 2

Therefore, there is a constant k > 0 such that e(it+1) = ke(it) , hence the quadratic convergence.

Non-linearities

149

The Newton-Raphson method therefore converges with, in general, relatively few it iterations. However, the assembly of it requires matrix [M ] as well as the solution of a new linear system of type [5.2] at each iteration. It is therefore very costly from a practical point of view and must be reserved for strongly non-linear problems for which it has been impossible to reach convergence with another method. (it)

Variants of this method, called modiﬁed Newton methods, keep the matrix [M ] constant during a certain number of iterations, before recalculating it. The assembly (it) of [M ] is then performed only for certain iterations. Moreover, using a direct method based upon the matrix triangulation to solve linear system [5.2] can appear to be particularly efﬁcient insofar as triangulation is performed only when the matrix is modiﬁed. The adverse effect is that the convergence domain of these methods is restricted compared with the Newton-Raphson method. In addition, these methods require more iterations to converge than the Newton-Raphson method does. 5.1.2.2. Substitution method In a large number of situations, the residual vector can be written as follows: {R(T )} = {F (T )} − [M (T )] · {T }

[5.4]

Such is the case of steady state or transient heat conduction problems, for which we have: (steady state) [M (T )] = [K(T )] [M (T )] = [K(T )] − 1 [C(T )] ν∆t

(transient state)

5 6 (it) The substitution method consists of taking matrix M (T (it) ) for matrix [M ] . By combining equations [5.2], [5.3] and [5.4], we demonstrate that this method (it+1) comes down to choosing {T } as follows: (it+1)

{T }

B C−1 ' ( = M (T (it) ) · F (T (it) )

[5.5]

The substitution method can also be considered as a ﬁxed point method. Sometimes it is also called the Picard iterative method. Like the Newton-Raphson method, it requires the assembly and solution of a linear system at each iteration and is therefore very costly. Its interest lies in the

150

Finite Element Simulation of Heat Transfer (it)

fact that the matrix [M ] calculated in this way is generally symmetric whereas the tangent matrix calculated in the Newton-Raphson method is not necessarily so, which generates additional storage and solution costs. 5.1.2.3. Quasi-Newton methods Quasi-Newton methods generalize to non-linear equation systems the secant iterative method well-known for solving a non-linear equation of the type f (x) = 0. The principle of these methods is to replace the tangent matrix with an approximation which can be obtained more simply. The aim is to obtain convergence properties close to those of the Newton-Raphson method while avoiding the assembly and possibly the triangulation of a linear system at each iteration. Generally speaking, rather than construct an approximate tangent matrix, it is preferable to directly seek an approximation of its inverse. Therefore, we will seek a matrix [S](it) = ([M ](it) )−1 satisfying the following (secant) condition: {T }

(it)

(it−1)

− {T }

(it)

= [S]

(it−1) (it) · {R} − {R}

[5.6]

(it−1)

and On the other hand, we prescribe that two consecutive matrices [M ] (it) (it−1) (it) (it−1) [M ] differ only in the direction {∆T } = {T } − {T } . Therefore, we have: ∀ {U } such that ∆T (it)

[M ]

(it−1)

Allow {∆R}

· {U } = [M ]

(it−1)

= {R} [S]

(it)

(it−1)

(it)

− {R}

· {U } = 0,

(it−1)

[5.7]

· {U } (it)

and seek [S]

as follows:

(it) (it−1) · [S] = [I] + [A]

[5.8]

where [I] is the identity matrix. It is easy to demonstrate that condition [5.7] then becomes: (it−1)

∀ {U } such that ∆T

· {U } = 0,

[A]

(it)

· {U } = {0}

[5.9]

Non-linearities

151

which leads us to write: (it)

(it−1)

= {W }

[A] (it−1)

where {W }

· ∆T

(it−1)

is a vector which is still to be determined.

Considering equation [5.6] again, we obtain: (it−1)

{∆T }

(it−1) (it−1) (it−1) (it−1) · [S] = [I] + {W } · ∆T · {∆R}

from which we immediately derive: {W }

(it−1)

{∆T }

=

(it−1)

− [S]

(it−1)

· [S]

∆T

(it−1)

(it−1)

(it−1)

· {∆R}

[5.10]

(it−1)

· {∆R}

(it)

The correction {∆T } to make to the nodal temperature vector {T } written as follows, considering [5.2] and [5.8]: (it)

{∆T }

(it)

(it−1) (it−1) (it−1) (it) · [S] = [I] + {W } · ∆T · {R}

is then

[5.11]

Among the various existing quasi-Newton methods, the BFGS method is known (it) to be one of the most efﬁcient. Matrices [S] are calculated so that: (it−1)

is symmetric, then so is [S]

(it−1)

is positive deﬁnite, then so is [S]

– if [S] – if [S]

(it)

; (it)

(1)

By assuming [S] follows:

(it)

[S]

. (it)

to be positive deﬁnite, matrices [S]

can be sought as

(it−1) (it−1) (it−1) · [S] = [I] + {W } V [5.12]

(it−1) (it−1) · [I] + {V } W If the same course of action as above is taken, it is then demonstrated that: (it−1)

{W }

=

{∆T } (it−1)

∆R

(it−1)

· {∆T }

(it−1)

(it−1)

(it−1)

{V }

(it)

= {R}

−

∆R

∆T

· {∆T }

(it−1)

(it−1)

(it−1)

· {R}

(it−1)

{R}

152

Finite Element Simulation of Heat Transfer

The temperature correction {∆T } {∆T }

(it)

=

(it)

1 + @

is written as follows: 0

[I] + {W }

(k)

V

,

(k)

k=it−1

·

- it−1 @+

· [S]

(1)

0

(k)

[I] + {V }

(k)

W

,

[5.13] (it)

· {R}

k=1

In the equation above, it is obvious that the calculation of the temperature correction does not require a linear system solution or even the explicit calculation of (it) matrices [S] . This makes the method particularly efﬁcient in practical cases. If the computer code includes the BFGS method, it is most certainly the method to use ﬁrst and foremost to solve any non-linear problem. 5.1.3. Line search method (it)

Once {∆T } has been obtained by means of equation [5.2], the line search method consists of trying to cancel the residual in that direction. By considering the (it) 1D vectorial space generated by direction {∆T } , this comes down to canceling the projection of the residual {R} in that space, i.e. canceling the scalar function f of the scalar variable α as follows: f (α) =

n

(it)

∆Ti

(it)

Ri

+

{T }

(it)

(it)

+ α {∆T }

,

i=1

Function f being a non-linear function with real values, we seek α such that f (α) = 0 by means of a secant iterative method. At each iteration (k), a new value of α is obtained by: α(k) − α(k−1) f α(k) α(k+1) = α(k) − (k) (k−1) −f α f α The process stops when |f (α(k) )| ≤ η|f (0)|, or when |α(k+1) − α(k) | ≤ δ, the coefﬁcients η and δ being thresholds deﬁned by the user. Note that this method is substantially more costly as it requires a new estimation of the residual vector at each iteration (k) of the line search problem solution. This method is often combined with quasi-Newton or modiﬁed Newton methods.

Non-linearities

153

A similar method, called the under-relaxation method, is sometimes used to solve non-linear systems. It consists of reducing the modiﬁcation of the unknown to a (it) fraction of the quantity {∆T } calculated by equation [5.2], so as to satisfy an additional condition (for instance, all the components of the unknown must remain positive) or to better control strong variations in the solution. 5.2. Traditional non-linearities As mentioned earlier on in this chapter, traditional non-linearities in a thermal conduction problem are: – physical properties depending on the temperature; – ﬂux boundary conditions or a volumetric heat source depending on the temperature. With regard to physical properties, we saw in previous chapters that the problem becomes non-linear as soon as the thermal conductivity λ, the volumetric mass ρ, or the mass heat C, of the material is a function of the temperature. In the case of ﬂux q boundary conditions or the volumetric heat source Q, the problem remains linear as long as these very functions remain linear according to the temperature. There are other causes for the non-linearity of the problem. For example, the case of the thermal contact processed with a contact resistance most often leads to a non-linear problem. In this section, we will restrict our study to classic non-linearities. 5.2.1. Physical properties The case of temperature-dependent physical properties is relatively simple. In the case of a temperature-dependent thermal conductivity, λ(T ), the term involved in the element residual expression for steady state or transient cases is: −−→ −−→T e (λ) Ri = − grad (Ni ) · λ(T ) · grad(T ) dv Ωe

=−

j

Ωe

−−→ e −−→T e grad (Ni ) · λ(T ) · grad(Nj ) dv Tje

which is no longer a linear function of the vector {T e } including the node temperatures of element Ωe .

154

Finite Element Simulation of Heat Transfer

If, for example, a Newton-Raphson method is used for the solution, we obtain a tangent matrix including the following term: −−→ −−→ gradT (Nie ) · λ · grad(Nje ) dv (λ) ∂R Ωe (λ) Mij = − i e = ∂Tj −−→T e dλ −−→ + · grad(T )Nje dv grad (Ni ) · dT e Ω

In addition to a part similar to that already encountered in the linear case, we can note in this expression the appearance of a second term due to the fact that the thermal conductivity changes with the temperature. This term is not symmetric to indices i and j and so as to avoid the necessity of processing non-symmetric matrices that require appropriate storage and solution methods, the term is neglected. This comes down to considering the part related to thermal conductivity as a “substitution type” matrix. This approximation is all the more reasonable as the non-linearity introduced by the temperature-dependent conductivity is generally weak.

In the case of a transient calculation, the volumetric mass ρ and the speciﬁc heat capacity C of the material can also depend on the temperature. In that case, the product ρC involved in the element residual expression is a function of the temperature and the term involved is: (ρC) Ri =− Nie ρ(T )C(T )T˙ dv Ωe

If a generalized trapezoidal rule is used to integrate the differential equation system obtained between instants t and t + ∆t, we obtain: 1 1−ν ˙ T (t + ∆t) − T (t) T˙ (t + ∆t) = − T (t) + ν ν∆t

with

ν ∈ ]0, 1]

It follows from this that the residual parts in which the product ρC is involved are written as follows (if we write T = T (t + ∆t)): 1−ν Nie ρ(T )C(T )T˙ (t) dv ν e Ω (ρC) Ri (t + ∆t) = 1 − Nie ρ(T )C(T ) T − T (t) dv ν∆t Ωe

Non-linearities

155

If, for instance, a Newton-Raphson method is used to solve the non-linear equation system obtained, a tangent matrix is calculated in which the contribution of the product ρC will be involved as follows: (ρC)

(ρC)

Mij

∂Ri ∂Tje 1 e e ν∆t e ρCNi Nj dv Ω = 1−ν ˙ 1 e ∂(ρC) − T (t) + T − T (t) Nje dv + N i ∂T ν ν∆t e Ω =−

Note that this expression is composed of two symmetric terms at i and j. In fact, these terms can be calculated in the computer codes. Indeed, they produce no dissymmetry in the tangent matrix. Moreover, speciﬁc temperature-dependent heat changes may be relatively signiﬁcant. Such is the case when phase change latent heat is represented via an equivalent speciﬁc heat capacity (see section 5.2.3). 5.2.2. Flux or volumetric heat source boundary conditions The surface ﬂux densities q on the boundary ∂Ωq of the solid, as well as the volumetric heat source term Q, may be temperature-dependent in a non-linear way. In that case, the problem to solve is non-linear as well. Among the problems to be considered are: – the interactions between a coolant and a material. In the case of steel water quenching, such interactions are often modeled by an exchange coefﬁcient varying strongly in a non-linear way according to the temperature of the piece surface. To illustrate this variation, we can say that this coefﬁcient is low when the surface temperature is high (calefaction state), then increases abruptly when the burnout ﬁlm is pierced by liquid droplets coming to the surface (nucleate boiling state), and ﬁnally decreases again when the surface temperature goes under the boiling temperature of the water (convection state); – the material radiation effects toward an external environment, often expressed by an equivalent exchange coefﬁcient varying strongly according to the temperature. These effects are important when the solid temperature is typically above 300°C. They are covered in detail in Chapter 7; – the Curie transition point during steel induction heating. Above that temperature, the induction effect on the material volume changes substantially, so that the volumetric heat source used becomes a strongly non-linear function of the temperature.

156

Finite Element Simulation of Heat Transfer

When the ﬂux density on the boundary ∂Ωq of the solid and/or the volumetric heat source become non-linear functions of the temperature, i.e. q(T ) and Q(T ), the corresponding terms in the element residual also become non-linear: (Q) Ri = Nie Q(T )dv Ωe

(q)

Ri

Nie q(T )ds

= ∂Ωe ∩∂Ω

q

Using a tangent matrix for the solution then leads to considering the variation of q and Q with the temperature. This is expressed as follows: (Q)

(Q)

Mij

=−

∂Ri =− ∂Tje (q)

(q) Mij

∂R = − ie = − ∂Tj

Ωe

Nie

∂Q e N dv ∂T j

∂Ωe ∩∂Ω

Nie q

∂q e N ds ∂T j

These expressions create no particular problem for the solution. They keep the tangent matrix symmetric. For instance, in the case of radiation exchange in an inﬁnite environment, we have 4 q(T ) = σ(T∞ − T 4 ), where is the wall emissivity and σ the Stefan constant. The tangent matrix is written as follows: (q) Mij = 4σT 3 Nie Nje ds ∂Ωe ∩∂Ωq

2 A substitution matrix can also be deﬁned by writing q(T ) = σ(T∞ + T 2 )(T∞ + T )(T∞ − T ). This comes down to deﬁning an equivalent exchange coefﬁcient 2 H(T ) = σ(T∞ + T 2 )(T∞ + T ). The substitution matrix is written as follows: 2 (q) Mij = σ T∞ + T 2 T∞ + T Nie Nje ds ∂Ωe ∩∂Ωq

Using such a matrix is of no signiﬁcant interest in this case, insofar as the tangent matrix, which is more efﬁcient, is already symmetric.

Non-linearities

157

5.2.3. Modeling state changes State change problems are traditional in thermal science. To make things clear, a solidiﬁcation problem will be examined hereafter. When dealing with a pure substance, starting with the liquid state, solidiﬁcation occurs at a given temperature expressed as Tf . Then a Stefan problem has to be treated; the equations to solve are written as follows: – in the domain occupied by the liquid phase Ωl ρl Cl

−−→ ∂T − div λl grad(T ) = 0 ∂t

– in the domain occupied by the solid phase Ωs ρs Cs

−−→ ∂T − div λs grad(T ) = 0 ∂t

– at the liquid-solid interface T l = Ts = Tf λs

∂T ∂ns/l

− λl s

∂T ∂ns/l

[5.14]

= (ρL)s/l Vs/l

[5.15]

l

In the above equations, indices l and s correspond respectively to the liquid phase and to the solid phase, ns/l represents the norm at the liquid/solid interface from the solid to the liquid, (ρL)s/l is the fusion/solidiﬁcation latent heat per volume unit and Vs/l the normal component of the liquid/solid interface moving speed. The numerical solution of such a problem requires front tracking methods. These can be combined with various numerical methods like the boundary integral equation method, which is particularly interesting since only the mobile interface is meshed, or the ﬁnite element method. The ﬁnite element formulation of such a problem requires remeshing techniques. At each moment, the mesh is composed of 2 sub-domains, one for the liquid and the other for the solid, interconnected by a set of nodes located on the mobile interface. For each sub-domain, in addition to the classic capacitance and conduction terms, the problem formulation reveals a transport term (see Chapter 6) related to the movements of the nodes located on the mobile interface. The interface position is updated by solving equations [5.14] and [5.15].

158

Finite Element Simulation of Heat Transfer

However, for most metal alloys used in industry, solidiﬁcation does not occur at a given temperature but on a range [Ts , Tl ], Ts and Tl representing respectively the solidus and liquidus temperatures. The interface between the two phases is no longer clear. The transformation takes place in an intermediate zone called a mushy zone. The phase change latent heat can then be introduced as a source term as follows: Q = (ρL)s/l

dfs dt

[5.16]

where fs represents the volumetric fraction of the solid. In the solidiﬁcation case, fs is generally assumed to depend on the temperature only, as follows: 1 if T ≤ Ts fs = T − Tl if Ts < T < Tl Ts − Tl 0 if T ≥ Tl

If we assume that we operate in a constant stress environment, we can introduce the volumetric enthalpy H of the medium with:

T

H(T ) = (ρL)s/l 1 − fs +

ρCdT

[5.17]

Tref

It is obvious that equation [5.17] immediately gives: dH ˙ H˙ = T = −(ρL)s/l f˙s + ρC T˙ dT so that the heat equation with the source term [5.16] can now be written: −−→ H˙ − div λ · grad(T ) = 0

[5.18]

5.2.3.1. Equivalent speciﬁc heat method Starting with equation [5.17], we can formally deﬁne an equivalent speciﬁc heat with: Ceq =

(ρL)s/l dfs 1 dH =C− ρ dT ρ dT

[5.19]

Non-linearities

H

Ceq

Ts

159

Tl

T

Figure 5.3. Enthalpy and equivalent speciﬁc heat

In fact the equivalent speciﬁc heat is the usual speciﬁc heat modiﬁed for Ts < T < Tl so as to include the state change latent heat. Equation [5.18] can now be written in the usual way: − → −−→ ρCeq T˙ − div λ · grad(T ) = 0

[5.20]

The advantage of this method is its simplicity. The phase change is treated as a non-linearity of physical properties. However, it may create some numerical difﬁculties related to the abrupt changes of Ceq at temperatures Ts and Tl , as illustrated in Figure 5.3. Besides the real consideration of the latent heat requires an accurate integration of Ceq (T ) and, consequently, imposes very small time steps. One way of reducing these inaccuracies is to smooth the speciﬁc heat on a wider temperature interval [Ts , Tl ]. In that case however, we have to ensure that:

Tl

Ts

ρCeq dT = H Tl − H Ts

The time step chosen for the analysis will have to be small enough to integrate the above equation accurately. To avoid the problems mentioned above, some authors ([COM 74, LEW 87]) operate in a different way. Starting with the fact that enthalpy is a more regular function of the temperature than the speciﬁc heat, they suggest we estimate an equivalent speciﬁc heat from the enthalpy and temperature space variations.

