International Communications in Heat and Mass Transfer 92 (2018) 107–111

Contents lists available at ScienceDirect

International Communications in Heat and Mass Transfer

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

Estimation of parameters of the dual-phase-lag model for heat conduction in T

metal-oxide-semiconductor field-effect transistors
⁎
Milena Vilar França, Helcio Rangel Barreto Orlande
Department of Mechanical Engineering, PEM/COPPE Federal University of Rio de Janeiro, UFRJ Caixa, 68503 Cidade Universitária, Rio de Janeiro, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: This communication deals with the solution of an inverse parameter estimation problem with the dual-phase-lag
Dual-phase-lag heat conduction model heat conduction model. The case considered involves the heating of a metal-oxide-semiconductor field-effect
Microelectronics transistor, for time and spatial scales where the validity of the classical heat conduction model based on Fourier's
Inverse problem Law, which considers an infinite speed of propagation of thermal waves, has been questioned. The Markov chain
Markov chain Monte Carlo method
Monte Carlo method is applied for the estimation of the parameters within the Bayesian framework of statistics,
Metropolis-Hastings algorithm
Bayesian statistics
by using simulated transient temperature measurements.

1. Introduction temperature, heat fluxes, radiation intensities, etc. [17–20]. Several

works can be found in the literature regarding inverse analysis with
Experimental evidences that Fourier's Law fails for small time and mathematical models for heat conduction where the speed of propa-
spatial scales, as well as for temperatures near the absolute zero (see, gation is considered as finite, including the dual-phase-lag model
e.g. [1]), lead to the development of new constitutive equations that [21–26]. In this communication, we apply the Markov chain Monte
relate the heat conduction vector to the temperature gradient, in which Carlo method for the estimation of parameters appearing in the dual-
the speed of propagation of thermal waves is finite [2–11]. The so- phase-lag model for heat conduction in MOSFET transistors, by using
called dual-phase-lag model [7,8,10] considers the non-instantaneous simulated transient temperature measurements. The parameters are
effects between the heat conduction vector and the temperature gra- estimated within the Bayesian framework, where the information
dient, in the form of relaxation times for these two quantities. De- available before the measurements are taken is used in the inverse
pending on the magnitudes of the relaxation times, the heat conduction analysis in the form of prior statistical distribution functions
vector and the temperature gradient can be synchronized, or either of [18–20,27,28]. The solution of the inverse problem within the Bayesian
one of these two vectors can be in advance with respect to the other. In framework is obtained from the statistics of the posterior distribution
general, when these newly developed constitutive equations are used, function, which considers the prior distribution and the likelihood
which consider the speed of propagation of thermal waves as finite, a (statistical model for the measurement errors) through Bayes´ theorem.
hyperbolic equation results for modelling heat conduction, instead of Samples of the posterior distribution were simulated in this commu-
the classical parabolic model obtained with Fourier's Law [2–11]. nication with the Metropolis-Hastings algorithm [18–20,27–29]. Such
The technological advancement towards micro and nanoscales in algorithm is not presented here for the sake of brevity, but can be
electronic devices, often with large concentrated power, demands readily found in several references, including [18–20,27–29].
thermal design and analysis where heat conduction plays a major role.
In particular, heat conduction in metal-oxide-semiconductor field-effect 2. Physical problem and mathematical formulation
(MOSFET) transistors has been modelled using the dual-phase-lag
constitutive equation [12–14], including boundary conditions that The physical problem considered here is based on that addressed in
consider a temperature jump to simulate the phonon diffusion at the [13]. Fig. 1 illustrates the two-dimensional domain, which consists of a
boundaries [15,16]. MOSFET transistor made of silicon (τq = 33.33 ps, τT = 1.66 ps,
Inverse problems of heat transfer deal with the estimation of para- k = 150 Wm−1 K−1, C = 1.5 × 106 Jm−3 K−1, α = 10−4 m2 s−1,
meters or functions appearing in the mathematical models of thermal Λ = 100 nm [13,16]), with dimensions L = 100 nm and l = 50 nm.
problems, by using measurements of dependent variables, like Heat is uniformly generated within the square region of side

⁎
Corresponding author.
E-mail addresses: [email protected] (M.V. França), [email protected] (H.R.B. Orlande).

https://doi.org/10.1016/j.icheatmasstransfer.2018.02.001

0735-1933/ © 2018 Elsevier Ltd. All rights reserved.

