Evidence-Based Multidisciplinary Design Optimization with the Active Global Kriging Model

State Key Laboratory of Mechanics and Control of Mechanical Structures, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China College of Aerospace Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China State Key Laboratory for Strength and Vibration of Mechanical Structures, Xi’an Jiaotong University, Xi’an 710049, China Department of Engineering Mechanics, Northwestern Polytechnincal University, Xi’an 710072, China Wuxi Hengding Supercomputing Center Ltd., Wuxi 214135, China


Introduction
e performance of structures in engineering system is often in uenced by uncertainties. e optimal solution obtained by conventional deterministic design frequently falls on the boundary of the feasible region [1]. e deterministic design may fall into the failure region if the system is disturbed by uncertainties. In the multidisciplinary coupling systems, the propagation of uncertainty factors between di erent subjects makes the search for an optimal design more complicated as well as di cult than the search for an optimal individual discipline design due to the coupling e ect [2,3]. e inuence of uncertainty factors must be considered to ensure the reliability of the optimized design results, and reliabilitybased multidisciplinary design optimization (RBMDO) needs to be performed [4,5].
Uncertainties can be categorized as aleatory and epistemic [6,7]. Aleatory or objective uncertainties arise from the inherent randomness of a system. Probability theory can be adopted to handle random uncertainties. However, su cient information is required to construct the distribution function of uncertain variables. Epistemic or subjective uncertainties stem from the lack of su cient information during modeling and optimization. Several theories, including probability box models, fuzzy sets, Bayesian approaches, and evidence theory, have been developed to deal with epistemic uncertainties [8][9][10][11][12][13][14][15]. Evidence theory provides a general modeling of epistemic uncertainty and can be reduced to other theories. When the interval of evidence theory is in nite, it is equivalent to traditional probability theory. When no con ict exists between uncertain pieces of information, the theory is equivalent to possibility theory. When the subinterval of the evidence variable is unique, the theory degenerates into convex model theory.
Several methods have been proposed for reliability analyses based on evidence theory, and these approaches include Cartesian product method (CPM) [16], Monte Carlo simulation (MCS) [17][18][19][20], polynomial chaos expansion [21], and surrogate-based models [22][23][24]. Design optimization based on evidence theory has also been studied [25][26][27][28]. Zhang et al. [29] adopted two-stage framework to handle the evidence-based design optimization problem, where the evidence variables were transformed into random variables and the sequential optimization and reliability assessment were used. He and Qu [30] gave an overview of possibility and evidence theory-based design optimization. However, limited research has been conducted on evidence-based multidisciplinary design optimization (EBMDO) [31][32][33]. Conventional EBMDO has a three-level loop architecture. e outer loop is multidisciplinary optimization, the middle loop is reliability analysis based on evidence theory, and the inner loop is discipline analysis. e computational cost of EBMDO is high, especially in the inner discipline analysis. Practical engineering problems require repeatedly calling the simulation model to evaluate limit state and objective functions, and this process is time consuming.
To address this issue, this study proposes a surrogate model-based method with an active global learning strategy, which constructs surrogate models in entire design space with intelligent sampling approach, to reduce the amount of computation [34]. e active learning Kriging [35] model is utilized as a surrogate model to evaluate the limit state function in evidence-based reliability analysis. Since the original active Kriging model is utilized for reliability analysis, a model in local region, here the Kriging model is constructed in the entire design space called active global Kriging model. en, two multidisciplinary integration frameworks, namely, Multidisciplinary Feasible (MDF) and Collaborative Optimization (CO), are adopted and combined with the Kriging model. Using the global Kriging model, the computational cost of EBMDO is reduced to the approximate equivalent in deterministic multidisciplinary design optimization (MDO). e rest of the article is organized as follows. MDF and CO strategies for EBMDO are briefly introduced in Section 2. An evidence-based reliability analysis using the global kriging model for MDO is presented in detail in Section 3. Numerical examples are provided and discussed in Section 4, and the conclusions are presented in Section 5.

