Comput. Methods Appl. Mech. Engrg.

213216 (2012) 93103

Contents lists available at SciVerse ScienceDirect

Comput. Methods Appl. Mech. Engrg.

journal homepage: www.elsevier.com/locate/cma

Non-linear nite element formulation applied to thermoelectric materials

under hyperbolic heat conduction model
R. Palma a,, J.L. Prez-Aparicio a, R.L. Taylor b
a
b

Mecnica de Medios Continuos y Teora de Estructuras, Universidad Politcnica de Valencia, Spain

Department of Civil and Environmental Engineering, University of California at Berkeley, CA, USA

a r t i c l e

i n f o

Article history:
Received 9 March 2011
Received in revised form 12 August 2011
Accepted 14 November 2011
Available online 25 November 2011
Keywords:
Thermoelectrics
Dynamics
Non-linear nite element
Second sound
HHT algorithm
Newmark-b algorithm

a b s t r a c t
In the present work, a three-dimensional, dynamic and non-linear nite element to simulate thermoelectric behavior under a hyperbolic heat conduction model is presented. The transport equations, which couple electric and thermal energies by the Seebeck, Peltier and Thomson effects, are analytically obtained
through extended non-equilibrium thermodynamics, since the local equilibrium hypothesis is not valid
under the hyperbolic model. In addition, unidimensional analytical solutions are obtained to validate
the nite element formulation. Numerically, isoparametric eight-node elements with two degrees of freedom (voltage and temperature) per node are used. Non-linearities due to the temperature-dependence
on the transport properties and the Joule effects are addressed with the NewtonRaphson algorithm.
For the dynamic problem, HHT and Newmark-b algorithms are compared to obtain accurate results, since
numerical oscillations (Gibbs phenomena) are present when the initial boundary conditions are discontinuous. The last algorithm, which is regularized by relating time steps and element sizes, provides the
best results. Finally, the nite element implementation is validated, comparing the analytical and the
numerical solutions, and a three-dimensional example is presented.
2011 Elsevier B.V. All rights reserved.

1. Introduction
Thermoelectric materials couple electric and thermal energies
by means of three separated transport effects: Seebeck, Peltier
and Thomson. In addition, the Ohm and Fourier laws that are inherent to electric and thermal uxes are also present. Thermoelectric
materials are used as heat pumps (heating and cooling) and generators, see [1], and in the last decade miniaturized thermoelectric
devices and high-frequency processes are increasingly applied to
cool micro-electronic devices, see [2].
The classical Fourier law leads to a parabolic heat propagation
problem incorrect from a physical point of view: especially applied
to micro-devices and under rapid transient effects such as micropulses. Cattaneo [3] proposed a modication to the Fourier law that
leads to a hyperbolic heat propagation problem, avoiding the classical paradox related with the innite temperature velocity of
propagation. The modication, named sometimes second sound,
is made through the introduction of a relaxation time s, an empirical parameter dened as the time-interval between two successive collisions at the microscopic level of either holes or electrons.
From a theoretical point of view, the classical non-equilibrium
thermodynamics [4] cannot lead to hyperbolic propagation since
the local equilibrium hypothesis is not valid. For this reason, ex Corresponding author.
E-mail address: ropalgue@upv.es (R. Palma).
0045-7825/$ - see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2011.11.011

tended non-equilibrium thermodynamics [4,5] must be applied

to obtain the correct transport equations.
Several nite element (FE) formulations applied to the thermoelectric problem using the Fourier (parabolic) model, see [6,7] have
been published. Analytically, in [8] the Laplace transform technique
was used to solve problems formulated using the Cattaneo (hyperbolic) model. For heat propagation (without thermoelectric coupling) analytical investigations were performed using the Laplace
transform technique under the Cattaneo model [911]. In [12] a
mixed FE formulation using the CrankNicolson scheme for timeintegration was presented and in [13] a nite difference scheme
was developed to compare implicit and explicit Euler algorithm
solutions. Several FE works in which the Cattaneo model was used
also have been published. For example, in [14,15] FEs were formulated to simulate convectiondiffusion (Ficks law) and thermomechanical coupling, respectively.
The aim of the current work is to develop a three-dimensional
(3D), transient, non-linear FE formulation to solve general thermoelectric problems with the hyperbolic model. For this purpose, a
non-equilibrium entropy depending on temperature and heat ux
is dened and the transport equations are formulated by applying
the Onsager relations. Transport and balance equations are then
transformed into a matrix form and implemented into the research
FE code FEAP [16]. The non-linearities due to: (i) Joule effect, (ii)
temperature-dependence of transport properties and (iii) thermoelectric coupling between electric and heat uxes, are solved with

94

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

the NewtonRaphson (NR) algorithm. Two discrete algorithms for

transient solutions, namely, HHT or a [17] and Newmark-b (N-b)
are compared. Both are regularized by means of a linear relationship between time steps and element lengths. In order to validate
the FE behavior, four unidimensional (1D) analytical solutions are
compared with the numerical results. When the initial boundary
conditions are discontinuous, strong oscillations (Gibbs phenom and c

ena) are observed and the N-b algorithm, with modied b
parameters, provides results that are consistent and accurate with
respect to the analytical solutions. Finally, a 3D case is simulated,
obtaining several conclusions for the design and optimization of
thermoelectric devices.

