Massive parameter extraction via coupled mode theory for enhanced intelligent metamaterial design

Metamaterials exhibit a diverse range of capabilities in electromagnetic manipulation. However, designing metamaterial structures for expected functions encounters an inverse problem. The advent of deep learning offers a data-driven approach to address this issue. To enhance the transparency of deep learning models, coupled mode theory (CMT) is incorporated into the design process. Despite this advancement, efficiently extracting coupled mode parameters for neural network training remains a significant challenge. To address this, we propose a simulation-free theory-driven pretraining neural network method for extracting coupled mode parameters. Using these extracted parameters, we perform both forward and inverse design of metamaterials, demonstrating the benefits of integrating CMT into intelligent design. By incorporating CMT, our approach transitions from a purely data-driven model to a "physics-driven" one, improving design accuracy. Compared to end-to-end data-driven methods, our CMT-based scheme offers superior design precision. The results, validated by simulation data, highlight that our method bridges the gap between intelligent and interpretable design.