On Advanced Surrogate Models And Application

Traditional engineering design and optimization is usually based on a small amount of data from physical experiments to explore the operating law of systems.Several shortcomings cannot be ignored,such as long design cycles and high experimental costs.With the rapid development of numerical methods,computer technologies have been significantly improved.and numerical simulation techniques have been gradually applied into engineering design and optimization.However,to ensure the reliability and accuracy of design and analysis results.simulation models are becoming more and more refined,and the fidelity and complexity of the model are simultaneously improved,resulting in the rapid development of computer technology that still cannot satisfy the needs of engineering design and optimization.Therefore,the surrogate model(SM),which is a numerical analysis method based on small sample points.came as demanded by the times,and because of its advantages such as short time,rapid response,and low cost it has been widely used in engineering problems.However,several basis problems in the surrogate technique still do not have good solutions,mainly focusing on three aspects including sequential sampling(SS)methods,hybrid surroatge(HS)models,and multi-fidelity surrogate(MFS)models.In this thesis,an adaptive hybrid surrogate model is developed for problems of weak robustness,poor prediction accuracy,and high computational complexity in existing HS models;to further improve the accuracy of SFS models and solve problems of poor sampling accuracy and low efficiency in DoEs,a hybrid SS strategy that balances local exploitation and global exploration is established;to address multiple fidelity problems,and take advantages of multi-fidelity information,an MFS model is proposed.Meanwhile,key factors,such as the effect of cost ratio of high-fidelity(HF)to low-fidelity(LF)models,and the effect of combinations of HF and LF samples,on the performance of MFS models are investigated.The primary contents are as follows:(1)Studying and analyzing the existing single-fidelity surrogate(SFS)techniques including classical individual and hybrid surrogate models,a two-stage adaptive hybrid surrogate(T-AHS)model with stronger robustness and higher prediction accuracy is developed.The modeling process mainly includes two stages:i)constructing the surrogate model library and ii)calculating of weight coefficients and blending individual surrogate models.Through 40 testing functions,comparative studies with five component individual surrogate models and four benchmark hybrid surrogate models are conducted.Results show that the proposed model is significantly better than other benchmark models in terms of prediction accuracy and robustness.In addition,the T-AHS model which is simple to build can save an average of about 82%of computational cost compared to other benchmark models.(2)To further improve the prediction performance of SFS models and save costs,a Go-inspired hybrid infilling(Go-HI)strategy is established by researching basic theories of DoEs and analyzing problems in DoEs.which aims at addressing the shortcomings of current sequential sampling methods,such as poor sampling accuracy,low sampling efficiency,and out of the capacity of balancing the local exploitation and global exploration,etc.The strategy mainly consists of three stages,including the construction of sequential sampling strategy library,the construction of tree-like structure,and the calculation of decision values for subtrees.The effect of the number of component sequential sampling methods and tree depths on the performance of the strategy is investigated,and results show that the number of component SS methods has greater impact on the strategy than the tree depths,and the performance of the strategy can not be improved no end with increase in the number of sampling methods,it is observed that the Go-HI strategy focuses on exploring the region of interest(Rol),when exploring the local area to some degree,it can jump out of the current Rol,identify the next Rol.and complete the global exploration task.The Go-HI strategy can achieve the task chain of local exploration and global exploitation from‘Local'to‘Global'to‘Local',and compared with other reference SS methods the Go-HI strategy can save about 16%-3 1%cost.(3)To solve the shortcomings that SFS models including individual and hybrid surrogate models can not settle the multi-fidelity information,and The correlation between HF and LF models is not taken into account in the construction of existing MFS models,resulting in HF and LF sample information is not fully used.Therefore,to further make use of the HF and LF sample information in the modeling process,the canonical correlation analysis(CCA)method is used to estimate the correlation between HF and LF models by using that between small HF and LF samples.The thought of RBF model is used to build the discrepancy function between HF and LF models,and then the least squares(LS)method is adopted to structure the objective optimizing the scaling factor and hyper-parameters in the discrepancy function,a regression MFS model is proposed,named CCA-MFS model.Comparisons with other benchmark MFS models with different amount of HF samples are conducted.It is seen that the CCA-MFS model outperforms other benchmark models in terms of global prediction accuracy and robustness.In addition,MFS models are not applicable when the number of HF samples is small.Moreover,the influence of key factors is also studied.It can be seen that the proposed model is less sensitive to the correlation HF/LF models.The cost ratio of HF to LF models plays a significant role.When the cost ratio increases within a certain range,the performance of the CCA-MFS model will be continuously improved,indicating that more LF information is likely to improve the performance of the CCA-MFS model.(4)In the pressure relief valve(PRV)analysis,performances of the proposed Go-HI stategy,the T-AHS model,and the CCA-MFS model were further verified.The performance of the Go-HI strategy is studied with the two-dimensional(2-D)computational fliud dynamics(CFD)simulation model of PRV.Results show that the Go-HI strategy is helpful to quickly construct a higher precision surrogate model and can save an average of about 37%of the computational cost.The T-AHS model is more accurate and robust than benchmark surrogate models.Using the physical test experiment of PRV as the HF model and the 2-D CFD model as the LF model,the CCA-MFS model is used to fuse the HF and LF model information.Results show that the CCA-MFS model outperforms the other benchmark MFS models in prediction accuracy,and the MFS model is not applicable when the amount of HF training points is small.