EBMDO.
e general EBMDO model is expressed as where d pertains to the deterministic variables, d s pertains to shared design variables that are common among all disciplines, d i pertains to the design variables of the i th discipline, X pertains to the uncertain variables, μ X denotes the mean of the X, which is treated as the uncertain design variables, X s pertains to uncertain shared evidence variables that are the common input variables for all disciplines, and X i pertains to the local evidence variables of the i th discipline. e shared variable X s and the input variable X are independent variables, K encompasses coupling variables, and f is the objective function. Pl[·] is the upper bound of failure probability which represents the worst-case and is called is the failure probability constraint of the i th discipline, and G i (·) is the limit state of the i th discipline. e failure model is defined by G i (·) < 0. β is the reliability index, and Φ(·) denotes the cumulative distribution function (CDF) of the standard normal distribution. e coupling variables between design and uncertain variables are expressed as where the first subscript j and the second subscript i of K ji denote the values from the j th discipline to the i th discipline. e meaning of K ji can be explained in detail that the coupling variables are inner shared variables in the j th discipline and the i th discipline, and K ji is the output values of the j th discipline and the input values for the i th discipline. For example, in fluid structure thermal coupling analysis, the pressure values from fluid analysis should be transferred to the structure discipline for finite element analysis. e conventional structure of EBMDO contains a triple loop, as shown in Figure 1 [36,37].
As shown in Figure 1, due to the nested architecture, failure plausibility must be calculated at every design point during optimization. erefore, the multidisciplinary analysis is repeatedly called to evaluate the limit state function. Furthermore, according to evidence theory, the optimization problem should be implemented to determine the upper bound of the limit state.
is nested structure makes EBMDO costly.

Multidisciplinary Feasible Method.
e framework of MDO aims to manage coupled variables and subsystems and improve calculation efficiency. e MDO framework has two main forms, namely, single-level and multilevel structures.
MDF is a traditional single-level optimization method. A systematic analysis is performed for each iteration during optimization, and a multidisciplinary analysis is performed repeatedly to obtain consistent solutions of the coupling variables among disciplines.
is optimization framework includes a system of integrated and subsystem analysis modules. Only one interface exists for the input and output between the system module and subsystem modules, as shown in Figure 2. e system module is an optimizer to search for the optimal solution. e subsystem modules are for multidisciplinary and reliability analyses. In each optimization iteration, the coupling variables should be consistent and the constraints should be also satisfied. erefore, when the optimal solution is obtained, the consistency constraints of coupling variables and reliability constraints are satisfied.

Collaborative Optimization
Method. CO is a bilevel framework of MDO. e framework of MDO is designed to 2 Complexity be top level (system level) and low level (discipline level) by CO. e top level is a system optimizer responsible for assigning target values for system-level state variables to disciplines. e low level is a parallel, distributed, multiplediscipline subsystem. In addition to satisfying the constraints of the subsystems, the objective function is the smallest difference between the coupling state variables of the top-level and low-level systems. After low-level system optimization, the objective function is fed back to the toplevel system, which constitutes the top-level consistency constraint. Optimization of the top-level system is performed to address the uncoordinated state variables among the subsystems. e framework for CO is defined as follows: where f is the objective function of system-level problems, d s_sl denotes the shared deterministic design variables of system-level problems, d sl denotes the deterministic design variables of discipline-level problems, X s_sl denotes the shared uncertain design variables of system-level problems, X sl denotes the uncertain design variables of the discipline-   Complexity 3 level problems, P sl denotes the global uncertain parameters, and J * is the constraint of the discipline-level problems. e i th discipline subproblem is where J i is the objective of the discipline-level problems, d sl denotes the shared design variables, d i denotes the design variables of the i th discipline, si denotes the shared uncertain design variables, and X i denotes the uncertain design variables of the i th discipline. e diagram of CO for EBMDO is shown in Figure 3.

Reliability Analysis Based on Evidence eory. e limit state function is denoted as
In the evidence analysis, the focal element is usually an interval. m Ej is the basic probability assignment (BPA) of each focal element. e joint focal element (JFE) is defined by CPM: e BPA associated with JFE is calculated by where n X is the number of JFEs. In the evidence space, failure probability P f is an interval. e lower and upper bounds of the interval are called belief (Bel) and plausibility (Pl), respectively. Bel means that the system is in the absolute failure state, and Pl means that the system is in the possible failure state. e belief and plausibility measures of failure probability are calculated by where I(·) is the indicator function.
e equivalent form of equations (7) and (8) are is method involves a combinatorial explosion problem because it needs to explore the extrema of G(x) in n E JFEs. According to equations (7) and (8), if n 1 variables and n 2 focal elements exist for each variable, then the total number of JFEs will be n n 1 2 .
e Monte Carlo method is adopted to randomly sample the focal elements and overcome the obstacle of dimensions. JFEs are formed by these focal elements. An extremum analysis is performed based on JFEs, and Bel and Pl can be estimated using statistical methods, shown as where N is the number of samples for the Monte Carlo simulation (MCS).

