Abstract: To compensate for long design cycles and high experimentation costs in industrial product development, surrogate model has become an indispensable technique for engineering applications. Ensemble modeling using multiple types of single surrogate models can improve the stability and prediction accuracy of the models. However, the ensemble modeling of single surrogate models still faces challenges such as high model complexity and poor generalization ability. In this context, an adaptive variable weight ensemble surrogate model based on correlation analysis is proposed. The model is divided into two parts: surrogate model library selection and adaptive weight calculation. In the model selection phase, the Spielman correlation coefficient and the cross-validation error are first combined as the benchmark for model selection, and then the global optimal model and the surrogate model library are obtained. Then, the adaptive weight coefficients are calculated for each component surrogate model in the model library using the the error inverse variance weight method. Finally, the component models are connected according to the ensemble surrogate model formula to obtain the final prediction model. The proposed surrogate model considers the global accuracy and local accuracy of each single surrogate model. The experimental results show that the proposed model has better performance compared with the ensemble surrogate models.

Key words: surrogate model, ensemble modeling, model selection, cross-validation, adaptive weight factor