Efficient Incremental Variable-Fidelity Machine-Learning-Assisted Hybrid Optimization and Its Application to Multiobjective Antenna Design

Abstract

Online-model-based machine-learning-assisted optimization (MLAO) methods are widely used to reduce the computational burden of complex electromagnetic (EM) optimization problems. Multidesign parameter and multiobjective EM problems are common in engineering practice. As the problem design dimensionality increases, the training time of the surrogate model in the optimization process becomes nonnegligible. The performance of optimization algorithms degrades for high design dimensions and multiple objectives, and many full-wave simulation calculations are required before convergence. In this work, an incremental variable-fidelity machine-learning-assisted hybrid optimization (IVF-MLAHO) algorithm is proposed to solve a multiobjective EM problem with medium-scale (i.e., 20–50) design variables. First, reliable variable-fidelity models are used for initial sampling to reduce the computational cost of sampling. Then, in the training process, incremental learning or retraining is adaptively selected to update the surrogate models, which reduces the training burden. Furthermore, a hybrid global multiobjective and local single-objective optimization algorithm is adopted to markedly improve the convergence performance. Finally, the superiority of the IVF-MLAHO algorithm is verified on a substrate-integrated waveguide (SIW) broadband millimeter-wave slot antenna array, in which the training time is greatly reduced.