Evidence Analysis with the Kriging Model.
Although the dimensionality challenge is addressed, the computational cost remains large for MCS. To further improve the computational efficiency, the Kriging model is proposed to evaluate the limit state function in a multidisciplinary coupling system. e Kriging model is briefly introduced as follows.
Suppose that a set of N samples called design of experiments (DoE) has the form (x, G), where x is an n-dimensional vector [x (1) , x (2) , . . . , x (n) ] T and G is the corresponding output [G(x (1) ), G(x (2) ), . . . , G(x (n) )] T . e predicted value G(x) and variance s 2 (x) are given as where 1 is an n-dimensional unit vector, r is the vector of the correlation function of x and points in DoE, R is the matrix of the correlation function of points in DoE, and β, σ 2 are the parameters to be determined. e methods of the fitting parameters are adopted as modified DIRECT algorithm [38]. Further details about the Kriging model can be found in Ref. [39].
For evidence-based design optimization, the Kriging model is used to approximate the limit state function in the entire design space. e idea here is that if the sign given by the Kriging model is the same to that of the true limit state function, then the result of MCS combined with the Kriging model can be considered trustworthy. e following 4 Complexity property can ensure that the sign of the extremum value by the Kriging model is consistent with that of the true limit state function.
Proof. Consider the case of the minimum. Suppose X * min is the minimum point of G(X), and X * min is the minimum point of G(X), namely, G(X * min ) � min(G(X)) and G(X * min ) � min((X)). We will discuss in two cases.
Case 2. If G(X * min ) > 0 because G(X * min ) � min(G(X)), then 0 < G(X * min ) ≤ G(X * min ) because sign (G(X)) � sign (G(X)), so G(X * min ) > 0. erefore, when the limit state function has a minimum value, the sign by the Kriging model and the true limit state function are the same. e case of the maximum value can be proven in a similar way. Suppose that the Kriging model has already been constructed to correctly estimate the sign of the limit state function, equations (11) and (12) can be written as To reduce the sample size, we use the expected risk function (ERF) [23,39] as learning function to adaptively select the samples.
where ϕ(·) and Φ(·) are the probability density function and CDF of standard normal distribution, respectively. e convergence criterion for the Kriging model is written as where X * is the value for maximum of ERF, X is the mean of the X, and ε is 10 − 6 .

Validation of Reliability Analysis with the Evidence
Variables. e following numerical example is adopted to validate the proposed reliability analysis method. e limit state function is written as [13] where α is a deterministic design variable and x 1 and x 2 are evidence variables. When g < 0, the system is in the failure region. e BPA structures of x 1 and x 2 are listed in Table 1. ere are 13 points selected to construct the Kriging model. en, 10 5 Monte Carlo simulation is performed to calculate the Pl of the failure probability. e results of the reliability analysis are shown in Table 2. As can be seen, the result is close to the result from literature [13], which means the proposed evidence analysis method has a high accuracy.

Summary of EBMDO with Kriging-Based Reliability
Analysis. First, a large population S that contains n MC points is generated. Points are randomly selected from population S. e limit state function is evaluated based on these points, thus forming the initial DoE, S DoE , for the Kriging model. e Kriging model is then constructed, and the values of ERF in population S are calculated. e point with the largest value of ERF is selected and added to S DoE . e Kriging model is updated by S DoE until convergence is achieved.
Second, a reliability constraint is applied during optimization by the Kriging model. e framework of MDF and CO is adopted to achieve an optimal design. e procedure of EBMDO is summarized as follows: Step 1. Generate n MC points S according to the range of the variables Step 2. Randomly select n 0 points from S, evaluate the limit state function, and form the initial S DoE

Step 3. Construct the Kriging model based on S DoE
Step 4. Calculate the ERF on each point in S, and pick out point x * with the maximum value of ERF Step 5. Evaluate the limit state function on the point x * , and add the sample in S DoE Step 6. If the Kriging model is converged, go to next step; otherwise, go to Step (3) Step 7. Set the initial value of the optimization Step 8. At each optimal point during optimization, evaluate the reliability of the system by the Kriging model Step 9. If the reliability constraint is satisfied and the objective function is converged, stop; otherwise, go to Step (8) During Kriging modeling, MDA should be performed to evaluate the limit state function. During optimization, MDA is not required for reliability analysis because the Kriging model is constructed to approximate the limit state function in the entire design space. erefore, computational efficiency is considerably improved.

