Dynamic weighted aggregation of surrogate models for engineering optimization

Surrogate models have emerged as a critical tool for reducing computational costs in complex engineering simulations. However, individual surrogate models often exhibit limitations in accuracy and robustness across diverse scenarios. This paper proposes a Point-Optimized Ensemble Method (POEM) that dynamically aggregates multiple surrogate models using localized weight optimization. The method partitions the design space into interpolation and extrapolation regions, assigning model weights based on proximity to training data and cross-validation errors. The case study on maximum stress for train and body bogie vibration frequency prediction demonstrates that POEM reduces prediction errors by 3.6% compared to single models when sufficient samples are available. Experimental results validate the method’s efficacy in balancing computational efficiency and prediction reliability, offering a versatile solution for high-dimensional nonlinear systems.