Neglecting qX, assumption valid for most thermoelectric applications [19], the charge and energy balances are

The extended non-equilibrium thermodynamics theory assumes the existence of a non-equilibrium entropy s that depends
on state variables and on uxes. Since the electric energy propagates much faster than the thermal one, the hyperbolic model is
applied only for the thermal eld, and the entropy balance is given
by Jou and Lebon [4] adding the Joule irreversibility

qs_ r 

2.1. Balance laws

Consider a thermodynamic universe composed of a system of
domain X, boundary C and its surroundings, see Fig. 1. Assuming
the validity of a continuum hypothesis, the balance laws of continuum physics are those of charge and energy. The rst is obtained
from a Maxwell equation, see [18]

q_ X r  j;

where a dot () denotes a time derivative, j is electric ux and qX is

charge density.
Energy balance is obtained by applying the rst law of
thermodynamics

qu_ r  1 j  E;

2.2. Transport equations

2. Thermodynamic formulation
The aim of this section is develop balance and transport equations for the hyperbolic thermoelectric problem. As mentioned
above, the hyperbolic problem cannot be obtained from classical
non-equilibrium thermodynamics. This is due to the lack of validity of the local equilibrium hypothesis that states the equality of
the local relations between thermodynamic quantities in a system
in and out of equilibrium [5].
This validity is closely related with the Deborah number De =
s/sM < 1, where sM is the macroscopic experimental time (effect
duration) and s the time interval between two successive collisions
between particles. Large Deborah numbers occur if: s is large, for
materials such as polymers, and/or sM is short, for very fast phenomena such as ultrasound propagation. The extended non-equilibrium thermodynamics [5] must be used for transport equations
when De P 1.

qu_ r  1 j  E:

r  j 0;

1
T

  

j  rV
1
s
 2 1_ ;
1 r
T
T
jT

where T, V, j denote temperature, voltage and thermal conductivity,

respectively. From the general form of the entropy balance, the entropy ux js and the rate of entropy production rs are

js

1
T

rs

  

j  rV
1
s
1 r
 2 1_ :
T
T
jT

The classical form of the entropy production for the thermoelectric

problem, see [20], is recovered if s = 0: the hyperbolic model adds a
term to the entropy production.
Finally, the transport equations are obtained by expressing the
entropy production in the form from [5] and by satisfying the
OnsagerCasimir reciprocal relations to give

j crV  acrT;
1 s1_ jrT aTj;

where a, c denote the Seebeck coefcient and thermal conductivity,

respectively. These phenomenological coefcients are denominated
thermoelectric properties in the following. Again if s = 0 in (6)-bottom, the classical thermoelectric transport equations are recovered.
Although linear relationships between uxes and gradients
have been used, the thermoelectric properties depend strongly
on T (material non-linearity), accordingly in the present work

aT 1:98  104 3:35  107 T  7:52  1010 T 2 ;

cT 1:09  105  5:59  102 T 2:49T 2 ;
jT 1:66  3:58  103 T 3:19  105 T 2

in which dimensions are: a [V/K], c [A/mV], j [W/mK] but T is introduced in Celsius (C) degrees, see [21].

_ 1 and E denote mass density, rate of internal energy

where q, u,
density, heat ux and electric eld, respectively. The last term on
the right-hand represents the rate of electromagnetic energy converted into thermal by the irreversible Joule effect.

2.3. Strong form

From the previous equations and taking into account the coupled nature of the thermoelectric problem, the strong form is
formed as a system of two coupled partial differential equations.
The electric strong form is obtained by replacing (6)-top into (3)left and considering the T-dependence of the thermoelectric properties (7)

cTr2 V @ T fcTgrT  rV cT@ T faTg aT@ T fcTg

aTcTr2 T:
8
Here and in the following, @ T implies partial derivation with respect
to the subscript or subscripts, in this equation the temperature T. To
obtain analytical solutions, from (6)-top we deduct

rV aTrT 
Fig. 1. Energy balance in a domain X: the three effects, reversible couplings (arrows
without border) and irreversible thermal and electric uxes (arrows with border).

cT

where the rst term in the right-hand represents the thermal energy converted into electric by the Seebeck effect. Due to restrictions

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

in analytical solutions (not in the numerical ones), in this article we

present cases in which only a is variable with T. The values of j, c
will be evaluated at a certain average T from (7).
The thermal strong form is obtained substituting (6)-bottom
into (3)-right and using (3)-left. Furthermore, the relations u = cT
(c, specic capacity), E = rV are introduced

95

rdV  jdX C dVjc dC 0;

R
R
X rdT  qdX X dTj  rVdX X dT s@ t fj  rV gdX

RX

RX
C

X
dT sqcTd

_ X
dT qcTd

13

dTqc dC 0:

3.2. Discretization

sqcT qcT_ jr2 T  th rT  j  s@ t fj  rVg;

c

10
The continuum domain X is discretized into elements Xe, see
nel
Fig. 2. Therefore, X [ei
Xe where nel is the total number of elements, each one delimited by npe = 8 nodes. In turn, each node relates to two degrees of freedom (voltage, temperature) per node.
For an element e, degrees of freedom and spatial coordinates x
are approximated using 3D isoparametric shape functions Ne

where th = T[@ a(T)/@T] is the Thomson coefcient. This expression is

an hyperbolic equation, notice the rst term on the left-hand. The
right-hand terms represent several effects: the rst classical heat
conduction (Fourier), the second Joule, the third Thomson and the
last dissipation related to the hyperbolic model.
Fig. 1 schematizes the energy balance of the thermodynamics
formulation. Electric and thermal energies are represented left
and right, respectively; sources top and the ux bottom. The electric energy is irreversibly converted into thermal by the Joule effect,
therefore the related arrow points towards the thermal energy
side. Inside the domain, electric energy is reversibly converted into
thermal and vice versa by three separated effects: Seebeck (thermal
into electric), Peltier (electric into thermal) and Thomson in both
ways.

n
o
n
o
where Ve V 1e ; . . . ; V ne pe , Te T 1e ; . . . ; T ne pe . Be is the discretized
gradient matrix of the element composed of npe submatrices with
the form:

3. FE formulation

BtA

The aim of this section is to develop a variational formulation

within the FE framework to permit the implementation of the
thermoelectric problem into a computer code. For this purpose,
consider a thermoelectric body of domain X and boundary C governed by balance and transport equations obtained in the previous
section. To facilitate the FE implementation, again the hyperbolic
ux s1_ from the thermal transport Eq. (6)-bottom is introduced
in the energy balance (3)-right, with the result that the expression
of the pure parabolic heat ux q is recovered in the transport
equation

sqcT qcT_ r  q  j  rV  s@ t fj  rVg

q jrT aTj

11

V  V h Ne V e ;

