2) INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL. 5, 65-82 (1972) DUAL ANALYSIS FOR HEAT CONDUCTION PROBLEMS BY FINITE ELEMENTS BM, FRAEUS DE VEUBEKEf AND M. A. HOGGE} Aerospace Engineering Laboratory, University of Lidge, Lidge, Belglum SUMMARY ‘An alternative approach to the usual finite clement treatment of steady-state temperature problems is presented, using approximations for the field of the dual variables. “The appropriate extremum principle is established and its minimization is discussed in connection with a plane ttiangular finite element process, Original heat flow elements are derived: in conjunction with temperature clements, they enable dual analysis of a given structure and an important estimate of the convergence to the true Solution by upper and lower bounds to the dissipation function, as illustrated by means of several examples. INTRODUCTION Steady-state heat conduction is one of the many field problems of engineering that can receive a variational formulation. The basic formulation’ is in terms of the temperature field alone and the discretization in finite elements makes use of parametric approximations of the temperature within each element with continuity secured across the interfaces. In the present paper an alternative formulation is presented in terms of the dual (vector) variable: the heat flow. In the corresponding finite element models the heat flow satisfies thermal equilibrium within each element and flux continuity is maintained at the interfaces. In this manner a dual analysis of a steady-state heat conduction problem extends to the benefit of the dissipation function a numerical estimate of the convergence by upper bounds (temperature models) and lower bounds (heat flow models) when no heat sources are prescribed. When the body is in contact with a uniform temperature bath taken as zero level, the role of bounds is reversed. The general case where both sources and non-uniform outside temperatures prevail can be treated by superposition iff convergence estimates are required. BASIC TEMPERATURE FORMULATION ‘The temperature functional in a simply connected domain D (1) fice a72%,T- enav+ | erase f, ) whT—Tykas o where $= d, D+a D+ay D, denotes its boundary, 27 the temperature gradient vector, ky a symmetrical heat conductivity tensor, © a prescribed volume heat source distribution, a prescribed surface heat sink, ‘ha convection coefficient and 7, the local prescribed external temperature, is stationary under small pertur- bations 87° of the temperature field subject to the boundary conditions T=T, 8T=0 on aD @ +f Professor of Continuum Mechanics. 4 Research Assistant Received I June 1971 (© 1972 by Jone Wiley & Sos, Lid. 3 6s6 3B. M, FRAEUS DE VEUBEKE AND M, A. HOGGE when the Euler equation A(kg2)T)+0=0 ®) and the natural boundary conditions (n, direction cosines of the outward normal) njkyQT+G=0 on aD O) ngkyy@jT+MT—T,) = 0 on 24D 6 are satisfied. The Euler equation is a statement of thermal equilibrium for the unit volume; under Fourier's law for the heat flow vector Gm Ky GT © it becomes = %q:+9=0 o Three types of boundary conditions are distinguished. The surface heat sink condition (4) on part & D of the boundary, that can also be written =nq¢+4=0 on &D @® ‘the convection loss condition (5) or =ngqy+M(T-T,) =0 on &D o and the prescribed surface temperature condition (2) that can be regarded as a limiting case h->00 of ‘equation (9). ‘The functional of this ‘principle of variation of temperatures’ is the sum of a ‘dissipation functional’ Foty= [ ¥y0707aD+ f, pPr—Tyts (20) of the whole system body plus surrounding,* ie. including the boundary heat convection effect, and a ‘potential energy of prescribed thermal loads” Pr) = ~[,orap+ | ares ay EXTENSION BY RECIPROCAL TRANSFORMATION ‘As was done in the case of elasticity theory® the basic formulation will now be extended through a reciprocal transformation. Fourier's law