Professional Documents
Culture Documents
A Kriging-Based Dynamic Adaptive Sampling Method For Uncertainty Quanti Fication
A Kriging-Based Dynamic Adaptive Sampling Method For Uncertainty Quanti Fication
A new method for dynamic sampling of Kriging surrogate models for uncertainty quantification is developed and
presented. The criterion for the dynamic adaptive sampling proposed is based on combining the expected uncertainty
of the fit and the gradient information resulting from the Kriging predictors, and an error-estimate term (based on the differ-
ence in the Kriging predictors with different correlation length scales). The Kriging-based dynamic adaptive sampling
method proposed is tested on two-dimensional analytic functions with smoothly and steeply varied responses in the quan-
tities of interest under normal uncertainty distributions. Compared with a classical polynomial chaos expansion method
based on the Gauss quadrature rule and a dynamic adaptive sampling method based only on the uncertainty of the Kriging
predictor fit, this new method shows superior performance for estimating the statistics of the quantity of interest in terms of
both accuracy and robustness, and regardless of either the choice of the initial set of samples or the smoothness of the
stochastic space.
137
Trans. Japan Soc. Aero. Space Sci., Vol. 62, No. 3, 2019
titative characterization and reduction of uncertainties in ap- SC employs multivariate interpolation polynomials (e.g., La-
plications of interest that include numerical simulations. A grange and Hermite). PCE estimates l for known orthogo-
conventional deterministic simulation considers the model nal polynomials, while SC forms interpolation functions
that has as an output a unique value of a solution fðxÞ at fixed for known l .
values of input parameters x. On the other hand, a probabil- Eldred and Burkardt4) compared the performance of PCE
istic simulation considers the model, which assumes that the and SC in simple algebraic test problems, and showed that, in
input parameters x are subject to a number of uncertainties principle, their performance is comparable and both ap-
and are, therefore, represented as random variables xðÞ. proaches result in very fast convergence relative to MC sam-
The output of the simulation fðxÞ is a distribution itself that pling. Hosder et al.5,6) applied PCE to computational fluid
depends on the parameter uncertainties , i.e., fðxðÞÞ ¼ dynamics (CFD) simulations: shock-wave and expansion-
fðÞ. Proper characterization of the relation between input wave problems with uncertain geometries and a transonic
and output uncertainties is a way to extract valuable informa- wing case with uncertain freestream conditions. The results
tion from our simulation models. computed using PCE agreed closely with the MC results at
UQ approaches are typically categorized into intrusive and a much lower computational cost. Moreover, PCE and SC
non-intrusive methods. Non-intrusive methods discussed in approaches have been applied and demonstrated in other
this paper only require multiple evaluations of the original UQ problems related to CFD simulations (these efforts are
(deterministic) model, i.e., users do not need to tamper with well summarized in Bijl et al.7)).
the details of the simulation codes. On the other hand, intru- However, when the stochastic behavior includes a steeply
sive methods require the formulation and solution of a sto- varied response, conventional PCE and SC approaches often
chastic version of the original model, i.e., users must spend suffer from approximation errors and deterioration in accu-
considerable time and effort modifying the simulation codes. racy and efficiency due to their reliance on a linear combina-
The most popular methods for non-intrusive UQ are sam- tion of polynomials. For example,
Fig. 1 shows the re-
pling methods, such as Monte Carlo (MC) sampling and Lat- sponses of fðÞ ¼ erf 2 þ 1 estimated using PCE (i.e.,
in hypercube sampling (LHS).1) To estimate the stochastic based on the Hermite polynomials suitable for a normally-
behavior of an output quantity of interest fðÞ resulting from distributed uncertainty). There are two types of approxima-
input parameter uncertainties , sampling methods evaluate tion errors; discretization errors appear around the steeply
the values of fðÞ, at many different locations sampled in varied response ( ¼ 1) due to a lack of sample points,
the stochastic space of , directly from repeated evaluation
of the simulation models. MC samples randomly in the sto-
chastic space, and it requires a large number of samples to Polynomial Discretization Polynomial
ensure the accuracy and convergence of the stochastic esti- Error Error Error
1.5
XP -1
and polynomial errors appear in the regions near the end- without unphysical local oscillations. In the same spirit, Wit-
points ( ¼ 4 and 4) because of Runge’s phenomenon. teveen and Iaccarino16) proposed a strategy for the adaptive
Figure 1 indicates that the discretization error can be reduced space-infilling sampling in SSC. It is based on an estimation
by additional samples close to the high-gradient region, but of the errors derived from hierarchical surpluses, polynomial
the polynomial error is amplified as the number of samples degree, and local extremum diminishing concepts. This pro-
and the polynomial order increase. cedure outperformed uniform refinement in terms of accu-
To improve the accuracy and efficiency of PCE and SC, racy. Palacios et al.17) utilized the adjoint method and similar
multiple researchers have been working on improved PCE ideas in a robust grid adaptation approach to CFD simula-
and SC approaches. Poëtte et al.8) proposed a new PCE meth- tions. This approach aims to minimize the numerical discre-
od based on the entropy variable to reduce the discretization tization error over small variations of the input parameters
errors in discontinuous responses. This approach bounds the around a baseline flow state. Similarly, Duraisamy and Chan-
oscillations in the vicinity of steeply varied responses (i.e., drashekar18) proposed a strategy for adaptive sampling in
discontinuities as the extreme) to a certain range through SSC based on an estimation of the error that relied on the
the entropy of the system and without the use of any adaptive use of a stochastic adjoint method.
