Chinese Journal of Aeronautics, (2022), 35(11): 59–73

Chinese Society of Aeronautics and Astronautics

& Beihang University
Chinese Journal of Aeronautics
[email protected]
www.sciencedirect.com

Selecting scale factor of Bayesian multi-fidelity

surrogate by minimizing posterior variance
Hongyan BU, Liming SONG, Zhendong GUO *, Jun LI

Institute of Turbomachinery, School of Energy & Power Engineering, Xi’an Jiaotong University, Xi’an 710049, China

Received 29 October 2021; revised 29 November 2021; accepted 7 January 2022

Available online 3 June 2022

KEYWORDS Abstract The Bayesian Multi-Fidelity Surrogate (MFS) proposed by Kennedy and O’Hagan
Co-Kriging; (KOH model) has been widely employed in engineering design, which builds the approximation
Gaussian process regression; by decomposing the high-fidelity function into a scaled low-fidelity model plus a discrepancy func-
Multi-fidelity surrogate; tion. The scale factor before the low-fidelity function, q, plays a crucial role in the KOH model. This
Optimization; scale factor is always tuned by the Maximum Likelihood Estimation (MLE). However, recent stud-
Scale factor ies reported that the MLE may sometimes result in MFS of bad accuracy. In this paper, we first
present a detailed analysis of why MLE sometimes can lead to MFS of bad accuracy. This is
because, the MLE overly emphasizes the variation of discrepancy function but ignores the function
waviness when selecting q. To address the above issue, we propose an alternative approach that
chooses q by minimizing the posterior variance of the discrepancy function. Through tests on a
one-dimensional function, two high-dimensional functions, and a turbine blade design problem,
the proposed approach shows better accuracy than or comparable accuracy to MLE, and the pro-
posed approach is more robust than MLE. Additionally, through a comparative test on the design
optimization of a turbine endwall cooling layout, the advantage of the proposed approach is further
validated.
Ó 2022 Chinese Society of Aeronautics and Astronautics. Production and hosting by Elsevier Ltd. This is
an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction process.1,2 These computations provide plenty of information

and could be employed for design optimization3 and sensitivity
In the past decades, computer simulations such as Computa- analysis.4,5 However, these simulations can be time-consuming
tional Fluid Dynamics (CFD) and Finite Element Analysis and excessive running of them is infeasible. To reduce the com-
(FEA) have been widely adopted in the aeronautical design putational cost, surrogate models, also known as metamodels
or response surface have drawn much attention as they are
* Corresponding author. algebraic expressions and cheap to evaluate. Several kinds of
E-mail address: [email protected] (Z. GUO). surrogates such as Radial Basis Function (RBF),6 Support
Peer review under responsibility of Editorial Committee of CJA. Vector Regression (SVR),7 and Gaussian Process (GP) Baye-
sian model (also known as Kriging)8 have been introduced
to the aeronautical design. Due to its stochastic structure
which gives both the prediction and the uncertainty, GP Baye-
Production and hosting by Elsevier sian model has become more and more popular.

https://doi.org/10.1016/j.cja.2022.05.012
1000-9361 Ó 2022 Chinese Society of Aeronautics and Astronautics. Production and hosting by Elsevier Ltd.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
60 H. BU et al.

In the design process, computer simulations could run at tion in MF modeling. The main contributions of the current
multiple levels of accuracy. For example, CFD simulations study are given as follows. It is the first time that the MLE’s
could run with variable mesh resolutions. The governing equa- problem in q selection is analyzed in detail. Besides, an alterna-
tions range from inviscid (Euler) to viscous (Navier-Stokes). tive approach for q selection is proposed to overcome this
The turbulence could be modeled as scale averaged (RANS, problem. The remainder of this paper is organized as follows.
URANS) or scale resolved (LES, DNS). Some of the above- Section 2 gives the background of Bayesian MFS. Section 3
mentioned High-Fidelity (HF) computations are so expensive provides a detailed illustration of the MLE’s problem. Sec-
that only limited samples could be evaluated, which means that tion 4 presents the formulation of the proposed approach. In
the accuracy of the obtained surrogate may not be satisfactory. Section 5, numerical examples and a turbine blade design
To achieve a better approximation of the HF simulations, problem are tested to show the effectiveness of the proposed
some proposed to employ the Low-Fidelity (LF) simulations approach and its applicability to real-world design problems.
as auxiliary information to build the Multi-Fidelity Surrogate Additionally, a comparative test on the design optimization
(MFS). The most popular MFS is Kennedy and O’Hagan’s of a turbine endwall cooling layout is carried out, which fur-
Bayesian MFS (KOH model).9,10 They developed a Bayesian ther validates the advantage of the proposed approach.
multi-fidelity framework in which the HF function is approxi-
mated as the sum of the scaled LF function and a discrepancy 2. Background of Bayesian MFS
function. Given the fact that there are enough LF samples to
obtain a fine approximation of the LF function, if the discrep- In this section, we will briefly introduce the classical Bayesian
ancy function is tuned to be simple and easy to predict, the HF multi-fidelity surrogate. The formulation of model prediction
function is expected to be predicted more accurately. In this and uncertainty are given and the MLE parameter tuning
MFS, the scale factor q is always selected with Maximum approach is introduced.
Likelihood Estimation (MLE). As will be presented in the fol- In Bayesian multi-fidelity surrogate, the high-fidelity func-
lowing sections, this paper’s investigations are conducted tion is modeled as a Gaussian Process (GP) which is the com-
under the framework of the KOH model. bination of low-fidelity and discrepancy GPs:
In the past decade, several variants of MFS have been pro-
posed to improve model performance. Han and Görtz11 pro- Yh ðxÞ ¼ qYl ðxÞ þ dðxÞ ð1Þ
posed a simple but decent MFS called Hierarchical Kriging where Yh ðxÞ,Yl ðxÞ, and dðxÞ denote the GPs of the high-
(HK). In this framework, the HF function is approximated fidelity, low-fidelity, and discrepancy function respectively. q
with a Kriging surrogate model in which another Kriging is the scale factor that scales the data of low fidelity to match
model built from LF data serves as the trend function. Several with high fidelity. Once the observations of high- and low-
favorable properties are obtained within this framework, such fidelity functions are obtained, the multi-fidelity prediction of
as a more reasonable formulation of the prediction uncer- unknown point x can be calculated from the mean of posterior
tainty, no requirement of the nested sample sets. Zhang et al.12 distribution9:
proposed a novel MFS based on single linear regression. In
^
this model, the discrepancy function is modeled with polyno- yh ðxÞ ¼ EðYh ðxÞjyh ; yl ; hl ; hd ; qÞ ð2Þ
mials and the low-fidelity model is considered as an additional
where yh and yl denote the observations of high- and low-
basis function. The unknown coefficients of the polynomials
fidelity functions respectively; hl and hd denote the hyper-
along with the LF model scale factor are obtained with a single
parameters of low-fidelity and discrepancy GPs respectively.
linear regression. Numerical results indicate that this approach
The uncertainty of prediction can be calculated from the vari-
is more robust with noisy data. Hao et al.13 incorporated the
ance of the posterior distribution:
gradient information into an MFS framework and proposed
an adaptive infill sampling criterion which achieved the opti- ^2h ðxÞ ¼ VarðYh ðxÞjyh ; yl ; hl ; hd ; qÞ
r ð3Þ
mum value with fewer function calls. Although in some newly
proposed MFS the scale factors could be obtained analytically, To be more specific, the GP models of low-fidelity and dis-
MLE is still widely adopted for the selection of q. Zhou et al.14 crepancy functions are defined as.