M.V. França, H.R.B. Orlande International Communications in Heat and Mass Transfer 92 (2018) 107–111

Fig. 1. Heat conduction in a MOSFET transistor.

Lh = 10 nm by Joule effect. The heat source is assumed to vary linearly 3. Inverse problem
from zero at time t = 0 to Qmax = 1019 W/m3 at t = 10 ps and kept
constant afterwards. The top and lateral surfaces are thermally in- The inverse problem of interest deals with the simultaneous esti-
sulated, while the bottom surface is subjected to a temperature jump mation of the vector of parameters
boundary condition to simulate the phonon diffusion at the boundary,
P = [α∗, τq∗ , τT∗ , Kn] (3)
where the surface is externally maintained at the temperature Tw. The
initial temperature is uniform and taken as T0 = 300 K. from the simulated temperature response at the point of maximum
The following dimensionless variables were utilized in this work: temperature in the substrate, located at x⁎ = y⁎ = 5 (P(5,5) in Fig. 1).
T − T0 tαref τ τq x The transient temperature readings contain errors, which are assumed
T∗ = , t∗ = , τT ∗ = T , τq∗ = , x∗ = , to be additive, uncorrelated and normally distributed, with zero mean
T0 Lh 2 τTref τqref Lh
y and known constant standard-deviation σ. The measurements were si-
y∗ = mulated by adding Gaussian errors to the solution of the direct problem
Lh
at the point P(5,5) obtained with the reference values and with stan-
k C α Λ Lh2
k∗ = , C∗ = , α∗ = , Kn = , Q∗ = Q dard deviations of 5% and 10% relative to the maximum temperature at
kref Cref αref Lh T0 kref this point.
(1.a–k)
where the subscript ref denotes reference values, Kn is the Knudsen 4. Results and discussions
number, Λ is the mean free path of phonons, τq is the relaxation time for
the heat flux vector and τT is the relaxation time for the temperature Before estimating the parameters, the reduced sensitivity coeffi-
gradient. The reference values for the thermophysical properties were cients are examined. The reference values are used for such local ana-
taken as those of silicon and τTref = τqref = Lh 2 / αref . lysis of the sensitivity coefficients, since the present estimation problem
The mathematical formulation of the physical problem in di- is nonlinear. The reduced sensitivity coefficients are obtained from the
mensionless form is given by: multiplication of the original sensitivity coefficients [17,19] by their
corresponding parameter values. Hence, the reduced sensitivity coeffi-
1 ∂T ∗ τq∗ ∂2T ∗ ∂ 2T ∗ ∂ 2T ∗ ⎞ ∂ ∂ 2T ∗ ∂ 2T ∗ ⎞ Q∗ τq∗ ∂Q∗
+ ∗ ∗2 = ⎛ ∗2 + ⎜ + τT∗ ∗ ⎛ ∗2 + ⎟ + ∗ + ∗ ∗
⎜ ⎟ cients for this problem can be compared to the temperature variation at
α∗ ∂t ∗ α ∂t ⎝ ∂ x ∂y ∗2
⎠ ∂t ⎝ ∂ x ∂ y ∗2
⎠ k k ∂t
the measurement position, in order to detect small magnitudes and
in 0 < x ∗ < L∗ , 0 < y∗ < l∗, for t ∗ > 0
linear dependence. For appropriate estimation of the parameters, the
(2a) sensitivity coefficients should be linearly independent and with large
∂ T∗ ∂ ⎛ ∂ T∗ magnitude [17,19]. The time variations of the sensitivity coefficients
+ τT∗ ⎞ = 0 at x ∗ = 0, 0 < y∗ < l∗, for t ∗ > 0
with respect to each parameter are presented by Fig. 2, together with
∂ x∗ ∂ t∗ ⎝ ∂ x∗ ⎠ (2b)
the transient temperature at the measurement location. This figure
∂ T∗ ∂ ⎛ ∂ T∗ shows that the sensitivity coefficients with respect to thermal diffusivity
+ τT∗ ⎞ = 0 at x ∗ = L∗, 0 < y∗ < l∗, for t ∗ > 0
∂ x∗ ∂ t∗ ⎝ ∂ x∗ ⎠ (2c) and to the relaxation times are of the same order of magnitude of the
temperature variation. On the other hand, the magnitude of the sensi-
k ∗ ∂T ∗ ∂ ∂T ∗ 1 ∂T ∗ 1 1
− ∗
− k ∗τT∗ ∗ ⎛ ∗ ⎞ + τq∗ ⎜ ⎟ + T∗ = Tw∗ tivity coefficient with respect to the Knudsen number is practically null.
d1 Kn ∂y ∂t ⎝ ∂y ⎠ d1 Kn ∂t ∗ d1 Kn d1 Kn
Such is the case because the Knudsen number appears in the tempera-
at y∗ = 0, 0 < x ∗ < L∗, for t ∗ > 0 (2d) ture jump boundary condition at the bottom surface (eq. 2d), which
does not significantly affect the temperatures at the measurement po-
∂ T∗ ∂ ∂ T∗
+ τT∗ ∗ ⎛ ∗ ⎞ = 0 at y∗ = l∗, 0 < x ∗ < L∗, for t ∗ > 0
⎜ ⎟
sition during the time range analysed. The analysis of Fig. 2 also reveals
∂ y∗ ∂t ⎝ ∂y ⎠ (2e)
that the sensitivity coefficients with respect to the thermal diffusivity
T∗ = 0 in 0 < x∗ < L∗, 0< y∗ < l ∗, for t∗ =0 (2f) and to the heat flux relaxation time tend towards linear dependence. On
the other hand, the sensitivity coefficients with respect to the relaxation
∂T ∗
= 0 in 0 < x ∗ < L∗, 0 < y∗ < l∗, for t ∗ = 0 times are not linearly dependent and can be simultaneously estimated
∂t ∗ (2g)
by using temperature measurements taken at the top surface of the
where the properties were supposed constant, d1 = 0.061 is the ad- transistor. In fact, the relaxation times are the quantities of most in-
T −T
justing coefficient used in [13] and Tw∗ = wT 0 . terest for this inverse analysis, because they cannot be measured by
0