is used to eliminate the temperature gradients in the dissipation function that becomes, FG = | HrntidndD+ | anr-THas 2) ‘The heat resistivity tensor r;,, is reciprocal to the heat conductivity one renkis= Bg 3) so that Fourier’s law, solved for the temperature gradients, reads int ~ Pinky T= OppT ‘Considered as differential constraints pT + tng = 0 «aay between the temperature and heat flow ficlds those relations are incorporated into a ‘dislocation functional?HEAT CONDUCTION PROBLEMS 6 javolving Lagrange multipliers Ay DG.T Ay) --J MalOn Trina dAD+ | oT—T)AS asy the second part of which also removes the necessity of satisfying a priori the boundary constraint (2). ‘The ‘three-feld’ variational principle FG.T)+D@TAW+P(D), stationary 6) has the following variational equations and natural boundary conditions: en(Qn— Aw) =O OF n= Ay a as variational equation for the heat flow field, Aq dn 9 = 0 as) = tg Ag+ =O on 8D as) =p Ant =0 on %D em inv T-T)=0 on 2D ey as variational equation and natural boundary conditions for the temperature field, and finally (14) and (2) for the variations on the Lagrange multipliers. Equations (17) and (19) identify those multipliers with the heat flow field and its normal component on @ D. When incorporated into the other equations they express all the thermal equilibrium conditions and boundary conditions of the problem. When incorporated ab initio into (16) they reduce this principle to the thermal analogue of the Hellinger-Reissner principle of elasticity e f, Gan T+ inti + OTVAD+ h, jit mass paras | ,IME-T)S, stationary @2) It contains only two independent fields, the temperature and the heat flow. Its corresponding variational equations are respectively (7) and (14), while (8), (9) and (2) all appear as natural boundary conditions, An integration by parts on the first term of (22) puts it into the equivalent form J,7@ntn-Oa0~ |) PintdndD~ |) tnt Td5+ | (4-ratadTAS +f GHT-T)—MadaTIAS, stationary. (3) a0 THE DUAL OR HEAT FLOW FUNCTIONAL “The presence of the temperature field can be completely eliminated from the functional (23) by assuming that we satisfy a priori the volume heat equilibrium equation (7) and the corresponding surface equation (8), whereby the first and fourth terms disappear. However, since the last term still contains T' on 2 D, equation (9) is solved for Tand back substituted. ‘This produces, after a change of sign the ‘principle of variation of heat flow’: I { PrntitndD +], ram Ma5+ |, Oude) Mada TS, sioner. — CA tis the thermal analogue of the complementary energy principle of elasticity. It is again the sum of the68 BM. FRAEUS DE VEUBEKE AND M. A, HOGGE dissipation functional (10) expressed however in terms of the heat flow through Fourier’s law and (9) aft 1 £G)= | SrentitmdD+ |. (ram dS es and a ‘dual or complementary potential energy’ 216) fap TetntndS 6) asap ‘The upper and lower bound character to the dissipation functional of approximations based on either 11) = F)+PC), minimum en o J@)=FG)+O@), minimum @ is briefly discussed in Appendix I. FINITE ELEMENT ANALYSIS The temperature models of finite elements have been fully described elsewhere? as based on a continuous, Piecewise differentiable temperature field with a finite number of degrees of freedom. Note, however, an essential difference in the treatment of the heat convection boundary condition (5). This case is no longer treated by additional generalized fluxes but results in essentially different elements because of the addition of the second term in the dissipation functional (10). This modification is necessary to obtain bounds to this functional. With temperature models, temperature boundary specifications 7, can be accounted for exactly iffexpressible in terms of the approximating shaping functions of the models. Heat flux boundary specifications are translated into equivalent gencralized fluxes, and boundary conditions such as (4) are only satisfied