proposed a generalized multi-fidelity Kriging model which Yl ðxÞ  GP fðxÞT bl ; r2l
does not require nested samples and is more robust when cor-
 ð4Þ
relations between fidelity levels are weak. Long et al.15 devel- dðxÞ  GP fðxÞT bd ; r2d
oped a multi-objective adaptive infill sampling approach for
a Bayesian MFS model and applied it to the optimization of where fðxÞT is the basis function vector, bl and bd are the coef-
geostationary orbit satellite systems. In these multi-fidelity sur- ficient vectors of the process mean function, and r2l and r2d are
rogates, the scale factors are also obtained with MLE. the process variances. The prediction and prediction’s uncer-
As introduced above, the key parameter of the Bayesian tainty of unknown point x are given by.
MFS is the scale factor q. As given in the work of Forrester ^
et al.16 the scale factor is often selected with MLE. However, yh ðxÞ ¼ f0 ðxÞT b0 þ tðxÞT C1 ðyS  Fb0 Þ ð5Þ
some studies reported that the MLE may fail to select suitable
^2

T
q. Guo et al .17 studied the MLE’s choice in the selection rh ðxÞ ¼ q2 r2l þ r2d  tðxÞT C1 tðxÞ þ f 0 ðxÞ  tðxÞT C1 F 
between HF only model (q ¼ 0) and Multi-Fidelity (MF)
 T 1 1
0 
model (q > 0). They found that even if the HF model is more F C F f ðxÞ  tðxÞT C1 F
accurate, the MLE frequently chooses the MF model. How-
ever, their study mainly focused on the model selection ð6Þ
between HF and MF and did not pay attention to the q selec-
Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 61

where f 0 ,F, b are the extended basis function and coefficient Nested sample sets are adopted, which are the same as
vectors (matrix) of the process mean function, tðxÞ is the those in the study of Han and Görtz.11 Samples of high- and
covariance vector between x and the sample points, C is the low-fidelity functions are Sh ¼ f0; 0:4; 0:6; 1:0g and
covariance matrix of the sample points, and yS is the observa- Sl ¼ f0; 0:1; 0:2; 0:3; 0:4; 0:5; 0:6; 0:7; 0:8; 0:9; 1:0g. The ranges
tion vector. More detailed formulations of the above- and tuning strategy of hyper-parameters can be found in the
mentioned quantities are listed in Appendix A. next section. The MF models built with MLE and manually
The hyper-parameters of Bayesian multi-fidelity surrogate selected parameters (q ¼ 1, hd ¼ 1) are compared. As shown
have always been estimated using MLE.16 The MLE chooses in Fig. 1, the MLE model’s prediction is not satisfactory, espe-
the hyper-parameters by minimizing the negative logarithm cially in the range of [0.6, 0.9], while the manually chosen
of the concentrated likelihood function: model presents accurate prediction in the whole range. q
8 selected by MLE is 1.46 which is not the best value, as the pre-
< hl ¼ arg min fLl ðhl Þg
hl dicted discrepancy function fails to approximate the real one.
ð7Þ
: ðq; hd Þ ¼ arg min fLd ðq; hd Þg However, as shown in Fig. 1(d), when q ¼ 1, the discrepancy
q;hd
function is well predicted. Hence, it is worth investigating that
where Ll and Ld are the negative logarithm likelihood function why MLE did not choose the best q. For clarity of compar-
of low-fidelity and discrepancy model respectively: ison, Table 1 gives the q and negative-logarithmic likelihood
nl  2  1 values of MLE and manually chosen model.
Ll ðhl Þ ¼ ln rl þ lnjRl j ð8Þ As mentioned above, the likelihood function is composed
2 2
of two terms: the variation term r2d and the waviness term
nh  2  1 jRd j. To analyze each term’s contribution to likelihood func-
Ld ðq; hd Þ ¼ ln rd þ lnjRd j ð9Þ tion, the likelihood function and each term’s value under dif-
2 2
ferent q and hd are presented in Fig. 2. As the likelihood and
where nl , nh are the number of LF and HF samples respec-
variation term are the functions of both hd and q, the values
tively, Rl and Rh are the correlation matrix of LF and HF sam-
are plotted at several different q. Those hd that minimize the
ples respectively.
likelihood under specific q are plotted with red crosses in
It should be noted that Ld is the function of q because the
Fig. 2(a). Those hd that minimize the variation term under
r2d term in Eq. (9) is determined by the observations of discrep-
specific q are plotted with red crosses in Fig. 2(b). These red
ancy yd ¼ yh  qyl , as given in Eq. (6). We will further discuss
crosses could be regarded as candidate hyper-parameter com-
Ld in the next section to show how the minimization of Ld may
binations ððq; hd ÞÞ for the selection of q. As shown in Fig. 2(a),
not lead to the best q.
among all the candidate points, the point on the q ¼ 1:5 curve
produces the lowest likelihood function value, which is consis-
3. Discussion on MLE’s effectiveness tent with the MLE result plotted with red dots. Comparing
Fig. 2(a) with Fig. 2(b), one can observe that the relative posi-
Although the MLE is widely adopted to select the scale factor tions of constant q curves and candidate points on them do not
q, it is not always effective. Guo et al.17 found that, in the change much. This indicates that the variation term’s effect in
selection between HF only model (q ¼ 0) and MF model the likelihood function is overwhelming so that the waviness
(q > 0), MLE frequently chose the MF model even if the HF term’s effect is nearly eliminated. This could also be observed
model is more accurate. However, their research mainly in Fig. 2(c). Although the waviness term’s absolute value
focused on the model selection rather than q tuning. In this increases greatly as hd approaches 0, it is relatively small in
section, we will analyze the MLE’s problem in q tuning in the region where the candidate points locate (hd > 5). Conse-
detail. To solve this problem, an alternative approach to select quently, the MLE overly emphasized the discrepancy model’s
the scale factor is proposed in the next section. variation while ignored the waviness. As shown in Fig. 1, this
As given in Eq. (9), the likelihood function employed by bias leads to a wrong estimation of q and corresponding poor
MLE to find q and hd is composed of two terms: the r2d term model accuracy.
and the jRd j term. These terms represent different properties This property of MLE could also be found in some other
of the discrepancy function. Guo et al.17 pointed out that the researches. Guo et al.17 studied the MLE’s selection between
r2d term reflects the variation of the discrepancy model and HF only model ðq ¼ 0Þ and MF model ðq > 0Þ. They found
the jRd j term is a representation of the model waviness. Park that the MLE tends to select the model with smaller discrep-
et al.10 further defined the combined effect of these two terms ancy function variation. As the MF model’s discrepancy func-
as an indicator of model bumpiness. In general, the MLE tries tion variation is often smaller than the HF function variation,
to reduce the discrepancy model’s variation and waviness
together, so that the real discrepancy function could be
approximated easier with limited observations. However, this
approach may not always be effective. Here, we illustrate this Table 1 q and negative-logarithmic likelihood values of
drawback with a one-dimensional example, the Forrester func- models.
tion in Ref.14:
8 Model q Ld
< yh ¼ ð6x  2Þ2 sin ½2ð6x  2Þ MLE 1.46 0.957
h i x 2 ½0; 1 ð10Þ
: yl ¼ yh  0:5 ð6x  2Þ2  6x Manually chosen 1 6.883
62 H. BU et al.