108
M.V. França, H.R.B. Orlande International Communications in Heat and Mass Transfer 92 (2018) 107–111

mean values, respectively, satisfying the positivity constraint (zero

probability for negative values). The priors for τq∗ and τT∗ were based on
the relaxation times for normal and umklapp scattering of phonons,
τN and τR, respectively, which range from 10−13 s to 10−10 s for semi-
conductors, respectively [30]. The relaxation times of phonon scat-
tering are related to those of the dual-phase-lag model by τN = 5τT/9
and τR = τq [30]. The priors for the relaxation times were then con-
sidered as uniform distributions in the intervals 10−13s ≤ τq ≤ 10−10s
and 1.8 × 10−13s ≤ τT ≤ 1.8 × 10−10s.
Table 1 summarizes the test cases examined in this work, which
involve different starting points for the Markov chains and experi-
mental errors with standard deviations of 5% or 10% of the maximum
temperature at the measurement location. For the results presented
below, 30,000 states were simulated in the Markov chains, with burn-in
periods of 5000 states. The chains were generated with random-walk
proposals with uniform distributions. Table 2 presents the results ob-
tained for the four test cases, given as the means, standard deviations
and 99% credibility intervals (0.5% and 99.5% quantiles) of the mar-
ginal posterior distributions obtained from the simulated samples of the
Fig. 2. Time variations of the temperature at the measurement location and of the sen-
sitivity coefficients.
Markov chains. In general, the estimated means are in excellent
agreement with the exact values of the parameters, despite the large
simulated measurement errors. Moreover, the credibility intervals were
Table 1 not significantly affected by either the starting points of the Markov
Test cases.
chains or by the standard deviations of the measurement errors. Table 2
Case Parameters Initial states Standard deviation of the simulated also shows that, relatively, the relaxation time for the temperature
measurements gradient is the parameter estimated with the largest variance, because
of its uniform prior and smaller magnitude of the sensitivity coefficient
1 τq∗ and τT∗ 10 and 10 0.05Tmax∗
α∗ and Kn 1 and 10
(see also Fig. 2).
2 τq∗ and τT∗ 10 and 10 0.1Tmax∗ Fig. 3 presents the Markov chains for the four parameters of eq. (3)
α∗ and Kn 1 and 10 for test case 3 (see Table 1), which are representative of the results
3 τq∗ and τT∗ 50 and 50 0.05Tmax∗ obtained for the other test cases. The fast convergence of the samples
α∗ and Kn 1 and 10
generated with the Metropolis-Hastings algorithm to the equilibrium
4 τq∗ and τT∗ 50 and 50 0.1Tmax∗
α∗ and Kn 1 and 10 marginal posterior distributions can be observed in this figure. The
histograms of the samples of the marginal posteriors are presented by
Fig. 4. As expected, the histograms approximate Gaussian distributions
standard techniques, like the thermal diffusivity or the thermal con- for all parameters, since the likelihood is Gaussian and the priors are
ductivity, and because they are quantities inherent to the dual-phase- either Gaussian or uniform.
lag model. The equilibrium Markov chains generated with the Metropolis-
The prior distributions for the parameters were then prescribed by Hastings algorithm also allow for the uncertainty quantification of the
considering the foregoing analysis of the sensitivity coefficients, as well direct problem solution under the effects of uncertainties in the para-
as the information available in the literature. While uniform priors with meters. The solution of the forward problem was then computed with
large variances were prescribed for the relaxation times, Gaussian the parameter samples at each state of the Markov chains after equili-
priors with small variances were prescribed for the thermal diffusivity brium was reached, for comparison with the simulated measurements
and for the Knudsen number. The priors for α∗ and Kn were centered at and with the exact temperature variation at the measurement position.
their mean values, with standard deviations of 5% and 1% of their Fig. 5 shows the curves that represent the mean and the 99% credibility

Table 2
Estimated parameters.

Case Dimensionless parameter Dimensionless exact value Estimated mean Standard deviation of the Markov chains Credibility interval (99%)

Lower limit Upper limit

∗
1 τq 33.33 32.92 1.19 31.04 35.00
τT∗ 1.66 1.71 0.28 1.24 2.17
α∗ 1.00 1.00 0.05 0.91 1.08
Kn 10.00 10.00 0.09 9.85 10.15
2 τq∗ 33.33 33.13 2.23 29.60 37.00
τT∗ 1.66 1.36 0.47 0.62 2.16
α∗ 1.00 0.99 0.05 0.91 1.07
Kn 10.00 10.00 0.10 9.83 10.17
3 τq∗ 33.33 34.94 1.36 32.76 37.22
τT∗ 1.66 1.69 0.29 1.21 2.18
α∗ 1.00 0.99 0.05 0.91 1.07
Kn 10.00 10.00 0.10 9.84 10.16
4 τq∗ 33.33 34.81 2.54 30.81 39.19
τT∗ 1.66 1.89 0.54 1.06 2.84
α∗ 1.00 1.00 0.05 0.92 1.08
Kn 10.00 10.00 0.10 9.83 10.16

109
M.V. França, H.R.B. Orlande International Communications in Heat and Mass Transfer 92 (2018) 107–111

Fig. 3. Markov chains generated with the Metropolis-Hastings algorithm.