Application Examples
e proposed methodology is validated with two examples of EBMDO. For the inequality constraints of failure probability, evidence theory is used to estimate the epistemic uncertainty in terms of failure plausibility.

Case 1.
e first EMBDO problem is a simple mathematical example. e problem is defined as where d 1 , d 2 are the design variables and x 1 , x 2 are the evidence variables, and x i � d + Δx i . e system utilizes two design variables and outputs two states. ey are not shared variables, thus the systems are uncoupled. e objective function is nonlinear with two constraints. P f is the constraint of failure probability. e BPA structure of the uncertain variables is shown in Table 3, and the corresponding bar chart of epistemic variables is illustrated in Figure 4. e CO architecture of this problem can be written as one system optimization and two discipline optimizations.
e system optimization is written as where d 1sl , d 2sl are the system-level design variables, d ij denotes the i th variables in the j th discipline, and J 1 , J 2 are two auxiliary variables and the objective function of the discipline optimization problem. e auxiliary variables aim to reduce the difference of the system and discipline variables to 10 − 5 . e optimization problem of discipline 1 is defined as e optimization problem of discipline 2 is defined as e initial value of the design variables are set to [1.6, 0.8], which is the result of deterministic optimization (DO). To verify the accuracy of the proposed approach, MCS combined with the MDF architecture is used to directly solve EBMDO. e nonlinear sequential quadratic programming (NLPQL) is adopted as the optimization algorithm for both system and discipline optimization. Firstly, the 10 5 candidate points are generated by Latin hypercube sampling. en, according to experience, 12 points are randomly selected and evaluated by executing the disciplinary analysis. e Kriging model is constructed and updated by the proposed active global learning method. Subsequently, the MDF and CO framework are performed based on the Kriging model to search the optimal solutions. e results are summarized and compared with those of the deterministic design and MCS in Table 4. e second row shows the deterministic results. EBMDO is performed with three failure probability constraints, namely, P f � 0.2, 0.1, 0.0013. e results show that the objective function of MDO is smaller than that of EBMDO. e constraints of failure probability are not satisfied. For a small failure probability constraint, the design variables and the objective function must be large to meet the requirement of failure plausibility. e obtained results are close to those of MCS, and only the second constraint is active. e outer optimization loop has 100 iterations, and for each design point, 10 5 times of MCS are performed in the evidence analysis. For each evidence analysis, the limit state function is called for about 10 times to obtain the minimum value. Hence, the limit state function is called for approximately 2 × 10 8 times.
e number of limit state function calls by the proposed method is 26 times for this example, which is only 1.3 × 10 − 7 of MCS.