160

Finite Element Simulation of Heat Transfer

A nodal approximation of enthalpy, similar in nature to that of temperature, is ﬁrst made from the enthalpy values at the nodes directly deduced from the temperature values: e

H=

n

→ Nie (− x )Hi

i=1

The relationship: ρCeq =

dH dT

is then approximated with one of the following formulae: 1 ∂H/∂xi ≈ 3 i=1 ∂T /∂xi 3

ρCeq

or otherwise, according to [LEM 81]: +;

3 i=1

ρCeq ≈ + ;

3 i=1

∂H/∂xi

∂T /∂xi

2 ,1/2

2 ,1/2

However, it has been demonstrated [MOR 78] that such an approximation was unnecessary and an estimation of ρCeq at instant n∆t with: ρCeq ≈

Hn − Hn−1 Tn − Tn−1

yields excellent results. Enthalpy solution methods are addressed here. In [DAL 86] a comparative study of the various methods for the solution of solidiﬁcation problems will be found. 5.2.3.2. Enthalpy solution method Actually, it is not necessary to use a speciﬁc heat. Rather than apply a time integration algorithm such as [4.26] or [4.27] to the term T˙ of equation [5.20], it is possible to treat the term H˙ of equation [5.18] directly. Equation [5.18] thus becomes, after applying an implicit Euler algorithm: − → −−→ H(T ) − H0 − div λ (T ) · grad(T ) = 0 ∆t

[5.21]

Non-linearities

161

an equation where H = H(T ) is the enthalpy at t+∆t and H0 = H(T0 ), the enthalpy at instant t. The ﬁnite element method applied to the solution of equation [5.21] then produces the following expression of the element residual: → −−→ −−→T e − Rie = − grad (Ni ) · λ (T ) · grad(T ) dv −

Ωe

Ωe

Nie

H(T ) − H0 dv ∆t

The enthalpy H is now directly involved in the expression of the element residual via the term: H(T ) − H0 (H) Ri = − dv [5.22] Nie ∆t e Ω Then the problem can be solved with one of the methods introduced in section 5.1.2. When a Newton-Raphson method is used, the contribution to the tangent matrix introduced by the enthalpy term is the following: (H)

(H)

Mij

=−

∂Ri 1 = ∂Tje ∆t

Ωe

Nie

dH(T ) e Nj dv dT

[5.23]

It is obvious that the difference between the enthalpy method and the equivalent ˙ The enthalpy speciﬁc heat method results from the discretization of the term H. change taken into consideration by the enthalpy method is accurate:

t+∆t

H˙ dt = H − H0

t

whereas it is only approximated with the equivalent speciﬁc heat method:

t+∆t

H˙ dt = Ceq (T ) T − T0

t

Therefore, the enthalpy method is more accurate. However, using an enthalpy approximation requires storing the enthalpy H, as well as H˙ when using an algorithm with 0 < ν < 1, at each integration point of

162

Finite Element Simulation of Heat Transfer

the elements. Besides, with traditional Gauss integration schemes, we may obtain oscillating solutions as the discrete maximum principle might not be satisﬁed. As seen in Chapter 4 concerning transient heat conduction, one way of eliminating these oscillations with ﬁrst order elements, is to use a lumped capacitance matrix. In the case of the enthalpy approximation, this comes down to use a scheme in which the integration points would be located at the element nodes to integrate terms [5.22] and [5.23]. We recommend the use of ﬁrst order ﬁnite elements and Euler implicit algorithm for state change problems. 5.3. A temperature-enthalpy formulation In fact, the methods presented previously do not allow us to deal with the isothermal phase change (Figure 5.4). Indeed, in this case, TM being the temperature of transformation, H(TM ) and dH dT (TM ) are not deﬁned. On the other hand, each enthalpy value has a corresponding unique temperature value. Let us denote by T = Θ(H) the function giving temperature from enthalpy. Therefore, the problem needs to be reformulated using enthalpy instead of temperature. With that aim, Droux ([DES 87, DRO 90]) proposes an explicit ﬁnite element procedure considering enthalpy as a nodal unknown: − −−→ → H − H0 − div λ (T0 ) · grad(T0 ) = 0 ∆t

[5.24]

where H0 and H represent enthalpy at time t and t+∆t respectively and T0 = Θ(H0 ).

H

TM Figure 5.4. Isothermal phase change

T

Non-linearities

163

The application of this explicit approach for modeling casting or welding processes leads to the use of a very large number of time steps for calculation, which can be reduced using an implicit algorithm: −−→ → dΘ − H − H0 − div λ (H) · grad(H) = 0 [5.25] ∆t dH The temperature does not appear in this formulation. Therefore, the boundary conditions need to be deﬁned from the enthalpy H. 5.3.1. Mathematical formulation To avoid the difﬁculties mentioned above, we suggest the use of formulation where both temperature and enthalpy are unknown ([FEU 07]). The problem is to ﬁnd the temperature ﬁeld and the enthalpy ﬁeld solution of: −−→ H˙ − div λ · grad(T ) = 0

in Ω

[5.26]

T − Θ(H) = 0 in Ω

[5.27]

with classical boundary conditions: T = Td T −λ

on ∂ΩT

∂T =q ∂n

[5.28]

on ∂Ωq

[5.29]

Temperature and enthalpy are expressed from a nodal approximation: T = N e · {T e }

'

H = N e · {H e } which also gives H˙ = N e · H˙ e

(

[5.30] [5.31]

Insofar as the same approximation scheme is used for both temperature and enthalpy and due to the strong non-linearity of the enthalpy-temperature relationship, equation [5.27] cannot be solved in a strong sense. Therefore, to obtain the ﬁnite element formulation of the problem, both equations [5.26] and [5.27] must be solved in a weak sense. Thus, both equations are classically ﬁrst multiplied by the shape functions and then integrated over the whole domain Ω. After integrating the ﬁrst equation by parts, including boundary conditions, and substituting the nodal

164

Finite Element Simulation of Heat Transfer

approximations into the Galerkin weighted residual form of [5.26] and [5.27], we obtain a non-linear semi-discrete system of equations giving the residual vector at each instant: ' ( {RT } = [K] · {T } + [M ] · H˙ − {F } = {0} [5.31a] {R } = [M ] · {T } − {G} = {0} H

{T } and {H} are the vectors containing the nodal values of temperature and enthalpy. Matrices [K], [M ] and vectors {F }, {G} are obtained from an assembly process of element quantities deﬁned as follows: ' −−→T e ( %−−→ e & [K e ] = grad (N ) λ grad(N ) dv Ωe

[M e ] =

Ωe

{N e } N e dv

{F } =

[5.32]

e

{N } q ds e

∂Ωe ∩∂Ωq

{Ge } =

Ωe

{N e } Θ(H)dv

Matrix [M ] is usually arbitrarily lumped at the nodes, resulting in a diagonal matrix [M ∗ ]. This can be achieved by applying a row-sum technique: ∗ = δij Mij

Mik

[5.33]

k

The non-linear semi-discrete equation system [5.31a] therefore becomes: ' ( {RT } = [K] · {T } + [M ∗ ] · H˙ − {F } = {0} {R } = [M ∗ ] · {T } − {G} = {0} H

[5.33a]

System [5.33a] is solved step by step in time. Let us apply for example an implicit Euler backward algorithm on the enthalpy: ˙ H = H0 + H∆t

[5.34]

Non-linearities

165

˙ the enthalpy rate at where H0 and H are the enthalpy at time t and t + ∆t and H, ˙ can time t + ∆t. At each time step t + ∆t, the nodal unknown values {T } and {H} be obtained using an iterative procedure like the Newton-Raphson method. Therefore, at each iteration it, we have to solve the following linear system of equations: .

(it) /(it)

(it) {∆T } {RT } ' ( · =− ∆H˙ [M ∗ ] −∆t [Q] {RH } [M ∗ ]

[K]

[5.35]

where [Q] results from the assembly of element matrices [Qe ] deﬁned by: ∂Θ e [Q ] = ({N e } N e )dv ∂H e Ω

[5.36]

The second equation of [5.35] gives: (it)

{∆T }

= [M ∗ ]

−1

, + ' ((it) (it) (it) · ∆t [Q] · ∆H˙ − {RH }

[5.37]

explicitly as a Because matrix [M ∗ ] is lumped, equation [5.37] gives {∆T } ˙ (it) . Taking account of equation [5.37], the ﬁrst equation of [5.35] function of {∆H} ˙ (it) as the solution of the following symmetric linear system of now gives {∆H} (it)

equations: + ∆t [K]

(it)

= [K]

−1

· [M ∗ ] (it)

(it)

· [Q] −1

· [M ∗ ]

, (it) + [M ∗ ] · ∆H˙ [5.38] (it)

· {RH }

− {RT }

(it)

˙ (it) has been determined, equation [5.37] gives {∆T } Once {∆H} solution is updated by setting:

(it+1)

{T } H˙

=

{T } H˙

(it) +

{∆T } ∆H˙

(it)

and the

(it) [5.39]

The iterative process is continued until convergence. One drawback of the methods using enthalpy as the unknown comes from the fact that enthalpy is not continuous at the boundary between two materials. Therefore, to perform calculations on multi-material models, which is necessary for example for

166

Finite Element Simulation of Heat Transfer

0

1

2

3

4

Ti me

0 -2 -4 -6 -8 -10 Analytical solution -12

∆t=0. 1

-14

∆t=0. 2

-16

∆t=0. 3

Te mper atur e

Figure 5.5. Temperature variation at x = 1 for different time steps (40 elements, from [FEU 07])

casting applications, it is essential to include thermal contact elements ensuring the continuity of temperature and an enthalpy jump at the interface between two different materials. 5.3.2. Example Figures 5.5 and 5.6 give the results of a non-dimensional example similar to that proposed by Morgan et al. ([MOR 78]) and consisting of a unidirectional isothermal phase change in a semi inﬁnite body. The speciﬁc heat (C), the mass density (ρ) and the thermal conductivity (λ) are set to 1. An isothermal phase change takes place at 0°C and the latent heat of transformation is assumed to be equal to 20. The length of the mesh, assumed to be equal to 4, has been chosen big enough to represent a semi-inﬁnite body for the ﬁrst seconds of simulation. Three meshes have been considered, consisting of a subdivision into 40, 20 and 8 8-noded hexahedral elements. The initial temperature and enthalpy at all the nodes are assumed to be equal to the temperature of transformation T = 0 and H = 20 respectively. For t > 0, the ∂Θ temperature at x = 0 is set equal to −20. To ensure a correct convergence, ∂H should not be less than or equal to zero. It is therefore assumed to be equal to = 10−4 for 0 ≤ H ≤ 20.

Non-linearities

0

0.5

1

1.5

2

2.5

167

x

0 -2 -4 -6 -8 -10 -12

Analytical profile

-14

Numerical results at nodes (40 elements)

-16

Numerical results at nodes (20 elements)

-18

Numerical results at nodes (8 elements)

-20

Temperature

Figure 5.6. Temperature proﬁle at t = 1 s for different meshes (∆t = 0.1 s, from [FEU 07])

The results obtained with the method presented above are compared with analytical results given in [CAR 59]. Figure 5.5 shows the inﬂuence of the time step on the computed temperature at x = 1 and Figure 5.6 demonstrates the inﬂuence of the mesh on the temperature proﬁle at time t = 1. The very good agreement between the analytical solution and the numerical results obtained with different time steps and mesh sizes shows the efﬁciency of the method.

This page intentionally left blank

Chapter 6

Transport Phenomena

6.1. Highlighting instabilities In this chapter, either a liquid medium or a mobile solid medium will be considered. The velocity ﬁeld in the material is assumed to be known. The case of a mobile solid, which is a new extension of Part 1 of this book, is frequent enough in practice. For instance, welding, cutting or surface thermal treatment processes may involve a heat source moving at a given velocity (laser beam, electron beam, etc.), or otherwise the whole solid is moving across heating and cooling areas (line thermal treatment, continuous casting, etc.). In all cases, the thermal problem will be addressed in a framework of reference related to the stresses (heat source, etc.). Therefore, the solid is moving. First, we will highlight the term to add to the thermal balance in order to take account for this movement. Then we will demonstrate by means of a simple 1D case that the traditional solution techniques produce spatial instabilities in the solution. In the last sections, a method commonly used in computer codes to solve these so-called convection-diffusion problems will be described. 6.1.1. Thermal balance → When the medium is moving, the temperature ﬁeld sought, T (− x , t), is written − → → − T ( x (t), t). Indeed, the position x of each material point is here a function of time. It follows that the temperature change over time is written as: − −−→ ∂T −−→ d→ x ∂T − T˙ = + grad(T ) · = +→ v · grad(T ) ∂t dt ∂t

170

Finite Element Simulation of Heat Transfer

→ In this equation, − v represents the velocity of the points of the solid. Of course, → this velocity is a priori a function of the position of point − x and time t. From this result, it is possible to gather the thermal balance which must be satisﬁed at any point of the solid Ω: −−→ −−→ ∂T → div λ · grad(T ) + Q − ρC − v · grad(T ) = ρC ∂t

in Ω

In this chapter, our study will be restricted to the calculation of the steady state. In the stress-related frame of reference (heat or cooling source), the steady state corresponds to the condition ∂T ∂t = 0. This ﬁnally produces the following equation: −−→ −−→ → div λ · grad(T ) + Q − ρC − v · grad(T ) = 0 in Ω

[6.1]

In this type of problem, the boundary conditions are similar to those of a steady state, where ﬂux q is prescribed on a boundary ∂Ωq of the solid, and temperature Td is known on the other portion ∂ΩT . The ﬁnite element method applied to this equation and these boundary conditions obviously leads to the calculation of an element residual {Re }. As in the transient state case, this residual will be linear according to temperature if: – the thermal conductivity λ and material product ρC are independent of temperature; – the volumetric heat source Q varies linearly according to the temperature: Q = aQ T + bQ ; – the surface ﬂux density q varies linearly according to the temperature: q = aq T + bq . In this case, the element residual is written {Re } = [K e ] . {T e } − {F e }, where the load vector {F e } and conductance matrix [K e ] are written as: −−→ e −−→T e (N ) · λ · grad(N ) dv − Nie aQ Nje dv grad i j e Kij =

Ωe

−

∂Ωq ∩Ωe

Nie aq Nje

Fie

= Ωe

Ωe

ds −

Ωe

→ Nie ρC − v

−−→ · grad(Nje ) dv

[6.2]

Nie bQ

dv + ∂Ωq ∩Ωe

Nie bq ds

[6.3]

Transport Phenomena

171

In these expressions, we can see that the element matrix [K e ] is not symmetric. The non-symmetric term is the transport term: −−→ → Tije = Nie ρC − v · grad(Nje ) dv Ωe

In a non-linear case, the tangent matrix involved would be non-symmetric too, because of the presence of the transport term. Moreover, this term cannot be neglected, as it is the only term taking into account the movement of the solid or ﬂuid. 6.1.2. Treating a simple case Applying the classic ﬁnite element method to the diffusion-convection problem produces instabilities. In order to highlight them, consider a ﬂuid with physical properties, λ, ρ and C, running at constant speed, v, in a pipe. The problem is assumed to be 1D, the ﬂowing direction is written x, and the temperature ﬁeld T (x) corresponding to the steady state is calculated. A temperature T = 0 is prescribed at the pipe entry (x = 0), and T = 1 at the exit (x = 1). The temperature ﬁeld T (x) is therefore the solution of the following boundary value problem: d2 T dT ρCv − λ =0 dx dx2

for 0 ≤ x ≤ 1

T (0) = 0 T (1) = 1 This problem has a simple analytical solution: T (x) =

1 − eαx 1 − eα

with α =

ρCv λ

[6.4]

