A harmonic domain regressor with dynamic task weighting strategy for multi-fidelity surrogate modeling in engineering design

Lin You, Songqing Xing, Jin Yi, Shujin Yuan, Jiangtao Yang, Huayan Pu, Jun Luo

Abstract: Surrogate modeling has become increasingly popular in engineering design optimization because it can construct cheap yet accurate models. Multi-fidelity surrogate models (MFSMs) are beneficial as they combine low- and high-fidelity data to improve prediction accuracy while reducing computational cost. However, constructing MFSMs is challenging due to the complex relations between low- and high-fidelity models. In addressing this challenge, we introduce an innovative approach for constructing MFSMs based on transfer learning with harmonic domain regressor and dynamic task weighting strategy. The proposed approach does not need pre-assumptions of the relations on low- and high-fidelity models and also avoids the difficult choice of the sharing layer of the finetune-based transfer learning methods. The efficacy of the proposed method has been successfully validated on 10 numerical problems up to 20 dimensions and a case study of the optimal design for a coaxial magnetic gear. Experimental results demonstrate that the proposed approach outperforms existing multi-fidelity fidelity modeling approaches in prediction accuracy, highlighting its potential for engineering design with low computational cost.