stochastic space discretization. For discontinuous responses, Unlike polynomial functions used in PCE and SC, the
Poëtte and Lucor9) also proposed an iterative generalized Kriging surrogate model19) approximates stochastic behav-
PCE method that recovers the same error bounds in the vicin- iors in UQ problems as simple algebraic functions based
ity of discontinuities as in smoother regions without the need on Bayesian statistics. This approach can be easily adapted
to increase the polynomial order. Lockwood and Mavriplis10) to work well with nonlinear functions. It estimates not only
examined a gradient-based UQ approach to accurately pre- the function values, but also an estimate of the uncertainty
dict the stochastic behaviors in hypersonic flow with a non- in the fit that is equivalent to a measure of the approximation
linear response. This approach couples PCE with gradient in- error. Although the Kriging model has been developed and
formation evaluated by the adjoint method. Witteveen and utilized for engineering design optimization problems that
Iaccarino11) proposed the simplex stochastic collocation require many simulations,20,21) this paper proposes to use
(SSC) approach to collocate samples in nonhypercube sto- the Kriging model to approximate the stochastic space and
chastic spaces for approximating discontinuous responses ef- determine the effective criteria for both adaptive sampling
ficiently. This approach discretizes the stochastic space using and detecting steeply varied responses, and thus reducing er-
a simplex tessellation of the sampling points. Witteveen and rors in UQ.
Iaccarino12) then focused on discontinuous/smooth-mixed The Kriging-based dynamic adaptive sampling considered
responses and introduced a stencil selection strategy based here is implemented as shown in Fig. 2. For N given initial
on the essentially non-oscillatory (ENO) scheme into the samples, the Kriging model is constructed and the location
SSC approach for sharper resolution of discontinuities and of a new N þ 1-th sample is determined based on a criterion
higher-order approximations in smooth regions. that is formulated by the Kriging predictors. For N þ 1 sam-
The approximation errors in PCE and SC are also affected ples updated with the new additional sample, the Kriging
by the choice of sample locations. The Gauss quadrature model is reconstructed, and thus the next sample is deter-
rule13,14) is helpful for the PCE method to mathematically de- mined and added one-by-one by iterating this process. It is
termine where sample points should be collocated in the sto- expected to collocate the minimum number of samples nec-
chastic space, such that unphysical oscillations will disap- essary to approximate the black-box stochastic space.
pear as much as possible. Nevertheless, this rule is Yamazaki22) has also applied the Kriging model to UQ,
sometimes not acceptable for practical use because it may re- but his results indicated that, without adaptive sampling,
quest the location of a sample where numerical simulation is the performance of the Kriging model is inferior to classical
not available (e.g., typical CFD solvers cannot cope with the PCE. Dwight and Han23) proposed a Kriging-based adaptive
Mach number of 1 due to its singular flow condition that a sampling method that is based on the product of the fit uncer-
shock wave can stand infinitely forward of an object), and tainty in the Kriging predictor and the distributions of the in-
it does not allow us to add new samples in a space-filling put parameter uncertainties. Bilionis and Zabaras24) em-
manner. In addition, Hosder et al.15) demonstrated that, with ployed a similar adaptive refinement criterion based on the
a given value of the size of the sample set, PCE based on
LHS samples leads to more accurate estimation of statistics Generate initial samples
than if we were to use brute-force MC samples. Unfortu-
Construct Kriging
nately, the cost of such approximations exhibits the so-called
surrogate models
curse of dimensionality; meaning that the number of samples
required for PCE and SC increases exponentially as the num- Find the location of
a new sample
ber of input parameter uncertainties (i.e., dimensions in the based on a criterion
stochastic space) increases.
For these reasons, this paper attempts to adaptively refine Add the new sample
the stochastic space using as few samples as possible to ap-
proximate the stochastic behavior of a quantity of interest Fig. 2. Flowchart of Kriging-based dynamic adaptive sampling.
fit uncertainty predicted in Gaussian process (GP) regres- realization of a stochastic process FðÞ, defined as
sion,25) which is another Bayesian-statistics-based surrogate
FðÞ ¼ þ ZðÞ; ð2Þ
model similar to the Kriging model. However, as we will
show in this paper, such criteria are not effective in speeding where, ® is the global model assumed to be constant (i.e., so-
up the convergence of the UQ problem when the responses called ordinary Kriging) and ZðÞ is the local model corre-
behave steeply. sponding to the deviation from ® at . ZðÞ is assumed to
A highly nonlinear Kriging model enhanced with gradient have the stochastic behaviors (i.e., mean E½ZðÞ and cova-
information (@fðÞ=@) has recently been investigated for riance Cov½ZðÞ; Zð0 Þ) defined as follows:
adaptive sampling in UQ. Ulaganathan et al.26) considered E½ZðÞ ¼ 0; ð3aÞ
nonlinearity appearing in the gradient-enhanced Kriging 0 2 0
model as a sampling criterion to detect steeply varied re- Cov½ZðÞ; Zð Þ ¼ ð; jÞ; ð3bÞ
0
sponses. For a similar purpose, Boopathy and Rumpfkeil27) where, 2 is the process variance and ð; jÞ is the correla-
proposed a sampling criterion based on deviating the gra- tion function (or kernel) between any two locations and 0 .