Fig. 4. Histograms for the marginal posterior distributions of each parameter.

intervals of the solutions obtained with such a Monte Carlo simulation 5. Conclusions
of the forward problem. The direct problem solution obtained with the
mean values of the parameters, the solution obtained with the exact In this communication we estimated parameters of the dual-phase-
parameters and the simulated measurements are also presented in lag heat conduction model, for a two dimensional problem in a
Fig. 5. An analysis of Fig. 5 reveals that the parameter samples obtained MOSFET transistor using a Bayesian Approach. Focus is given on the
from the Markov chains provide Monte Carlo solutions of the direct estimation of the relaxation times for the heat conduction flux vector
problem with a mean in excellent agreement with the exact solution and for the temperature gradient, which are parameters inherent to this
and with variances much smaller than those for the measurements. model and cannot be estimated with other standard methods. The
Therefore, uncertainties in the direct problem solution resulting from parameters were estimated with simulated measurements containing
uncertainties in the parameters (represented by their marginal posterior large additive Gaussian errors and four different test cases were ex-
distributions) are small for the heat conduction problem with the dual- amined here. The results obtained with the Markov chain Monte Carlo
phase-lag model, in MOSFET transistors. method, implemented via the Metropolis-Hastings algorithm, show that
the estimated parameters are not sensitive to the initial states of the