In order to obtain a ﬁnite element solution, the weighted residual method is applied, at ﬁrst by introducing weighting functions ψ, then a partial integration involving a temperature second derivative so as to end up with the weak formulation of the continuous problem: Find T (x) such that T (0) = 0, T (1) = 1, and for any ψ: 1 1 dT dψ dT λ dx + ρCv ψ(x) dx = 0 dx 0 dx dx 0

172

Finite Element Simulation of Heat Transfer

Consider a mesh obtained by the regular division of the segment [0, 1] into n linear interpolation elements, of length h = 1/n. On each element Ωe = [xe , xe+1 ] (with e = 1, . . . , n) two linear interpolation functions, N1e = (xe+1 − x)/h (corresponding to the node with x-axis xe ) and N2e = (x − xe )/h (corresponding to the node with x-axis xe+1 ) are deﬁned. The element quantity calculation is then relatively simple. We obtain: dNje dNie dNje e Kij dx + ρCv dx =λ Nie (x) dx Ωe dx dx Ωe F e = 0 i

By analytically calculating the integrals involved, the following element quantities are obtained: . . / / λ 1 −1 ρCv −1 1 e + [K ] = h −1 1 2 −1 1

0 {F e } = 0 Likewise, the assembly is relatively simple in this case and produces the following matrices: 1 −1 0 −1 1 0 −1 2 −1 ρCv −1 0 1 λ .. .. [K] = + . . h 2 −1 2 −1 −1 0 1 0

−1

1

0

−1

1

and: 0 0 .. {F } = . 0 0 The generic equation at a node i such that 2 ≤ i ≤ n is written as: ∀i = 2, . . . , n,

ρCv λ − Ti−1 + 2Ti − Ti+1 + − Ti−1 + Ti+1 = 0 h 2

[6.5]

Transport Phenomena

173

This equation clearly reveals a Peclet number P related to an element: P =

ρCvh λ

[6.6]

Indeed, the solution of this generic equation can be written as: Ti+1 − Ti = A Ti − Ti−1 = · · · = Ai−1 T2 − T1

with

A=

1+ 1−

P 2 P 2

Coefﬁcient A deﬁned by this equation tends to inﬁnity for a Peclet number P = 2, and changes signs on either side of this value. Therefore, physically realistic solutions will be obtained only for values of A positive or equal to zero, i.e. for Peclet numbers P ≤ 2. For Peclet numbers greater than 2, the solution will produce spatial oscillations. By consecutively applying the conditions T1 = 0 and Tn+1 = 1 to the generic solution, we obtain: 1 − Ai−1 T2 Ti = Ai−2 + Ai−3 + · · · + A + 1 T2 = 1−A Tn+1 =

1 − An 1−A T2 = 1 =⇒ T2 = 1−A 1 − An

The ﬁnite element solution can ﬁnally be written in a general way according to Peclet number P : ∀i = 1, . . . , n + 1,

Ti =

1 − Ai−1 1 − An

with

A=

1+ 1−

P 2 P 2

[6.7]

Figure 6.1 illustrates the inﬂuence of Peclet number P on the solution of this example. To do so, a value α = ρCv = 30 was used and the variation of P was λ obtained by modifying the discretization parameter h by the consecutive use of 30, 15, then 10 elements. We can see that for values P < 2, the ﬁnite element approximate solution is relatively close to the analytical solution. On the contrary, when Peclet number P becomes greater than 2, the approximate solution produces spatial oscillations, due to the sign change of the coefﬁcient A. If P = 2, coefﬁcient A tends to inﬁnity. In this case, a complete solution produces temperatures equal to zero at all the mesh nodes, except for node number n + 1, where the condition Tn+1 = 1 is satisﬁed.

174

Finite Element Simulation of Heat Transfer

Figure 6.1. Temperature proﬁles obtained with P = 1, P = 2 and P = 3

6.2. Resolution techniques With the classic ﬁnite element resolution, the solution obtained is only valid for low Peclet numbers P . Now, this number depends on the material physical properties (λ, ρ and C), the solid or ﬂuid velocity v, and the length of the element h in its movement direction (equation [6.6]). For a given material and mesh, the classic ﬁnite element method is applicable only if the solid or ﬂuid velocity v is low enough. If not, ﬁner elements have to be used to reduce h, hence Peclet number P . Speciﬁc resolution techniques for this type of problem have been implemented. In the case of ﬁnite differences, it is the upwind technique, which is described in section 6.2.1. This technique numerically solves a diffusion-convection problem with a high Peclet number. To adapt it to the ﬁnite element method, it is possible to use a Petrov-Galerkin variational formulation. This method is termed the Streamline-Upwind-Petrov-Galerkin (SUPG) method. This formulation is described in section 6.2.2.

Transport Phenomena

175

6.2.1. Upwind technique Numerical mathematicians using the ﬁnite difference method are very familiar with the spatial oscillation phenomenon in diffusion-convection. In fact, in the example above, the equations obtained by means of this method with centered difference schemes and a similar discretization are: T1 = 0 ρCv λ Ti+1 − Ti−1 = 0 ∀i = 2, . . . , n, − 2 Ti+1 − 2Ti + Ti−1 + h 2h Tn+1 = 1 These equations are strictly identical to those obtained by the ﬁnite element method and therefore produce the same spatial oscillation problems for high Peclet numbers P . In order to rectify this, it is possible to write an off-centered difference scheme that estimates the ﬁrst and second derivatives involved in the physical model: dT Ti+1 − Ti = dx x=xi h - 0 Ti+2 − 2Ti+1 + Ti 1 dT dT d2 T = = − dx2 x=xi h dx x=xi+1 dx x=xi h2 Using this off-centered scheme on nodes i = 1, . . . , n − 1 produces the following equations: T1 = 0 ρCv λ T T + =0 ∀i = 1, . . . , n − 1, − − 2T + T − T i+2 i+1 i i+1 i 2 h h Tn+1 = 1 The solution is written as: ∀i = 1, . . . , n + 1,

Ti =

1 − Ai−1 1 − An

with A = 1 + P = 1 +

ρCvh λ

[6.8]

Figure 6.2 shows the solutions obtained with this method for Peclet numbers P equal to 1, 2 or 3 obtained as before (Figure 6.1). It is possible to verify that the solution no longer oscillates for the high values of P .

176

Finite Element Simulation of Heat Transfer

Figure 6.2. Temperature proﬁles obtained with an off-centered scheme

In order to compare the centered scheme (equation [6.7]) with the off-centered scheme (equation [6.8]), we analyze the generic equations Eic and Eid derived in both cases. We obtain: Eic =

1 λ − Ti−1 + 2Ti − Ti+1 + ρCv − Ti−1 + Ti+1 = 0 h 2

Eid =

λ − Ti−1 + 2Ti − Ti+1 + ρCv − Ti−1 + Ti = 0 h

The ﬁrst terms of Eic and Eid correspond to the discretizations of the conduction term, and the second terms to those of the convection term. If these expressions are analyzed in-depth, we can note that: 1 Eid = Eic + ρCv − Ti−1 + 2Ti − Ti+1 2 The off-centered scheme therefore corresponds to a centered scheme in which more weight would have been given to conduction. Indeed, we obtain Eid again by

Transport Phenomena

177

replacing inside Eic the thermal conduction λ with an equivalent conduction λ such that: P ρCvh =λ 1+ λ =λ 1+ 2λ 2 This result is used to establish the ﬁnite element formulations corresponding to the off-centered scheme. 6.2.2. SUPG method In order to apply the upwind technique to ﬁnite elements, it is possible to use a Petrov-Galerkin method, which uses discontinuous weighting functions on the element boundaries. In this section, the Petrov-Galerkin method will be illustrated on the simple 1D example of the previous sections. To do so, the weighted residual method is applied to the problem to solve to obtain: 1 0.387 for elements with 9 integration points and cl > 0.583 with elements with 4 integration points. However, this method does not enable us to accurately calculate of the view factor between two plane surfaces with a common edge (c = 0 in 7.5).

Radiation Exchanges in a Chamber

197

The calculation of the view factor between two spheres makes it possible to validate this technique in the case of a non-convex cavity. Consider an inner sphere of radius 1 and an outer sphere of radius 2, each sphere being meshed with 96 quadrangles with 4 integration points and 32 triangles with 1 integration point. The sum of the view factors being calculated, related to each element, remains in the ;f range of 0.9978 ≤ b=1 Fab ≤ 1.014, which produces an error of less than 1.4%. Moreover, the view factor sought, from the inner sphere to the outer sphere, is equal to 0.25024, compared with a theoretical factor equal to 0.25, which is very satisfactory. In practice, above all in the case of a non-convex cavity, it is generally preferable to rectify the view factors after calculation, so as to satisfy equation [7.1] for each of the plane surfaces. It is therefore advisable to take the following iterative (k) option. If Fab are the view factors obtained after k iterations, the following sequence is deﬁned: (k+1)

Fab

=

1 1 1 (k) Fab + (k) 2 φ(k) φb a

with

φ(k) a =

f

(k)

Fab

b=1

It is demonstrated that this sequence converges, and at the convergence, for any plane surface a, we ﬁnd that φa = 1. Other view factor calculation techniques may be used (contour integrals, Monte Carlo method, etc.). These calculations are often parts of specialized programs [CHU 82, SHA 85]. The reader will ﬁnd in [EME 91] a comparative study of the various calculation methods available. 7.1.3. Diffusion-radiation coupling Assume a cavity surface meshed with f skin ﬁnite elements, each element a representing a plane surface Sa . In order to take the action of heat radiation inside the cavity into consideration, it is sufﬁcient to add to the residual expression the contribution of the corresponding heat ﬂux density qa . This contribution is written as: f T {Rr } = [Aa ] · {Rra } with {Rra } = {N a } qa ds [7.9] Sa

a=1 a

a

where [A ] represents the location matrix and N the shape functions corresponding to element a. We could stop at this point. However, heat radiation problems produce signiﬁcant non-linearities and it is wise to insert a contribution due to these effects into the

198

Finite Element Simulation of Heat Transfer (it)

iterative matrix [M ] corresponding to the non-linear system to solve. The additional term {Rr } inﬂuences only the nodes attached to the cavity surface and only depends (it) on these node temperatures. The contribution [Mr ] to the matrix [M ] is therefore a square matrix of dimension nr × nr , where nr represents the cavity’s number of nodes. Consider {Tr } the temperature vector of the nodes attached to the cavity surface. In the case of radiative exchanges, the term {Rr } depends on all the temperatures of the cavity surface nodes, and hence on all the components of {Tr }. It follows that matrix [Mr ] is full and its assembly will signiﬁcantly increase the bandwidth of global matrix (it) [M ] . Another method often used is to assemble only the terms of [Mr ] included in (it) the band structure of matrix [M ] corresponding to diffusion phenomena. (it)

Two types of radiation matrix [Mr ] to assemble in [M ] iteration can be considered:

at each resolution

– a real tangent matrix: 1 [Mr ] = −

∂Rr ∂Tr

2

– a substitution matrix: 1

δRr [Mr ] = − δTr

2

7.1.3.1. Tangent matrix Reconsidering the expression of the heat ﬂux density qa given by equation [7.7], we obtain: - f 0 f ∂ 1 4 4 T [Mr ] = [Aa ] · {N a } Dab σ T a − T b ds ∂ {Tr } a=1 Sa Sa b=1, b =a

Then the mean temperature of a plane surface is deﬁned with the ﬁnite element method nodal approximation: D E 1 1 a Ta = T ds = N ds · {T a } [7.10] S a Sa Sa Sa

Radiation Exchanges in a Chamber

199

2 is 5 It 6 then easily demonstrated that the matrix [Mr ] consists of the element blocks ab Mr inﬂuencing only the nodes of the plane surfaces a and b:

[Mr ] =

f

6 5 6 T 5 [Aa ] · Mrab · Ab

with

a,b=1

3 4σT a aa [M ] = r Sa2

0 3

f

4 D

Sa

b=1, b =a

E

N a ds

Dab

N a ds

[7.11]

Sa

3 4 D E 3 5 ab 6 4σT b Dab a b N ds N ds Mr = − Sa Sb Sa Sb

for b = a

The tangent radiation matrix obtained by matrix assembly has the drawback of being non-symmetric. 7.1.3.2. Substitution matrix Expression [7.9] of {Rr } can be written taking into consideration expression [7.7]: f 1 {N a } {Rr } = − S a S a a=1

, + 2 2 Dab σ T a + T b T a + T b T a − T b ds ! " b=1, b =a

hab

The quantity hab in the previous equation is similar to an exchange coefﬁcient between two plane surfaces a and b. Given equation [7.10] deﬁning a plane surface mean temperature, {Rr } can be written as {Rr } = − [Mr ] · {Tr }, where the matrix [Mr ] is a substitution matrix which is written as: [Mr ] =

f

6 5 6 T 5 [Aa ] · Mrab · Ab

with

a,b=1

- f 0 3 4 D E 1 a a aa hab N ds N ds [Mr ] = Sa2 Sa Sa b =a

3 4 D E 5 ab 6 hab a b N ds N ds Mr = Sa Sb Sa Sb

for b = a

2. The possible temperature-dependence of heat emissivities is neglected for the calculation of the tangent operator.

200

Finite Element Simulation of Heat Transfer

6 5 As above, the matrices Mrab are element blocks inﬂuencing only the nodes attached to the plane surfaces a and b. It is easy to demonstrate that the symmetry of the Dab leads to that of the hab , and consequently the symmetry of [Mr ] thus deﬁned. This is a major advantage of this method. 7.2. Examples Two application examples are addressed in this section: – radiation between two walls, which will validate the method presented in the section above; – cylinder quenching, which uses symmetry relationships to simplify the global problem [BER 01]. 7.2.1. Radiation between two walls Consider the abstract 1D example in Figure 7.6. The walls are assumed to have a thermal conductivity λ = 1.5 W/m/K. The analytical solution is obtained, in the steady state, by equalizing the set of ﬂux densities involved, i.e. the conduction between the surfaces A and B, the conduction between the surfaces C and D, and the radiation between the surfaces B and C. If φr is the radiated ﬂux density, we obtain: −λ

TB − TA T D − TC = φr = −λ xB − xA xD − xC

It is easily demonstrated that if F represents the view factor between the surfaces B and C, then the radiation ﬂux density between these two surfaces is written: φr = Kσ TB4 − TA4 with K =

1−

F2

F B C 1 − B 1 − C

The mesh used for the calculations is regularly divided, every 0.05 m, in the direction x. Five solution strategies are examined: – NC (no radiation contribution) method: the conduction matrix alone is used to make the iterations; – PS (partial substitution) method: the radiation contribution is introduced as a substitution matrix, with consideration of only those terms belonging to the band structure corresponding to the diffusion ﬁnite elements; – TS (total substitution) method: the radiation contribution is taken into consideration by means of the symmetric substitution matrix;

Radiation Exchanges in a Chamber

201

face B εB face A θA=0°C

face D θD=1,000°C

h=2m

face C εC

xA = 0

x B = 0.5 m x C = 5.5 m x D = 5.75 m

Figure 7.6. Radiation exchange between two walls

– TM (tangent matrix) method: the radiation contribution is taken into consideration by means of the tangent matrix (non-symmetric); – BFGS method: the BFGS method is used. In all cases, the iterative calculation is interrupted as soon as the largest residual component (in absolute value) becomes less than 0.01 W, or as soon as that of the temperature difference is less than 0.01°C. Different couples (A , B ) are tested; all the results are in Table 7.1. As could be expected, it is with the real tangent matrix that convergence is fastest (in number of iterations). The substitution matrix produces convergence too, without using a non-symmetric solver. In practice, these methods very signiﬁcantly increase the bandwidth of the ﬁrst member matrix used to make the iterations and, as a consequence, the analysis cost. The conduction matrix alone does not produce convergence generally. Introducing a radiation partial contribution, without destroying the matrix band structure, has produced convergence in all cases (often with a large number of iterations, if truth be told). The BFGS method seems to be the most efﬁcient.

202

Finite Element Simulation of Heat Transfer

B

C

0.1

0.3

0.4

0.01

1

0.2

1

0.2

1

0.2

+0.0008T +0.0008T

TB (°C) TC (°C) method iterations 204.4

581.2

19.06

630.9

599.5

897.8

709.4

990.5

684.6

700.3

PC

14

SP

9

ST

4

MT

3

PC

∞

SP

41

ST

6

MT

4

PC

4

SP

4

ST

3

MT

3

PC

∞

SP

93

ST

5

MT

4

BFGS

12

PC

∞

SP

45

ST

7

MT

6

BFGS

13

Table 7.1. Work example – number of iterations at the convergence

The example considered here is much more difﬁcult than it ﬁrst appears. On the one hand, the calculation of a steady state is much more difﬁcult to make than that of a transient state where the phenomena are regularized over time. On the other hand, the example deals with two conductive regions which are only coupled by the radiative exchanges. This coupling is only inserted into the ﬁrst member matrix when a radiation contribution by means of a substitution matrix or a tangent matrix is considered. That is the reason why the methods using them converge in fewer iterations.

