You are on page 1of 7

Microelectronics Reliability 76–77 (2017) 619–625

Contents lists available at ScienceDirect

Microelectronics Reliability

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

Application of multiplicative dimensional reduction method for


uncertainty quantification and sensitivity analysis of MEMS
electrostatic actuators
Pham Luu Trung Duong a, Trong Toan Tran b, Nagarajan Raghavan a,⁎
a
Engineering Product Development (EPD) Pillar, Singapore University of Technology and Design, 487 372, Singapore
b
Faculty of Electronic Technology, Industrial University of Ho Chi Minh City, Vietnam

a r t i c l e i n f o a b s t r a c t

Article history: The effect of electrostatic actuation allows MEMS devices to have specific physical movements. They have several
Received 19 May 2017 advantages including low power, low cost, and small size and are used widely in variable capacitors, micro-acceler-
Received in revised form 24 June 2017 ometers, etc. In this study, we consider a MEMS actuator consisting of a moveable plate and a fixed plate in the pres-
Accepted 26 July 2017
ence of an applied electric field. The gap between the two plates can normally be changed by voltage control. It is
Available online 3 August 2017
known that as the gap reduces to two thirds of the original gap, the so-called “pull-in effect” tends to occur, causing
the plates to collide (resulting in dielectric breakdown and actuator failure). It is therefore important to predict the
onset of the “pull-in effect”. As it is practically impossible to obtain the model parameters precisely, this prediction
should account for the presence of uncertainties. Sampling methods such as Monte Carlo and Quasi Monte Carlo are
easy to use with the caveat of low accuracy and high computational cost. The other popular method is polynomial
chaos. It has high accuracy and low computational cost under smoothness assumption for problems with small
number of uncertain parameters. In this study, we consider a two-stage approach to quantify the parametric uncer-
tainty of MEMS electrostatic actuators with a moderate number of causal stochastic factors. In the first stage, a mul-
tiplicative dimensional reduction method is used to approximate the variance-based global sensitivity measures in
order to simplify the model for the uncertainty quantification stage. The second stage involves the use of the gen-
eralized polynomial chaos (gPC) approach to quantify uncertainty of the simplified model from the first stage.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction systems [5–9]. The method however, is only effective for cases with
small number of uncertain factors.
MEMS electrostatic actuators have a wide range of applications Our study here proposes a two stage approach for uncertainty
owing to their low power consumption, low cost, and small size. They quantification of double plate MEMS actuators with large number of
have been frequently used in variable capacitors, micro-accelerometers uncertain quantities. At the first stage, a multiplicative dimensional re-
etc. [1,2]. This work focuses on a double plate MEMS electrostatic actu- duction method is used for sensitivity analysis (SA). Sensitivity analysis
ator, which is known to have an unstable drive region after the gap is is a study of how “uncertainty in the output of a model (numerical or
decreased to two thirds of the original gap. It is therefore necessary to otherwise) can be apportioned to different sources of uncertainty in
predict when the actuator will fail to work in the unstable region. Un- the model input” [10]. It can help one to get a better understanding of
certainties in performance are unavoidable in system models due to the model and simplify it, if needed. Perhaps, the most popular SA is
modeling errors, and environmental changes etc. Thus, a confident pre- based on variances of system output. Different approaches such as:
diction of the range of lifetime can only be made when uncertainties are MC, random balance design (RBD), Fourier amplitude sensitivity test
taken into account. One possibility to propagate stochastic variations (FAST) [10,11] can be used to calculate variance based quantitative indi-
and make predictions is to use Monte Carlo/Quasi Monte Carlo based ces known as Sobol indices, which can be used to determine how a
methods (MC/QMC) [3,4]. However, MC/QMC based approaches are model output variable of interest is influenced by individual or subsets
very computationally demanding since they require large sample sizes of uncertain parameters.
to get accurate results. Recently, generalized polynomial chaos (gPC) This work is organized as follows. Section 2 provides an overview
expansion has been applied successfully to uncertainty quantification, about variance based sensitivity analysis using multiplicative dimen-
prediction, propagation and decision making for different engineering sional reduction method and explains how it can be used to simplify
the model for the next step. In Section 3, the polynomial chaos approach
⁎ Corresponding author. method for UQ is described. The method is then demonstrated for UQ
E-mail address: nagarajan@sutd.edu.sg (N. Raghavan). and SA of a double plate MEMS electrostatic actuator in Section 4.