110
M.V. França, H.R.B. Orlande International Communications in Heat and Mass Transfer 92 (2018) 107–111

[7] M.N. Ozisik, D.Y. Tzou, On the wave theory in heat conduction, Int. J. Heat Mass
Transf. 116 (1994) 526–535.
[8] D.Y. Tzou, R.J. Chiffelle, M.N. Ozisik, The lattice temperature in the microscopic
two step model, ASME J. Heat Transf. 116 (1994) 1034–1038.
[9] T.Q. Qiu, C.L. Tien, Heat transfer mechanisms during short-pulse laser heating of
metals, ASME J. Heat Transf. 115 (1993) 835–841.
[10] D.Y. Tzou, Macro-to-Micro-Scale Heat Transfer: The Lagging Behavior, second
edition, Taylor & Francis, Bristol, 2015.
[11] L. Wang, X. Zhou, X. Wei, Heat Conduction: Mathematical Models and Analytical
Solutions, Springer, New York, 2010.
[12] J. Ghazanfarian, A. Abbassi, Effect of boundary phonon scattering on dual-phase-
lag model to simulate micro- and nano-scale heat conduction, Int. J. Heat Mass
Transf. 52 (2009) 3706–3711.
[13] J. Ghazanfarian, A. Abbassi, Investigation of 2D transient heat transfer under the
effect of dual-phase-lag model in a nanoscale geometry equation, Int. J.
Thermophys. 33 (2012) 552–566.
[14] J. Ghazanfarian, Z. Shomalli, Investigation of dual-phase-lag heat conduction model
in a nanoscale metal-oxide-semiconductor field-effect transistor, Int. J. Heat Mass
Transf. 55 (2012) 6231–6237.
[15] G. Chen, Ballistic-diffusive equation for transient heat conduction from nano to
macroscale, ASME J. Heat Transf. 124 (2002) 320–328.
[16] R. Yang, G. Chen, L. Marine, Y. Taur, Simulation of nanoscale multidimensional
transient heat conduction problems using ballistic-diffusive equations and phonon
Boltzmann equation, ASME J. Heat Transf. 127 (2005) 298–306.
Fig. 5. Dimensionless temperature at the measurement location. [17] M.N. Ozisik, H.R.B. Orlande, Inverse Heat Transfer: Fundamentals and
Applications, Taylor and Francis, New York, 2000.
[18] H.R.B. Orlande, Inverse problems in heat transfer: new trends on solution meth-
Markov chains, neither by the uncertainties in the measurements. The odologies and applications, ASME J. Heat Transf. 134 (2012) (031011-031011-13).
[19] H.R.B. Orlande, O. Fudym, D. Maillet, R.M. Cotta, Thermal Measurements and
samples of the marginal posterior distributions for the parameters,
Inverse Techniques, CRC Press, Boca Raton, 2011.
which were generated with the Metropolis-Hastings algorithm, allow [20] H.R.B. Orlande, O. Fudym, F.A. Kulacki (Ed.), Thermophysical Properties
for the simulation of the direct problem under uncertainties in these Measurement and Identification, Chapter in Handbook of Thermal Science and
quantities. The results obtained reveal that the mean of the Monte Carlo Engineering, Springer International Publishing AG, 2017, , http://dx.doi.org/10.