Fig. 1 Comparison of models of MLE and manually selected parameters.

the MLE tends to choose the MF model even if it is inferior to reflects the complexity of the discrepancy model and can be
the HF only model. Park et al.10 pointed out that MLE tries to employed to select suitable q. Here, we use the Integrated
find q that minimizes the bumpiness of the discrepancy func- Mean Square Error (IMSE) to measure the overall estimation
tion which is the combined effect of variation and waviness. uncertainty, as adopted by several studies regarding the
They defined and calculated the bumpiness, variation, and sequential design of Bayesian surrogate models.20,21 The IMSE
waviness numerically. For the Hartmann 6 function that they integrates the posterior variance over the design space and is
studied, the bumpiness correlated with the variation strongly. computationally expensive:
Although the waviness presented an opposite trend, the contri- Z
bution of waviness was overwhelmed by the variation. IMSE ¼ r2d ðxÞdx ð11Þ
x2D
Although this problem of MLE has been noticed by some
researchers, few proposed alternative approaches to select q. where D is the full design space. To obtain a cheap approxima-
Shu et al.18 proposed to minimize the product of jyd j and tion of Eq. (11), we consider the average MSE at all the LF
jhd j, where j  j denotes the L2 norm of a vector, with h obtained sample points:
with MLE. This indicator is also a combination of variation Xnl
(jyd j) and waviness (jhd j). Numerical tests have shown the effec- d ¼1
IMSE r2 ðxi Þ ð12Þ
tiveness of this approach for some problems. However, this nl i¼1 d
approach is designed for the simple discrepancy MFS frame- When selecting q, we minimize the modified IMSE:
work19 and may not be suitable for the Bayesian MFS. In n
o
the next section, we’ll propose an alternative approach for d q; b
q ¼ arg min IMSE h ð13Þ
q
the selection of scale factors in Bayesian MFS.
Note that, in Eq. (13), b
h is determined by minimizing the
4. Proposed approach likelihood Eq. (9) given a specific q, as also adopted in
Ref.18. The full MFS building process using the current
Recall the one-dimensional case in the previous section, and approach is illustrated with a flowchart in Fig. 3. As the pro-
the MLE produces a model of poor accuracy, as shown in posed approach tries to find a q that minimizes the posterior
Fig. 1(a). Besides, the prediction uncertainty of the MLE variance of the model, we denote the proposed approach with
model is remarkably large compared with the manually Minimum Posterior Variance (MPV).
selected model, as shown in Fig. 1. This inspires us to consider Fig. 4 presents the MF model of the Forrester function
whether the prediction uncertainty (the posterior variance) obtained with the proposed approach. As shown in the figure,
Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 63

Fig. 3 MFS building process of current approach.

Fig. 2 Likelihood function and each term’s value.

the model accuracy has been greatly improved compared with

the MLE model in Fig. 1(a). q found by the proposed
approach is 1.08. Fig. 5 compares the MF model’s predicted
d with its true Mean Square Error (MSE) with the change
IMSE
of q. Although the predicted and true error’s magnitudes are Fig. 4 MF model of proposed approach.
different, they share the common minimum point approxi-
mately, near q ¼ 1. This indicates that the modified IMSE is
an effective indicator for the selection of the scale factor. In proposed approach with MLE. Further, two high-
the next section, more numerical examples will be employed dimensional analytical functions are tested to validate the pro-
to illustrate the effectiveness of the proposed approach. posed method. Further, a seven-dimensional turbine blade
design problem is employed to illustrate the applicability of
5. Numerical examples the current approach to real-world design problems. Finally,
a five-dimensional turbine endwall cooling layout design opti-
In this section, a one-dimensional test function with variable mization is carried out to validate the advantage of the current
high-to-low fidelity correlation is employed to compare the approach in engineering optimization.
64 H. BU et al.

Fig. 5 d and true MSE.

Comparison of IMSE Fig. 6 Relation between R2 andA.

To evaluate the accuracy of the obtained multi-fidelity sur-

rogate model, two error metrics are employed: Root Mean R2 gradually decreases to 0.146. Fig. 7 shows the plot of
Square Error (RMSE) and Max Absolute Error (MAE). These high-fidelity function with several low-fidelity functions of dif-
two metrics represent the global and local accuracy of the ferent A.
multi-fidelity surrogate model respectively. Definitions of In this case, nested samples are generated as required by the
RMSE and MAE are given as follows: Bayesian multi-fidelity surrogate. The sample points of high-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi and low-fidelity functions are Sh ¼ f0; 0:4; 0:6; 1:0g
X
m
andSl ¼ f0; 0:1; 0:2; 0:3; 0:4; 0:5; 0:6; 0:7; 0:8; 0:9; 1:0g, which
RMSE ¼ ½yðxi Þ  y^ðxi Þ=m ð14Þ
are the same as those in Ref.11. An adaptive differential evolu-
i¼1
tion algorithm JADE22 is employed to tune the Hyper Param-
MAE ¼ max jyðxi Þ  y^ðxi Þj i ¼ 1; 2;    ; m ð15Þ eters (HP). In the HP tuning process, 1000 evaluations of the
likelihood function are taken to find a more global optimum
where m is the number of evaluation points, yðxi Þ and y^ðxi Þ are value. The scale factor q and h (including hl and hd ) are in
the real function value and model’s prediction value at xi the range of ½0:2; 2:5 and ½0:001; 100 respectively. 100 addi-
respectively. tional samples were uniformly generated to evaluate the error
metrics of the multi-fidelity surrogate.
5.1. One-dimensional example with variable correlation Fig. 8 presents the RMSE and MAE of MF surrogate mod-
els obtained by MLE and the proposed approach (MPV) along
The one-dimensional example with variable high-to-low fide- with the model built with HF samples only, under different
lity level correlation is an alternative version of the Forrester correlations. As the figures depict, the MF model obtained
function taken from Ref.14. The high-fidelity and low-fidelity using MPV is almost always more accurate than MLE.
functions are defined as follows: Besides, when the correlation changes, the model accuracy of
8 the MPV approach only decreases slightly while the MLE
>
> y ¼ ð6x  2Þ2 sin ½2ð6x  2Þ model’s accuracy fluctuates drastically. When A ¼ 1 and the
< h
x 2 ½0; 1; A 2 ½0; 1 correlation coefficient is at its lowest value (0.146), the MPV
>
> h i
: y ¼ y  ðA þ 0:5Þ ð6x  2Þ2  6x approach still generates a MF model with reasonable accuracy
l h
(RMSE  1), while the model produced by MLE is not satis-
ð16Þ
where A changes the form of low-fidelity function and the cor-
responding correlation between high- and low-fidelity func-
tion. The correlation between high- and low-fidelity function
can be measured by the correlation coefficient defined as
follows:
2 32
P