T  T h Ne Te ;

rV  rV Be Ve ;
T_  T_ h Ne T_ e ;

rT  rT h Be Te ;

e ;
T  T h Ne T

14

x  xh Ne xe ;

@N A
@x1

@N A
@x2

@N A
@x3

i
;

15

where ()t denotes matrix transpose. Similarly, the variations are

discretized as

dV  dV h Ne dVe ; dT  dT h Ne dTe ;
rdV  rdV h Be dVe ; rdT  rdT h Be dTe :

16

Introducing these discretizations into the transport equations (6)top and (11)-bottom

je cBe Ve  acBe Te ;
qe jBe Te  acNe Te Be Ve  a2 cNe Be T2e :

17

Introducing now (14)(16) into (13) the Galerkin forms are

obtained

The governing equations are complemented by Dirichlet (on

boundaries CV, CT) and Neumann (on Cj, Cq) boundary conditions
for electric and thermal elds

VV

on CV ;

j  n jc

on Cj ;

TT

on CT ;

q  n qc

on Cq ;

where V, T, jc and qc are prescribed voltage, temperature, electric

ux and thermal ux, respectively, and the outward normal is denoted by n. Boundaries satisfy the conditions CV [ Cj = C,
CV \ Cj = 0 and CT [ Cq = C, CT \ Cq = 0.
3.1. Weak forms
According to standard variational methods [22], the weak forms
are obtained by multiplying the balance Eq. (3)-left and (11)-top by
variations of the degrees of freedom, dV, dT and integrating over X

R
R

dV r  jdX 0;
R
R
_
X dT r  qdX X dTj  rVdX X dT qc TdX
R
R

X dT s@ t fj  rVg dX X dT sqcTdX 0:
X

12

The s and q-terms are new with respect to the formulation developed in [7]. The divergence theorem is now applied to the rst term
of both equations, and Dirichlet boundary conditions are introduced

Fig. 2. Domain discretization in nite elements Xe. Isoparametric three-dimensional and interface ux elements. Two degrees of freedom per node.

96

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103




R
R

Xe

Bte je dXe

Xe

Bte qe dXe

R
R

Nte

Ce

Nte jc dCe 0;

V
K VV
AB @ V B fRA g

Nte je Be Ve dXe Xe sNte @ t fje Be Ve gdXe

R
qcNe T e dXe Xe Nte qcNe T_ e dXe
Xe

Xe

Ce

Nte qc dCe 0:

18

3.3. Non-linear transient solution form

In the present work a non-linear, transient hyperbolic problem
is studied. As mentioned several non-linearities are present, and
the transient behavior is relevant only for the thermal eld, since
qX in (1) is assumed to be zero.
In order to solve the transient problem, the time interval of
interest is divided into small increments, Dt, and the N-b or HHT
method is introduced to replace time derivatives by discrete forms.
Then for each time point the NR method is used to solve the
resulting non-linear algebraic problem. These transient algorithms
have been widely used in structural and solid mechanics (see [22])
and will not be described in detail here. Basically, the assembled
non-linear FE equations are written in a residual form R and linearized by

k
@RA 
RkA 
dgkA
g 

19

where A, B are the global numbering of two nodes, k the NR iteration counter and gA derivatives of the degrees of freedom at node A:
_ A fV_ A ; T_ A gt , and second
value as UA = {VA, TA}t, rst derivative as U
t

derivative as UA f0; T A g . The algorithms for time integration are

taken as

k
@RA 

c1 KAB c2 CAB c3 MAB :
@gB 

20

 c
 for
In Table 1 the parameters c1, c2 and c3 are given in terms of b,
 for the HHT algorithm.
the N-b algorithm and a
The consistent tangent, capacity and mass matrices, are derived
for each iteration

KAB 

@RA
;
@UB

CAB 

@RA
;
@ U_ B

MAB 

@RA
:
B
@U

21

Finally, the solution is updated by gk1

gkB dgkB . Note that to enB
sure a correct derivation of the tangent matrix, the NR should exhibit a quadratic asymptotic rate of convergence.
From (18), the assembled residuals for each eld at each global
node are formulated as

RVA 
RTA

BtA jA dX

C N A jc dC;
R
t
B
q
d
X

X A A
X N A j A BA V A dX
R
R
X NA s@ t fjA BA V A gdX X NA sqcT A dX
R
R
NA qcT_ A dX NA q dC:

V
K VT
AB @ T B fRA g

K TV
AB

From (21)-left and (22), the tangent stiffness matrices are

BtA @ V B fjgdX;

BtA @ T B fjgdX;

23

Similar tangent matrices were obtained in [7] for the steady-state

problem, situation for which both theories (parabolic and hyperbolic) are equivalent. Using now (17) and the chain rule

@ V B fjg cBB ;
@ T B fjg acBB  @ T fagNB cBB T B
@ T fcgBB V B aBB T B NB ;

24

@ V B fqg aNB T B @ V B fjg;

@ T B fqg aNB T B @ T B fjg ajB NB  jBB
@ T fagNB T B jB NB  @ T fkgBB T B NB :
From (21)-middle and (22), the capacity matrices are

n o
V
C VV
0;
AB @ V_ B RA
C TT
AB

n o
V
C VT
0;
AB @ T_ B RA

R
NA @ tV_ B fjgBB V B dX  s X NA jB BB dX;
R
R
@ T_ B fRTA g s X NA @ tT_ B fjgBB V B dX  X NA qcN B dX;

T
C TV
AB @ V_ B fRA g s

25
where the derivatives are again obtained applying the chain rule to
the discretized transport Eq. (17)

@ tV_ B fjg cBB ;

@ tT_ B fjg acBB :

26

As mentioned the symbol @ tV_ B implies second partial derivative with

respect to the time and the temporal derivative of V. From (21)right and (22), the mass matrices are obtained by
V
MVV
fRA g 0;
AB @ V
B

V
M VT
AB @ T B fRA g 0;

T
MTV
fRA g 0;
AB @ V
B
R
T
MTT
AB @ T B fRA g  X N A sqcN B dX:

27

The nal assembled matrix problem is given by

"

c1 KTT c2 CTT c3 MTT

T
V

22

X BtA @ V B fqgdX
R
 X NA @ V B fjgBB V B dX  X N A jA BB dX
R
R
 X NA s@ V B fjgBB V_ B dX  X NA s@ t fjgBB dX;
R t
T
K TT
AB @ T B fRA g X BA @ T B fqgdX
R
R
 X NA @ T B fjgBB V B dX  X NA s@ T B fjgBB V_ B dX:

@ V B fRTA g

k1

c1 KTV c2 CTV

c1 KVT
c1 KVV

k

k
T
dT

:
V
dV

dT
dV

RT
RV

)k
;

28
The present formulation is implemented as a user element in the
non-linear nite element code FEAP, see [16]. Each element uses
a standard, in the nite element method (FEM) sense, eight-node
isoparametric element with two degrees of freedom per node A:
voltage VA and temperature TA.
4. Validations

Table 1
Tangent matrix parameters for Newmark and HHT algorithms.
Method

Parameters
c1

c2

c3

N-b

HHT

a

c
 Dt
b
a c
 Dt
b

1
 Dt 2
b
1
 Dt 2
b

The aim of this section is to validate the FE developed in Section

3 by means of four cases for which analytical solutions are found in
the literature or modied and developed here. These solutions are
1D simplications of the coupled partial differential equations (9)
and (10) and are summarized in Table 2. Due to the difculty in
developing analytical solutions, the T-dependence of the electric
and thermal conductivities is not included in the validations. However, for the steady-state problem, this dependence has been veried in [21] using experimental results.

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103
Table 2
Unidimensional validation cases, simplications and considered effects: F - Fourier, S Seebeck, J - Joule, Th - Thomson, ss - second sound.
Case

Simplications

Effects

I
II
III
IV

s = 0; j = 0; a, c, j = cte
s = 0; j 0; a, c, j = cte
s = 0; j 0; a(T); c, j = cte
s 0; j 0; a(T); c, j = cte

F,
F,
F,
F,

S
S, J
S, J, Th
S, J, Th, ss

For all cases, a p-type thermoelement (TE) device manufactured

by MELCOR [23] with properties described by (7) is modeled. The
TE is a parallelepiped of dimensions 1.4  1.4  1.14 [mm]. For a
p-type TE, the electric ux direction is co-linear with the heat ux.
The boundary conditions are discontinuous: initially the temperature is T = 0 everywhere. For t > 0 the boundary temperatures are
set to Tc = 30 on the cold face and to Th = 50 C on the hot face.
The boundary voltage is set to V = 0 [V] on the cold face. Fig. 3
shows the geometric dimensions and the boundary conditions.
For cases II, III and IV, an electric ux in the x3 direction is prescribed using the special interface element developed by the
authors in [21], see Fig. 2. The applied intensity 5.2 [A] is an average of the ones specied by the manufacturer and corresponds to
an electric ux j3 = 2.65  106 [A/m2].
The constant parameters for a, c, j are obtained from (7) using
an average temperature Tm = (Th + Tc)/2. For all cases, the electric
strong form is directly given by (9), however, the thermal strong
form for each case must be obtained by introducing the simplications from Table 2 into (10)

j@ xx fTg qc@ t fTg;

j@ xx fTg qc@ t fTg  j2 =c;
2
III : j@ xx fTg qc@ t fTg  j =c t h j@ x fTg;
2
IV : j@ xx fTg qc@ t fTg  j =c t h j@ x fTg
s@ t fj@ x fVgg sqc@ tt fTg:

I:

II :

29

The fourth and fth terms of the right side of case IV (second line)
represent the second sound effects: irreversibility, (5)-right and
hyperbolicity, respectively.
4.1. Cases IIII
For these cases, s = 0 and the thermal strong forms (29) are diffusive: second order parabolic partial differential equations. For I,
the situation is linear and homogeneous; for II linear and nonhomogeneous and for III non-linear and non-homogeneous.
Analytical solutions for the thermal eld in cases I and II are given in [24], while the electrical elds can be calculated from (9),
see the Appendix. For III, the thermal and electrical solutions are
given in [25].
Numerical solutions are obtained using a structured and coarse
mesh of 11 elements in the x3 direction. Only one element is used

97

in the x1 and x2 directions since the problem is fundamentally 1D.

A time step of Dt = 0.1 [s] and the standard N-b parameters
 0:25, c
 0:5 are used.
b
Fig. 4 shows the voltage (left) and temperature (right) distributions along the x3 direction for case I (top), and for coincident cases
II and III (bottom). Solutions at t = 0.05, 0.3, 5 [s] are represented:
the analytical ones with lines and the FEM results with circles.
For case I, the T distributions are quadratic in nature during the initial transient response due to the boundary conditions. At near
steady-state (e.g., t = 5 [s]) they become nearly linear, since this
problem is of the Laplace-type: the lack of electric current implies
the absence of a Joule effect. The V distributions are proportional to
T, see (9).
For case II, near quadratic distributions appear for the steadystate due to the Joule effect: this is a Poisson-type problem. Now,
V distributions are not proportional to T, since the potential drop
increases due to electric energy being converted into thermal energy. Results for case III are very similar, since the Thomson effect
is not relevant under the applied intensity, see [7]. As expected
for these simple cases, the agreement between analytical and
FEM distributions is very good.
Fig. 5 shows the vertical thermal ux vs. length for cases I (left)
and coincident II, III (right). Analytical and FE results differ slightly
at the edges, since the mesh is not highly rened to capture the
Newmann boundary condition. For short times, the ux distribution shapes are very similar in the three cases, since the T distributions have the also the same shape. Near steady-state, a constant
distribution is obtained for case I (from the quasi-linear T distribution) and a linear for case II (from a quadratic T). In addition, for
case II the heat ux changes sign due to the second term in the
right-side of (11)-bottom prevailing over the rst one. Note that
the prescribed electric ux is negative, according to the ux direction into the p-type TE.
4.2. Case IV
Case IV is a non-linear and non-homogeneous hyperbolic problem. The analytical solution for the temperature is given in [8] and
is based on a Laplace transform solution technique. However, several differences between this solution and our numerical results
are found, thus, a new corrected comparison solution is deduced
and summarized in the Appendix. Furthermore, in the present
work the heat ux is calculated using a semi-analytical procedure
that combines the analytical temperature solution with nite differences to obtain the gradient in (11)-bottom. No analytical results for V are generated due to the necessity of using numerical
techniques in all steps.
The response in this case is going to be hyperbolic at the beginning (transient state) and parabolic after some transitory time
(permanent state). The rst behavior is evident at early times
where the V, T distributions appear as piecewise functions with
sharp discontinuities near the edges. This occurs since both energies travel as nite velocity waves, see Fig. 6. In addition, strong
changes of T occur when discontinuous initial boundary conditions
are used, inducing a Gibbs phenomena.
For larger times, the parabolic behavior dominates and, therefore, the numerical algorithm must be:
 robust and efcient to automatically solve both parabolic and
hyperbolic problems depending on the instant;
 able to have controllable numerical dissipation to mitigate the
Gibbs phenomena.