http://dx.doi.org/10.1016/j.microrel.2017.07.091
0026-2714/© 2017 Elsevier Ltd. All rights reserved.
620 P.L.T. Duong et al. / Microelectronics Reliability 76–77 (2017) 619–625

Finally, Section 5 concludes with a summary of the findings, providing Finally, the total Sobol indices quantify the total impact of a given pa-
suggestions to pursue further investigations along this line. rameter zi (including all of the interactions with the other parameters):
X
2. Variance based sensitivity analysis using multiplicative T i ðt Þ ¼ SA ðt Þ ð7Þ
dimensional reduction method i∈A

Sensitivity analysis (SA) aims at identifying the model parameters Remarks: If Ti is close to zero, that implies that zi is non-influential
that are relatively more important than the others in scientific and engi- and can be fixed anywhere in its distribution without affecting the var-
neering applications. The methods for SA can be classified into two P
iance of the output [11]. The difference 1− ni¼1 Si is an indicator of the
types: local SA and global. In the local approach, the partial derivative presence of interactions in the model.
of model output with respect to model input parameters are calculated. There are several approaches for calculation of Sobol indices such as:
It is suitable to study the local variability of the model only. On the other FAST, RBD [11], Monte Carlo [12] and multiplicative dimensional reduc-
hand, global SA using Sobol indices [10–13], which is also known as a tion method (M-DRM) [13]. Among these methods, M-DRM provides a
variance based method, deals with the global variations in the output good tradeoff between accuracy and computational cost. The main idea
for the given entire range of uncertainties. Thus, it can be useful to sim- of M-DRM is to approximate a multivariate function as a product of sev-
plify the uncertain model by simply fixing the non-influential parame- eral univariate functions. The algorithm for M-DRM is described below.
ters at their nominal values. Taking logarithmic transformation of the model output in Eq. (1),
We consider a dynamical system described by a computational one can obtain:
model
  φðt; Z Þ ¼ logðjyðt; Z ÞjÞ ð8Þ
2 2
yðt; Z Þ ¼ Μ t; Z; dy=dt; dy =dt ; … ð1Þ

The conventional dimensional reduction method [14] can be used to


where t is the time variable, Z = {z1, … , zn} is the vector of uncertain get an approximation of φ(t, Z) with a cut point c =(c1, … , cn):
model parameters and y is the model output of interest. The uncertain
parameters are independent random variables with density X
n
functions,pi(zi) : Γi → R+. Assuming that y has a finite variance, it can be φðt; Z Þ≃ φðt; zi ; c−i Þ−ðn−1Þφ0 ð9Þ
decomposed into the sum of increasing dimension terms: i¼1

X
n X n X n   where, the terms are related to those in the original space as follows:
yðt; Z Þ ¼ y0 ðt Þ þ yi ðzi ; t Þ þ yij t; z1 ; z j þ … þ yi1 …in ðz1 ; …; zn ; t Þ
i¼1 X i¼1 ib j
φ0 ðt Þ ¼ logðjyðt; cÞjÞ
¼ y0 ðt Þ þ yA ðt; Z A Þ ð10Þ
A⊂f1; …; ng
φi ðt; zi ; c−i Þ ¼ logðjyðt; c1 ; …; ci−1 ; zi ; ciþ1 ; …; cn ÞjÞ
A≠∅