on the average, The same is true for thermal equilibrium in volume, equation (3) and there are discontinuities in heat flux transmission between adjacent elements. This behaviour is analogous to the behaviour of ap- proximate stresses in conforming displacement models for structural analysis. The heat flow models based on the dual principle (28) must, by contrast, satisfy in detail the thermal equilibrium conditions (7) and (8). In general, the finite degree of freedom approximations on the heat flow do not satisfy the integrability conditions for the temperature field required by equations (14). Similarly, temperature boundary conditions such as (2) and (9) are only averaged. This averaging process results automatically from the definitions of generalized temperature co-ordinates as weighted averages along the boundaries of the element or in its interior, so that no restrictive conditions exist on the prescribed tempera- tures input. This situation is again completely analogous to the one governing the theory of equilibrium models in elasticity Particular emphasis will be given here on the heat flow models and, for brevity, only internally unloaded (G = 0) elements will be considered, HEAT FLOW ELEMENTS GENERATION Let E denote the simply connected domain occupied by the element and 9E its boundary. The boundary consists of several parts 0, E (¢ = 1, 2, ...8) corresponding to various adjacent elements or to pieces of the boundary of the complete structure, The heat flow vector field in E is taken as a linear superposition of ‘modes’ My(x), each in thermal equi- librium without internal source 9 = Myx) (29) &Mg=0 (7=1,2,.1m) G0) Consider now the normal heat flux generated along part 2, E of the boundary by each of the heat flow modes L(2) =n) Mo), xE8,E enHEAT CONDUCTION PROBLEMS CF Determine a maximal set of independent vectors jm (mt = 1,2,...4M,) such that Yinf{2) = 0, x€0,E G2) Then, any complementary set of vectors Mp (P= 1,2,» indexing the columns, (*) \"p, is non-singular, will generate a set of non-zero and independent boundary flux modes along @,.E iph{2) = byl), KEBLE (P= 1,2 00 Pa) G3) Introducing the matrix notations 4; My+P, =n) such that the nxn matrix, J HO). fiC), x€8,E x) bp), ¥E0,E (32) and (33) are equivalent to the matrix equation 6.t.0)= [fol eek @o Inversion of this produces £,() = BTb,(), x2, E 5) where BY, which is a part of G4, isa nx P, matrix of rank P,. Thus, along @, £, the general parametric heat flow will generate a boundary flux of type s7E62) = (B,a)"b,(a), xe0,E 06) ‘The set of ‘generalized flux co-ordinates’ for @,E is defined by t= Bea 67 so that (36) becomes equivalent to fi) = Eph), XEREE (38) The generalized flux co-ordinates are thus intensities of the boundary heat flux modes; their knowledge determines the complete heat flux distribution along @, E so that the continuity of heat fluxes at interfaces ‘can be secured simply by reciprocity of corresponding generalized flux co-ordinates. Any predetermined set of boundary heat flux modes may be replaced by a new set of P, linearly independent combinations of the elements of the old set. Advantage can be taken of this freedom to give simple physical significance to the generalized flux co-ordinates, either as local values or moments of various order. The set of relations (37) for all parts of the boundary @£ results in a global relation =Ba 69), Where the boundary flux co-ordinates are listed in g in some conventional order and B correspondingly partitioned. As a determines g, we should have conversely that a compatible set g determines a uniquely. ‘The condition therefore is that Ba = 0 possesses only the trivial solution a = 0, If this is not true there are “bubble heat flow modes’ that were not taken into consideration. Bubble heat flow modes are linear combi- nations of the modes M(x) that produce no heat flux along any part of the boundary 2B and are therefore riot accounted