Fig. 3. p-type thermoelement geometry, applied boundary conditions, dimensions

uxes and coordinates.

For a linear problem and according to [13], no numerical oscillations will appear if the dynamic algorithm is regularized by the
relationship Dt 6 Ch/v, where C is the Courant number [26] and

98

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

Fig. 4. Voltage (left) and temperature (right) distributions along thermoelement length for cases I (top row) and coincident II, III (bottom row) at three times in [s], see Table
2. Analytical results with lines, nite element with circles.

Fig. 5. Thermal uxes for cases I (left) and coincident II, III (right) at three times in [s], see Table 2. Analytical results with lines, nite element with circles.

v j=qcs,

the linear wave velocity. The cases considered in

the present work are highly non-linear, therefore the spectrum
of the algorithm is not enough evident for a good estimation of
the ratio Dt/h. In our analyses a structured mesh of 200 elements
(h = 8.1429  106 [m]) in the x3 direction is used. Since a precise
spectral analysis is not an objective of this paper, a value C = 1/6
is chosen, selection based on a series of numerical tests. If smaller values of C (e.g., 1/10 and consequently smaller Dt) are used
the oscillation appears again, as seen in Fig. 6 bottom. Larger
values of C will articially smear the distribution at the front
itself.
In order to evaluate the performance of the time integration
scheme, N-b and HHT algorithms are compared using parameters
 , see [27] and Table 3.
that produce different spectral radii q
 should stay close to unity for small to intermeAccording to [28], q
b ! 1,
  0:5 only when Dt= T
diate time steps and decrease to q
b is the undamped natural period.
where T
Fig. 6 top shows T vs. length x3 for an assumed s = 0.02 [s] (as in
most of this section) at t = 0.06 [s]; for shorter times the comparison analytical solution does not converge since it involves Riemman inverses; the numerical one does although with increasing
Gibbs noise. The analytical solution (solid line) and the FE (dashed)
are obtained with N-b and HHT algorithms (these are unconditionally stable for the parameters tested when applied to linear
problems).

For N-b, the use of standard parameters (test a) results in

numerical oscillations since this algorithm is non-dissipative. The
HHT algorithm (test c) slightly oscillates due to its low numerical
damping. Therefore, in the remainder of the present work the
parameters of test b will be used, since according to [27] this
choice results in the highest numerical dissipation.
Fig. 7 shows the V (left) and T (right) distributions along the TE
at several instants. In the rst, it can be observed the transition
from very small times and hyperbolic solution to higher ones and
parabolic, the last with almost linear distribution between boundary values. Notice in the curve for t = 0.06 [s] the electric wave
fronts from left and right at around 0.28 and 0.8 [mm], with a
non-linear distribution inside the wave and linear between the
fronts. For t = 0.13 [s] both waves almost collide at around
0.63 [mm], point shifted to the right. This is so since the T
boundary condition is higher at the left: the energy and consequently velocity are also higher for the left wave.
In the T distribution a similar transition from hyperbolic to parabolic can be seen. The collision between the thermal waves is
again clear for t = 0.13 [s] at the same TE point, a logical result since
the thermal and electric waves are coupled. For t = 0.06 [s] it can be
appreciated that inside the wave zone the slope of the distribution
is higher than the obtained with a parabolic model, with the consequence of a higher energy conned in a smaller volume. The
agreement between analytical and numerical results is very good,

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

99

Fig. 8. Semi-analytical and numerical comparison of front wave thermal ux vs.

thermoelement length at t = 0.06 [s] for s = 0.02 [s].

Fig. 6. Analytical temperature solution for t = 0.06, s = 0.02 [s], C = 1/6 (top).
Newmark-b and different HHT parameters to study Gibbs phenomena (middle),
see Table 3. Adjusting by trial and error the Courant number in Test b to avoid
oscillations (bottom).

Table 3
Newmark-b and HHT parameters used for numerical testing.
Test

Algorithm


q


b

c

a

N-b
N-b
N-b
N-b

1
3/2
1/2
0

1/4
9/14
4/9
1

1/2
16/49
5/6
3/2

HHT
HHT

3/2
1/2

9/14
4/9

16/49
5/6

6/7
2/3

except for a small region around the wave front at t = 0.06 [s]; this
is due to the intrinsic oscillations from the discrete time integration scheme and is very difcult to avoid.
Fig. 8 compares the semi-analytical and FEM thermal uxes for
t = 0.06 [s]. Two peaks due to the discontinuous boundary conditions at the wave front are evident, representing the propagation