6 m ðiÞ
h yðl iÞ  yl 7
6 i¼1 yh  y 7
R2 ¼ 6 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
6 m
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 ð17Þ
4 P ðiÞ 2 P m
2 7 5
ðiÞ
yh  yh yl  yl
i¼1 i¼1

 
where i is the index of the test points to compute R2 , yh and yl
are the mean value of the HF and LF test points.
Fig. 6 presents the relation between R2 and A. As shown in
Fig. 6, the correlation between high- and low-fidelity function
decreases as A increases. When A ¼ 0, the correlation coeffi- Fig. 7 Plot of HF and several LF functions.
cient R2 ¼ 0:882, which is the highest. As A increases to 1,
Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 65

Fig. 8 Model errors under different correlations. Fig. 9 MF models when A ¼ 1.

factory as the model error (RMSE  5) is approaching the one

obtained with HF samples only. That is to say, when the cor-
relation between high- and low-fidelity function is not ideal,
MLE would be on the brink of failure, generating a MF model
even inferior to the HF only model.
Further, we present a more detailed comparison of the
models of different approaches. Fig. 9 shows the model
obtained from MLE and MPV when A ¼ 1. It could be
observed from the figure that the MPV model produces a finer
approximation of the HF function in the whole range. Besides,
the estimation error is smaller than the MLE model.
Next, the failure of MLE is illustrated quantitatively.
Fig. 10 presents the negative-logarithmic likelihood function
of MF models under different correlations. As the picture
depicts, the likelihoods of MLE models are larger than those
of the MPV models. However, this does not guarantee that Fig. 10 Likelihood values of MF models.
the scale factor found by MLE is the best one in terms of
model accuracy, as shown in Fig. 8. Interestingly, the likeli-
hood function values obtained from the two approaches are
very close when A ¼ 1, while the models are of contrasting mance of MLE and the proposed approach. Besides, a five-
accuracy, as shown in Fig. 9. The scale factor obtained by dimensional turbine endwall cooling layout design optimiza-
MLE and MPV approaches are 0.2 and 1.15 respectively. This tion is carried out to validate the advantage of the current
phenomenon further reveals the incompetence of MLE. approach in engineering optimization. The Hartmann 3 func-
tion and Borehole 8 function are directly taken from previous
5.2. Multi-dimensional examples studies.14,23 For all the problems, the number of HF sample
points ranges from d to 6d and the high-to-low sample ratio
In this section, two classical numerical functions and a turbine ranges from 0.3 to 0.6, where d denotes the problem
blade design problem are employed to compare the perfor- dimension.
66 H. BU et al.

5.2.1. Hartmann 3
The Hartmann 3 function is defined as follows:
8 " #
>
> P4 P 3  2
> y
> h
> ¼  ci exp  a ij x j  pij
>
> i¼1 j¼1
>
>
>
>
>
> yl ¼ yh þ 7:6ð0:585  0:324x1  0:379x2  0:431x3  0:208x1 x2
>
>
>
> 
>
> þ0:326x1 x3 þ 0:193x2 x3 þ 0:225x21 þ 0:263x22 þ 0:274x23
>
>
>
>
>
> 2 3 2 3
>
> 3 10 30 1
>
>
>
> 6 7 6 7
>
> 6 0:1 10 35 7 6 1:2 7
>
<a ¼ 6 7 6 7
6 7; c ¼ 6 7
6 3 10 30 7 6 3 7
> 4 5 4 5
>
>
>
> 0:1 10 35 3:2
>
>
>
>
>
> 2 3
>
> 0:3689 0:1170 0:2673
>
>
>
> 6 7
>
> 6 0:4699 0:4387 0:7470 7
>
> 6 7
>
> p ¼ 6 7
>
> 6 0:1091 0:8732 0:5547 7
>
> 4 5
>
>
>
> 0:03815 0:5743 0:8828
>
>
>
:
xj 2 ½0; 1 j ¼ 1; 2; 3
ð18Þ
A nested sample set with 18 HF samples and 60 LF samples
is generated using Latin Hypercube Sampling (LHS) and Gra-
tiet’s nearest neighbor method. The hyper-parameters q and h
(including hl and hd ) are in the range of ½0:2; 2:5 and
½0:001; 100 respectively. Hyper-parameters are tuned using
the same strategy as in the previous subsection. To eliminate
the impact of random sampling, 100 standalone experiments
Fig. 11 Error statistics of MF models, Hartmann 3.
are conducted. For each experiment, 1000 additional samples
are generated using LHS to evaluate the accuracy of the built
models. A nested sample set containing 24 HF samples and 80 LF
Table 2 gives the median errors of the two approaches. As samples is generated using the same approach as in the Hart-
shown in the table, the MPV approach produces more accu- mann 3 case. The hyper-parameter tuning strategy and model
rate models in terms of both local and global error. Fig. 11 pre- accuracy evaluation method are kept the same as those in the
sents the error statistics of all numerical experiments. As could previous subsection. The experiment is repeated 100 times to
be seen from the figure, the MPV approach produces a statis- obtain the performance statistics.
tically more accurate MF model. The error medians (red line) Table 3 gives the median errors of the two approaches. As
of the MPV approach are all lower than the MLE approach, shown in the table, the MPV approach produces a statistically
and there are fewer extreme cases with poor accuracy gener- more accurate model. The global and local accuracies of MPV
ated by the MPV approach. models are all better than those of MLE models. Fig. 12 pre-
sents the error statistics of all numerical experiments. As
5.2.2. Borehole 8 shown in the figure, the MPV approach is steadier, generating
The Borehole 8 function is defined as follows: fewer extreme cases with poor accuracy.
8
3 ðx4 x6 Þ
>
> yh ¼ ln ðx =x Þ 1þ2x 2px
> 2 1 ½ 7 x4 =ðln ðx2 =x1 Þx1 x8 Þþx3 =x5  5.2.3. Impact of high-to-low sample ratio
2
>
>
>
>
< yl ¼ 0:4yh þ 0:07x1 x8 þ x1 x7 =x3 þ x1 x6 =x2 þ x21 x4
2 Additional experiments are conducted for the Hartmann 3
x1 2 ½0:05; 0:15; x2 2 ½100; 50000; x3 2 ½63070; 115660 function to compare the two approaches’ performance under
>
>
>
> different high-to-low sample ratios. The sampling ratios tested
>
> x 4 2 ½990; 1110; x5 2 ½63:1; 116; x6 2 ½700; 820
>
: here are nh =nl ¼ 0:3; 0:4; 0:6, each employing 60, 45, 30 LF
x7 2 ½1120; 1680; x8 2 ½9855; 12045 samples respectively. Ranges and tuning strategy of hyper-
ð19Þ parameters are kept the same as those in the previous subsec-