for by any of the generalized boundary heat flux co-ordinates. This happens only for models involving a large number of parameters; if the My(x) form a complete set of polynomials of degree Ms, 5) — (55g) (74) IOS. IS)— 8, 0) 45,8) — (6) (75) resulting from the fact that the minima of (68) and (69) are reached by the exact solution. Eliminating in ‘them (5,8) through (73) and combining with (71) and (72) there comes (hye 6sho- ABs Fo) 6) This upper and lower bound characteristic for the linear functional (oho €) = (hd — (6.0) = (+e he) (hehe) chd—(se) = [rape Terman dS~ | Teds can be used directly in solving a mixed boundary value problem but does not guarantee convergence in norm. + Complete analogous statements to equations (27) and (28) are in fact: 1, Principle of variation of temperatures: Heartem c9*+ 0 fi) — Hho Coy = Heatescotd—(6, hy, if we drop the constant («4 2. Principle of variation of heat flow: Hoth ey hy he) Hy Co hee), minima heh hot #)— Ch, 6), if we drop the constant (fy hy)HEAT CONDUCTION PROBLEMS ” To obtain a measure of convergence in norm it is necessary to solve the problem in two steps by super- position. First step. c=, the prescribed temperatures are set equal to zero, only the prescribed heat sources are taken into account. Then cf = c* and the previous result becomes (cf) < (5) < (Eh) However from expressions (71)-(73) also Cet) = (Cha) (5,8) = (hg) (ASL 1B) = (hd a) ‘so that we obtain upper and lower bounds for the quadratic dissipation functional (e*c*) < (5,5) <3, 3) Second step. hy =0, the flow rate of the prescribed heat sourees is set equal to zero but account is taken of the prescribed temperatures. Then ht = h* and (76) becomes ($60) > (5,66) (H%, 60) Furthermore from expressions (71)-(73) (hed) = (CBB) (EI = G8) Oa) = HY and the dissipation functional of that case is now bracketed in the reverse way hed) 26.9200) APPENDIX II Constant heat flow element Fw 3 Coste ow Seen tent naa, Ma() JB Gad xy “Yn YulLhy ull 1 =Onrg Barer rary F Tas 6 Xiiy BE a Yye Oa Xe Ta leceoss80 B. M, FRAEUS DE VEUBEKE AND M. A. HOGGE where Lay= Length of edge i-j, xy =X Yy=Y-Yy 87 = (Bie Sm Sa) t™ (a fn a) Ja %a vk ie SYM ° a hasly v-[om ae }, mea 2), me Ya Xa ie Ey Shes Dik; where A= Area of the clement, hy = Uniform transfer coefficient for edge i—j =2if y= Oy APPENDIX IIL Linear heat flow element Ge = HH MX+ OY, = at mx), 1 Ty OH Xa Xs We HHH PH) a fe Ppa? OF B= Fe x) m3) ajar? on ate 9where where HEAT CONDUCTION PROBLEMS B, -3(* ye My Mn ava) My xy Xu WK AW IY 1 (eatesret MX Xyt IIE ARI a) & Doxa Wye ta Xeyt 19 X5 Xeg + MAY; Vset OLR; Yye Yj Xe) 2 BaF ([ mey-ods sa) Bees Jig fen gu fo Sa gn 8) a(h2 fa fe fe fa fs) Ja Me Ae a name Ja Ma 2 Jen name Jan Se Xa%en Jaen adn —Jaten Jor Ma Xa%en Jaen FY a%en Jas As Aa%1 YsVia adhe Ia%a Ja An AX Ke Xe ike 0 Aly SYM 0 Salky —Ialky Syke 0 0 Ike Sclke —Sylky —Jaylky Jaylka Tale Iylky a= [ax Ss m xydsidy, ete, f, dy, Se free fore dy, Sey I. wdxdy, et Rand, = pet x a pct ah P= BE Ry Ram BO Pm Ry Ryn ECEAAD Ap poh utah Bh FO ShgLy Oty) Iie 1 BOD Ra Thal2 B. M, FRAEUS DE VEUBEKE AND M. A. HOGGE 2X, y F XVp + 4 Vet DNV) Rep Ve Re Chal (ict a) ye Ot yD Xie Ra OT fe (sc 9) xy yy~ Out 9) ay Rigg = BE thm Ras 2 gly Re 2b bey ta) xu Cxoretxoy tact ep) aty Rant hy ly sd Ayn BOWEL Dri Ob ortyD su |p, 7 Ghyly Ry = BOLD. Sit + ey Doh Crt ny +X Yet 2H) YI Se isla REFERENCES |. 0. C, Zienkiewiee and ¥. K. Cheung, The Finite Element Method in Structural and Continuum Mechanics, McGraw-Hill, London, 1967, Chap. 10. 2. M. A. Biot, “New methods in heat low analysis with application to fight structures’ J. aeronaut. Sei. 24, 857-873 (1957) 3. B. M. Fraels de Veubeke, ‘Displacement and equilibrium models’ in Stress Analysts, Wiley, London, 1965, Chap. 9. R. Courant and D. Hilbert, Methods of Mathematical Physies, Vol. 1, Wiley-Intestience, New York, 1953. 5. PJ. Schneider, Conduction Heat Transfer, Addison-Wesley, Reading, 1957.