dient-enhanced Kriging model from another local surrogate ð; 0 jÞ is defined as a function of the Euclidean distance
model (i.e., multivariate interpolation and regression). These h ¼ j 0 j with a set of constants called hyperparameters .
approaches are promising, but not always practical due to the ð; 0 jÞ satisfies the following conditions:
difficulty of providing the gradient information as the sample
lim ð; 0 jÞ ¼ 1; ð4aÞ
data set. h!0
The present study is focused on developing a new dynam- lim ð; 0 jÞ ¼ 0; ð4bÞ
ic adaptive sampling method based on an ordinary (i.e., gra- h!1
dient-free) Kriging model for effective non-intrusive UQ (a specific choice for ð; 0 jÞ will be given later). This paper
when the response of an output solution behaves either assumes ð; 0 jÞ to be stationary, which implies that the hy-
smoothly or steeply in the stochastic space. We first show perparameters are invariant to . Although a non-stationary
that the existing Kriging-based method23) is plagued by per- correlation function is expected to be more effective for non-
sistent errors when the response behaves steeply, and then linear function approximation,29–31) it is out of scope here
develop a method that eliminates these errors. The method and we focus on the stationary correlation function for its
proposed is then tested on two-dimensional analytic func- simple formulation.
tions with different smoothnesses, each of which models dif- Next, consider that the true value of fðÞ is given at N sam-
ferent flow physics related to aeronautics and astronautics. ple points ð1Þ ; ð2Þ ; ; ðNÞ . The Kriging model attempts to
The robustness of the method proposed is also investigated derive a posterior distribution over fðÞ, as illustrated in
by evaluating statistics using 30 different sets of initial sam-
ples generated at random. The Kriging-based method pro-
posed is compared with the existing Kriging-based method
as well as a classical PCE method based on the Gauss quad-
rature rule.
Function f()
[
[ ^s ([ )
[
- ^s ()~
ð1Þ ; ð2Þ ; ; ðNÞ in the stochastic space and evaluate their ð21Þ
output function values fðð1Þ Þ; fðð2Þ Þ; ; fððNÞ Þ. Criterion 2 is expected to add new samples where fit un-
: Step 2: Construct Kriging models fðÞ ^ and sðÞ
^ using the certainty is large and the gradient is also large. Thus, it
N samples. works for a balance between sampling evenly in the whole
: Step 3: Find the location of a new sample point
ðNþ1Þ
, stochastic space and capturing steeply varied responses
where the value of a criterion (i.e., CritðÞ to be defined where output solutions change markedly according to in-
later) is maximized in the stochastic space, and evaluate put parameter uncertainties. This is considered to reduce
its output function value fððNþ1Þ Þ. the discretization errors, as seen in Fig. 1.
: Step 4: Add fð
ðNþ1Þ
Þ to the N samples, set N N þ 1, : Criterion 3 (proposed in this paper)
and go back to Step 2. Criterion 2 still seems insufficient because new solutions
In Step 1, this study employs LHS1) to generate initial sam- ^
will not be sampled at the locations where j@fðÞ=@j ¼0
ples. In Step 3, for simplicity, this study carries out random ^ is large. This implies that this criterion may not
even if sðÞ
search, in which 104 samples are generated using MC sam- be able to suppress the polynomial errors as seen in the re-
pling, to find the location where CritðÞ is maximum; note gions near the endpoints in Fig. 1, where the true response
that optimizers such as genetic algorithms are also available is smooth. Therefore, Criterion 2 is now modified into the
for better findings. This paper considers the following three following Criterion 3, so that the locations where
criteria to be used in Step 3 for comparing existing sampling ^
j@fðÞ=@j ¼ 0 will still be sampled for new solutions.
^ !
methods to investigate effective criteria for UQ. @fðÞ
CritðÞ ¼
ðÞ þ D ^ ðÞ sðÞPDFðÞ
^ ð22Þ
@
: Criterion 1 (existing in Ref. 23))
f
^
CritðÞ ¼ sðÞPDFðÞ; ð18Þ ^
j@fðÞ=@j is now multiplied by
ðÞ to keep the consis-
^ is the Kriging model fit uncertainty given by
where, sðÞ tency of units with an extra term Df^ ðÞ.
