Approximate Features of Electromagnetic Responses by Coarse Deep Learning Model to Optimize Metasurfaces

Le résumé fourni par la source

Training a deep learning (DL) model with data collected from full-wave simulations to accurately estimate electromagnetic responses has become a prevalent approach for metasurface optimization. Several methods have been proposed to reduce the size of the dataset to train the model while maintain the model’s performance so that the overall optimization efficiency can be improved. A new optimization method is proposed, sacrificing the performance of the DL model to further reduce the reliance on dataset size. To make the use of the coarse model, a feature-oriented optimization procedure is adopted. Firstly, a coarse DL model, trained with a smaller dataset, is employed to identify a preliminary solution within the parameter space. Secondly, the cost function is reformulated to be represented by the features of the electromagnetic response, which are extracted and approximated using the coarse DL model. Finally, a trust region optimizer is adopted, complemented by a random pick and cascaded finetuning strategy to ensure the stability of the optimization process. The effectiveness of the proposed method is demonstrated through optimization of a single-layer metasurface with a complex pattern for ultrabroadband absorption, as well as a three-layer metasurface aiming for a wide stopband above the low passband. Numerical results indicate that high-performance metasurfaces can be achieved with significantly fewer full-wave simulations during the optimization. A prototype of the optimized metasurface absorber was fabricated and measured, with the results exhibiting satisfactory agreement with the design target.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.