ð2Þ Thus, the original model output can be rewritten as:

where A ¼ fi1 ; …; is g⊂f1; ::; ng is a subset of indices, and Z A is a subset n


eðφðt;Z ÞÞ ≃eð1−nÞφ0 e∑i¼1 φi ðc1 ;…;ci−1 ;zi ;ciþ1 ;…;cn Þ ð11Þ
of Z that contains only parameters that are indexed by A. The terms in
Eq. (2) can be computed recursively as:
By using Eqs. (10) and (11), the model output can be approximated
y0 ðt Þ ¼ E½yðt; Z Þ in the form of:
yi ðt; zi Þ ¼ E½yðt; Z Þjzi −y0 ðtÞ  
yij t; zi ; z j ¼ E yðt; Z Þjzi ; z j −yi ðt; zi Þ−y j t; z j ð3Þ Y
n
−y0 ðt Þ yðt; Z Þ≃yðt; cÞ1−n yðt; c1 ; …; ci−1 ; zi ; ciþ1 ; …; cn Þ
… i¼1
ð12Þ
1−n
Y
n
¼ h0 ðt Þ yðt; zi ; c−i Þ
where E denotes the expectation operator, and EðZjZ A Þ is the conditional i¼1
expectation. It should be noted that the functions defined in Eq. (3) are
orthogonal to each other. Utilizing this orthogonality property, the var- where c−i = (c−1,c−2, … ,ci −1, ci+1, … , cn).
iance of the model output can be decomposed into: This approximation is known as the univariate M-DRM.
X Let us denote the mean and mean square of the ith dimensional
V y ðt Þ ¼ Var ðyA ðt; Z A ÞÞ ð4Þ function as:
A⊂f1; …; ng
A≠∅
μ i ðt Þ ¼ E½hyðt; zi ; c−i Þ ¼i E½hihðzi ; t Þ i
2 ð13Þ
σ i ðt Þ ¼ E yðt; zi ; c−i Þ2 ¼ E ðhi ðzi ; t ÞÞ
First order Sobol indices (S) quantify the portion of the total variance
that can be apportioned to the sole parameter zi, as shown by Eq. (5):
Thus, the mean and mean square of output can be obtained as:
V i ðt Þ
Si ðt Þ ¼ ð5Þ
V y ðt Þ Y
n
_ 1−n
μ y ðt Þ≃h0 μ i ðt Þ
i¼1
where Vi denote the variance of yi(t, zi).   Y
n ð14Þ
The Sobol index for each subset of parameters is defined by Eq. (6), 2−2n
E yðt; Z Þ2 ≃h0 σ i ðt Þ
which quantifies the joint effect of subset parameters Z A . i¼1

V A ðt Þ For convenience in notation, the dependence on the time variable


SA ðt Þ ¼ ð6Þ
V y ðt Þ will be dropped from now.
P.L.T. Duong et al. / Microelectronics Reliability 76–77 (2017) 619–625 621

Utilizing M-DRM approximation, the first and second conditional


moments can be obtained as:

1−n
Y
n
E½yðZjzi Þ≃h0 yi ðzi Þ μk
k¼1;k≠i
h i Y
n ð15Þ
2−2n _
E ðyðZjzi ÞÞ2 ¼ h0 σi μ k 2 ¼ μ y σ i =μ i 2
k¼1;k≠i

Fig. 1. Electro-mechanical schematic model of an electrostatic actuator.


Hence, the primary variance can be computed as:
_  
Di ≃μ y σ i =μ i 2 −1 ð16Þ dimensional quadrature. Thus, for calculating the coefficients expan-
sion, one needs to evaluate the computational model qm times. Once
The first order Sobol indices can then be computed as: the coefficients have been computed, we practically have an analytical
relationship between the model output and the input parameters.
σ =μ 2 −1 The mean and variance of the output model can be computed
Si ¼ Di =Dy ≃ Qn i i 2  ð17Þ
i¼k σ k =μ k −1 through the coefficients, αk:

Utilizing M-DRM and after some algebraic manipulation, the total X


NP
Sobol index for parameter zi can be approximated as: μ Y ¼ α 1 ; Dy ¼ ðα k Þ2 ð23Þ
k¼2
1−μ 2 =σ i
Ti≃ Qn i  ð18Þ
1− k¼1 μ k 2 =σ k where the bar on top of the mean and variance is used to denote an ap-
proximate quantity. The output density function can be obtained easily
A Gaussian quadrature with q nodes {wil, zil}ql=1 can be used to calcu- by sampling the cheap-to-evaluate analytical model in Eq. (20). More
late the moments of the dimensional function in Eq. (13) as details on the polynomial chaos methodology can be found in Refs.
[14–16].
q
μ i ¼ ∑l¼1 wil yðt; c1 ; …; ci−1 ; zil ; ciþ1 ; …; cn Þ
q ð19Þ
σ i ¼ ∑l¼1 wil ðyðt; c1 ; …; ci−1 ; zil ; ciþ1 ; …; cn ÞÞ2 4. Results and discussions

As for the M-DRM method, only nq computational model evalua- In this section, the proposed two stage approach is used for UQ and
tions are needed for estimation of the sensitivity indices. On the other SA for a one degree of freedom MEMS actuator. The MEMS actuator con-
hand, the MC method requires (n + 2) Q computational model evalua- sists of one moveable plate and one fixed plate, subject to an electric
tions to estimate these indices. Note that only a small number of nodes field. The electromechanical model of the system is shown in Fig. 1.
q = 5–10 is normally used in M-DRM while Q = 5000–10,000 samples The dynamics of the electrostatic actuator in Fig. 1 can be described
are normally used in the MC method [12]. physically by the following system of first order differential equations:

8
3. Uncertainty quantification using polynomial chaos >
> x_ 1 ¼ x2 
>
> 2
< x_ ¼ − 1 bx þ kx − x3
2 2 1
The SA step in the previous section can be used to detect m b n im- m a 2εA ð24Þ
>
> x3 ðg 0 −x1 Þ
>
> 1
portant parameters, Zm. These influential parameters will then be used : x_ 3 ¼ V in −
in the polynomial chaos method for UQ. The others (non-influential) R εA
are fixed at their nominal values. Hence, the model output can be ex-
panded into a series of m-variate pth order multivariate polynomials as where x1 is the displacement, x2 is the velocity of the movable plate and
x3is the capacitor charge. We consider the case where each of the pa-
X
NP
ðm þ pÞ! rameters {A, ma, b, k, R} is perturbed by a relative uncertainty equivalent
yðt; Z m Þ≃ α i ðt ÞΨi ðz1 ; ::; zm Þ Np ¼ ð20Þ
m!p! to ρ = 20% around the nominal value, as reported in Table 1. The nom-
i¼1
inal value of the parameters are taken from Ref. [17]. The output of in-
terest is the displacement. It is known that when the displacement
The multivariate polynomials in Eq. (20) are calculated as products
becomes one third of the at-rest gap [2], the “pull in” phenomenon
of univariate orthonormal polynomials, as shown in Eq. (21):
tends to occur and the moving plate immediately crashes on to the
Y
m fixed plate. Hence, it is important to predict the probability of the dis-
ði Þ
Ψk ðz1 ; ::; zm Þ ¼ Ψk ðZ m Þ ¼ ψ ðz Þ ; k∈f1; ::; NP g ð21Þ placement exceeding its limit value, given the inherent uncertainties
I k ðiÞ i
i¼1
in the parameter values.
Pm ð jÞ
such that j¼1 I k ≤NP . The univariate orthonormal polynomial
Table 1
needs to be orthonormal with respect to the density function, Parameters of the physical model describing the kinetics of the MEMS actuator.
pi(zi) , i = 1 , … , m.
The coefficients αk can be found using the relationship: Parameter Symbol Value