Radiation Exchanges in a Chamber

203

Generally speaking, it is recommended to treat diffusion-radiation problems with the BFGS method. Using a substitution matrix will be restricted to cases in which the latter does not converge. Using the tangent matrix, in addition to the hopeless cases in which none of the methods above would have produced a solution, will be recommended in cases where the ﬁrst member matrix is already non-symmetric in nature, for example, as is the case of diffusion-convection problems. 7.2.2. Cylinder quenching Consider here the case of ﬁve cylinders at the initial temperature of 900°C, which are cooled in a chamber whose walls and top surface are assumed to be at the constant temperature of 200°C, with an emissivity of 1 (black body), and whose bottom surface is at 60°C, with an emissivity of 0.6. The cylinder emissivity is assumed to be equal to 0.9. We added to the radiative exchanges convection exchanges on the cylinder surfaces, with a convection coefﬁcient h = 10 Wm−2 K−1 , and a temperature equal to that of the chamber walls, 200°C. The thermal properties of the cylinders are λ = 40 Wm−1 K−1 and ρC = 4.7.106 Jm−3 K−1 . The numerical simulation of this problem takes the symmetries into consideration by means of the formulation described in detail in [BER 01]. In order to illustrate the inﬂuence of symmetries on calculation times, we have made two calculations with two meshes presented in Figure 7.7. The ﬁrst image represents the complete structure, without the chamber walls. It altogether consists of 2, 761 nodes, 2, 400 3D isoparametric elements. Furthermore, 1, 720 2D elements were added to illustrate the radiative exchanges in the chamber. In the second image, the problem symmetries lead us to model only one-eighth of the whole structure. The mesh used for the calculation then includes only 437 nodes, 300 3D elements and 215 2D elements. The view factors Fab between each surface and the chamber walls are calculated only once at the beginning of the analysis. The transient calculation is made with a time step of 20 s between the initial moment and 60 s, then a step of 60 s between moments 60 s and 600 s. The solution was calculated with a tangent matrix restricted to the assembly of the terms [Mraa ] in equation [7.11]. The convergence was set at 10−5 W on the residual maximum component (in absolute value), or at 10−2 °C on the maximum temperature variation (in absolute value) between two iterations. Figure 7.8 gives the temperature distribution on the cylinder surfaces at t = 240 s. We see that the reduced model gives the same distribution as the full model, whereas the calculation times were as follows:

204

Finite Element Simulation of Heat Transfer

Figure 7.7. Mesh used for the calculation, from [BER 01]

Figure 7.8. Temperature distribution at t = 240 s, excerpt from [BER 01]

Radiation Exchanges in a Chamber

205

– on the full model, 669 s for the view factors, 25 s for the calculation of the radiation contribution to the tangent matrix, and 54 s for the transient thermal calculation, – on the reduced model (one eighth on account of symmetries), 79 s for the view factors, 0.07 s for the calculation of the radiation contribution to the tangent matrix, and 5 s for the transient thermal calculation. These results show how important it is to use the symmetries of a problem when it is modeled, with a view to its numerical simulation. Additionally, using a partial tangent matrix (radiation contribution restricted to the terms [Mraa ] in equation [7.11]) was not very prejudicial, for the convergence of the solution process (at each time step) was obtained with a maximum of ﬁve iterations. Finally, the following points can be noted: – as expected, the calculation of the view factors takes nine times as long as when the problem symmetries are not considered. Indeed, it is demonstrated that the method used here to take the symmetries into consideration reduces the time for the calculation of the view factors by a ratio Nsym + 1 where Nsym is the number of times when the simulation model has to be duplicated to obtain the real model (here, Nsym = 8); – in both cases, the calculation time of the radiation contribution to the tangent matrix is negligible compared with that of the view factor calculation; – of course, the transient thermal calculation time is signiﬁcantly reduced when the problem symmetries are used.

This page intentionally left blank

Chapter 8

Fluid-Structure Coupling in a Pipe

8.1. Modeling the ß uid Many industrial applications are related to heat exchanges. Modeling such devices requires the correct consideration of the heat exchanges between the heat-carrying ﬂuid and the wall. This problem is all the more difﬁcult since the temperature of the ﬂuid, because of its exchanges with the wall, changes as it moves along. Modeling must therefore make it possible to calculate these temperature changes and take into consideration the ﬂuid local temperature in the complex heat exchanges occurring with the wall. This is a presentation of the physical interaction model and its mathematical formulation, leading axiomatically to a use of the ﬁnite element method. Finally, we will deal with the elements creating the ﬂuid-structure coupling [BER 98]. 8.1.1. Physical model and mathematical formulation Modeling consists of representing the heat-carrying ﬂuid with a set of 1D elements and solving a diffusion-convection problem in these elements by means of the technique introduced in Chapter 6. The heat exchanges between the ﬂuid and the wall are supported by 2D elements and involve the ﬂuid temperature and wall temperature. The pipe is therefore subdivided into macro-elements. Each macro-element combines: – a 2-noded 1D ﬂuid element; – a set of 2D elements representing the structure wall and supporting the heat exchange conditions between the ﬂuid and the wall (exchange elements).

208

Finite Element Simulation of Heat Transfer

2D element supporting heat exchange

fluid element macro-element

Figure 8.1. Description of the pipe

The macro-element mathematical formulation is based upon the following assumptions: – A1: the ﬂuid ﬂow is assumed to be steady and 1D. The ﬂow direction is given by the direction of the elements representing the ﬂuid; – A2: the ﬂuid mass ﬂow m ˙ is assumed to be constant along the pipe, but may possibly vary over time; – A3: inside each macro-element, heat exchanges can only occur between the ﬂuid element and the set of exchange elements corresponding to the macro-element. The heat ﬂux density received locally by the wall can thus be written as: φ T p , T f = h T f − Tp

[8.1]

where Tp , T f and h represent respectively the wall temperature, the ﬂuid mean temperature (in the ﬂuid element) and the exchange coefﬁcient between ﬂuid and wall. In practice, the heat ﬂux density received locally by the wall will be calculated at each integration point of the exchange elements. The wall temperature under consideration will therefore be the one calculated at each integration point. Moreover, the expression of h depends on the heat exchange conditions (natural convection, forced convection or mixed convection) and may depend on Tp and T f . The expression of h results from a Nusselt number by means of various empirical relationships between numbers without dimension, which are functions of the convection type (natural, forced, mixed, etc.), and possible phase changes in the heat-carrying ﬂuid (vaporization, etc.), among others.

Fluid-Structure Coupling in a Pipe

The heat equation, considering assumptions A1 and A2, is written as: ∂ ∂Tf ∂Tf m ˙ − λf −q =0 Cf Sf ∂l ∂l ∂l

209

[8.2]

In this equation, l represents the curvilinear x-axis, Sf the pipe section, Cf the speciﬁc heat, λf the ﬂuid thermal conductivity, and q the heat ﬂux received per ﬂuid volume unit. The quantity q is assumed to be constant in each ﬂuid element. As seen in Chapter 6, using a classic Galerkin method to formulate the ﬁnite elements from equation [8.2] produces spatial oscillations of the solution when the Peclet number is greater than a critical value (2 for ﬁrst order elements). It is obvious that such a limitation is not acceptable for modeling heat exchangers. To avoid these oscillations, a SUPG method can be used. This approach produces the following equation system for the ﬂuid:

[Kf ] + [Uf ] + [Df ] · {Tf } = {Ff }

[8.3]

In this equation, [Kf ] represents the classic conductance matrix (obtained as assembly of the element conductance matrices in the ﬂuid), {Tf } represents the ﬂuid nodal temperature vector, whereas [Uf ], [Df ] and {Ff } are obtained by assembling on the ﬂuid elements Ωef the following element quantities:

Ufe ij Dfe ij Ffe i

∂Nje dl Nie + Hie mC ˙ f ∂l Ωef ∂Nje ∂ λf dl = −Sf Hie ∂l ∂l Ωef e Ni + Hie q dl = Sf =

Ωef

[8.4]

[8.5] [8.6]

The functions Nie and Hie appearing in these equations represent respectively the shape function and the disturbance function corresponding to node i of element e. It can be noted that, for ﬁrst order elements and a constant thermal conductivity, [Df ] = 0. Considering assumption A3 and equation [8.1], the thermal balance inside a macro-element is written as follows: qVf + h T f − Tp ds = 0 exchange elements

element

210

Finite Element Simulation of Heat Transfer

The ﬁrst term of this equation represents the heat ﬂux received by the ﬂuid (Vf is the ﬂuid element volume). The second term represents the heat ﬂux received by the structure. From this equation the expression of the quantity q is directly derived: 1 q=− [8.7] h T f − Tp ds Vf element exchange elements

8.1.2. Modeling the coupling The equation system to solve for the solid structure can, in the steady state, be written: [Ks ] · {Ts } = {Fs }

[8.8]

In this equation, [Ks ] is the traditional thermal conductance matrix of the solid, {Ts } is the unknown (temperature) vector corresponding to the solid structure and {Fs } is the load vector. If, for convenience, it is assumed that the solid structure exchanges heat only with the ﬂuid ﬂowing in the pipe, it is possible to write: {Fs } =

5 m−e 6T m−e A · Fs

[8.9]

macro-elements m − e

In this equation, matrix [Am−e ] is the assembly matrix described in the ﬁrst chapter of this book, applied to the macro-element m − e. Load vector {Fs } therefore results from the contributions of all the macro-elements m − e, contributions which can be written as follows: m−e T Fs = [Ae ] {Qes } [8.10] exchange elements e

Again, in this equation, assembly matrix [Ae ] is that described in Chapter 1. The components Qesi of element term {Qes } are ﬁnally written: e [8.11] Qsi = Nie h T f − Tp ds Ωe

Inside each exchange element, Tp is given by the classic ﬁnite element approximation, where Tje is the temperature of node number j of element e (local numbering): e

− Tp (→ x)=

n j=1

→ Nje (− x )Tje

Fluid-Structure Coupling in a Pipe

211

T f is the ﬂuid’s mean temperature. It is obtained, like the average of temperatures Tf 1 and Tf 2 , at the two nodes of the ﬂuid element: Tf =

T f 1 + Tf 2 2

The non-linear problem composed of equations [8.3] and [8.8] coupled by equations [8.6], [8.7] and [8.11] can be solved by means of a Newton-Raphson iterative method. Equations [8.3] and [8.8] are written as: /

. Fs Ks Ts Rs 0 = − · [8.12] {R} = Rf Ff Tf 0 Kf + Uf + Df At each resolution iteration (k), a better solution is obtained by solving a linear system (see Chapter 5). The process is repeated until the norm (Euclidian, for instance) of the residual {R} becomes less than the desired accuracy. Equation [8.12] can also be written: . /

5 m−e 6T m−e Ks 0 Ts {R} = A R − · T 0 0 f macro-elements In this equation, {Rm−e } represents the contribution of the macro-element m − e to the residual:

. /

m−e Fsm−e 0 0 Tsm−e R = − · [8.13] Ffm−e Tfm−e 0 Kfm−e + Ufm−e + Dfm−e In this expression, {Fsm−e } is deﬁned by equation [8.10], whereas {Ffm−e }, [Kfm−e ], [Ufm−e ] and [Dfm−e ] are the load vector and element matrices of the only ﬂuid element corresponding to the macro-element m − e. The terms [Ufm−e ], [Dfm−e ] and {Ffm−e } are given by equations [8.4], [8.5] and [8.6]. Finally, the terms {Tsm−e } and {Tfm−e } are the restrictions of the vectors {Ts } and {Tf } at the only nodes connected to the macro-element m − e. The contribution [K m−e ] of each macro-element to the tangent matrix [KT ] of the global system is directly derived from equation [8.13]:

∂Fsm−e − m−e 5 m−e 6 ∂Ts = KT m−e ∂Ff − m−e ∂Ts

−

∂Fsm−e ∂Tfm−e

Kfm−e + Ufm−e + Dfm−e −

∂Ffm−e ∂Tfm−e

212

Finite Element Simulation of Heat Transfer

6 5 In this equation, it can be noted that the matrix KTm−e is not symmetric. 8.2. Example The example given here is that of the auxiliary sprinkling line of a vapor generator in a 1,300 MW PWR nuclear plant. It is a security system designed to remedy a failure in the primary pumps, by injecting water to keep the pressure constant and equal to 17.13 bars in the circuit. We ﬁrst describe the physical and geometric modeling used, then analyze a few results. 8.2.1. Physical and geometric modeling The geometry is very simple here. It is a 51.1 m long pipe modeled with 1,200 volumetric elements. The exchange with the ﬂuid is represented by 100 ﬂuid-structure coupling macro-elements (Figure 8.2). The geometric and physical speciﬁcations used for the simulation are the following, considering that the pipe is assumed to be adiabatic on the outer wall: – pipe dimensions: 51.1 m long, 60.3 mm in internal diameter, 8.7 mm thick, – initial temperature: 20°C,

51.1 m

Figure 8.2. Sprinkling pipe mesh

Fluid-Structure Coupling in a Pipe

213

– pipe speciﬁcations: conductivity 15 W/m/°C, volumetric mass 7,800 Kg/m3 , speciﬁc heat capacity 500 J/Kg/°C, – water speciﬁcations: conductivity 0.603 W/m/°C, volumetric mass 1,000 Kg/m3 , speciﬁc heat capacity 4,180 J/Kg/°C, – mass ﬂow rate: 12 tonnes/hour, – exchange coefﬁcient: 10,400 W/m2 /°C. 8.2.2. Results

Temperature (°C)

Figure 8.3 gives the water temperature variation in the pipe, over time, at the entry and exit. It can be noted that the water injected under pressure at the entry reaches at the exit a temperature of 280°C after about 150 s.

in

out

time (s)

Figure 8.3. Temperature variation over time, at the sprinkler entry and exit

This page intentionally left blank

Chapter 9

Thermometallurgical Coupling

9.1. Modeling phase changes This chapter deals with the interactions between heat exchanges and phase changes in a solid material. The ﬁrst section is a description of the most common models used to describe the variations of phase proportions in a material undergoing thermal cycles [LEB 84, FER 85, INO 85]. To do so, differential equations governing each phase change are ﬁgured ﬁrst, followed by the method to integrate them numerically along any thermal path. Eventually, the case of multiple transformations which may occur simultaneously will be dealt with. 9.1.1. Rate of phase changes Solid state phase change models usually distinguish two types of transformation: – those governed by diffusion, described by Avrami kinetics [AVR 40a, AVR 40b, AVR 41], – those of the martensitic type, represented by the Koistinen-Marburger law [KOI 59]. 9.1.1.1. Avrami kinetics Avrami kinetics used to represent diffusion phase changes result from the extended volume concept [AVR 40a, AVR 40b, AVR 41]. In Figure 9.1, the extended volume Ve is obtained by considering that the new phase germination may occur anywhere in the material. Its variation velocity is therefore proportional to the new phase germ creation rate, N˙ , and the mean volume of these germs, vg . The variation velocity of the phase

216

Finite Element Simulation of Heat Transfer

Figure 9.1. Extended volume concept

proportion formed is in turn proportional to the extended volume variation velocity V˙ e , as well as to the phase proportion left for transformation. Thus, we obtain: p˙ = (P − p)V˙ e

In this equation, P designates the maximum phase proportion that can be formed, i.e. that which will be produced at the end of the transformation process.