Table 2 Error medians of MF models, Hartmann 3. Table 3 Error medians of MF models, Borehole 8.
Model MLE MPV
Model MLE MPV
RMSE 0.91731 0.84574
RMSE 0.34057 0.33543
MAE 6.27474 5.87569
MAE 1.48780 1.42233
Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 67

Table 4 gives the median errors of the two approaches

under different high-to-low sample ratios. As shown in the
table, the MPV approach produces more accurate models in
terms of both local and global error. The RMSE and MAE
of MPV models are lower than MLE models under all the
tested high-to-low sample ratios.
Fig. 13 presents the error statistics of all numerical experi-
ments. As could be seen from the figure, for all the high-to-low
sample ratios, the MPV approach produces a statistically more
accurate MF model. The error medians (red line) of the MPV
approach are all lower than those of the MLE approach, and
there are fewer extreme cases with poor accuracy generated by
the MPV approach.

5.2.4. Turbine blade design problem

To validate the applicability of the proposed approach to real-
world design problems, a turbine blade parametric design
problem is chosen to compare the models built with MLE
and the current approach. The first stage turbine guide vane
of GE’s Energy Efficient Engine24 is selected as the baseline
blade profile to be redesigned. The blade profile is defined with
11 geometric parameters and 7 of them are chosen as design

Table 4 Error medians under different sample ratios, Hart-

mann 3.
Case nh =nl Error metrics MLE MPV
1 0.6 RMSE 0.34057 0.33543
Fig. 12 Error statistics of MF models, Borehole 8. MAE 1.48780 1.42233
2 0.4 RMSE 0.24890 0.21557
tion. For each sample ratio, 100 standalone experiments are MAE 1.03883 1.00745
conducted. For each experiment, 1000 additional samples are 3 0.3 RMSE 0.20563 0.18601
generated using LHS to evaluate the accuracy of the built MAE 0.88381 0.83896
models.

Fig. 13 Error statistics under different sample ratios, Hartmann 3.

68 H. BU et al.

variables. Table 5 gives the description and range of these where Pexit and Pexit denote the static and total pressure at the
variables. cascade exit respectively, Pinlet denotes the total pressure at the
To illustrate the parameterization method clearly, Fig. 14 cascade inlet, and c is the heat capacity ratio. In house soft-
presents the definition and influence of several key parameters. ware is employed to compute the energy loss coefficient.
The axial chord controls the width of the blade in the axial Table 6 gives the boundary condition of the computation.
direction. The connecting angle controls the installation angle As widely adopted in aeronautical multi-fidelity design, the
of the blade. The full blade profile consists of five curves con- high- and low-fidelity aerodynamic data are obtained from com-
nected end to end smoothly with one another, including the putations of fine and coarse meshes. A preliminary mesh depen-
leading- and trailing-edge arcs and the Bézier curves on blade dent study was carried out to choose a suitable mesh resolution.
pressure and suction sides. As shown in Fig. 14(a), the full Fig. 15 gives the computed energy loss coefficient obtained from
blade suction side curve is composed of two Bézier curve seg- meshes of different resolutions. As shown in the figure, the
ments. When the throat size is known, the junction position of energy loss coefficient does not change much when the cell num-
these two segments is controlled by the outlet deflection angle. ber is larger than 1 105 . The coarsest mesh that can produce a
As shown with the red curve in Fig. 14(a), the shape of each similar result compared with the fine mesh is of 1:214 104 cells.
Bézier curve segment is controlled with two points (B, D) con- Hence, the meshes of 1:007 105 and 1:214 104 cells are cho-
strained within the tangential direction. The positions of these sen for the computations of high and low fidelity. Fig. 16 pre-
two points are defined with two control coefficients. In this sents a detailed view of the employed meshes, with every
problem, the control coefficients of the suction side inlet Bézier fourth grid point displayed for clarity.
segment are modified to have more control of this critical For this problem, 21 HF samples and 42 LF samples were
region. Fig. 14(b) shows the influence of three variables includ- generated to build the MF model using the same method as in
ing axial chord, deflection angle, and control coefficient 1 with the previous sections. To validate the built models, 21 addi-
each variable perturbed 0:25 times the deviation range. tional HF samples were generated. The normalized sample
To measure the aerodynamic performance of the designed points and observations are listed in the Appendix B (see
blade profile, an energy loss coefficient is defined as follows: Tables B1 B3). The hyper-parameter tuning method is kept
"  c1 # "  c1 #
Pexit c Pexit c the same with that in previous subsections. Fig. 17 presents
Cpt ¼ 1  1  = 1 ð20Þ the Cpt predictions of the MF models against the validation
Pexit Pinlet
data, where Cpt and C^pt are the real and predicted values
respectively. As shown in the figure, the MPV approach
Table 5 Input variables of turbine blade design problem.
No. Description Baseline Range
1 Axial chord (mm) 33.85 [31.85, 35.85] Table 6 Boundary conditions of computation.
2 Connecting angle (°) 59.8 [55.8, 61.8] Parameter Value
3 Inlet up deviation angle (°) 69 [54, 72]
Inlet total temperature (K) 982
4 Outlet deflect angle (°) 4.5 [3, 9]
Inlet total pressure (kPa) 344.74
5 Relation coefficient 0.35 [0.30, 0.45]
Inlet flow angle (°) 90
6 Bézier control coefficient 1 0.4 [0.25, 0.55]
Mach number at cascade exit 0.878
7 Bézier control coefficient 2 0.54 [0.39, 0.69]

Fig. 14 Blade profile parameterization method.

Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 69

Fig. 15 Mesh dependence study.

Fig. 17 Comparison of MLE and MPV models for prediction of
Cpt .

Table 7 Error of MF models, turbine blade design problem.

