Pre-trained deep learning space mapping with intermittent model enhancement for fast multi-fidelity microwave design optimization

Abstract

Parameter tuning is essential in the development of microwave devices. In recent years, a growing interest in formal optimization methods has emerged, driven by their ability to simultaneously adjust multiple decision variables, even under constraints. Their disadvantage is their high computational cost, which is a serious obstacle to the optimization of electromagnetic (EM) models. This difficulty can be alleviated using multi-fidelity simulations. Problem-independent approaches rely on low-fidelity models constructed through coarse-discretization EM analysis (in contrast to problem-specific equivalent network representations), where a critical factor is the appropriate model correction strategy. This paper introduces a versatile deep-learning space mapping (DLSM) strategy that leverages a reusable pre-trained neural network surrogate. The non-parametric DLSM model implements multi-point response correction, applied independently to the real and imaginary components of relevant S-parameter responses. Furthermore, it is trained as a function of the problem's decision variables and response interrelations. It is embedded in a gradient-based optimization loop, where it is locally retrained using intermittent high-fidelity simulations and sample weighting to place greater emphasis on the neighborhood of the current solution. Comprehensive verification of the procedure involving three planar circuits underscores its competitive efficacy, with a mean running cost of 16 high-fidelity simulations and relative savings over the baseline algorithm up to 87%. Meanwhile, consistent results obtained for multiple scenarios targeting diverse performance specifications corroborate DLSM reusability and applicability across broad ranges of operating conditions.