Z Area A (m2) 1.2e−6


Permittivity ε (F/m) 8.85e−12
α k ðt Þ ¼ yðt; Z m ÞΨk ðZ m ÞdZ m ð22Þ
At-rest gap g0 (m) 8e−6
Γ Mass ma (kg) 15.4e−6
Damping constant B (Ns/m) 1.7e−3
where Γ is the support of the uncertain parameters. Spring constant K (N/m) 4.6
Normally, the integral in Eq. (22) is calculated using Gaussian Resistance R (Ω) 1
Control voltage Vin (V) 10
cubature. The cubature can be obtained by tensorizing the one
622 P.L.T. Duong et al. / Microelectronics Reliability 76–77 (2017) 619–625

random variables, i.e., the mass ma and plate area A, were detected as
being the most critical due to their high total sensitivity index {TA,
Tma} values among others. Since the total sensitivity index values for
R, b and k, {TR,Tb,Tk} are close to zero, these random parameters are
non-influential factors and can be omitted for UQ. For comparison pur-
pose, we also include the local sensitivity analysis of this model, utilizing
the complex step derivative. By definition, the local SA index may be
expressed by:

∂Mðt; Z Þ
Li ðt; zi Þ ¼ ð25Þ
∂zi

Recently, the complex step method [19] has been used as an efficient
approach for calculating the local sensitivity indices in Eq. (25). The idea
of using a complex variable was first proposed in Ref. [20]. It has been
shown to be very accurate and robust in addition to the ease of imple-
mentation. The idea is based on an approximation of the analytic func-
tion as:

∂Mðt; Z Þ Im½M ðt; Z þ j  hei Þ  


2
≈ þΟ h ð26Þ
∂zi h
Fig. 2. Total (dynamic) global sensitivity indices plotted as a function of time, obtained by
the M-DRM approach.
where j is an imaginary unit, h is the step size and ei is the ith unit vector
in Rn. Readers interested in knowing more about the advantages of the
According to M-DRM, a set of 50 Gaussian quadrature nodes, which complex step method over the finite difference approach are encour-
was generated using the MATLAB codes from the orthogonal polynomi- aged to refer to Refs. [19,20]. As can be seen from Fig. 3, while the
al toolbox [18], was passed on to the physical model, represented in Eq. local method correctly indicates A to be the most important parameter,
(24). The Sobol sensitivity indices, which can be used to identify param- it also points to {ma, R, k} being equally important. This is however mis-
eters of the model that have the largest influence on the prediction, leading, when compared to our results in Fig. 2.
were computed as in Section 2. The total value of the indices are plotted At the second stage, the polynomial chaos of order ten is constructed
in Fig. 2 as a function of time. As can be seen from this figure, the two and used to predict the time dependent density function of the

Fig. 3. Local sensitivity indices plotted as a function of time, obtained using the complex step derivative.
P.L.T. Duong et al. / Microelectronics Reliability 76–77 (2017) 619–625 623

displacement with the two important uncertain parameters, {A, ma}.