ðÞ is set to be
Eq. (13), and PDFðÞ is the probability density function the distance from to the most adjacent sample point; that
of input parameter uncertainties . This criterion has been is,
proposed and investigated in a previous study by Dwight
ðÞ ¼ min j ðiÞ j: ð23Þ
and Han.23) Criterion 1 is expected to sample output solu- i¼1;2;;N
tions evenly in the stochastic space (i.e., from Eq. (14b), The extra term Df^ ðÞ in Eq. (22) is assumed to estimate
new samples will never duplicate existing ones), as well the polynomial errors. This paper formulates Df^ ðÞ as
as focus the solutions on locations where there are signifi-
^
Df^ ðÞ ¼ jfðjÞ ^
fðj2Þj; ð24Þ
cant effects on the evaluation of statistics (i.e., solutions
where PDFðÞ is higher will be sampled more frequently). ^
where, fðjÞ is the Kriging predictor modeled with the or-
: Criterion 2 dinary values of hyperparameters , which are determined
^ ^
by the MLE, and fðj2Þ is the predictor modeled with the
@fðÞ
CritðÞ ¼ sðÞPDFðÞ;
^ ð19Þ doubled values of hyperparameters 2.
@ For example, Fig. 4 compares the Kriging predictors
^
with different correlation length scales (fðj Þ, ^
fðj2 Þ,
where, ^ ^
fðj4 Þ, and fðj1=4 Þ) for fðÞ ¼ erf 2 þ 1 (i.e.,
T same example as Fig. 1). The predictor with a smaller cor-
^
@fðÞ=@ ^
¼ @fðÞ=@ ^ ^
1 ; @fðÞ=@2 ; ; @fðÞ=@n ^
relation length scale (fðj1=4 Þ) yields an unrealistic and
^ given by
is a vector of the Kriging predictor gradients fðÞ bumpy response that is excessively attracted by the given
Eq. (12) with respect to input parameter uncertainties . samples. On the other hand, the predictor with a larger cor-
^ consists of differentiable corre-
As the vector rðÞ in fðÞ ^
relation length scale (fðj2 Þ) slightly fluctuates from the
^
lation elements, @fðÞ=@ can be derived analytically as ^
ordinary predictor fðj Þ in regions where the true re-
sponse is smooth (4 2 and 0 4). Note
0.5
f ( [)
-0.5
true
Kriging (T)
-1 Kriging (2T)
Kriging (4T)
Kriging (1/4T)
samples
-1.5
-4 -3 -2 -1 0 1 2 3 4
[
Fig. 4. Examples of the Kriging Fig. 5. True response fð1 ; 2 Þ for the log-sin function.
predictors with different correlation
length scales (fðÞ ¼ erf 2 þ 1 ).
1
0.1
This indicates that the consideration of probabilistic signifi-
cance enhances the performance of dynamic adaptive sam-
pling. 0.01
4. Numerical Tests
0.001
Kriging (Crit. 1, 30 trials)
This study investigates the capability of the dynamic adap- Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials)
tive sampling criteria in the following two-dimensional UQ Kriging (LHS, 30 trials)
PCE (1 trial)
test problems. 0.0001
5 10 20 50 100
4.1. Log-Sin function N
4.1.1. Problem definition (b)ࠉsf
This function was considered in Hosder et al.15) and
Fig. 6. Errors in the statistics (trial averages) estimated for the log-sin
Yamazaki,22) and is formulated as
function.
fð1 ; 2 Þ ¼ ln 1 þ x1 2 sin 5x2 ; ð25aÞ
xi ¼ 2:0 þ 0:4i ði ¼ 1; 2Þ: ð25bÞ 4 i 4 (i ¼ 1; 2) using LHS. To ensure the robustness
The true response of the log-sin function is shown in Fig. 5. of results, the present test averages the results among 30 trials
This function has a multimodal response, which can be seen of dynamic adaptive sampling, each of which starts from a
in aeroacoustics considering coherent waves, for example. different LHS initial sample set.
Here, it is assumed that input variables 1 and 2 are uncer- For the log-sin function, the statistics (i.e., mean f and
tain and normally distributed as i Normð0; 12 Þ (i ¼ 1; 2). standard deviation f ) of the response fð1 ; 2 Þ against the
4.1.2. Results input uncertainties in 1 and 2 are estimated using the Krig-
The present test starts dynamic adaptive sampling based ing models updated from N ¼ 5 to 100 with Criteria 1–3, re-
on either Criteria 1, 2, or 3 from initial samples with N ¼ 5, spectively. Note that there is an empirical rule suggesting the
which are randomly generated within the range of number of initial samples as N ¼ 11n 1 for Kriging-surro-
1 0.45
6
MC (N=1u10 , 1 trial)
0.4
Kriging (Crit. 1, 30 trials)
Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials)
0.1 0.35 PCE (1 trial)
error in mf (trial std. dev.)