Model MLE MPV
3
RMSE (10 ) 1.4275 1.0077
MAE (103) 2.6543 2.6175

our previous study25 to compare the efficiency of design opti-

mization using MLE and the current approach. The cooling
layout of the turbine endwall consists of two rows of film cool-
ing holes, namely the circumferential row and the axial row, as
shown in Fig. 18. The positions of these two rows of holes are
controlled by five parameters: vc1 , vc5 , va1 , va5 , ua5 . For detailed
explanation of each parameter’s meaning, please refer to our
previous paper.25 Table 8 gives the ranges of these five design
parameters. The objective of design optimization is to mini-
mize the high-temperature area on the turbine endwall Ah .
In the multi-fidelity design optimization, the high- and low-
fidelity data were obtained from CFD simulations of fine and
coarse meshes. Fig. 19 presents the employed meshes of the
high- and low-fidelity simulations. For the details of the simu-
lation including the boundary conditions and validation of tur-
bulence model, please refer to our previous paper.25
In the multi-fidelity design optimization, 23 LF samples
and 7 HF samples were generated using LHS method to build
the initial model. Then optimizations were carried out based
on the sequential sampling using the criterion of maximized
expected improvement16:

Fig. 16 Meshes for high- and low-fidelity computations with

every fourth grid point displayed.

produced a model slightly more accurately than MLE. Table 7

compares the prediction errors of the MF models. The model
obtained with the MPV approach is better than MLE in terms
of both global and local accuracy.

5.2.5. Turbine endwall cooling layout design problem

To further validate the effectiveness of the proposed approach, Fig. 18 Parametric design method of turbine endwall cooling
a turbine endwall cooling layout design problem is taken from layout.
70 H. BU et al.

Besides, the proposed approach is more robust than the MLE

Table 8 Ranges of design parameters.
approach with a narrower deviation band. These superior per-
Variable Range formances can be owning to the more accurate selection of q,
vc1 [0.05,0.30] which is most evident when there are fewer samples (Steps 5–
vc5 [0.70,0.93] 20) and the estimation of q is very challenging.
va1 [0.05,0.93]
va5 [0.05,0.93] 6. Conclusions
ua5 [0.55,0.9]

In this study, a detailed numerical analysis was conducted to

explain why MLE failed to choose the best scale factor of
Bayesian multi-fidelity surrogate. It is shown that MLE may
overly emphasize the discrepancy function’s variation and
ignore its waviness.
To overcome this drawback of MLE, an alternative
approach is proposed, which selects the scale factor by mini-
mizing the posterior variance of the surrogate model. The pro-
posed method has been validated with several numerical
examples, a turbine blade design problem and a turbine end-
wall cooling layout design optimization problem. For the
one-dimensional test function with variable correlation
between fidelity levels, the proposed approach produced a
more reasonable scale factor and a corresponding more accu-
rate surrogate model. Besides, the proposed approach is less
sensitive to the correlation between fidelity levels, producing
better approximation of the real function in a wide range of
correlation. Numerical tests conducted on high-dimensional
functions also present the potential of the current approach
in accurate surrogate modeling. For the classic three-
dimensional Hartmann function and eight-dimensional Bore-
hole function, the proposed approach produced comparable
Fig. 19 Meshes for high- and low-fidelity computations. or better model accuracy at different high-to-low sample
ratios. Applicability of the current method to real-world design
was also confirmed with better prediction in a turbine blade
design problem and superior efficiency of design optimization
in a turbine endwall cooling design problem.
Further research is needed to investigate the applied range
of the proposed approach and the impact of high-to-low fide-
lity level correlation on high dimensional problem modeling.

Declaration of Competing Interest

The authors declare that they have no known competing

financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgements
Fig. 20 Convergence histories of MLE and proposed approach.
The authors would like to acknowledge the financial support
from the National Science and Technology Major Project,
Z 1 h i
^  ^
^ China (No. 2019-II-0008-0028) and Key Program of National
E½IðxÞ ¼ max min ðyh Þ  yh ðxÞ; 0 / yh ðxÞ dyh ðxÞ Natural Science Foundation of China (No. 51936008).
1
ð21Þ
Appendix A. Detailed formulation of multi-fidelity predictor
where / is the probability density function of the normal dis-
tribution. To reduce the influence of the initial sampling on the The detailed formulation of quantities in Eqs. (5) and (2) are
performance of optimization, 10 standalone sampling and given as follows:
optimization experiments were carried out. Fig. 20 presents 8
  
the convergence histories of the MLE and the proposed < f 0 ðxÞT ¼ qfðxÞT ; fðxÞT ; b0 ¼ bTl ; bTd
approach (MPV), with the mean and standard deviation of : yT ¼ yT ; yT ; tðxÞT ¼ qr2 R ðx; X Þ; q2 r2 R ðx; X Þ þ r2 R ðx; X Þ
ten experiments plotted. As shown clearly in the figure, the S l h l l l l l h d d h

proposed approach converged faster than the MLE approach. ðA1Þ

Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 71
8 0
T 1 8

0 0 T 0
>
>
> B f xl
ð1Þ
0 C < Rl ðx; x Þ ¼ exp ðx  x Þ diagðhl Þðx  x Þ
>
>
>
>
> B
 C
 ðA3Þ
>
> B T C >
: R ðx; x0 Þ ¼ exp ðx  x0 ÞT diagðh Þðx  x0 Þ
>
> B f xðl 2Þ 0 C
>
> B C d d
>
> B C
>
>
>
B
B .. .. C
C 8

>
> B
.  . C bl ¼ fðXl ÞT R1 T 1
l fðXl Þ fðXl Þ Rl yl
>
> B C >
>
>
> B f xðnl Þ
T
C >
>
>
> B 0 C >
>
>
>
l >
>

>
< F¼B
B
ð1Þ T
T C C >
>
< bd ¼ fðXd Þ R1
> T T 1
B qf x f xh
ð1Þ C d fðXd Þ fðXd Þ Rd yd
B h C ðA2Þ ðA4Þ
>
> B

T C >
>
> B T C >
>
> B qf xð2Þ f xh
ð2Þ C > T 1
> > rl ¼ nl ðyl  fðXl Þbl Þ Rl ðyl  fðXl Þbl Þ
2 1
>
> B h C >
>
> B C >
>
>
> B .. .. C >
>
>
> B C > 2
:
>
> B . . C rd ¼ n1h ðyd  fðXd Þbd ÞT R1
d ðyd  fðXd Þbd Þ
>
> @
T
T A
>
> ðn Þ ðnh Þ
> qf xh h f xh
>
>
> where Xh and Xl are the full HF and LF design matrix
>
>  
>
>
> r2l Rl ðXl Þ qr2l Rl ðXl ; Xh Þ respectively.
:C ¼
qrl Rl ðXh ; Xl Þ qrl Rl ðXh Þ þ r2d Rd ðXh Þ
2 2
Appendix B Samples and validation data of engineering problem

Table B1 High-fidelity samples and observations.