of this wave. Even if a nil ux at the TE center could be expected,

it has a constant negative value. The reason is found in (11) bottom: the rst right-term is zero due to a constant T distribution
where the wave is not present (see Fig. 7). This distribution is
due to the electric coupling, see the last term in (11) top. Also notice that the second term in (11) bottom will be negative due to the
prescribed j. The agreement is again very good, even at the peak
maximums, although it has to be considered that the two maximums are affected by the respective choices of the numerical
parameters.
As mentioned before, the behavior is hyperbolic in the initial instants and becomes parabolic later. This is clearly seen in Fig. 9 top,
where the heat ux is plotted for several times. For t < 0.06 [s] the
peaks described in the previous gure are visible, but for larger
times they become smoother. At t = 0.13 [s] both waves collide,
and after that the ux will be linear as in Fig. 5. The mathematical
explanation of this attenuation can be seen in (11)-top: the rst
term in the left hand side is hyperbolic, the second parabolic and
the last in the right dispersive. In the initial instants the hyperbolic
term is dominant; physically this can be interpreted as a ballistic
motion of either electrons or holes. The wave is constantly attenuated by the dissipation introduced by the last term up to a time in
which the inuence of the s terms vanish, and the rst time derivative is the dominant term.
The last attenuation is obviously affected by the value of the
constant s, see Fig. 9 bottom, where it can be seen that for larger
s the behavior will be more hyperbolic (more difcult to simulate)
and vice versa. This gure is a test for the robustness of the nite
element performance, but even for s = 0.06 [s] the simulation is
satisfactory as long as the inequality Dt 6 Ch/v is used.
5. A complete three-dimensional simulation
There are in the literature several studies aimed to the optimization of TE geometry, see [29,30] for example. However, these
studies use 1D models (at most 2D), not being able to observe
3D effects. The aim of this section is to study the 3D temperature
propagation inside a TE by means of the hyperbolic model validated in the previous section. Note that the propagation of V is proportional to that of T through the thermoelectric coupling, and
therefore will not be included here.

Fig. 7. Voltage (left) and temperature (right) vs. thermoelement length at three times in [s]. For voltage, only nite element results; for temperature, analytical results with
lines and nite element with circles.

100

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

Fig. 9. Front wave thermal ux vs. thermoelement length at several times for s = 0.02 (top) and relaxation time inuence in the wave shape (bottom) at t = 0.06, both in [s].
Only nite element results represented.

The function of a cooling thermoelectric device is to transport

heat from the hot to the cold faces. As mentioned before, inside
the device there are three heat uxes (Peltier, Joule, Fourier), the
rst being the most interesting from a performance standpoint;
Joule and Fourier uxes are bulk effects while Peltier is a surface
one. For this reason, Hoyos et al. [29] recommends to optimize
the TE geometry widening the hot face section, to increase the Peltier ux and thus the device performance.
To study the validity of this solution, this section analyzes a
truncated pyramid TE, see Fig. 10. Sections at hot and cold faces
are 1.96 and 1.58 [mm2] respectively, the length L = 0.28 [mm]
and the prescribed vertical ux j = 2.65  106 [A/m2]. The temperatures at the hot and cold faces are xed to Th = 50, Tc = 30 [C],
respectively. Initially (t = 0), the temperature is T0 = 0 C everywhere. From an FE point of view, the element sizes and time steps
are the ones optimized in the previous section.
Fig. 11 shows the T propagation using the parabolic (left, s = 0)
and hyperbolic (right, s = 0.02 [s]) models for four representative
instants. The T propagation velocity is larger from the hot face than
that from the cold, since in the former the prescribed T and the area
are both larger and therefore the applied energy is higher.
There are signicant differences between both models. For parabolic, a T diffusion from hot to cold faces is observed that could
easily be simulated by a 1D model, except for small, non-relevant

Fig. 10. Truncated pyramid thermoelement dimensions for three-dimensional

simulation. On hot face surface 1.96 [mm2] and Th = 50 C; on cold 1.58, Tc = 30 with
constant prescribed vertical ux j = 2.65  106 [A/m2].

variations at the edges. For hyperbolic, many non-uniformities at

several sections are observed, requiring a 3D model. The main differences between both models are:
 At initial t = 7.74  103 [s]. For parabolic, the velocity of propagation is innite, instantly increasing T by 3.47C in all the
domain except the faces. In addition, the vertical distribution
is linear and constant at every section. For hyperbolic, the wavefront increases T only in the wave front, while the rest of the
domain stays at the initial T0. Furthermore, T at the inclined
edges increases due to wave reection at these edges, reaching
a value of 53.3 C. This increase will reduce the performance of
the truncated geometry TE.
 At medium t = 3.44  102 [s]. For parabolic, the lower half
increases its T and the rest reaches 25 C all due to conduction.
The vertical distribution is still constant at every section, except
near the edges where concentrations occur. For hyperbolic, the
wavefront from the hot face progresses upwards to more than
half the TE, while the wavefront from the cold face does only
one quarter. At this instant, the edge overheating is progressing
with the front, and it also extends across the bottom of the TE
due to the constant supply of energy. In the rest of the front section the T decreases to fulll the conservation of total energy, see
(2).
 At medium t = 6.51  102 [s]. For parabolic, the system has
clearly reached the steady-state with an almost linear distribution between Th and Tc. For hyperbolic, both wavefronts have
already met. The concentration at the edges is now more evident, due to the multiple bounces that occur in the inclined
edge.
 For larger times and hyperbolic model, the wavefront reected
at the cold face returns to the hot one, reaching the steady-state
T distribution, that as expected is very similar to that predicted
by the parabolic.
Although the geometry of this example is meso, these results
could be relevant in the design and optimization of thermoelectric

101

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

PARABOLIC MODEL

Time = 7.74E-03

Time = 3.44E-02

Time = 6.51E-02

Time = 1.29E-01

3.47E+00

HYPERBOLIC MODEL

-8.43E-09

2.30E+01

2.30E+01

2.90E+01

2.90E+01

3.50E+01

3.50E+01

4.10E+01

4.10E+01

4.70E+01

4.70E+01

5.00E+01

5.33E+01

2.78E+01

5.93E-01

2.30E+01

2.30E+01

2.90E+01

2.90E+01

3.50E+01

3.50E+01

4.10E+01

4.10E+01

4.70E+01

4.70E+01

5.00E+01

5.28E+01

3.00E+01

3.00E+01

2.30E+01

2.30E+01

2.90E+01

2.90E+01

3.50E+01

3.50E+01

4.10E+01

4.10E+01

4.70E+01

4.70E+01

5.00E+01

5.34E+01

3.00E+01

3.00E+01

2.30E+01

2.30E+01

2.90E+01

2.90E+01

3.50E+01

3.50E+01

4.10E+01

4.10E+01

4.70E+01

4.70E+01

5.00E+01

5.00E+01

Fig. 11. Temperature distribution inside the truncated pyramid thermoelement for the parabolic (left) and hyperbolic (right) models at several instants. For parabolic model
s = 0, for hyperbolic s = 0.02 [s] assumed. Boundary conditions and geometry shown in Fig. 10.