Once the gPC model has been constructed, we practically have an ana-
lytical relationship between the displacement and the parameters. The
density functions of displacement at any time moment can then be sam-
pled using this analytical gPC model without any computational burden.
Fig. 4 shows the density functions of displacement predicted by the gPC
with two random parameters at different time moments. For compari-
son purpose, the density functions of displacement predicted by the
QMC method with 5 and 2 random parameters are also shown in this
figure. It can be seen that there is very good agreement between these
density functions. Thus, the M-DRM + gPC approach accurately detects
the two influential parameters and gives good result when predicting
the density functions of displacement. On the other hand, the local SA
method only identifies A correctly as the most critical parameter and
fails to predict ma to be the second most critical one. The computational
time for 10,000 QMC simulations (with Halton sequence) was 467.85 s.
The proposed M-DRM + gPC method required (100 + 50) simulations
with the Gaussian quadrature nodes. The computational time of the
proposed method was only 5.13 s. All the computations were performed
using an Intel Core i5 6400T 2.2 GHz computer with an 8 GB RAM. The
proposed M-DRM + gPC indeed gives good results with the computa-
tional time reduced by two orders of magnitude in comparison to the
QMC method. It can be seen from Fig. 4 that after four milliseconds,
the actuator operates mostly in the unstable region (pull-in mode). Fig. 5. Comparison of the statistics of displacement obtained from the QMC method and
While for time lower than three milliseconds, the actuator is still mostly our proposed method. Statistical features from the QMC method are computed from
in the stable region. Fig. 5 compares the displacement statistics (first 10,000 simulations with a Halton Sequence. As for our proposed method, the statistical
results are computed from just 160 simulations with Gauss quadrature.
and second order moments) obtained by QMC/proposed method. It is
observed from Fig. 5 that the proposed method exhibits good agree-
ment with the QMC method for statistics of displacement. Fig. 6 com- combination with the proposed method to force the system output to
pares the density function predicted at t = 4 ms by sampling the be satisfied in tracking the objective and constraints even under the ef-
analytical surrogate gPC model with the standard Gaussian approxima- fect of uncertainties. However, it is clear that the QMC method is not
tion based on the mean and variance from Eq. (23). It should be noted suitable to be incorporated with the model predictive control frame-
from Figs. 5 and 6 that a reliability analysis of the system model utilizing work due to the high computational cost.
Gaussian approximation with mean and variance from Eq. (23) may be
misleading. 5. Conclusions of the study
The case above was for a step voltage of Vin. We present here an ad-
ditional short case study where the input Vin is a pulse with magnitude In this paper, the UQ and global SA methods were adopted for a dou-
7 V and length 3 ms. The proposed method was used to predict the den- ble plate MEMS electrostatic actuator. The purpose of UQ and global SA
sity function of the displacement. The results are shown in Fig. 7. It can involves testing of the robustness/reliability of the model and identifica-
be seen that the MEMS always operates in the stable region for this case. tion of the key parameters that influence the model outputs. We also
However, this is just a simple open loop control strategy. In future work, make comparisons of the global and local SA for the model. It has
the model predictive control framework [21] should be considered in been shown that the global SA can correctly identify the key parameters.

Fig. 4. Predicted density functions of displacement at t = {1, 2, 3, 4 milliseconds} using the QMC/M-DRM + gPC techniques (proposed). The solid line shows the start of pull-in mode.
10,000 Halton points were used in the QMC simulations with 2 and 5 random parameters.
624 P.L.T. Duong et al. / Microelectronics Reliability 76–77 (2017) 619–625

with the standard QMC method, which has a computational cost of


the order of 104.

Acknowledgement

This study is supported by the Temasek Seed Research Grant No.


IGDSS1604021 and the SUTD-ZJU Research Collaboration Grant No.
ZJURP1500103. The authors would like to thank these two agencies
for provision of research funding.