0.01 0.25
0.2
0.001 0.15
0.1
0.0001 0.05
Kriging (Crit. 1, 30 trials)
Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials) 0
-3 -2 -1 0 1 2 3
1e-05
Kriging (LHS, 30 trials) f( x1, x2)
5 10 20 50 100
N (a)ࠉN = 25
(a)ࠉmf 0.45
6
MC (N=1u10 , 1 trial)
1 0.4
Kriging (Crit. 1, 30 trials)
Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials)
0.35 PCE (1 trial)
0.25
0.01 0.2
0.15
0.001 0.1
0.05
0.0001 0
Kriging (Crit. 1, 30 trials)
-3 -2 -1 0 1 2 3
Kriging (Crit. 2, 30 trials) f( x1, x2)
Kriging (Crit. 3, 30 trials)
1e-05
Kriging (LHS, 30 trials) (b)ࠉN = 49
5 10 20 50 100
N 0.45
6
MC (N=1u10 , 1 trial)
(b)ࠉsf 0.4
Kriging (Crit. 1, 30 trials)
Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials)
0.35 PCE (1 trial)
Fig. 7. Errors in the statistics (trial standard deviations) estimated for the
PDF (trial average)
0.25
0.2
(a)ࠉKriging (Criterion 1)
(b)ࠉKriging (Criterion 2)
(c)ࠉKriging (Criterion 3)
(d)ࠉPCE
Fig. 9. Estimated responses for the log-sin function (left: N ¼ 25; middle: N ¼ 49; right: N ¼ 100).
f and f estimated by each method over 30 trials. Com- functions of 30 trials estimated using the Kriging models up-
pared to the pure Kriging, the Kriging-based dynamic adap- dated by Criteria 1–3 and one trial using the PCE method
tive sampling, especially based on Criterion 3, can quickly with N ¼ 25, 49, and 100 compared to one trial with
reduce the standard deviations of the errors in both f and N ¼ 1 106 estimated using the MC sampling method.
f . It means that the proposed Criterion 3 performs most ro- With N ¼ 25 and 49, the PCE method does not agree with
bustly regardless of the choice of initial samples. the MC sampling method although its error in f is well re-
Figure 8 shows the trial-averaged probability density duced with N ¼ 49, as seen in Fig. 6(a). This is a proof that
with N ¼ 25, 49, and 100. The sample points used are also
depicted as white circles in Fig. 9. Compared to the true re-
X
n1
fðÞ ¼ ui 2 þ vðn Þ; ð26aÞ
i¼1 Kriging (Crit. 1, 30 trials)
Kriging (Crit. 2, 30 trials)
where, ui is defined implicitly by Kriging (Crit. 3, 30 trials)
Kriging (Crit. 3', 30 trials)
2 Kriging (LHS, 30 trials)
n þ 1 ui þ ui 3 i ¼ 0 ði ¼ 1; 2; ; n 1Þ; ð26bÞ 0.01
PCE (1 trial)
5 10 20 50 100
while vðn Þ is explicitly given by N
(b)ࠉsf
vðn Þ ¼ a xn 4 þ ðxn 1Þ2 b ; ð26cÞ
xn ¼ 0:05n þ 0:25 ð26dÞ Fig. 11. Errors in the statistics (trial averages) estimated for the sandtimer
function.
1 1.4
MC (N=1u106, 1 trial)
Kriging (Crit. 1, 30 trials)
1.2 Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials)
PCE
error in mf (trial std. dev.)
0.1
0.8
0.6
0.01 0.4
0.8
0.1
0.6 0.2
0.4
0.01
0.2 0.1
1 1.1 1.2 1.3 1.4 1.5
Kriging (Crit. 1, 30 trials)
Kriging (Crit. 2, 30 trials) 0
Kriging (Crit. 3, 30 trials) -1 0 1 2 3 4 5
Kriging (Crit. 3', 30 trials) f ( x1, x2)
Kriging (LHS, 30 trials)
0.001
5 10 20 50 100
(b)ࠉN = 49 (with the enlarged view)
N
1.4
(b)ࠉsf MC ( N=1u106, 1 trial)
Kriging (Crit. 1, 30 trials)
1.2 Kriging (Crit. 2, 30 trials)
Kriging (Crit. 3, 30 trials)
Fig. 12. Errors in the statistics (trial standard deviations) estimated for the PCE (1 trial)
PDF (trial average)
0.8
1.2
ing pure Kriging LHS sampling sets and one trial using the 0.6
Gauss quadrature rule. The pure Kriging method also seems (c)ࠉN = 100 (with the enlarged view)
stagnant in error reduction for estimating f . This indicate a
significant advantage of Kriging-based dynamic adaptive Fig. 13. Probability density functions (trial averages) of fð1 ; 2 Þ esti-
sampling for effective UQ. For the dynamic adaptive sam- mated for the sandtimer function.
pling, Criterion 3 shows the fastest and continuous error re-
duction until N ¼ 100. The error in f is almost saturated by which eliminates Df^ ðÞ from Criterion 3.
Criteria 1 and 2, even with N > 20. Hence Criterion 3 seems Figures 11 and 12 indicate that Criterion 3A performs well
most promising for efficient dynamic adaptive sampling. In in Kriging-based dynamic adaptive sampling. Nevertheless,
Fig. 12, moreover, Criterion 3 indicates another advantage Criterion 3 outperforms Criterion 3A in the later stage (i.e.,
that performance is less sensitive to the choice of initial sam- N > 35) to estimate f . This is because, as explained in
ples than Criteria 1 and 2. Fig. 4, Df^ ðÞ in Criterion 3 is capable of detecting small
To confirm the contribution of the term Df^ ðÞ in Criteri- fluctuations caused by different correlation length scales,
on 3 (Eq. (22)), Figs. 11 and 12 compare the following Cri- which affect the accuracy to estimate a smooth response.
terion 3A as well. Hence, consideration of Df^ ðÞ is essential to push through
: Criterion 3¤ the dynamic adaptive sampling toward comprehensive UQ
^ approximation.