The classic form of Avrami kinetics is obtained with the previous equation. For this purpose, it has to be integrated at a constant temperature (temperature contour path), and use a power law for the extended volume variation velocity: n−1 n

V˙ e = nf Ve

n =⇒ p = P 1 − e−(f t)

Here, f and n are parameters of the phase change kinetics, at the temperature under consideration. If the same expression is kept for V˙e , it is also possible to deﬁne with the previous equations the variation velocity p˙ of the phase proportion p. This variation speed is written as follows, considering that the parameters P , f and n may be temperature-dependent: 1 p˙ = nf (P − p) ln

P P −p

2 n−1 n [9.1]

Thermometallurgical Coupling

217

9.1.1.2. Martensitic kinetics The martensitic transformation occurs instantly. Martensitic kinetics therefore involve only the temperature, and are represented by the Koistinen-Marburger law [KOI 59]. According to this law, the (martensitic) phase proportion p formed at temperature T is given by: p(T ) =

0 1 − eb(T −Ms )

for T > Ms for T ≤ Ms

Here, b is a homogenous kinetic factor, unlike temperature, and Ms (martensite start) is the temperature at the beginning of the transformation. These coefﬁcients may depend on the material. For instance, in the case of steel, Ms is a function of the transformed austenite carbon content. As with Avrami kinetics, martensitic transformation kinetics can be ﬁgured in the differential form. This is done by deriving the previous equation with respect to time. We obtain: 0 for T > Ms or T˙ ≥ 0 p˙ = [9.2] −b(1 − p)T˙ otherwise It can be noted that this equation can be written like equation [9.1]. To do so, it is sufﬁcient to choose in this equation n = 1, P = 1, and a frequency f depending on the thermal path: 0 when T > Ms or T˙ ≥ 0 f= −bT˙ otherwise 9.1.2. Numerical integration Different methods can be used to integrate equations [9.1] and [9.2]. For instance, in the case of Avrami kinetics, the thermal path can be divided into a series of temperature contour plateaus, of length ∆t, along which it is possible to apply an integrated form of the kinetics (Figure 9.2 [PUM 48, FER 85]). However, it is also possible to perform a classic numerical integration, with time steps ∆t, with a second order Runge-Kutta method (Figure 9.3 [LEB 84]). In the case of Figure 9.2, a dummy time t∗ is introduced to take into consideration the phase transformation already formed at the beginning of the time increment. The

218

Finite Element Simulation of Heat Transfer isothermal evolution p(t) phase proportion (p) time (t)

t

3

4

t+∆t

2

1

t*

t*+∆t thermal path

T(t) T(t+∆t) temperature (T)

Figure 9.2. Numerical integration of a phase proportion along a thermal path with the dummy time concept

value of t∗ is the time that would have been required to form a phase proportion p(t) from zero, at temperature T (t + ∆t) of the temperature contour plateau (point 2 of Figure 9.2). This dummy time is obtained in the following form, all the parameters being estimated at temperature T (t + ∆t): ∗ n p(t) = P 1 − e(f t ) or

t∗ =

1 2 n1 P 1 ln f P − p(t)

The dummy time t∗ thus obtained is then used as the starting point of a temperature contour transformation occurring at a temperature T (t + ∆t). The ﬁnal phase proportion is therefore obtained analytically: ∗ n p(t + ∆t) = P 1 − e(f (t +∆t)) In the case of Figure 9.3, intermediate points are introduced by the classic second order Runge-Kutta scheme to integrate equations [9.1] or [9.2] along the thermal path T (t). Point 2 is obtained by choosing a constant variation velocity of the phase proportion, on a temperature contour variation curve at temperature T (t). This point is used to construct point 2 , which represents a phase proportion similar to that of point 2, but on the temperature contour variation curve of p corresponding to temperature T (t + ∆t). Point 3 is then obtained like point 2, but with the variation velocity of p obtained at point 2 . Finally, the last point, 4, is obtained by calculating the mean value of the phase proportions obtained at points 2 and 3. Note that in Figure 9.3 this point is not located on the phase proportion or temperature contour variation curves. For

Thermometallurgical Coupling

219

isothermal evolution p(t) phase proportion (p) time (t) 2 2’ t

3

t+∆t

4

1 thermal path

T(t) T(t+∆t) temperature (T)

Figure 9.3. Numerical integration of a phase proportion along a thermal path with a second order Runge-Kutta scheme

instance, let p(p, ˙ t) be the phase proportion variation velocity obtained by equation [9.1], for a common proportion p, and with parameters P , f and n estimated at instant t of the thermal path (at temperature T (t)). The second order Runge-Kutta method then gives the phase proportion at instant t + ∆t by estimating consecutively the following quantities: p˙1 = p(p(t), ˙ t) p2 = p(t) + p˙1 ∆t ˙ 2 , t + ∆t) p˙2 = p(p p3 = p(t) + p˙2 ∆t

then p(t + ∆t) =

1 p2 + p3 2

This can be written globally as: p(t + ∆t) = p(t) +

1 p(p(t), ˙ t) + p(p(t) ˙ + p(p(t), ˙ t)∆t, t + ∆t) 2

The Runge-Kutta method makes it possible to integrate, with the same method, the Avrami and martensitic kinetics, providing that the latter are formulated like Avrami kinetics with n = 1, P = 1 and a frequency f depending on the thermal path. Hereafter, therefore, general formulation [9.1] will be used for all possible phase changes. If it is of the Avrami kinetics type, the parameters n, P and f of this formulation will be temperature-dependent. If it is the martensitic type kinetics, we will choose n = 1, P = 1, and a frequency f depending on the temperature variation velocity (heating or cooling) and the value of Ms (see section 9.1.1).

220

Finite Element Simulation of Heat Transfer

9.1.3. The case of several phase changes Actually, several different phases may be present in the material. For instance, in low-carbon steel, we can ﬁnd austenite, ferrite, perlite, bainite and martensite, i.e. ﬁve phases, and transformations between these phases may occur in a similar volume element. In that case, if pi designates the phase’s common proportion i, then it is possible to deﬁne a driving force Eij of the transformation i → j of the form Eij = Pij (pi + pj ) − pj , where Pij is the maximal proportion of phase j which can be obtained from phase i. It should be noted at this point that, if the driving force Eij is positive, then the driving force Eji (corresponding to the transformation j → i) is negative. These driving forces therefore indicate the transformation direction between phases i and j. The problem of transformations between several phases can now be solved [LEB 84, FOR 00b]: – for each transformation i → j, by substituting p into equation [9.1],

pj pi +pj

in differential

– by considering that the variation of the phase proportion j, p˙j , results from all the possible transformations i → j and j → i, which are assumed to be independent of one another. With this method, it is possible to write: p˙ j = Aij

[9.3]

i =j

where Aij can be considered to be the reaction activity between phases i and j. This activity is obtained from equation [9.1] as: 1 2 nnij −1 ij p j if Eij > 0 (reaction i → j) nij fij Eij ln 1 + Eij 1 2 nnji −1 Aij = [9.4] ji pi ln 1 + f E if E > 0 (reaction j → i) −n ji ji ji ji Eji 0 otherwise

The parameters Pij , nij and fij of this equation depend on the transformation under consideration. Note ﬁnally that the activities Aij given by equation [9.4] make up an antisymmetric matrix. Indeed, as Aij represents the phase proportion i which is transformed into phase j per time unit, we obtain Aij + Aji = 0. One of the

Thermometallurgical Coupling

221

consequences is that the sum of all the matrix terms is zero, which expresses the conservation of the phase proportions since the sum of their variation velocities is equal to zero. Another model was suggested for steel and inserted into various calculation software [WAE 94]. By making p1 , p2 and p3 the common proportions of ferrite, perlite and bainite, and Ms the temperature of initial martensite transformation, for these phases the model equations are written as: p˙1 = f1 T, T˙ , p1 , p2 , p3 , Ms , d p˙2 = f2 T, T˙ , p1 , p2 , p3 , Ms , d p˙3 = f3 T, T˙ , p1 , p2 , p3 , Ms , d M = M p , p , p , d s s 1 2 3 Martensitic transformation is described by equation [9.2]. Functions f1 , f2 , f3 and Ms are not made explicit, but calculated point after point from the states (T, T˙ , p1 , p2 , p3 , Ms , d) corresponding to the continuous cooling transformation diagram (or CCT diagram) of the shade treated. As a consequence, the model is very close to this diagram and is no longer very accurate when the cooling conditions depart from those of the diagram. Its major advantage, however, is that it is easy to identify. 9.1.4. Modeling the coupling Strong coupling between thermal and metallurgical phenomena is relatively well-known. Phase changes are time- and temperature-dependent, and produce in turn temperature changes (exothermic and endothermic reactions). Generally speaking, the heat equation must be solved to obtain the temperature ﬁeld in a domain Ω subjected to temperature or ﬂux density thermal boundary conditions. This problem can be formulated as follows: −−→ ˙ div λ · grad(T ) + Q − ρH = 0 in Ω [9.5] T = Td on ∂ΩT with ∂Ω = ∂ΩT ∪ ∂Ωq − −−→ → λ · grad(T ) · n = q on ∂Ω q

In this equation, Q is an internal heat source, q a ﬂux surface density entering Ω, and λ, ρ and H represent respectively the thermal conductivity, the volumetric

222

Finite Element Simulation of Heat Transfer

mass and the speciﬁc enthalpy of the material. The material properties λ, ρ and H are temperature- and phase proportion-dependent. They are generally obtained from linear mixture laws on the corresponding properties of each phase i, λi , ρi and Hi which are temperature-dependent only. If N is the total number of phases, we therefore write: N pi λi (T ) λ p1 , . . . , pN , T = i=1 N ρ p1 , . . . , pN , T = pi ρi (T ) i=1 N pi Hi (T ) H p1 , . . . , pN , T =

[9.6]

i=1

It follows that the enthalpy variation velocity involved in heat equation [9.5] is written: H˙ = T˙

N i=1

pi Ci +

N i=1

p˙i Hi

with

Ci =

dHi dT

[9.7]

In this equation, Ci represents the heat capacity of phase i, and the terms p˙i can be obtained from equations [9.3] and [9.4]. The ﬁrst term of equation [9.7] involves ;N the phase mixture heat capacity, i=1 pi Ci . The second term represents the latent heat effects. Indeed, the curves Hi (T ) giving the enthalpy of the phases i according to the temperature, are different, not only in their slopes, representing the phase heat capacities, but also in their y-axes at the origin, which includes these latent heat effects. This is illustrated by Figure 9.4. Inserting equation [9.7] into [9.5] makes it possible to treat heat exchange problems, including phase changes, on complex geometries. For this purpose, the ﬁnite element method can be used. Before showing a few examples, we will deal with the case of phase transformation diagrams, which can be simulated to validate the parameters included in the expression of the activities of equation [9.4]. 9.2. Examples In this section, we give a few examples of numerical process simulations in which thermal and metallurgical aspects are strongly coupled. The initial focus will be on phase transformation diagrams, which are used to obtain the parameters of the phase change laws used. Later, we give the example of the simulation of steel quenching.

Thermometallurgical Coupling

223

H (105 J/kg) λ (W/m/°C) γ→α

8

transformation

60

γ phase

6 α phase

40

γ phase

20

4 2

300 600 900 1200 1500 T (°C)

α phase

latent heat due to γ→α transformation

300 600 900 1200 1500 T (°C)

Figure 9.4. Typical variation of the thermal conductivity λ and the enthalpy H of phases α (ferrite) and phase γ (austenite) in low carbon-steel, according to temperature

9.2.1. Phase transformation diagrams Metal and alloy transformation diagrams give the boundaries of the stability and phase change domains. Time is on the x-axis and temperature on the y-axis. There are two major types of metal and alloy phase transformation diagrams. IT diagrams (isothermal transformation diagram) are obtained by keeping a test tube at high temperature, called austenitizing temperature, for a certain time, called austenitizing time, before cooling it suddenly and keeping it at a constant temperature. The variation of austenite transformation is measured from the constant temperature. CCT diagrams are obtained with controlled thermal cycles. The test tube is kept at a high temperature, also called austenitizing temperature, for a certain time, also called austenitizing time, then continuously cooled down to ambient temperature. The variation of austenite transformation is measured during the cooling process. Figures 9.5 and 9.6 give examples of IT and CCT diagrams obtained on steel with 0.35% carbon, 1% chromium, and 0.2% molybdenum. The austenitizing temperature is 850°C. The austenitizing time is 30 mn. In the IT diagram, there is a dotted line. It represents the temperature under which a martensitic transformation independent of time occurs. In the CCT diagram, note that, on each cooling curve, the percentage of transformed austenite is mentioned after each passage into a phase change domain. At the end of the cooling process, the test tube hardness is given. The IT and CCT diagrams must be used to obtain the coefﬁcients of phase change laws used during numerical simulation. For this purpose, setting methods were developed [PON 94], and some simple transformations can be set manually:

224

Finite Element Simulation of Heat Transfer

Figure 9.5. Example of IT diagram obtained on steel with 0.35% carbon, 1% chromium and 0.2% molybdenum, from [BÉN 84]

– martensitic transformation only requires the initial transformation temperature Ms and the kinetic parameter b (see equation [9.2]). The temperature Ms is directly read in the IT or CTT diagrams. For instance we obtain 335°C in Figure 9.5. The kinetic parameter b does not depend greatly on the steel chemical composition. A value of 0.011°C−1 is often used; – with an IT diagram, the transformation kinetics given by equation [9.1] can be integrated at a constant temperature to obtain the classic Avrami relationship [AVR 40a, AVR 40b, AVR 41]: n p = P 1 − e−(f t)

Thermometallurgical Coupling

225

Figure 9.6. Example of CCT diagram obtained on steel with 0.35% carbon, 1% chromium and 0.2% molybdenum, from [BÉN 84]

In this equation, P , n and f are three parameters which can be determined at any temperature with the IT diagram. P is the phase proportion obtained after an inﬁnite time, and this relationship gives: P ln ln = n ln f + n ln t P −p It follows that by plotting at each temperature the quantity ln(ln(P/(P − p))) according to ln(t), we obtain a straight line whose slope corresponds to the parameter n and the y-axis at the origin gives the parameter f .

226

Finite Element Simulation of Heat Transfer

Figure 9.5 illustrates how to obtain the perlite transformation parameters at different temperatures with an IT diagram. From equation [9.1] we ﬁnd P = 1 and n = 3 at any temperature and a frequency f varying signiﬁcantly according to the temperature, from 0 to 0.45 s−1 . The variation of this frequency is given by the right-hand diagram in Figure 9.5. The left-hand diagram in this ﬁgure gives the IT diagram used for the setting, and particularly the initial and ﬁnal transformation times at temperatures 400°C, 500°C, 600°C and 700°C. It is also possible to obtain the parameters of the transformation described in Figure 9.6 from the CCT diagram of steel. For this purpose, it is possible for instance to assume a particular form of the frequency f of the transformation according to the temperature. Figure 9.6 suggests the choice of a Gaussian function: f=

2

if T ≤ T0

−((T −T0 )/∆Ts )2

if T ≥ T0

f0 e−((T −T0 )/∆Ti ) f0 e

In this function, the parameters f0 , T0 , ∆Ti and ∆Ts must be determined. Then, assuming P and n to be constant, a setting algorithm based on the Levenberg-Marquardt method [LEV 44] is applied. Figures 9.7 and 9.8 give the results obtained by perlitic transformation in steel with 0.8% carbon. The left-hand CCT diagram is the experimental diagram, with initial and ﬁnal transformation times, and the perlite quantity formed, for four cooling velocities, 80°C/s, 50°C/s, 8°C/s and 0.8°C/s. The transformation parameters obtained are P =1, f0 = 0.347 s−1 , T0 = 593°C, ∆Ti = 122°C, ∆Ts = 67°C and n = 2.78. The right-hand CTT diagram was recalculated with these parameters. Note that it is fairly close to the experimental diagram, represented by the dotted lines.

T (°C) 800 700

f (s-1) 0.45 0.4

9 0.55

4.7

600

75

0.35

6.8

500 400

0.3 60

0.8

0.25 0.2

6.9

300

0.15 0.1

200

0.05

100 0,1

1

10

102

103

t (s)

0

0 100 200 300 400 500 600 700 800 T (°C)

Figure 9.7. Obtaining perlite transformation parameters in steel with 0.8% carbon from an IT diagram, from [PON 94]

Thermometallurgical Coupling T (°C) 800

T (°C) 800 8.9

700

700

2.5

600

88

600 500

24

5.5

500

400

400

300

300

200

100% 58%

100 0.1

227

1

200

100%

100%

10

102

100% 57.5%

100 103

t (s)

0.1

1

10

100%

100%

102

103

t (s)

Figure 9.8. Obtaining perlite transformation parameters in steel with 0.8% carbon from a CTT diagram, from [PON 94]

9.2.2. Steel quenching This section examines an example of steel distortion tracking during cooling [CLA 02]. Figure 9.9 gives the geometry of the piece used. The cylindrical symmetry makes it possible to perform calculations on a meridian section of the piece in the

Figure 9.9. Geometry and mesh used for tracking distortions in steel quenching (sizes in mm)

228

Finite Element Simulation of Heat Transfer

5 bars nitrogen

(°C)

900

Temperature Te mpérature (°C

1000

700

800

600 500 400 300 200 100 0 0

100

200

300 Temps (s)

400

500

600

Time (s)

Figure 9.10. Thermocouple position and theoretical and experimental variation of the local temperatures in the cylinder during the cooling process

axisymmetric calculation conﬁguration. Three bores with increasing diameters were regularly distributed on the cylinder height. They enable us to create different local thermal kinetics, and hence generate local distortion variations. The cylinder of Figure 9.9 undergoes gas quenching with nitrogen injected vertically at a pressure of 5 bars. Figure 9.10 illustrates the quenching and the different thermal kinetics in agreement with the bores in the cylinder. Four temperatures are measured by thermocouple (T 1 to T 4). They were used to set the thermal exchange coefﬁcient representing the heat extracted by the gas moving along the sample. The results of numerical simulations are also given in this ﬁgure. They were carried out on austenitic stainless steel, with no phase change during cooling. Note the good calculation-experiment agreement with an error always less than 10°C. When the piece is cooling down, a camera is used to ﬁlm its shape variation. It is thus possible to plot the shape variation of the austenitic stainless steel cylinder during the gas quenching operation (Figure 9.11). Note once again that experiment and calculations are in good agreement. This is mainly due to the good representation of the local thermal paths and the absence of phase changes. In order to test the inﬂuence of phase changes, experiments and simulations were performed on steel 30CrNiMo8 with phase changes. The same heat exchange coefﬁcient as for stainless steel was used. The thermal paths followed at the top and bottom of the piece were plotted on the CCT diagram of that steel. Figure 9.12 gives these thermal paths and the microstructure obtained. Note that the only phase transformation taking place is the austenite-martensite transformation. Figure 9.13 represents the shape variation of a 30CrNiMo8 steel cylinder during cooling. It is easy to note the inﬂuence of the phase change on the diameter. Indeed,

Thermometallurgical Coupling

229

simulation experiment

50s

35s

110s 185s 335s end

D0

0

0.1 0.2 0.3 0.4 0.5 0.6 0.7 D-D0 (mm)

Figure 9.11. Shape variation of an austenitic stainless steel cylinder during cooling

thermal path and microstructure: top of the cylinder

thermal path and microstructure: bottom of the cylinder

austenitegrain size: heated at 875°C during 30 mn

Figure 9.12. CCT diagram of steel 30CrNiMo8 – extremity thermal paths in the cylinder and corresponding ﬁnal micro-structures

the structure of martensite being less compact than that of austenite, a local expansion occurs during the phase change, together with plastiﬁcation called transformation plasticity [LEB 89b].