References

[1] J. Seeger, Charge control of parallel-plate, electrostatic actuators and the tip-in insta-
bility, J. MEMS 2 (5) (Oct. 2003) 656–671.
[2] S.D. Senturia, Microsystems Design, Kluwer Academic Publishers, 2002.
[3] J.S. Liu, Sequential Monte Carlo in Action, Monte Carlo Strategies in Scientific Com-
puting, Springer, New York, 2004.
[4] H. Niederreiter, P. Hellekalek, G. Larcher, P. Zinterhof, Monte Carlo and Quasi-Monte
Carlo Methods, Springer-Verlag, Berlin, 1996.
[5] P.L.T. Duong, M. Lee, Robust PID controller design for process with stochastic para-
metric uncertainties, J. Process Control 22 (Oct. 2012) 1559–1566.
[6] P.L.T. Duong, A. Wahid, E. Kwok, M. Lee, Uncertainty quantification and global sen-
sitivity analysis of complex chemical process using a generalized polynomial chaos,
Comput. Chem. Eng. 90 (July 2016) 23–30.
[7] G. Kewlani, J. Crawfordb, K. Iagnemma, A polynomial chaos approach to the analysis
of vehicle dynamics under uncertainty, Veh. Dyn. 50 (2) (May 2012) 749–774.
Fig. 6. Comparison of the density functions of displacement at t = 4 ms using the proposed
[8] W. Huberts, W.P. Donders, T. Delhaas, F.N. van de Vosse, Applicability of the polyno-
method and the Gaussian approximation from mean and variance. The probability density mial chaos expansion method for personalization of a cardiovascular pulse wave
function obtained from the proposed method is obtained by sampling the analytical gPC propagation model, Int. J. Numer. Methods Biomed. Eng. 30 (12) (December
surrogate model. The Gaussian approximation refers to the Gaussian density function 2014) 1679–1704.
with mean and variance computed using Eq. (23). [9] M. Dodson, G.T. Parks, Robust aerodynamic design optimization using polynomial
chaos, J. Aircr. 46 (2) (March–April 2009) 635–646.
[10] A. Saltelli, S. Tarantola, F. Campolongo, M. Ratto, Global Sensitivity Analysis for Im-
portance Assessment, Sensitivity Analysis in Practice, John Wiley & Sons, 2004.
To overcome the computational burden of the standard QMC method [11] A. Saltelli, M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M. Saisana, S.
for UQ and global SA, a two-stage approach is proposed here. Firstly, a Tarantola, Global Sensitivity Analysis: The Primer, John Wiley & Sons, 2008.
[12] A. Satelli, Making best use of model evaluations to compute sensitivity indices,
multiplicative dimensional model reduction method is used to identify-
Comput. Phys. Commun. 45 (2) (May 2002) 280–297.
ing key parameters from Sobol indices. Secondly, the gPC approach is [13] X. Zhang, M.D. Pandey, An effective approximation for variance-based global sensi-
used for UQ of the simplified model with influential parameters only. tivity analysis, Reliab. Eng. Syst. Saf. 121 (2014) 164–174.
The proposed approach shows superior performance in terms of com- [14] R.G. Ghanem, P.D. Spanos, Stochastic Finite Elements: A Spectral Approach, Dover
pubications, 2003.
putational burden when many uncertain parameters are to be consid- [15] D. Xiu, Numerical Methods for Stochastic Computations: A Spectral Method Ap-
ered. The results from the proposed approach show good agreement proach, Princeton University Press, 2010.

Fig. 7. Predicted density functions of displacement at t = {1, 2, 3, 4 milliseconds} using the M-DRM + gPC technique, when a pulse, Vin, of magnitude 7 V and length of 3 ms is applied.
P.L.T. Duong et al. / Microelectronics Reliability 76–77 (2017) 619–625 625

[16] K.A. Puvkov, N.D. Egupov, A.M. Makarenkov, A.I. Trofimov, Theory and Numerical [20] H.T. Banks, K.B. Maxwell, L. Bociu, M. Noorman, K. Tillman, The complex-step meth-
Methods for Studying Stochastic Systems, Fizmatlits, Moscow, 2003 (in Russian). od for sensitivity analysis of non-smooth problems arising in biology, Scientifice Re-
[17] L. Castaner, Understanding MEMS: Principles and Applications, John Wiley & Sons, port of Center for Research in Scientific Computation, October 2015https://www.
2016. ncsu.edu/crsc/reports/ftp/pdf/crsc-tr15-11.pdf.
[18] W. Gautschi, Orthogonal Polynomials: Computation and Approximation, Oxford [21] L. Grüne, J. Pannek, Nonlinear Model Predictive Control: Theory and Algorithms, 2nd
University Press, 2004https://www.cs.purdue.edu/homes/wxg/. edition Springer, 2016.
[19] J.N. Lyness, C.B. Moler, Numerical differentiation of analytic functions, SIAM J. Num.
Anal. 4 (1967) 202–210.

You might also like