Fig. 12. Analytical temperature distribution at t = 0.06 [s] for cases IIV vs. thermoelement length. Boundary conditions and prescribed values displayed in Fig. 3.

micro-devices under fast operation modes. For instance, an electronphonon thermal mismatch due to different thermal and electrical carriers was reported in [31] for cooling microthermoelectrics. This non-equilibrium has until now been modeled
by molecular dynamics or Monte Carlo techniques; however, the
present combined parabolichyperbolic model could simulate
these effects through the introduction of an empirical value of s.

6. Conclusions
This article has presented a non-linear and transient nite element formulation to simulate thermoelectric behavior under parabolic and hyperbolic heat conduction. The formulation has been
implemented into the computational code FEAP. Numerically,
non-linearities and transient hyperbolicity have been addressed

102

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103

by NewtonRaphson and by Newmark-b and HHT algorithms,

respectively.
Analytically, four 1D solutions have been developed to validate
the nite element results; the rst three solve parabolic problems
and both numerical and analytical results agree very well, using
standard parameters of Newmark-b. The fourth validation consists
on a hyperbolic problem with discontinuous initial boundary conditions. In this example, V and T distributions present sharp discontinuities, and the numerical results numerical oscillations. To
mitigate this Gibbs phenomena three numerical test have been
tested: Newmark-b with standard parameters, HHT and New 1, c
 3=2. The latter provides the best results
mark-b with b
since this parametrization gives the highest numerical dissipation.
In addition, the time steps and element sizes have been regularized
using a linear relationship with a Courant number C = 1/6. This C
has been chosen by trial and error, since the problem is highly
non-linear and a spectral analysis is not one of the current work
objectives. Finally a complete three-dimensional example is
shown, with conclusions related with the relevance of thermoelectric pellet shapes under wave propagation.
Physically, the obtained results show a T wave propagation, of
thermal ux and of V (due to the thermoelectric coupling). This
numerical tool could be applied to the design of micro-devices
for novel applications with fast processes, where the second sound
effect could be relevant.
Acknowledgments
This research was partially supported by the Spanish Ministry
of Education through Grant No. FPU AP-2006-02372 and also from
Grants MICINN BIA-2008-00522, CSD2008-00037 Canfranc Underground Physics, Excelencia Junta Andaluca P08-TEP-03641 and
Ayudas Investigacin from UPV. The authors would also like to
thank Prof. Guillermo Rus for his valuable contributions.
Appendix A. Analytical solutions
The analytical solution of case III is directly given in [25]. In this
appendix specially developed solutions for I, II, IV are presented.
The T distribution along the TE for I, II are reported in [24]. The V
solution is calculated with (9), the thermal ux one by (11)bottom.
Case I



1
P
Vx3 ; t a DLT x3  DT
cn SxeKt ;
n1

1
P

DT
L

Tx3 ; t T c x3
cn SxeKt ;
n1


1
P
q3 x3 ; t j DLT
cn Cx nLp eKt ;

A:1

n1

where DT = Th  Tc, K = b(np/L)2, b = j/qc, Sx  sin(npx3/L),

Cx  cos(npx3/L) and L is the TE length. In addition

2
cn
L


DT
x3 Sxdx3 :
Tc 
L

A:2

Case II

Vx3 ; t jc3 Lx a
A2
L

DT
L


1

P
 A1 x3 Lx
cn SxeKt ;
n1
1
P

x3 T C 

cn SxeKt ;
n1


1
P
q3 x3 ; t j AL2  2A3 x3
cn Cx nLp eKt aTx3 ; tj3 ;

Tx3 ; t

A3 x23

n1

A:3

where Lx = L  x3, A1 = j3/2jc, A2 DT j3 L2 =2bqcc, A3 j3 =2bqcc

and

cn

2
L



A2
x3 T c  A3 x23 Sx dx3 :
L

A:4

Case IV
The T distribution along the TE was reported in [8] using the
Laplace transform technique. However, we believe there are
several errors in this work, for which the result is developed
again. From the reference, we use the dimensionless parameters
n = x3/L and g = tj/ qcL2 but (10) is rewritten in an amenable
form to be solved by the Laplace technique

@ nn fTg F 1 @ n fTg F 2  @ g fTg  sF 3 @ gg fTg

n
o
sF 4 @ g F 2 ALj2  tajLA2 F 1 @ n fTg 0;

A:5

Tn; 0 0;

T0; g T h ;

T1; g T c :

Note that the boundary conditions have been included.

Regarding [8], F1 = thIL/jA, F2 = I2L2/ jcA2 have been modied
and F3 = j/qc L2, F4 = 1/ qcA are added. Applying the Laplace
transform

Wn; s P11 ek11 n P12 ek12 n ssF Fs22 s ;

3

Wn 0;

W1 Tsh ;

W0 Tsc ;

A:6

where s is time in the Laplace domain. The boundary conditions in

(A.6) are different from those in [8]: in this reference they are not
transformed. Then, the following parameters are modied
Th
k12
s e

P11

F2
T c
ssF 3 s2 s s

F2

sF3 s2 s

ek11 ek12

P12 Tsc  P11  ssF Fs22 s ;

3


q
k11
12 A1 A21  4B1 ;
k12
A1 F 1 t h L2 AajssF 4 ;

A:7

B1 t 2h L4 s F 3 s2 s:

Finally, (A.6) are inverted using the Riemann-sum approximation as

in [8].
In order to compare the different analytical solutions, Fig. 12
shows the T distributions obtained at t = 0.06 [s], using the geometry and boundary conditions shown in Fig. 3. Only T distributions
are shown since no analytical solutions for V and heat ux distributions for case IV are available. Results for cases II, III are very similar, since the Thomson effect is not relevant under the applied
electric intensity. For cases III, the distributions are quadratic
due to the boundary conditions. The small T increase for cases II,
III with respect to I is due to the Joule effect, which acts as a heat
source. For case IV, the T propagates as a wave, conning all energy
into the wavefront; the T wave has not reached the rest of the TE
and remains at the initial 0 C.
References
[1] S. Riffat, X. Ma, Thermoelectrics: a review of present and potential applications,
Appl. Therm. Engrg. 23 (2003) 913935.
[2] G. Snyder, J. Fleurial, T. Caillat, Supercooling of Peltier cooler using a current
pulse, J. Appl. Phys. 92 (3) (2002) 15641569.
[3] C. Cattaneo, Sulla conduzione del calore, Atti Seminario Mat. Fis. Univ. Modena
3 (1938) 83101.
[4] D. Jou, G. Lebon, Extended Irreversible Thermodynamics, Springer-Verlag,
Berlin, Heidelberg, 1996.
[5] G. Lebon, D. Jou, J. Casas-Vzquez, Understanding Non-equilibrium
Thermodynamics, Springer-Verlag, 2008.
[6] E. Antonova, D. Looman, Finite elements for thermoelectric device analysis in
ANSYS, in: International Conference on Thermoelectrics, 2005.
[7] J. Prez-Aparicio, R. Taylor, D. Gavela, Finite element analysis of nonlinear fully
coupled thermoelectric materials, Comput. Mech. 40 (2007) 3545.

R. Palma et al. / Comput. Methods Appl. Mech. Engrg. 213216 (2012) 93103
[8] M. Alata, M. Al-Nimr, M. Naji, Transient behaviour of a thermoelectric device
under the hyperbolic heat conduction model, Int. J. Thermophys. 24 (6) (2003)
17531768.
[9] J. Gembarovic, V. Majernik, Non-fourier propagation of heat pulses in nite
medium, Int. J. Heat Mass Transfer 31 (5) (1988) 10731080.
[10] H. Chen, J. Lin, Numerical analysis for hyperbolic heat conduction, Int. J. Heat
Mass Transfer 36 (11) (1993) 28912898.
[11] A. Barletta, E. Zanchini, Hyperbolic heat conduction and local equilibrium: a
second law analysis, Int. J. Heat Mass Transfer 40 (5) (1997) 10071016.
[12] M. Manzari, M. Manzari, On numerical solution of hyperbolic heat conduction,
Commun. Numer. Methods Engrg. 15 (1999) 853866.
[13] R. Ciegis, Numerical solution of hyperbolic heat conduction equation, Math.
Modell. Anal. 14 (1) (2009) 1124.
[14] H. Gmez, I. Colominas, F. Navarrina, M. Casteleiro, A discontinuous Galerkin
method for a hyperbolic model for convectiondiffusion problems in CFD, Int.
J. Numer. Methods Engrg. 71 (2007) 13421364.
[15] S. Bargmann, P. Steinmann, Theoretical and computational aspects of nonclassical thermoelasticity, Comput. Methods Appl. Mech. Engrg. 196 (2006)
516527.
[16] R. Taylor, FEAP, A Finite Element Analysis Program: User Manual, University of
California, Berkeley, 2010. <http://www.ce.berkeley.edu/feap>.
[17] H. Hilber, T. Hughes, R. Taylor, Improved numerical dissipation for time
integration algorithms in structural dynamics, Earthquake Engrg. Struct. Dyn.
5 (1977) 282292.
[18] J. Jackson, Classical Electrodynamics, John Wiley and Sons Inc., 1962.
[19] A. Eringen, G. Maugin, Electrodynamics of Continua I, Springer-Verlag New
York, Inc., 1990.

103

[20] A. Chakraborty, K. Ng, Thermodynamic formulation of temperature-entropy

diagram for the transient operation of a pulsed thermoelectric cooler, Int. J.
Heat Mass Transfer 49 (2006) 18451850.
[21] J.L. Prez-Aparicio, R. Palma, R.L. Taylor, Finite element analysis and material
sensitivity of Peltier thermoelectric cells coolers, Int. J. Heat Mass Transfer,
2011, doi:10.1016/j.ijheatmasstransfer.2011.08.031.
[22] O. Zienkiewicz, R. Taylor, J. Zhu, The Finite Element Method: The Basis,
Elsevier/Butterworth-Heinemann, 2005.
[23] MELCOR, Application notes for thermoelectric devices, Tech. Rep., Laird
Technologies, Melcor. <http://www.lairdtech.com>.
[24] A. Polyanin, Handbook of Linear Partial Differential Equations for Engineers
and Scientist, Chapman & Hall CRC, 2002.
[25] W. Seiffert, M. Ueltzen, E. Muller, One-dimensional modelling of
thermoelectric cooling, Phys. Stat. Sol. A 194 (1) (2002) 277290.
[26] K.F.R. Courant, H. Lewy, Uber die partiellen differenzengleichungen der
mathematischen physik, Math. Annal. 100 (2007) 3274.
[27] T. Hughes, The Finite Element Method, Linear Static and Dynamic Finite
Element Analysis, Prentice-Hall, Inc., 1987.
[28] T. Fung, Numerical dissipation in time-step integration algorithms for
structural dynamic analysis, Prog. Struct. Engrg. Mater. 5 (2003) 167180.
[29] G. Hoyos, K. Rao, D. Jerger, Fast transient response of novel Peltier junctions,
Energy Convers. 17 (1977) 2329.
[30] Y. Cheng, W. Lin, Geometric optimization of thermoelectric coolers in a conned
volume using genetic algorithms, Appl. Therm. Engrg. 25 (2005) 29832997.
[31] L. da Silva, M. Kaviany, Micro-thermoelectric cooler: interfacial effects on
thermal and electrical transport, Int. J. Heat Mass Transfer 47 (2004) 2417
2435.