Surrogate Model Study for Curvilinear Fibre Variable Stiffness Composite Laminates: Based on Dynamic Hybrid Sampling and Adaptive Surrogate Model Hyper-Parameters Co-Optimization Strategy

Variable stiffness composite laminates with curvilinear fibres have demonstrated significant capability in lightweight structural design, particularly regarding buckling resistance and stiffness enhancement. However, directly applying optimization algorithms often faces challenges such as high computational cost and slow convergence during the optimization design process. Consequently, the incorporation of surrogate models prior to employing optimization algorithms is necessary to simplify computations and accelerate convergence. Manual testing is a conventional approach for hyper-parameter (HP) tuning and continues to be widely used in research. However, manual tuning is suboptimal and time-consuming for many problems. Additionally, the effectiveness of these surrogate models largely depends on the training samples. Therefore, a dynamic hybrid sampling and adaptive surrogate model HP co-optimization strategy is proposed for the optimization design of the variable stiffness composite laminate with curvilinear fibre. In the numerical results, the performance of different surrogate models, comprising Support Vector Regression (SVR), Radial Basis Function Neural Networks (RBFNN), and Back Propagation Neural Networks (BPNN), is systematically compared under varying sample set sizes. Neural results show significant differences in accuracy and efficiency among these three models under varying sample set sizes. SVR demonstrates optimal generalization ability in small sample scenarios, RBFNN strikes a balance between accuracy and efficiency with medium sample size, while BPNN exhibits superior overall predictive performance under large sample condition. The proposed cooptimization strategy overcomes the limitations of traditional single strategy through the closed-loop interaction between dynamic sampling and Bayesian hyper-parameter optimization (HPO). This approach not only significantly improves the predictive accuracy of surrogate models but also greatly reduces the computational cost during the optimization process, making it suitable for computational mechanics problems with high nonlinearity and high-dimensional features. This study provides theoretical foundations and practical guidance for the selection and application of surrogate models in composite material structural optimization, contributing to improved design process efficiency and reliability.