@fðÞ
CritðÞ ¼
ðÞsðÞPDFðÞ;
^ ð27Þ Figure 13 compares the trial-average probability density
@ functions using Kriging models updated by 30 trials applying
(a)ࠉKriging (Criterion 1)
(b)ࠉKriging (Criterion 2)
(c)ࠉKriging (Criterion 3)
(d)ࠉPCE
^ 1 ; 2 Þ for the sandtimer function (left: N ¼ 25; middle: N ¼ 49; right: N ¼ 100).
Fig. 14. Estimated responses fð
Criteria 1–3, respectively, one trial of the PCE method using agree with the MC sampling method. This indicates that
N ¼ 25, 49 and 100, and one trial of the MC sampling meth- Kriging-based dynamic adaptive sampling is effective to re-
od using N ¼ 1 106 . There is still a discrepancy between duce the computational cost for UQ. With N ¼ 100, Crite-
the PCE method and the MC sampling method even with ria 2 and 3 can catch up with the true peak of probability den-
N ¼ 100. On the other hand, the Kriging models with sity around fð1 ; 2 Þ ¼ 0:23, while Criterion 1 approximates
N ¼ 100 obtain the probability density functions that almost it as two local peaks around fð1 ; 2 Þ ¼ 0:17 and 0.35. This
can be seen in the enlarged view in Fig. 13(c). Here, we that the criterion named Criterion 2 in this paper, which is
argue that it is very important to predict PDF in an engineer- based on information about both the fit uncertainty sðÞ^ and
ing sense. Even a slight change in the number and location of ^
gradient @fðÞ=@ estimated using the Kriging surrogate
peak(s) in an estimated PDF, which appears in Fig. 13(c), model, showed better performance for effective adaptive
may affect the reliability of engineering design (e.g., Kawai sampling than the existing criterion (i.e., Criterion 1) based
and Shimoyama38) showed that a transonic airfoil involves only on Kriging model fit uncertainty sðÞ.^ In addition, the
multiple peaks in the PDF of local surface pressure due to polynomial errors appearing near the endpoints in the sto-
a moving shock wave under an uncertain flight Mach num- chastic space can be reduced by adding an extra term, Df^ ðÞ,
ber, and an accurate estimate of these peaks is relevant to ver- which represents the difference in Kriging predictors based
ification and validation between CFD and wind tunnel ex- on different correlation length scales, to the gradient term
periments). In addition, with N ¼ 49, Criterion 3 can also in the adaptive sampling criterion. Finally, the criterion pro-
capture a smooth trend of probability density in the range posed (i.e., Criterion 3) showed superior performance in
of 1 fð1 ; 2 Þ 1:5, while Criterion 2 causes slight oscil- terms of accuracy and robustness for estimating the statistics
lations, which can be seen in the enlarged view in Fig. 13(b). of the output solution in UQ regardless of the choice of initial
It is worth noting that the range of 1 fð1 ; 2 Þ 1:5 corre- samples and smoothness of the stochastic space. Further-
sponds to the smooth region seen in the true response more, the Kriging-based dynamic adaptive sampling pro-
(Fig. 10). This demonstrates that Criterion 3 is effective for posed outperformed a classical polynomial chaos expansion
reducing the polynomial errors in smooth regions by consid- (PCE) method based on the Gauss quadrature rule in both
ering the extra error-reduction term Df^ ðÞ in Eq. (22). This test functions.
effectiveness is also be observed from the sample collocation This paper focused on analytical test functions that model
(Fig. 14). Hence, this suggests that the Kriging-based dy- typical flow physics seen in the fields of aeronautics and as-
namic adaptive sampling using Criterion 3 is capable of es- tronautics. Our future work will test the Kriging-based dy-
timating both smoothly and steeply varied stochastic behav- namic adaptive sampling proposed by applying it to more
iors of a solution. realistic UQ problems coupled with CFD to solve Navier-
^ 1 ; 2 Þ using the Kriging mod-
The responses estimated fð Stokes equations. As a matter of fact, that work is now under-
els and PCE method are compared in Fig. 14. The PCE way.
method yields a polynomial error due to the oscillations at
2 5, which do not appear in the true response Acknowledgments
(Fig. 10). For the Kriging model, Criterion 1 produces the
samples, which are almost normally distributed correspond- The authors would like to thank Juan J. Alonso of Stanford Uni-
ing to the given PDFð1 ; 2 Þ regardless of the smoothness of versity for his valuable comments, which helped to improve the dy-
^ 1 ; 2 Þ. Hence Criterion 1 does not concentrate on the
fð namic adaptive sampling method, and the Sandia National Labora-
tories for Dakota. In addition, the authors were supported in part by
sandtimer region with a steep gradient around 2 ¼ 0 to be the Grant-in-Aid for Scientific Research (B) No. H1503600 and the
captured for accurate estimation of f and f . Criterion 2 Grant-in-Aid for Scientific Research (B) No. 18H01620 adminis-
produces more sample points around 2 ¼ 0 than Criterion 1 tered by the Japan Society for the Promotion of Science (JSPS).
in the sandtimer region as N increases. However, Criterion 2
seems careless to add sample points in the smooth region far
away from 2 ¼ 0. Therefore, it cannot reduce the polyno- References
mial errors that cause slight oscillations in the probability
1) McKay, M. D., Beckman, R. J., and Conover, W. J.: A Comparison of
density functions, as seen at 1 fð1 ; 2 Þ 1:5 in Three Methods for Selecting Values of Input Variables in the Analysis
Fig. 13(b). Compared to Criteria 1 and 2, Criterion 3 can of Output from a Computer Code, Technometrics, 21, 2 (1979),
keep a balance in sample collocation between the sandtimer pp. 239–245.