230

Finite Element Simulation of Heat Transfer simulation experiment

35s

110s

55s

end

185s 335s

D0

-0.1

0

0.1

0.2

0.3

D-D0 (mm)

Figure 9.13. Shape variation of a 30CrNiMo8 steel cylinder during cooling

In Figure 9.13, experiment and simulation are in fairly good agreement. In particular, it is possible to make a good prediction of the ﬁnal shape of the cylinder after quenching. If, without a phase change, it comes back to its initial shape when the thermal gradients are not sufﬁcient to plasticize the material, when phase changes occur, the cylinder’s initial shape is no longer rectilinear on the outside, even without any traditional plastiﬁcation. This illustrates the presence of transformation plasticity. The plastic deformation of materials (traditional during phase changes) is not addressed in this book. The reader may refer to specialized books or articles [HIL 78, LEB 86, LEB 89b, LEB 89a, DEN 97, BES 01].

Chapter 10

Thermochemical Coupling

10.1. Finite element simulation of simultaneous diffusion and precipitation Thermochemical treatments are widely used to improve the mechanical properties of metallic components. For example, in the case of steel, high temperature diffusion of carbon and/or nitrogen leads to an increased hardness near the surface, as well as to local compressive residual stresses. Both contribute to an increased component lifetime. Modeling diffusion can be considered relatively straightforward, since the governing equations are well-established, and are very similar to those described in the previous chapters of this book. However, we must account for the fact that during diffusion, precipitation may occur and modify the kinetics of diffusion. For example, in the case of steel, carbides and/or nitrides can form during the diffusion of carbon and/or nitrogen. This precipitation contributes to the improvement in the mechanical properties of the treated component. The aim of this chapter is to describe a model aimed at predicting the local amount of chemical elements in a metallic component, during a thermochemical treatment. This model takes into account the diffusion and precipitation of these elements during the treatment, and has been incorporated into a commercial ﬁnite element computer code [FOR 95, FOR 00a, BER 03]. The application of the ﬁnite element method to simulate simultaneous diffusion and precipitation in metals is ﬁrst recalled. Then we focus on the method used to incorporate precipitation into diffusion equations, and on the numerical scheme used to solve the global problem. Finally, numerical applications which illustrate the approach proposed are expounded.

232

Finite Element Simulation of Heat Transfer

10.1.1. Governing equations We consider N chemical elements that can diffuse in a single matrix phase. For steels, the matrix consists of one major chemical element, iron, while alloying elements that can diffuse are mainly carbon, nitrogen, titanium or aluminum. These elements can combine together or with the matrix so as to form M precipitates. The chemical composition of the precipitates is characterized by stoichiometric coefﬁcients siα , each giving the number of atoms of element i contained in precipitate α. Throughout this chapter, Latin indices refer to chemical elements (i = 1, . . . , N ), whereas Greek subscripts denote precipitates (α = 1, . . . , M ). The total fraction of element i, in all its forms (dissolved in the matrix or enclosed in a precipitate), is denoted by fiT , while the part dissolved in the matrix is fiD . If pα denotes the fraction of precipitate α, the conservation of any element’s atoms i can be expressed as: fiT = fiD +

M

siα pα

[10.1]

α=1

It should be noted that we could choose to deal with mass fractions instead of molar fractions. In this case, we must introduce the atomic mass mi of element i and the molecular mass mα of precipitate α. Then the fiT , fiD and pα in equation [10.1] and the sequel must be replaced by fiT /mi , fiD /mi and pα /mα respectively. The activity ai of element i in the matrix is now introduced as a function of its dissolved fraction fiD as follows: ai = γi fiD

[10.2]

In this equation, γi denotes the thermodynamic coefﬁcient of element i. In the case of a dilute solution, i.e. for small fractions fiD , a ﬁrst order approximation of these coefﬁcients can be used: N ln γi = ln γi0 + eij fjD

[10.3]

j=1

In this equation, γi0 is Henry’s coefﬁcient and the eij the Wagner interaction parameters. Henry’s coefﬁcient can be made temperature-dependent using an Arrhenius law. According to equations [10.2] and [10.3], thermodynamic coefﬁcients can be calculated from given activities, by a Newton-like method, together with their partial derivatives with respect to these activities. Diffusion is assumed to take place only in the matrix (diffusion inside precipitates is neglected). According to Fick’s ﬁrst

Thermochemical Coupling

233

→ − law, the ﬂux density φ i of element i is assumed to be proportional to the gradient of its activity: − → Di −−→ φ i = − grad(ai ) γi

[10.4]

In this equation, Di is the diffusion coefﬁcient of element i inside the matrix. It depends on the temperature through an Arrhenius function. It should be noted that equation [10.4] does not imply that cross-diffusion is neglected. Since the thermodynamic coefﬁcients γi depend on the activities (equation [10.3]), the ﬂux → − density φ i can be expressed as a function of the dissolved fractions by: − → −−→ φi = −Dij grad(fjD ) N

with Dij = Di δij + fiD eij

[10.5]

j=1

Thus, cross-diffusion is governed by the Wagner interaction coefﬁcients eij introduced in equation [10.3]. According to the above equations, Fick’s second law relates the change rate of the total fraction of each element i to its activity as follows: Di −−→ [10.6] f˙iT = div grad(ai ) γi In order to integrate equation [10.6] in a given domain Ω, it is necessary to deﬁne boundary conditions (on ∂Ω) for the activities. A general expression for these boundary conditions is for each element i: Di −−→ → n = Ji + hi aei − ai grad(ai ) · − γi

[10.7]

→ In this equation, − n is the outward normal unit vector of the boundary, Ji is a prescribed ﬂux density, aei is an external activity and hi is a ﬁlm transfer coefﬁcient. Using this equation, we can obtain the following boundary conditions: – a prescribed ﬂux density Ji , with hi = 0; – a prescribed activity aei , with Ji = 0 and a very high value of hi ; – a ﬁlm transfer with Ji = 0. When N chemical elements are considered in Ω, equations [10.6] and [10.7] form a set of differential equations, with boundary conditions, which can be solved by using different methods. Among these methods, the ﬁnite element technique appears to be

234

Finite Element Simulation of Heat Transfer

the most suitable for complex shapes of the domain Ω, whereas a ﬁnite difference approximation is sufﬁcient for 1D calculations (i.e. when diffusion is assumed to occur in only one direction). In this book, we focus on the ﬁnite element approach used to approximate the solution of equations [10.6] and [10.7], which appear to be very similar to those obtained in thermal analyses. 10.1.2. Finite element formulation The ﬁnite element method applies to a weak formulation of the boundary value problem [10.6]-[10.7], which can be written for each chemical element. According to this formulation, it is necessary to ﬁnd ai : t ∈ ]0, T [ → ai (x, y, z, t) ∈ V such that ai (·, t = 0) = a0i and: ∀ψi ∈ V, ψi Ji + hi aei − ai ds ∂Ω

−

−−→ Di −−→ grad(ψi ) grad(ai ) dv − γi Ω

[10.8] ψi f˙iT dv = 0

Ω

In this formulation, ]0, T [ is a given time interval, a0i is an initial condition and V is a suitable functional space containing the test functions ψi . In order to build a ﬁnite element approximation for ai , the domain Ω is divided into elements Ωe connected by nodes. In each element Ωe , a spatial approximation of ai and its gradient is used: → − → Npe (− x )aip ai ( x ) = p → ∀− x ∈ Ωe , [10.9] −−→ −−→ → − → e − ( x )) = ( x ))a grad(a grad(N i ip p p

In this equation, aip is the activity of element i at node p of element Ωe , and is the shape function associated with node p of element Ωe . Equation [10.9] can also be applied to the test functions ψi . This leads to an approximation of the functional space. Then, by applying equation [10.8], we must ﬁnd {a} : t ∈ ]0, T [ → {a(t)} such that: → Npe (− x)

{R(a(t), t)} =

m

[Ae ] · {Re (a(t), t)} = {0}

[10.10]

e=1

At any instant t, the residual vector {R(a(t), t)} is a function of the unknown vector {a(t)}. Both have N × n components, where n is the number of nodes of the

Thermochemical Coupling

235

whole structure. {R(a(t), t)} is obtained by assembling all element residual vectors {Re (a(t), t)}, the component of which reads, for chemical element i and node p: e Rip = Npe Ji + hi aei − ai ds ∂Ωe ∩∂Ω

−−→ e Di −−→ − grad(Np ) grad(ai ) dv − γi e Ω

[10.11] Npe f˙iT

dv

Ω

It can be seen in equation [10.11] that the element residual vector {Re } involves the time derivative of the total fraction of each chemical element. We approximate this time derivative using the generalized trapezoidal rule. Therefore, it is assumed that for each interval of integration [t, t + ∆t]: fiT (t + ∆t) − fiT (t) = (1 − ν)f˙iT (t) ∆t + ν f˙T (t + ∆t) i

[10.12] with

0≤ν≤1

A value ν = 1/2 in equation [10.12] corresponds to the trapezoidal rule, which is the most accurate algorithm. Nevertheless, in the case of discontinuous time derivatives, the implicit Euler scheme (ν = 1) is generally preferred. According to this equation, the component of an element residual vector at node p, for chemical element i, and at time t + ∆t, reads: e −−→ e Di −−→ e e Rip = Np Ji + hi ai − ai ds − grad(Np ) grad(ai ) dv γi e e ∂Ω ∩∂Ω Ω [10.13] 1−ν T 1 T e T ˙ f (t + ∆t) − fi (t) − Np − fi (t) dv ν∆t i ν Ωe Starting from known quantities at time t, the activities of all elements at all the nodes of the structure, at time t + ∆t, result simultaneously from the solution of equation [10.10]. This solution is found by means of an iterative Newton procedure, which involves a global tangent matrix [K] obtained by assembling tangent element matrices [K e ]. The components of these tangent element matrices are: e =− Kipjq

e ∂Rip =− ∂ajq

Ωe

Di ∂γi e −−→ e −−→ N grad(Np )grad(ai ) dv γi2 ∂aj q

Ωe

Di −−→ e −−→ e δij grad(Np )grad(Nq ) dv γi

+

236

Finite Element Simulation of Heat Transfer

+ Ωe

1 ∂fiT e e N N dv ν∆t ∂aj p q

+ ∂Ω∩∂Ωe

hi δij Npe Nqe ds

[10.14]

In this equation, the partial derivatives of the thermodynamic coefﬁcient γi and the total fraction fiT of element i (at time t + ∆t), with respect to the activity aj of element j, have to be estimated. The former is derived from equation [10.3], whereas the latter is approximated by neglecting the variation of the precipitate fraction with respect to the activities. According to equations [10.1] and [10.2], this gives: ∂fiT 1 ∂γi δij − fiD ≈ ∂aj γi ∂aj

[10.15]

The main problem when solving the above equations is to estimate the total fraction of a chemical element as a function of its dissolved fraction (or its activity). This is done using equation [10.1], which involves the fractions of precipitates pα . The aim of the next section is to describe the procedure used to incorporate precipitation into this equation. 10.2. Calculation of precipitation 10.2.1. Mathematical formulation Starting from instant t, and thus assuming the fractions of precipitates at this time to be known, we want to estimate these fractions at time t + ∆t. For this purpose, we ﬁrst calculate an “equilibrium fraction” of precipitate, Pα , that would form if no diffusion were present for a large period of time. Then we use a kinetic law, which represents the time-dependent evolution of the precipitate fraction from its value at time t, to its value at time t + ∆t, towards the equilibrium value Pα . The equilibrium fractions Pα are calculated by applying the law of mass action, which can be formulated as follows: N @ = 0 if Asi iα < Kα P α i=1 [10.16] N @ siα Ai = Kα Pα ≥ 0 if i=1

Thermochemical Coupling

237

According to this equation, precipitation occurs only when the solubility product Kα is reached by the “equilibrium activities” of the elements Ai . These activities are obtained from the total fraction fiT of element i by the following relationship: M F D = f T − siα Pα i i Ai = Γi FiD with [10.17] α=1 ;N D Γ = γ 0 e j=1 eij Fj i i The solubility products Kα and the stoichiometric coefﬁcients siα in equation [10.16] are given as characteristics of the precipitates. The solubility products can be made temperature-dependent using the Arrhenius law. The Γi and FiD terms appearing in equation [10.17] are respectively the thermodynamic coefﬁcient and the dissolved fraction of element i, both expressed in equilibrium, i.e. without diffusion during a large period of time. For convenience, the law of mass action [10.16] can be formulated for all the precipitates that may be produced in the material as: P ≥0 α 0 N M 0 T ln γi + eij fj − sjβ Pβ N j=1 β=1 ≥0 Qα = ln Kα − 0 M i=1 T siβ Pβ + ln fi − β=1 M Qα Pα = 0

[10.18]

α=1

Equation [10.18] must be solved for a given set of total fractions fiT , and for given material characteristics γi0 , eij , siα and Kα . It turns out that in these equations Qα is a non-linear function of the Pα . Equation [10.18] can thus be considered as deﬁning a non-linear complementarity problem [KAR 72, LEM 80]. Starting with the Pα values calculated at time t, and from Pα = 0 (α = 1, . . . , M ) at t = 0, the following steps are used to solve this problem at time t + ∆t: – Qα (P1 , . . . , PM ) = 0 is solved for each precipitate for which Pα = 0. This leads to a new set of Pα values. A Newton method is used, which veriﬁes the condition that the FiD terms must remain positive. – The Qα (P1 , . . . , PM ) are calculated for each precipitate for which Pα = 0. This leads to a new set of Qα values.

238

Finite Element Simulation of Heat Transfer

– If all the Pα and Qα values obtained from the previous steps are non-negative, then the solution is reached, otherwise (i) the negative Pα are set to zero, (ii) the Pα corresponding to a negative Qα are set to a non-vanishing value, and we go back to the ﬁrst step. The kinetic law introduced to account for the time-dependence of precipitation is based on a Johnson-Mehl-Avrami-Kolmogorov law, so that the precipitate fraction is given by: n pα = Pα 1 − e−(f t)

[10.19]

In this equation, n and f are two parameters which characterize this kinetics. Thus, knowing pα (t), the precipitate fraction at time t + ∆t is: n Pα 1 − e−(f (t0 +δt)) 1 21/n Pα 1 ln pα (t + ∆t) = with t0 = f Pα − pα (t) n pα (t) − pα (t) − Pα 1 − e−(f t)

if pα (t) < Pα

[10.20]

if pα (t) ≥ Pα

It should be noted that when the calculated equilibrium value Pα (t) is less than the current value pα (t), the precipitate is currently being dissolved in the matrix and the exponent n of the kinetics is set to 1 (no nucleation period). 10.2.2. Numerical scheme At any instant t + ∆t, iterations are performed using a Newton-like method to determine the local activity ai of each chemical element i in the domain. If x(k) denotes the value of any variable x obtained at the end of the k th iteration, then the following scheme is used at each iteration (k) to incorporate precipitation into the diffusion model: T (k)

– The total fraction of chemical element i, fi , is calculated by means of (k) equations [10.1] and [10.2] with the current values of the activities ai and (k) thermodynamic coefﬁcients γi , and with the precipitate fractions obtained at the (k−1) (for the ﬁrst iteration, the precipitate fractions are end of the previous iteration pα taken from the previous time step). (k)

– The equilibrium fractions of precipitates, Pα , are calculated by means of T (k) equation [10.18] with the total fractions fi . Kinetics laws [10.20] then give the (k) precipitate fractions pα .