Case 2.
e second example contains two disciplines with three design and two coupled variables. e constraint function is nonlinear. e EBMDO problem is defined by where d i (i � 1, 2, 3) are design variables, y i (i � 1, 2) are coupled variables, x i (i � 1, 2) are uncertainty variables, and P f is the constraint of failure probability, P f � 0.01. In this example, the coupled variables make the problem highly complex. Probability theory is a special case of evidence theory. To compare the EBMDO design with the RBMDO design, the BPA for each interval of the uncertain variables is considered to be similar to the area under the probability density function used in RBMDO, as shown in Figure 5.
In RBMDO, x i pertains to normally distributed random parameters with x i ∼ N(d i , 0.5 2 ). e BPA structure of the uncertain variables for EBMDO is shown in Table 5. e samples are generated in the entire design space, namely, [2,8]. Subsequently, 12 points are randomly selected to construct the initial DoE. en, the learning strategy mentioned in Section 3 is utilized to select the points for DoE. In each selected point, the multidisciplinary analysis must be performed to obtain the response of the limit state function. e training history curve of the Kriging model is shown in Figure 6.
e Kriging model converges in 9 and 13 iterations for the first and second limit state function, respectively. To verify the accuracy of the Kriging model, the signs of 10 5 samples that are randomly generated in the design space are compared with that of the true limit state function. e histogram of the sign is illustrated in Figure 7. e predicted signs have a good match with that of the true function. e number of signs that are incorrectly predicted is only four, while the relative error is 4 × 10 − 5 .
en, the problem is solved using the CO strategy. e EBMDO problem can be decomposed into one system Table 3: BPA structure of epistemic variables of Case 1. problem and two corresponding discipline problems because of the presence of two disciplines. e system optimization problem is defined as where d isl (i � 1, 2, 3) contains the system design variables and d 1i , d 2i , d 3i denotes the design variables of the i th discipline. e coupled variables, y 1 , y 2 , are considered the design variables to decouple the disciplines. e optimization problem of the first discipline is defined by s.t. Pl g 1 < 0 < P f , g 1 � y 11 − 3.16, e optimization problem of the second discipline is defined by e data flowchart is presented in Figure 8 to show the solving architecture of cooperative optimization. e blue and red lines indicate the input and output, respectively. e system variables are passed to the discipline analysis, and the objective functions of disciplines are returned to system optimization.
e objective functions of disciplines are the constraints of system optimization. e reliability constraints are evaluated by evidence analysis with the global Kriging models. Table 6 compares the DO, EBMDO by MDF, EBMDO by CO, RBMDO, and MCS results. e conclusions are similar to those in the previous example. e evaluation number of the limit state function by the proposed approach is 46, which is about 2.3 × 10 − 7 that by MCS. is result indicates that the proposed method can ensure accuracy while having a great advantage over MCS.

Case 3.
e third case is multidisciplinary design optimization for cooling blade which considers the aerodynamic, heat transfer, and strength analysis, revised from [40]. e geometry model of the cooling turbine blade and the design parameters are shown in Figure 9, where x i (i � 1, . . . , 4)is the thickness of ribs and x i (i � 5, . . . , 10) is the thickness of blade wall at different blade profile. e EBMDO problem is defined by find x � x 1 , · · · , x 10 , s.t. T max ≤ 1500 K, Complexity where x is the uncertainty design variables, f is the objective function, and D is the damage of cooling turbine blade, which is the function of the design variables. In this case, the damage D is the whole damage, which is constituted by two parts, namely, the creep damage and the fatigue damage. e creep damage is predicted by the Larson-Miller equation [41] and the fatigue damage is calculated by the nominal stress method [42]. e fluid-thermal-solid interaction analysis is carried out to obtain the stress and the temperature in MDF framework. Also, in the CO framework, the three disciplines are decoupled and performed independently. T ave is the average temperature, w 1 � 3000 and w 2 � 1 are the weight factor, T max is the maximum temperature of cooling turbine blade, σ max is the maximum stress, P is the reliability, and x is the uncertainty design variables including x 1 , x 2 , . . . , x 10 . e design variables are considered as epistemic parameters. e lower and upper limit of variables are shown in Table 7. e thickness difference Δx obeys threeparameter Weibull distribution, with the location parameter, the shape parameter, and the scale parameter as − 0.1540, 2.2979, and 0.2923, respectively. According to the Weibull distribution and Monte Carlo sampling, the BPA of the design variables are shown in Table 8.
e EBMDO results compared with the RBMDO and MDO by [40] are listed in Table 9. e multidisciplinary feasible approach is adopted in this case. As can be seen, the result obtained by EBMDO is very close to RBMDO, which indicates the proposed method is effective.

Conclusions
is article presents a novel approach for EBMDO that combines the active global Kriging model, MDF, and CO. By   x 1 x 2 x 3 x 4 x 5,8 x 6,9 x 7,10

Rib4
Rib3 Rib2 Rib1  introducing a learning function, the global Kriging model is adaptively constructed to replace the limit state function. en, the evidence analysis is performed by MCS and the Kriging model. e computational cost of EBMDO is reduced to approximately that of DO. ree examples are provided to illustrate the proposed approach. Compared with MCS, the proposed method can obtain accurate results and hold a significant advantage in terms of computational efficiency. erefore, the proposed method is expected to be of great value in engineering applications.
As part of further work, some other architectures of multidisciplinary optimization can be taken into account in EBMDO, such as concurrent subspace optimization (CSSO) and bi-level integrated system synthesis (BLISS).

Data Availability
All relevant data used to support the findings of this study are included within the article.

Conflicts of Interest
e authors declare that they have no conflicts of interest.