2) Xiu, D. and Karniadakis, G. E.: The Wiener-Askey Polynomial Chaos
region and the other smooth region. Therefore, the present
for Stochastic Differential Equations, SIAM J. Sci. Comput., 24, 2
test demonstrates that the proposed criterion, Criterion 3, is (2002), pp. 619–644.
the best for dynamic adaptive sampling to make efficient 3) Xiu, D. and Hesthaven, J. S.: High-Order Collocation Methods for Dif-
UQ regardless of the choice of initial samples and the ferential Equations with Random Inputs, SIAM J. Sci. Comput., 27, 3
(2005), pp. 1118–1139.
smoothness of the response of an output solution in the sto-
4) Eldred, M. S. and Burkardt, J.: Comparison of Non-Intrusive Polyno-
chastic space. mial Chaos and Stochastic Collocation Methods for Uncertainty Quan-
tification, 47th AIAA Aerospace Sciences Meeting including The New
5. Conclusions Horizons Forum and Aerospace Exposition, AIAA Paper 2009-976,
AIAA, Reston, VA, January 2009.
5) Hosder, S., Walters, R., and Perez, R.: A Non-Intrusive Polynomial
This paper proposed a new Kriging-based criterion for dy- Chaos Method for Uncertainty Propagation in CFD Simulations,
namic adaptive sampling in uncertainty quantification (UQ), 44th AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper
and compared its performance in predicting stochastic be- 2006-891, AIAA, Reston, VA, January 2006.
6) Hosder, S., Walters, R. W., and Balch, M.: Point-Collocation Nonin-
haviors of an output solution in two-dimensional test func-
trusive Polynomial Chaos Method for Stochastic Computational Fluid
tions (i.e., log-sin function and sandtimer function) under Dynamics, AIAA J., 48, 12 (2010), pp. 409–435.
normally distributed uncertainties. The results demonstrated 7) Bijl, H., Lucor, D., Mishra, S., and Schwab, C. (eds.).: Uncertainty
Quantification in Computational Fluid Dynamics, Lecture Notes in Phys., 231, 17 (2012), pp. 5718–5746.
Computational Science and Engineering, No. 92, Springer, Heidelberg 25) Rasmussen, C. E. and Williams, C. K. I.: Gaussian Processes for Ma-
New York Dordrecht London, 2013. chine Learning, The MIT Press, Cambridge, MA, 2006.
8) Poëtte, G., Després, B., and Lucor, D.: Uncertainty Quantification for 26) Ulaganathan, S., Couckuyt, I., Dhaene, T., Degroote, J., and Laermans,
Systems of Conservation Laws, J. Comput. Phys., 228, 7 (2009), E.: High Dimensional Kriging Metamodelling Utilising Gradient In-
pp. 2443–2467. formation, Appl. Math. Modelling, 40, 9-10 (2016), pp. 5260–5270.
9) Poëtte, G. and Lucor, D.: Non Intrusive Iterative Stochastic Spectral 27) Boopathy, K. and Rumpfkeil, M. P.: Unified Framework for Training
Representation with Application to Compressible Gas Dynamics, Point Selection and Error Estimation for Surrogate Models, AIAA J.,
J. Comput. Phys., 231, 23 (2012), pp. 3587–3609. 53, 1 (2015), pp. 215–234.
10) Lockwood, B. and Mavriplis, D.: Gradient-based Method for Uncer- 28) Lukaczyk, T. W., Constantine, P., Palacios, F., and Alonso, J. J.: Ac-
tainty Quantification in Hypersonic Flows, Comput. Fluids, 85 tive Subspaces for Shape Optimization, 10th AIAA Multidisciplinary
(2013), pp. 27–38. Design Optimization Conference, AIAA Paper 2014-1171, AIAA, Re-
11) Witteveen, J. A. S. and Iaccarino, G.: Simplex Stochastic Collocation ston, VA, January 2014.
with Random Sampling and Extrapolation for Nonhypercube Proba- 29) Xiong, Y., Chen, W., Apley, D., and Ding, X.: A Non-Stationary Co-
bility Spaces, SIAM J. Sci. Comput., 34, 2 (2012), pp. A814–A838. variance-Based Kriging Method for Metamodelling in Engineering
12) Witteveen, J. A. S. and Iaccarino, G.: Simplex Stochastic Collocation Design, Int. J. Numer. Methods Eng., 71, 6 (2007), pp. 733–756.
with ENO-type Stencil Selection for Robust Uncertainty Quantifica- 30) Toal, D. J. J. and Keane, A. J.: Non-Stationary Kriging for Design Op-
tion, J. Comput. Phys., 239 (2013), pp. 1–21. timization, Eng. Optim., 44, 6 (2012), pp. 741–765.