No. x1 x2 x3 x4 x5 x6 x7 Observation
1 0.155903 0.432802 0.892364 0.665756 0.474023 0.346397 0.996273 0.0313
2 0.859857 0.201156 0.078528 0.537074 0.439040 0.430184 0.738785 0.0296
3 0.830629 0.016821 0.736668 0.390675 0.214375 0.533410 0.956335 0.0325
4 0.447376 0.940508 0.237033 0.589741 0.186224 0.293832 0.638848 0.0318
5 0.848328 0.587044 0.024788 0.462394 0.205292 0.168899 0.021912 0.0295
6 0.415288 0.039277 0.335105 0.944101 0.350889 0.693403 0.395542 0.0285
7 0.097153 0.920154 0.298569 0.861428 0.867762 0.815784 0.322596 0.0319
8 0.340715 0.963652 0.814918 0.278368 0.545562 0.011389 0.441503 0.0346
9 0.083899 0.380879 0.158022 0.837456 0.790812 0.723793 0.811957 0.0288
10 0.782664 0.068037 0.935661 0.140153 0.253581 0.779255 0.090008 0.0359
11 0.492997 0.991521 0.050902 0.885275 0.966277 0.132381 0.156874 0.0323
12 0.631120 0.530650 0.544891 0.447860 0.997185 0.500922 0.897187 0.0352
13 0.457329 0.140445 0.249948 0.105041 0.832526 0.932634 0.264618 0.0408
14 0.950309 0.811968 0.142117 0.059043 0.092998 0.051286 0.871325 0.0377
15 0.560670 0.777656 0.694228 0.966118 0.304568 0.856461 0.481774 0.0303
16 0.231035 0.835749 0.856084 0.143756 0.570140 0.913020 0.916746 0.0396
17 0.201421 0.477097 0.875669 0.195474 0.014324 0.112057 0.783141 0.0327
18 0.903631 0.878909 0.685615 0.026345 0.487207 0.232823 0.108714 0.0387
19 0.617432 0.083308 0.906580 0.489151 0.142941 0.213074 0.030675 0.0294
20 0.647982 0.104777 0.963874 0.978186 0.900958 0.383566 0.646650 0.0304
21 0.958130 0.460714 0.113077 0.079200 0.038858 0.883094 0.464165 0.0356

Table B2 Low-fidelity samples and observations.

No. x1 x2 x3 x4 x5 x6 x7 Observation
1 0.036853 0.340424 0.441353 0.254437 0.125312 0.283430 0.800666 0.0380
2 0.155903 0.432802 0.892364 0.665756 0.474023 0.346397 0.996273 0.0364
3 0.859857 0.201156 0.078528 0.537074 0.439040 0.430184 0.738785 0.0361
4 0.830629 0.016821 0.736668 0.390675 0.214375 0.533410 0.956335 0.0391
5 0.294303 0.181500 0.509969 0.762136 0.375840 0.418035 0.683124 0.0343
6 0.447376 0.940508 0.237033 0.589741 0.186224 0.293832 0.638848 0.0362
7 0.189751 0.615742 0.318362 0.172079 0.273317 0.636786 0.712555 0.0396
8 0.848328 0.587044 0.024788 0.462394 0.205292 0.168899 0.021912 0.0356
9 0.415288 0.039277 0.335105 0.944101 0.350889 0.693403 0.395542 0.0340
10 0.097153 0.920154 0.298569 0.861428 0.867762 0.815784 0.322596 0.0364
11 0.268304 0.396340 0.614467 0.318359 0.951001 0.670207 0.191118 0.0385
(continued on next page)
72 H. BU et al.

Table B2 (continued)
No. x1 x2 x3 x4 x5 x6 x7 Observation
12 0.594384 0.230590 0.748617 0.910226 0.771823 0.604569 0.065129 0.0345
13 0.401252 0.310364 0.186475 0.505991 0.665890 0.323853 0.526711 0.0362
14 0.313049 0.428469 0.998402 0.004000 0.670894 0.793081 0.732440 0.0489
15 0.921281 0.550723 0.479898 0.616961 0.855798 0.379506 0.250072 0.0362
16 0.258910 0.502983 0.200647 0.634875 0.600816 0.973113 0.340674 0.0360
17 0.977338 0.647921 0.425619 0.821700 0.330635 0.082301 0.229683 0.0348
18 0.340715 0.963652 0.814918 0.278368 0.545562 0.011389 0.441503 0.0385
19 0.083899 0.380879 0.158022 0.837456 0.790812 0.723793 0.811957 0.0360
20 0.782664 0.068037 0.935661 0.140153 0.253581 0.779255 0.090008 0.0411
21 0.492997 0.991521 0.050902 0.885275 0.966277 0.132381 0.156874 0.0374
22 0.631120 0.530650 0.544891 0.447860 0.997185 0.500922 0.897187 0.0391
23 0.699195 0.788697 0.283866 0.758497 0.753325 0.482189 0.121797 0.0356
24 0.543050 0.740033 0.021675 0.792794 0.572767 0.750337 0.581835 0.0357
25 0.126962 0.687085 0.658533 0.228745 0.111651 0.644878 0.940834 0.0399
26 0.683685 0.722353 0.785529 0.569685 0.924249 0.036797 0.408021 0.0374
27 0.721172 0.300778 0.375823 0.418744 0.627234 0.162397 0.522394 0.0363
28 0.457329 0.140445 0.249948 0.105041 0.832526 0.932634 0.264618 0.0459
29 0.011430 0.254736 0.805589 0.363539 0.734596 0.259428 0.548784 0.0383
30 0.511819 0.161992 0.565683 0.691572 0.520795 0.986560 0.288417 0.0353
31 0.950309 0.811968 0.142117 0.059043 0.092998 0.051286 0.871325 0.0413
32 0.560670 0.777656 0.694228 0.966118 0.304568 0.856461 0.481774 0.0352
33 0.231035 0.835749 0.856084 0.143756 0.570140 0.913020 0.916746 0.0430
34 0.059161 0.896607 0.641268 0.720989 0.423540 0.574063 0.185759 0.0355
35 0.787832 0.635058 0.382324 0.344716 0.402545 0.865873 0.839966 0.0395
36 0.201421 0.477097 0.875669 0.195474 0.014324 0.112057 0.783141 0.0393
37 0.365782 0.712652 0.589246 0.296842 0.069858 0.554533 0.369243 0.0374
38 0.903631 0.878909 0.685615 0.026345 0.487207 0.232823 0.108714 0.0427
39 0.751070 0.266025 0.460175 0.673440 0.703524 0.469243 0.604342 0.0360
40 0.617432 0.083308 0.906580 0.489151 0.142941 0.213074 0.030675 0.0358
41 0.647982 0.104777 0.963874 0.978186 0.900958 0.383566 0.646650 0.0365
42 0.958130 0.460714 0.113077 0.079200 0.038858 0.883094 0.464165 0.0406

Table B3 Validation samples and observations.

