End-to-end machine-learning framework for electromagnetic inverse design: From practical constraints to optimized structures

Existing machine learning (ML)-based inverse design methods for electromagnetic (EM) structures encounter two major challenges: impractical input requirements and nonuniqueness effects. This article introduces an end-to-end EM design framework based on a modified conditional variational autoencoder (MCVAE), which directly maps practical design constraints to optimized structure configurations. The proposed framework incorporates a modified decoder that interprets practical constraints, such as desired operating frequency, bandwidth (BW), maximum allowable value for reflection coefficients, and transmission phase, into detailed EM responses over the target frequency band. A modified encoder mitigates the effects of nonuniqueness common in inverse EM design, where multiple structures yield similar responses, by integrating Gaussian noise for robust latent space exploration and using a forward-model-based loss function to enforce structural accuracy, thereby enhancing output reliability and model performance. To validate the effectiveness of the proposed method, multiple practical implementations are presented: a linear-to-circular polarization converter design, a Fourier-phased metasurface design, a slotline filter design, a loop polarization converter design, and a polarization converter design with double- and triple-expanded parameter ranges. Compared with traditional optimization-based methods, the proposed method significantly improves design efficiency while maintaining high accuracy. This method offers a generalized framework for end-to-end EM design, bridging practical constraints to optimized structure realizations. Copyright © 2025 The Authors.