Thermochemical Coupling

239

T (k)

– The total fraction of chemical element i, fi , is modiﬁed according to equation (k) [10.1], with the new precipitate fractions pα , before estimating the residual vector of equation [10.10] and eventually going back to a new iteration. According to the scheme presented above, convergence of the global Newton method cannot be achieved if the precipitate fractions calculated present a large variation between two successive iterations. Thus precipitation is treated numerically by a so-called ﬁxed-point method. 10.3. Examples 10.3.1. Calculation of a phase diagram In order to validate the method used to estimate the equilibrium fractions of precipitates Pα (equation [10.18]), the phase diagram of a micro-alloyed steel containing T i, M n, S and C as alloying elements can be estimated. Within this steel, the precipitates that may appear and form phases at 1, 200°C are T iS, T iC, T i4 C2 S2 and M nS. Table 10.1 gives the data used for the chemical elements, whereas Table 10.2 gives those associated with the precipitates. Ti

Element Atomic mass

Mn

S

C

47.900 54.938 32.064 12.011

Henry’s coefﬁcient

1.

1.

1.

1.

Wagner parameters

0.

0.

0.

0.

Total mass fraction (%) 0 to 0.1

0.2

0 to 0.01 0.003

Table 10.1. Data used for the chemical elements involved in the phase diagram

Precipitate

T iS

T iC

T i4 C2 S2

M nS

Molecular mass

79.964

59.911

279.75

87.002

Solubility product 4.835310−5 5.355110−3 3.9510−17 6.403910−4 Table 10.2. Characteristics of the precipitates involved in the phase diagram

Figure 10.1 gives the phase diagram calculated from the data in Tables 10.1 and 10.2. This diagram was obtained by solving equations [10.18] at various points, and by detecting the domains where precipitates appear. It can be seen in this ﬁgure that

240

Finite Element Simulation of Heat Transfer

MnS Ti S

Ti 4 C 2 S 2

Figure 10.1. Recalculated phase diagram of a micro-alloyed steel

the T iC precipitate does not appear within the variation range used for the chemical elements. Moreover, the number of precipitates meets the Gibbs law. Finally, this diagram is in good agreement with the experiment. This validates equations [10.18], together with the numerical scheme used to solve them. 10.3.2. Carbon diffusion in a titanium steel In order to validate the method used to model simultaneous diffusion and precipitation, we can calculate carbon diffusion in a steel containing titanium. The data used for the calculation are given in Table 10.3. They apply to a cylinder with a 10 mm radius. Diffusion of carbon and titanium is assumed to take place only in the radial direction, and the possibility of the appearance of a T iC precipitate is taken into account. It can be seen in this table that the calculations were performed with the following assumptions: – Henry’s coefﬁcients of the elements are set to 1, and Wagner’s coefﬁcients vanish. This means that we assume that the activities of the elements and their dissolved fractions are equal. Moreover, it is assumed that the elements do not interact during diffusion. – The initial mass fractions of carbon and titanium are set to 0.2% and 0.01% respectively. Then the steel specimen is placed into a furnace. The carbon activity in the atmosphere of the furnace is set to 1, and a ﬁlm transfer coefﬁcient of 0.0001 is used to represent the carbon ﬂux density entering into the specimen during the treatment.

Thermochemical Coupling

Element Atomic mass

C

Ti

12.011 47.900

Wagner parameters

0.

0.

Initial activity (%)

0.2

0.01

Film transfer coefﬁcient

0.0001

0.

Activity in the furnace (%)

1.

0.

241

T iC

Precipitate

59, 911

Molecular mass −1

Frequency factor (s

)

Exponent

0.1 1.

Diffusion coefﬁcient (mm2 /s) 4.8710−7 510−7 Table 10.3. Data used for the simulation of carbon diffusion in a titanium steel

– A kinetic law with a frequency factor of 0.1 s−1 and an exponent of 1 is used for the T iC precipitate. Figure 10.2 gives the distribution of the total, dissolved and precipitated, fractions of T i along the radius, near the surface of the cylinder. It can be seen in this ﬁgure that carbon diffusion has led to the precipitation of T iC near the surface. Since the precipitate is made of T i and C, the dissolved fraction of T i is lowered, leading to a gradient of its activity, and thus to diffusion of this element. This diffusion then leads to an increase in the fraction of the T iC precipitate, so that in this case precipitation is governed by the mobility of titanium in the steel matrix, as well as by the kinetics of precipitation.

Figure 10.2. Variation of the fraction of T i along the cylinder radius, after 4 hours’ treatment

This page intentionally left blank

Chapter 11

Electrothermal Coupling

11.1. Electrokinetic modeling Here we deal with the case of a high-intensity current applied to a material, causing the heating of that material by means of the Joule effect, thus coupling it with thermal aspects. In resistance welding applications, given the current frequency (50Hz), the dynamic effects (induced currents) are negligible and the electric currents can be calculated by means of an electrokinetic model. Here we give the equations governing the electrokinetic phenomena involved in the material, their resolution by the ﬁnite element method and their coupling with thermal phenomena. 11.1.1. Weak formulation → − In a domain Ω, applying an electric ﬁeld E produces an electric current whose → − → − → − density termed J is given by the Ohm law, J = σ · E , where the symmetric tensor σ represents the electric conductivity of the material. When this conductivity is identical in all directions, then σ = σI, where I is the identity tensor and the → − scalar σ represents this conductivity. In electrokinetics, the electric ﬁeld E is such → → − −→ − that rot( E ) = 0 . It is therefore derived from a scalar potential V , for instance of the −−→ − → form E = −grad(V ). The current density in the material is thus proportional to the electric potential gradient: −−→ − → J = −σ.grad(V )

244

Finite Element Simulation of Heat Transfer

This equation is the material’s electrokinetic behavior law. It is similar to Fourier’s law of thermal transfer. If the current conservation inside any domain Ω is then written, without any possible volume current production, it is simple to determine the electrokinetic conservation equation: − → div J = 0 at any point of Ω Using these two equations, and applying current density or potential boundary conditions produces the following system to solve, which is very similar to the one encountered in steady state heating in Chapter 1:

→ Find V (− x ) such that:

−−→ div σ · grad(V ) = 0

→ for any − x ∈Ω

V = Vd → − → −−→ → −J ·− n = σ · grad(V ) · − n =j

on ∂ΩV

[11.1]

on ∂Ωj

→ We deﬁne − n as the outward norm to the domain Ω, ∂ΩV as the boundary of Ω on which the potential V is known, and ∂Ωj as the boundary of Ω where the entry current density is known. Note that, as in thermal science, it is necessary to know at any point of the boundary ∂Ω of the domain Ω either the exiting current potential or density, so that ∂ΩV ∪ ∂Ωj = ∂Ω and ∂ΩV ∩ ∂Ωj = ∅. If the same procedure as that described in Chapter 1 of this book for steady state heating is followed, it is possible to give a weak formulation of problem [11.1]: Find V ∈ E such that for any V ∗ ∈ E ∗ : V V [11.2] −−→ −−→T ∗ V ∗ j ds − grad (V ) · σ · grad(V ) dv = 0 ∂Ωj Ω We insert into this system the set EV of solutions satisfying the essential boundary conditions V = Vd on ∂ΩV , and the set EV ∗ of test functions V ∗ equal to zero on that same boundary. 11.1.2. Modeling the coupling The boundary value problem governing electrothermal coupling in a domain Ω consists of juxtaposing the boundary value problem related to heat diffusion with the problem related to electrokinetic phenomena. Both problems are coupled by: – the power dissipated by the Joule effect, incorporated as an internal source term in the heat equation; – the temperature dependence of electrical properties.

Electrothermal Coupling

245

The power dissipated by the Joule effect per material volume unit is of the form → − − → P = E · J . It follows that the volumetric source term Q involved in the heat problem → can be written directly according to the potential ﬁeld V (− x ), the solution of the electrokinetic problem: −−→ −−→ Q = gradT (V ) · σ · grad(V )

[11.3]

→ → The temperature ﬁeld T (− x ) and the electric potential ﬁeld V (− x ) are therefore obtained by coupling the two related weak formulations. The thermal calculation being performed in the transient state, and hence with a time-dependent temperature ﬁeld, the electrokinetic potential will also be time-dependent. The temperature ﬁeld → → T (− x , t) and the potential ﬁeld V (− x , t) will be obtained at any instant t between 0 and tf by solving the following problem: → Find T : t ∈ [0, tf [ −→ T (− x , t) ∈ ET , Find V : t ∈ [0, t [ −→ V (− → x , t) ∈ EV such that: f → → → → at t = 0: T (− x , 0) = T0 (− x ) and V (− x , 0) = V0 (− x) for t ∈ ]0, tf [, for any ψ ∈ Eψ and for any V ∗ ∈ EV ∗ : −−→ −−→T ˙ ψ(Q − ρH)dv + ψq ds − grad (ψ) · λ · grad(T )dv = 0 Ω ∂Ωq Ω −−→ −−→T ∗ V ∗ j ds − grad (V ) · σ · grad(V ) dv = 0 ∂Ωj

[11.4]

Ω

Over time, the natural boundary conditions provide the heat ﬂux density q entering via ∂Ωq and the current density j entering via ∂Ωj . Similarly, the essential boundary conditions are those making it possible to deﬁne the sets ET and EV according to the temperatures Td and the potentials Vd prescribed, as well as the sets Eψ and EV ∗ of the weighting functions: ET = T ∈ H 1 (Ω)/T = Td on ∂ΩT Eψ = ψ ∈ H 1 (Ω)/ψ = 0 on ∂ΩT EV = V ∈ H 1 (Ω)/V = Vd on ∂ΩV EV ∗ = V ∗ ∈ H 1 (Ω)/V ∗ = 0 on ∂ΩV

246

Finite Element Simulation of Heat Transfer

In problem [11.4], it must be noted that: – the volumetric mass ρ, the thermal conductivity (symmetric) tensor λ and the electrical conductivity (symmetric) tensor σ depend on temperature T ; – the enthalpy H is a function of temperature and time. This term also makes it possible to insert latent heat effects related to phase changes (see Chapter 9); – the volumetric source term Q is given by relationship [11.3]. It represents the power dissipated by the Joule effect and depends at each instant on the electric potential ﬁeld; – as was seen in the previous parts of this book, the heat ﬂux density q can take different forms to express a convection or radiation exchange with the external environment; – the current density j can be expressed from the electric resistance R of an interface, possibly a function of the temperature, with a medium of potential V∞ in the form J = (1/R)(V∞ − V ). 11.1.3. Solving the coupled problem In order to solve coupled problem [11.4], we could use a staggered approach where the electrokinetic and thermal problems would be treated independently. However, the importance of the coupling between such phenomena implies us to prefer an approach where the problems are treated together. To obtain the ﬁnite element formulation → of this problem, a Galerkin approach is used, with the temperatures T (− x , t) and → potentials V (− x , t) as unknown ﬁelds. Consider a single mesh for the electrokinetic and thermal problems, including n nodes and m elements Ωe . Inside each element, the temperature and the electric potential use the same approximation scheme: − ∀→ x ∈ Ωe , ∀t ∈ [0, tf ],

→ → T (− x , t) = N e (− x ) · {T e (t)} → → V (− x , t) = N e (− x ) · {V e (t)}

[11.5]

The problem now has 2n unknowns: the temperature and the electric potential at all the mesh nodes. In this instance an implicit Euler algorithm to discretize the time derivative of the enthalpy H of [11.4]. The ﬁnite element method then leads us to solve at each time step a system of 2n equations with 2n unknowns, which can be written as:

{RT } {0} = [11.6] {RV } {0}

Electrothermal Coupling

247

In the above equation, {RT } represents the residual vector at instant t related to the thermal problem and {RV } the vector related to the electrokinetic problem. They are both obtained by assembling element residuals: {RT } =

m

T

[Ae ] · {RTe }

e=1

{RV } =

m

T

[Ae ] · {RVe }

e=1

The element residuals are easily obtained from the results of previous chapters, they are written: −−→ e e −−→T {RT } = {N } grad (V ) · σ · grad(V ) dv + {N e } q ds Ωe

∂Ωe ∩∂Ωq

'−−→ ( −−→ gradT (N e ) · λ · grad(T ) dv

H − H0 dv − − {N } ρ ∆t e Ω Ωe ' −−→ −−→T e ( {RVe } = {N e } j ds − grad (N ) · σ · grad(V )dv e

∂Ωe ∩∂Ωj

[11.7]

[11.8]

Ωe

The resolution of system [11.6] can be performed with a Newton-Raphston method. This method involves a tangent matrix, where the different sub-matrices are obtained by assembling element matrices, as follows: m T ] = [Ae ] · [KTe T ] · [Ae ] [K TT e=1 m T ] = [Ae ] · [KTe V ] · [Ae ] [K . / T V [KT T ] [KT V ] e=1 {KT } = with m [KV T ] [KV V ] T [K ] = [Ae ] · [KVe T ] · [Ae ] VT e=1 m T ] = [Ae ] · [KVe V ] · [Ae ] [K V V e=1

Note that the global matrix [KT ] is non-symmetric. This is practically always the case when several coupled problems are treated together. The components of the element matrices from the previous equations are written:

248

Finite Element Simulation of Heat Transfer

KTe T ij

dσ −−→ dq e e · grad(V )Nj dv − N ds =− )· Nie dT dT j Ωe ∂Ωe ∩∂Ωq −−→T e dλ −−→ 1 d(ρH) e Nj dv + · grad(T )Nje dv Nie + grad (Ni ) · ∆t dT dT Ωe Ωe −−→ −−→T e + grad (Ni ) · λ · grad(Nje ) dv −−→ Nie gradT (V

Ωe

e KT V ij = −2 e KV T ij = −

Ωe

−−→ −−→ Nie gradT (V ) · σ · grad(Nje ) dv

−−→T e dσ −−→ · grad(V )Nje dv grad (Ni ) · dT ∂Ωe ∩∂Ωj Ωe e −−→ −−→T e e ∂j e KV V ij = − Nj ds + Ni grad (Ni ) · σ · grad(Nje ) dv ∂V e e ∂Ω ∩∂Ωj Ω Nie

∂j e N ds + ∂T j

11.2. Resistance welding The passage of electric currents through a conductive material generates power dissipation via the Joule effect. This principle is particularly used in resistance welding processes. They are very commonly used in the car industry because of their capacity to assemble thin metal sheets. Figure 11.1 introduces the general principle of spot resistance welding. Two copper electrodes are placed on both sides of the sheets to be assembled and are strongly pressed on them to facilitate contact. A high intensity current melts the metal at the sheet interface. The current used is generally a 50 Hz alternative current. As a result, i

i

Figure 11.1. Spot resistance welding – principle diagram

Electrothermal Coupling

249

the dynamic effects (induced currents) are negligible and the electric currents can be calculated by means of an electrokinetic model. The simulation of spot resistance welding is of particular interest to study the size of the molten core according to the operational procedure or analyze the electrode lifetime [FEU 06]. Resistance welding involves strongly coupled mechanical, thermometallurgical and electrokinetic phenomena. The electrodes for instance help hold together the sheets being assembled during the solidiﬁcation phase of the liquid core. The thermometallurgical aspects are covered in Chapter 9. Our study is here restricted to the thermal and electrokinetic aspects; treating the mechanical problems would require developments outside the scope of this book. First, the model implementation for simulating this process is described. A thermal and electric contact management method is given. Then the results obtained are analyzed to demonstrate the capability of the ﬁnite element method to simulate this type of process and thus offer an insight into the local temperatures reached, the material structure after welding, etc. 11.2.1. Implementing the model Figure 11.2 illustrates the mesh used to simulate the welding of three metal sheets 1.2 mm, 0.8 mm and 0.8 mm thick. Note the presence of a large number of contacts between each electrode and a metal sheet on the one hand, and the three metal sheets welded together on the other hand.