1007/978-3-319-32003-8_5-1.
simulations is in excellent agreement with the exact solution, and that [21] H.R.B. Orlande, M.N. Ozisik, D.Y. Tzou, Inverse analysis for estimating the elec-
the variances are much smaller than those of the measurements. tron–phonon coupling factor, J. Appl. Phys. 78 (1995) 1843–1899.
[22] P.H.A. Nóbrega, H.R.B. Orlande, J.-L. Battaglia, Bayesian estimation of thermo-
physical parameters of thin metal films heated by fast laser pulses, Int. Commun.
Acknowledgements Heat Mass Transf. 38 (2011) 1172–1177.
[23] J.-L. Battaglia, V. Schick, C. Rossignol, O. Fudym, H. Orlande, P. Affonso, Global
This work was supported by CAPES, CNPq and FAPERJ. estimation of thermal parameters from a picoseconds thermoreflectometry experi-
ment, Int. J. Thermal Sci. 57 (2012) 17–24.
[24] D.W. Tang, N. Araki, An inverse analysis to estimate relaxation parameters and
References thermal diffusivity with a universal heat conduction equation, Int. J. Thermophys.
21 (2) (2001) 553–561.
[25] T.-S. Wu, H.-L. Lee, W.-J. Chang, Y.-C. Yang, An inverse hyperbolic heat conduction
[1] S.I. Anisomov, B.L. Kapeliovich, T.L. Perel'man, Electron emission from metal sur-
problem in estimating pulse heat flux with a dual-phase-lag model, Int. Commun.
faces exposed to ultrashort laser pulses, Sov. Phys. JETP 39 (1975) 375–377.
Heat Mass Transf. 60 (2015) 1–8.
[2] M.C. Cattaneo, Sur une forme de l'equation de la chaleur eliminant le paradox d'une
[26] Ching-Yu Yang, Estimation of the heat source of laser pulses by a dual-phase-lag
propagation instantanee, C. R. Hebd. Seances Acad. Sci. 247 (1958) 431–433.
model, Int. J. Heat Mass Transf. 73 (2014) 358–364.
[3] P. Vernotte, les paradoxes de la theorie continue de l'equation de la chaleaur, C. R.
[27] J. Kaipio, E. Somersalo, Statistical and Computational Inverse Problems, Applied
Acad. Sci. Paris 246 (1958) 3154–3155.
Mathematical Sciences, 160, Springer-Verlag, New York, NY, 2004.
[4] T.Q. Qiu, C.L. Tien, Short pulse laser heating on metals, Int. J. Heat Mass Transf. 37
[28] J. Kaipio, C. Fox, The bayesian framework for inverse problems in heat transfer,
(1992) 719–726.
Heat Transfer Eng. 32 (9) (2010) 718–753.
[5] T.Q. Qiu, T. Juhasz, C. Suarez, W.E. Bron, C.L. Tien, Femtosecond laser heating of
[29] D. Gamerman, H.F. Lopes, Markov Chain Monte Carlo: Stochastic Simulation for
multi-layered metals - II. Experiments, Int. J. Heat Mass Transf. 37 (1994)
Bayesian Inference, Second edition, Chapman & Hall/CRC, Boca Raton, FL, 2006.
2799–2808.
[30] A. Vedavarz, S. Kumar, M.K. Moallemi, Significance of non-Fourier heat waves, J.
[6] T.Q. Qiu, C.L. Tien, Femtosecond laser heating of multi-layered metals - I. Analysis,
Heat Transf. 116 (1994) 221–224.
Int. J. Heat Mass Transf. 37 (1994) 2789–2797.

111