13) Stroud, A. H. and Secrest, D.: Gaussian Quadrature Formulas, Pren- 31) Clark, D. L., Jr., Bae, H.-R., Gobal, K., and Penmetsa, R.: Engineering
tice-Hall, Englewood Cliffs, NJ, 1966. Design Exploration Using Locally Optimized Covariance Kriging,
14) Stoer, J. and Bulwirsch, R.: Introduction to Numerical Analysis, Texts AIAA J., 54, 10 (2016), pp. 3160–3175.
in Applied Mathematics, Vol. 12, Springer, New York, NY, 2002. 32) Goldberg, D. E.: Genetic Algorithms in Search, Optimization and Ma-
15) Hosder, S., Walters, R., and Balch, M.: Efficient Sampling for Non- chine Learning, Addison-Wesley Publishing Company, Reading, MA,
Intrusive Polynomial Chaos Applications with Multiple Uncertain In- 1989.
put Variables, 9th AIAA Non-Deterministic Approaches Conference, 33) Wendland, H.: Piecewise Polynomial, Positive Definite and Com-
AIAA Paper 2007-1939, AIAA, Reston, VA, April 2007. pactly Supported Radial Functions of Minimal Degree, Adv.
16) Witteveen, J. A. S. and Iaccarino, G.: Refinement Criteria for Simplex Comput. Math., 4, 1 (1995), pp. 389–396.
Stochastic Collocation with Local Extremum Diminishing Ro- 34) Mehmani, A., Chowdhury, S., Zhang, J., and Messac, A.: Surrogate-
bustness, SIAM J. Sci. Comput., 34, 3 (2012), pp. A1522–A1543. based Design Optimization with Smart Sequential Sampling, 12th
17) Palacios, F., Duraisamy, K., Alonso, J. J., and Zuazua, E.: Robust Grid AIAA Aviation Technology, Integration, and Operations (ATIO) Con-
Adaptation for Efficient Uncertainty Quantification, AIAA J., 50, 7 ference and 14th AIAA/ISSMO Multidisciplinary Analysis and Opti-
(2012), pp. 1538–1546. mization Conference, AIAA Paper 2012-5707, AIAA, Reston, VA,
18) Duraisamy, K. and Chandrashekar, P.: Goal-Oriented Uncertainty September 2012.
Propagation Using Stochastic Adjoints, Comput. Fluids, 66, (2012), 35) Shimoyama, K., Kawai, S., and Alonso, J. J.: Dynamic Adaptive Sam-
pp. 10–20. pling Based on Kriging Surrogate Models for Efficient Uncertainty
19) Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. P.: Design and Quantification, 15th AIAA Non-Deterministic Approaches Confer-
Analysis of Computer Experiments, Statist. Sci., 4, 4 (1989), ence, AIAA 2013-1470, AIAA, Reston, VA, April 2013.
pp. 409–435. 36) Knowles, J. and Hughes, E. J.: Multiobjective Optimization on a Bud-
20) Jones, D. R., Schonlau, M., and Welch, W. J.: Efficient Global Opti- get of 250 Evaluations, Coello Coello, C. A., Aguirre, A. H., and Zit-
mization of Expensive Black-Box Function, J. Global Optimization, zler, E. (eds.), Evolutionary Multi-Criterion Optimization: Third Inter-
13 (1998), pp. 455–492. national Conference, EMO 2005, Lecture Notes in Computer Science,
21) Jeong, S., Minemura, Y., and Obayashi, S.: Optimization of Com- Vol. 3410, Springer, Berlin Heidelberg, 2005, pp. 176–190.
bustion Chamber for Diesel Engine Using Kriging Model, J. Fluid 37) Dakota Ver: 6.1, Sandia National Laboratories, Albuquerque, NM,
Sci. Technol., 1, 2 (2006), pp. 138–146. USA, November 2014, https://dakota.sandia.gov [cited 17 January
22) Yamazaki, W.: Stochastic Drag Analysis via Polynomial Chaos Un- 2015].
certainty Quantification, Trans. Jpn. Soc. Aeronaut. Space Sci., 58 38) Kawai, S. and Shimoyama, K.: Kriging-Model-Based Uncertainty
(2015), pp. 89–99. Quantification in Computational Fluid Dynamics, 32nd AIAA Applied
23) Dwight, R. and Han, Z.-H.: Efficient Uncertainty Quantification Us- Aerodynamics Conference, AIAA Paper 2014-2737, AIAA, Reston,
ing Gradient-Enhanced Kriging, 11th AIAA Non-Deterministic Ap- VA, June 2014.
proaches Conference, AIAA Paper 2009-2276, AIAA, Reston, VA,
May 2009.
Nobuyuki Tsuboi
24) Bilionis, I. and Zabaras, N.: Multi-Output Local Gaussian Process Re-
gression: Applications to Uncertainty Quantification, J. Comput. Associate Editor