Figure 11.2. Mesh used

250

Finite Element Simulation of Heat Transfer

q(i), j(i)

integration point (i)

T(i), V(i) subsurface w (i)

T (i), V (i)

Figure 11.3. Contact management

Figure 11.3 illustrates the method used to manage the electric and thermal contacts. It is based upon a Gauss numerical integration scheme [BAT 96] applied to contact elements [JAO 98, BEL 99]. Consider two meshed surfaces in contact and evaluate the current density and the heat ﬂux density entering locally (i.e. at each integration point of the elements) into each of the surfaces. Consider therefore an integration point (i) of an element belonging to one of the opposite surfaces, and term w(i) , T (i) and V (i) , respectively the weight of the integration point (i) in the Gauss scheme, the temperature and (i)

(i)

the electric potential at that point (Figure 11.3). Moreover, T and V represent respectively the temperature and the electric potential on the element opposite the integration point (belonging to the other surface). The current density j (i) entering through the subsurface w(i) related to the integration point (i) can be written according to the contact’s electric resistance R as: j (i) =

1 (i) V − V (i) R

The thermal ﬂux density q (i) received by the subsurface w(i) related to the integration point (i) can be decomposed into two parts: (i) q (i) = K T − T (i) + p(i) The ﬁrst term uses a thermal exchange coefﬁcient K (inverse of a contact thermal resistance) to illustrate the thermal exchanges at the interface. Of course, coefﬁcient

Electrothermal Coupling

251

K depends on the properties of that interface, and in that expression it is assumed that the interface has no thermal capacitance. The second term of the previous equation represents the thermal power dissipated by the Joule effect at the interface. Given the expression of j (i) , the total power (i) dissipated at the interface will be P (i) = (V (i) − V )2 /R, and each opposite surface will receive only part of that power. If f (i) is the fraction received by the subsurface w(i) , then the thermal ﬂux density received by that surface is written: f (i) (i) 2 (i) q (i) = K T − T (i) + V − V (i) R The expressions of q (i) and j (i) above are then inserted into the expressions of the global residuals to cancel to solve the coupled problem. The Gaussian method makes it possible to take into good consideration the thermal ﬂux q and current j local densities. However, corrective terms are inserted to satisfy the global ﬂux conservation [FEU 04]. 11.2.2. Results Figure 11.4 gives the temperature ﬁelds calculated at the end of the heating process in the case of three soft steel sheets welded together. Note the presence of a liquid ISOVALUES Temperature Time 0.34s Deformed x 1

Figure 11.4. Plotted temperature contour curves

252

Finite Element Simulation of Heat Transfer

melte d zone

heat affected zone

Figure 11.5. Comparing experiment with calculation

core in the center, which is in good agreement with the experiment, as illustrated by Figure 11.5, in which the temperature contour corresponding to the steel liquidus was superimposed on top of a micrographic section. During the calculation process, thermometallurgical coupling was performed with the method described in Chapter 9. In Figure 11.5, the temperature contour corresponding to the steel transformation point Ac1 (initial austenitizing temperature) was superimposed on top of the micrographic section. Once again there is a fairly good agreement with the experiment. This ﬁnite element numerical simulation of a complex process, involving thermometallurgical phenomena and contact problems, demonstrates that the method is highly appropriate for this type of approach. Numerical simulations make it possible on the one hand to make progress in the command of the process, i.e. the knowledge of relevant parameters, and on the other hand, to perform tests that enable us to optimize the heating sequence according to the liquid core desired.

Chapter 12

Magnetothermal Coupling

12.1. Introduction Any electric conductor placed in a time-varying magnetic ﬁeld is the base of eddy-currents that dissipate power through the Joule effect. This power is concentrated at the surface of the component, in a very small thickness which depends on the electric conductivity and magnetic permeability of the material, but also on the frequency of the currents. Induction hardening processes take advantage of this effect because it is therefore possible to control the thickness of the treated layer. Induction hardening of steel components is performed in two steps. First, an induction heating stage puts the surface of the piece in an austenitic state. Then a cooling stage involving heat conduction through the piece and an external shower enables metallurgical transformations in the heat affected zone in order to form hard structures. The numerical simulation of this process is of major interest to control and quantify the thickness of the treated layer, the superﬁcial hardness and the residual distortions in order to avoid a ﬁnal machining which could annihilate the beneﬁts of the treatment. Moreover, such processes often lead to compressive residual stresses at the surface which improve the fatigue behavior of the mechanical component. The complete modeling of the induction hardening process rests upon the solution of coupled electromagnetic, thermal, metallurgical and mechanical problems, as shown in Figure 12.1. In what follows, we will focus on the modeling of induction heating.

254

Finite Element Simulation of Heat Transfer

Electromagnetism power loss (Joule effect)

temperatures

temperatures

Heat transfer

Metallurgy

conduction, c onvection, radiation latent heat phase proportions

Phase proportions

stresses

Plastic dissipation

temperatures

Mechanics Figure 12.1. Couplings between physical phenomena involved by the induction hardening process

The modeling of induction heating rests upon the resolution of a magnetodynamic problem and a problem of heat transfer coupled by: – the power dissipated by the Joule effect on the one hand; – the dependence of the electromagnetic properties with temperature on the other hand. 12.2. Magnetic vector potential formulation for magnetodynamics Magnetodynamic phenomena are governed by the following Maxwell’s equations: −−→ − → → − curl( H ) = J

(Maxwell-Ampère’s equation)

[12.1]

− → −−→ − → ∂B curl( E ) = − ∂t − → div B = 0

(Maxwell-Faraday’s equation)

[12.2]

(Maxwell-Gauss’s equation)

[12.3]

with the following constitutive equations: − → → − → − H = ν( B , T ) B

(Magnetization law)

[12.4]

− → → − J = σ(T ) E

(Ohm’s law)

[12.5]

→ − − → − → → − In the equations above, H , J , E and B represent the magnetic ﬁeld, the electric current density, the electric ﬁeld and the magnetic ﬂux density vector respectively. In equation [12.4], ν is the magnetic reluctivity (the inverse of the magnetic

Magnetothermal Coupling

255

Figure 12.2. Typical magnetization curves B(H, T ) for steel

permeability). Taking account of the magnetic ﬁeld magnitude involved by induction heating applications, hysteresis loops are neglected. In equation [12.5], σ is the electric conductivity. We can note in Figures 12.2 and 12.3 that the electromagnetic properties strongly depend on temperature T . → − From equation [12.3], let us now introduce the magnetic vector potential A such that: −−→ − − → → B = curl( A )

[12.6]

To ensure the uniqueness of the magnetic vector potential, an additional equation − → on div A must be introduced, for example: − → div A = 0

(Coulomb gauge)

[12.7]

Equation [12.2] then gives: − → − → ∂ A −−→ − grad(V0 ) E =− ∂t where V0 is the electric scalar potential.

[12.8]

256

Finite Element Simulation of Heat Transfer

σ Ω

Figure 12.3. Typical temperature-dependent electric conductivity σ(T ) for steel

Here again, in order to deﬁne V0 uniquely, we must introduce the following additional equation derived from equation [12.1]: − → div J = 0

[12.9]

−−→ which, with equations [12.5], [12.8] and [12.7], gives: div (σ grad(V0 )) = 0. The vector potential formulation supposes that equation [12.9] is satisﬁed a priori. It can be shown that this equation is automatically veriﬁed for 2D and axisymmetric cases as deﬁned below. Therefore, using equations [12.4], [12.5], [12.6] and [12.8], equation [12.1] ﬁnally becomes: − → −−→ ∂ A −−→ −−→ σ + curl(ν curl(A)) + σ grad(V0 ) = 0 ∂t

[12.10]

So, the magnetic vector potential must be the solution of equations [12.7] and [12.10]. We must note that equations [12.7] and [12.10] hold in the whole space. Due to air, the spatial domain on which the solution of both equations must be found is not bounded except for symmetry reasons. Let us now consider the case of a geometry obtained by the translation of a cross-section along the z-axis direction and suppose

Magnetothermal Coupling

257

−−→ − → that the source current (J0 = −σ grad(V0 )) is also directed along the z-axis and only depends on x and y coordinates: − → → J0 = j0 (x, y, t)− ez

[12.11]

In this case, it can be shown that the problem is reduced to a 2D problem because the vector potential now has only one non-zero component: − → → A = a(x, y, t)− ez

[12.12]

Similarly, in the case of axisymmetric geometries and a circumferential source current: − → → eθ J0 = j0 (r, z, t)−

[12.13]

the problem is reduced to an axisymmetric problem and the vector potential can be written: − → → A = a(r, z, t)− eθ

[12.14]

We can note that the gauge condition (equation [12.7]) is automatically satisﬁed in both cases. Therefore, the magnetic vector potential now has to be the solution of equation [12.10] only [MEU 08].

12.3. Coupled Þ nite element-boundary element method The idea of coupling the ﬁnite element method (FEM) and the boundary element method (BEM) to beneﬁt from the advantages of both methods is rather old [ZIE 77, ATL 78]. It ﬁnds here a rather natural application [FET 97, PAS 03b]. Let us subdivide the physical space into a domain Ω1 made up of conductors (possibly traversed by sources of current) and a domain Ω2 constituted by air (Figure 12.4). Let us call Γ the common border between Ω1 and Ω2 . The problem to be solved in the air (Ω2 ) is reduced to: −−→ −−→ − → curl(ν0 curl( A )) = 0

[12.15]

− → div A = 0

[12.16]

258

Finite Element Simulation of Heat Transfer

Γ∞

Γ (FEM-BEM boundary)

FEM domain BEM domain

Figure 12.4. Coupling between FEM and BEM domains

with the following boundary conditions: − → A = 0 to inﬁnity on Γ∞

[12.17]

→ → − − continuity of A and div A on Γ

[12.18]

→ − − → → → − → → − n 1 = − h 2 = −H × − n 2 on Γ h1 = H ×−

[12.19]

→ → n 2 are the units normal to Γ outside Ω1 and Ω2 respectively and where − n 1 and − 1 ν0 = µ0 .

Magnetothermal Coupling

259

Consider a mesh of Ω1 including n nodes and a compatible mesh of the border Γ → − including b nodes. Moreover, suppose that h 1 in equation [12.19] is deﬁned from nodal values {H1 } (b components) by using the shape functions of the elements located on the border Γ.

12.3.1. Finite element formulation The weak formulation associated with equations [12.10] and [12.7] is obtained by − → using the following relation, where the A∗ vectors are weighting functions associated → − with A : − → −−→ −−→ − → −−→ − −−→ − → → A∗ · curl(ν curl( A )) dv = curl(A∗ ) · ν curl( A ) dv Ω1

Ω1

+

− → → −−→ − → n 1 ∧ ν curl( A ) ds A∗ · −

[12.20]

Γ

The following weak formulation can thus be easily derived: → − →→ − Find A : t ∈ ]0, T [ −→ A (− x , t) ∈ V such that − → − − → − →→ A (→ x , 0) = A0 (− x ) and for any A∗ ∈ V :

− → − → ∂A → −−→ − −−→ − → dv + A∗ · σ curl(A∗ ) · ν curl( A ) dv ∂t Ω1 Ω1 − → − − → − → → = A∗ · J0 dv + A∗ · h 1 ds Ω1

[12.21]

Γ

In the above equation, ]0, T [ is a given time interval (one or more periods of the − →→ source current), A0 (− x ) an initial condition and V the functional space of admissible functions i.e. satisfying equation [12.7] and the conditions of regularity necessary to give a sense to equation [12.21]. Let us now consider axisymmetric geometries and the circumferential source current. In this case, the source current and the magnetic vector potential are given by equations [12.13] and [12.14] respectively. We can note that we also have → − → − → → h 1 = h1 − eθ and h 2 = h2 − eθ . The gauge condition is automatically satisﬁed and → − → 1 −−→ we obtain curl A = r grad(ra) × − eθ . This formula encourages us to seek a nodal

260

Finite Element Simulation of Heat Transfer

approximation of ra rather than a. The ﬁnite element discretization of variational equation [12.21] then leads to the following system of equations: ' ( [12.22] [C] · A˙ + {R} − [T ] · {H1 } = {J0 } {A} contains the nodal values of ra, [C] is a n × n matrix, {R} is a n-component vector, [T ] is a n × b rectangle matrix and {J0 } is a n-component vector containing the nodal values equivalent to the source current. [C], {R}, [T ] and {J0 } are obtained from an assembly process from the following element vectors and matrices: σ [C e ] = {N e } · N e dv 2 Ωe r ' −−→T e ( %−−→ e & {Re } = curl (N ) · ν curl(N ) · {Ae } dv Ωe

[12.23]

e

[T ] = ∂Ωe ∩Γ

{J0e } =

Ωe

1 {N e } · N e ds r

J0 {N e } dv r

$ # −−→ −−→ −−→ → → eθ , . . . , 1r grad(Nnee ) × − eθ is the matrix giving where curl(N e ) = 1r grad(N1e ) × − −−→ − → curl( A ) from nodal values of ra. Note that {Re } can be a non-linear function of {A} → − → − since ν is B -dependent, and hence depends on A via equation [12.4]. 12.3.2. Boundary element formulation Equation [12.15] multiplied by a scalar weight function ψ(r, z), then integrated over Ω2 leads, after some mathematical handling and taking account of equations [12.16] and [12.17], to the following integral equation: −−→ −−→ − → − → → − → →− ν0 grad(ψ) · − ν0 ∆ψ A dv = n 2 A − ψν0 curl( A ) × n2 ds Ω2

Γ

−

−−→ − → →−−→ → → − ν0 grad(ψ) · A − n2 grad(ψ) ds n2 − ν0 A · −

[12.24]

Γ

−−→ → − →→ − Considering the axisymmetric case, we obtain grad(ψ). A = 0 and A .− n2 = 0. Moreover, taking equation [12.19] into account, equation [12.24] thus becomes: −−→ →a − ψh ds ν0 grad(ψ) · − [12.25] ν0 ∆ψa dv = n 2 2 Ω2

Γ

Magnetothermal Coupling

261

Considering weighting functions ψp as the solution of Poisson’s equation for a unit →: ∆ψ + δ(− → →) = 0 (δ Dirac’s distribution) with ψ = 0 load located at point − x x −− x p p p p at inﬁnity ([BRE 92]), equation [12.25] becomes at each node i of the border Γ: ci ∂ψi ν0 ai = a ds [12.26] ψi h2 ds − ν0 2π ∂n Γ Γ In this equation, ci is the inside angle at node i (ci = π for a very smooth boundary). Using the mesh of Γ previously introduced for the discretization of equation [12.26], we ﬁnally obtain: [H] · {A2 } − [G] · {H2 } = 0

[12.27]

ci δij + where [H] and [G] are b × b component matrices with Hij = ν0 2πr i < < ∂ψi 1 e e ν N ds and G = ψ N ds, and where {A } is a b-component vector ij 2 Γ 0 ∂n r j Γ i j containing the nodal values of ra at the nodes of Γ.

12.3.3. FEM-BEM coupling Equation [12.19] gives: {H1 } = − {H2 }

[12.28]

Equation [12.22] combined with equations [12.27] and [12.28] becomes: ' ( −1 [12.29] [C] · A˙ + {R} + [T ] · [G] · [H] · {A2 } = {J0 } −1

The non-symmetric matrix [T ] · [G] · [H] appears as the contribution of air to the system of equations [12.29]. The system of differential equations [12.29] can thus be solved by a method of direct integration (implicit Euler or central differences for example) together with a Newton-Raphson method to ﬁnd the solution at each time-step. 12.4. A harmonic balance method for the magnetodynamic problem In practice, the source current is always sinusoidal. However, due to the non-linearity of the magnetization law, induced electromagnetic ﬁelds contain higher order harmonics. Rather than solve equation [12.29] by a direct method step by step over time, a harmonic balance method can be used [YAM 88a, YAM 88b, LU 95,

262

Finite Element Simulation of Heat Transfer

TER 84]. This method consists of calculating the ﬁrst terms of the development in a Fourier series of the magnetic vector potential: −−→ →− − −−→ → → A (→ x , t) = x ) cos(kωt) + Aks (− x ) sin(kωt) [12.30] Akc (− k=1,3,...

where ω =

2π τ

represents the pulsation of the source current and τ , the period.

Equation [12.10] is then projected on to the basis of chosen functions of time. The equations to solve are now written: − → τ − → ∂ A −−→ −−→ + curl(ν curl(A)) − J0 cos(lωt)dt = 0 σ ∂t 0 for l = 1, 3, . . . [12.31] − → τ − − → − − → − → ∂ A + curl(ν curl(A)) − J0 sin(lωt)dt = 0 σ ∂t 0 −−→ → We obtain a new problem, where the unknowns are now the ﬁelds Akc (− x ) and −−→ − Aks (→ x ) for k = 1, 3, . . . , m, m being the order of the last harmonic considered. The FEM applied to equations [12.31] in the axisymmetric cases leads us to solve the following system of non-linear equations: .. . {R ({A } , {A } , . . .)} lc lc ls =0 {Rls ({Alc } , {Als } , . . .)} .. .

[12.32]

In these equations, the element terms are calculated as follows for each element e of Ω1 : e e lσω 1 e {Rlc {N e } J0e lc dv (· · · )} = − {N e } 2 N e · {Aels } dv + r r Ω Ω ( %−−→ & e '−−→ − curlT (N e ) · νlc,kc · curl(N e ) · {Aekc } dv k=1,3,...

−

Ω

k=1,3,...

e

'−−→ ( %−−→ & curlT (N e ) · νlc,ks · curl(N e ) · {Aeks } dv

Ω

[12.33]

Magnetothermal Coupling

e

e {Rls (· · · )} =

263

e lσω 1 e e {N e } J0e ls dv N · {A } dv + lc 2 r Ω r (T %−−→ & e '−−→ · νls,kc · curl(N e ) · {Aekc } dv curlT (N e )

{N e } Ω

−

k=1,3,...

−

Ω

k=1,3,...

e

'−−→ (T %−−→ & · νls,ks · curl(N e ) · {Aeks } dv curlT (N e )

Ω

[12.34]