No. x1 x2 x3 x4 x5 x6 x7 Observation
1 0.052534 0.557459 0.212479 0.757389 0.023063 0.708906 0.898104 0.0282
2 0.159547 0.890069 0.540561 0.689242 0.819807 0.954164 0.461925 0.0339
3 0.940292 0.075782 0.755374 0.317860 0.767540 0.630396 0.007049 0.0333
4 0.824153 0.703713 0.447558 0.910543 0.037599 0.753329 0.113684 0.0302
5 0.879585 0.971695 0.274869 0.884326 0.190084 0.764879 0.946776 0.0317
6 0.976409 0.825991 0.600872 0.954607 0.907681 0.492076 0.561282 0.0316
7 0.408760 0.249364 0.172223 0.784206 0.942800 0.401834 0.129667 0.0293
8 0.001992 0.988076 0.372022 0.491501 0.131700 0.896042 0.479294 0.0327
9 0.628683 0.864595 0.072393 0.713781 0.466903 0.798870 0.390684 0.0317
10 0.491455 0.924230 0.502678 0.162943 0.304858 0.417637 0.802124 0.0375
11 0.274847 0.547439 0.909339 0.946083 0.413239 0.986028 0.763389 0.0309
12 0.680529 0.064381 0.109480 0.792202 0.364850 0.950837 0.714598 0.0290
13 0.234348 0.671675 0.594016 0.424484 0.956144 0.173571 0.962291 0.0343
14 0.039661 0.310087 0.764266 0.076567 0.314211 0.103304 0.624748 0.0366
15 0.567107 0.768928 0.956444 0.645349 0.199017 0.087077 0.365109 0.0301
16 0.925833 0.149964 0.556671 0.041419 0.228461 0.210105 0.994103 0.0411
17 0.329567 0.491579 0.038350 0.850830 0.092574 0.502521 0.180051 0.0281
18 0.077049 0.941455 0.388968 0.465193 0.858964 0.006023 0.286827 0.0337
19 0.290184 0.507562 0.792422 0.116260 0.998268 0.721498 0.680380 0.0401
20 0.589969 0.139436 0.994966 0.563692 0.282809 0.365672 0.277157 0.0295
21 0.964820 0.612431 0.334225 0.253352 0.069851 0.147880 0.221952 0.0323
Selecting scale factor of Bayesian multi-fidelity surrogate by minimizing posterior variance 73

References 14. Zhou Q, Wu YD, Guo ZD, et al. A generalized hierarchical co-
Kriging model for multi-fidelity data fusion. Struct Multidiscip
1. Tyacke J, Vadlamani NR, Trojak W, et al. Turbomachinery Optim 2020;62(4):1885–904.
simulation challenges and the future. Prog Aerosp Sci 15. Shi RH, Liu L, Long T, et al. Multi-fidelity modeling and adaptive
2019;110:100554. co-Kriging-based optimization for all-electric geostationary orbit
2. Tucker PG, Wang ZN. Eddy resolving strategies in turbomachin- satellite systems. J Mech Des 2020;142(2) 021404.
ery and peripheral components. J Turbomach 2021;143(1) 010801. 16. Forrester AIJ, Sóbester A, Keane AJ. Multi-fidelity optimization
3. Queipo NV, Haftka RT, Shyy W, et al. Surrogate-based analysis via surrogate modelling. Proc R Soc A 2007;463(2088):3251–69.
and optimization. Prog Aerosp Sci 2005;41(1):1–28. 17. Guo ZD, Song LM, Park C, et al. Analysis of dataset selection for
4. Oakley JE, O’Hagan A. Probabilistic sensitivity analysis of multi-fidelity surrogates for a turbine problem. Struct Multidiscip
complex models: A Bayesian approach. J Royal Stat Soc Ser B Optim 2018;57(6):2127–42.
Stat Methodol 2004;66(3):751–69. 18. Shu LS, Jiang P, Song XG, et al. Novel approach for selecting
5. Bu HY, Guo ZD, Song LM, et al. Effects of cooling configura- low-fidelity scale factor in multifidelity metamodeling. AIAA J
tions on the aerothermal performance of a turbine endwall with jet 2019;57(12):5320–30.
impingement and film cooling. J Turbomach 2021;143(6) 061013. 19. Park C, Haftka RT, Kim NH. Remarks on multi-fidelity
6. Dyn N, Levin D, Rippa S. Numerical procedures for surface surrogates. Struct Multidiscip Optim 2017;55(3):1029–50.
fitting of scattered data by radial functions. SIAM J Sci And Stat 20. Bates RA, Buck RJ, Riccomagno E, et al. Experimental design
Comput 1986;7(2):639–59. and observation for large systems. J Royal Stat Soc Ser B
7. Clarke SM, Griebsch JH, Simpson TW. Analysis of support vector Methodol 1996;58(1):77–94.
regression for approximation of complex engineering analyses. J 21. Picheny V, Ginsbourger D, Roustant O, et al. Adaptive designs of
Mech Des 2005;127(6):1077–87. experiments for accurate approximation of a target region. J Mech
8. Jones DR, Schonlau M, Welch WJ. Efficient global optimization Des 2010;132(7) 071008.
of expensive black-box functions. J Global Optim 1998;13 22. Zhang JQ, Sanderson AC. JADE: Adaptive differential evolution
(4):455–92. with optional external archive. IEEE Trans Evol Comput 2009;13
9. Kennedy M, O’Hagan A. Predicting the output from a complex (5):945–58.
computer code when fast approximations are available. Biome- 23. Ruan XF, Jiang P, Zhou Q, et al. Variable-fidelity probability of
trika 2000;87(1):1–13. improvement method for efficient global optimization of expensive
10. Park C, Haftka RT, Kim NH. Low-fidelity scale factor improves black-box problems. Struct Multidiscip Optim 2020;62(6):3021–52.
Bayesian multi-fidelity prediction by reducing bumpiness of 24. Timko LP. Energy Efficient Engine high pressure turbine compo-
discrepancy function. Struct Multidiscip Optim 2018;58 nent test performance report. Washington, D.C.: NASA; 1984.
(2):399–414. Report No.: NASA CR-168289.
11. Han ZH, Görtz S. Hierarchical Kriging model for variable-fidelity 25. Bu HY, Yang YF, Song LM, et al. Improving the film cooling
surrogate modeling. AIAA J 2012;50(9):1885–96. performance of a turbine endwall with multi-fidelity modeling
12. Zhang YM, Kim NH, Park C, et al. Multifidelity surrogate based considering conjugate heat transfer. J Turbomach 2022;144(1)
on single linear regression. AIAA J 2018;56(12):4944–52. 011011.
13. Hao P, Feng S, Li YW, et al. Adaptive infill sampling criterion for
multi-fidelity gradient-enhanced Kriging model. Struct Multidiscip
Optim 2020;62(1):353–73.