Controlled Synthesis of Space–Time Modulated Metamaterial for Enhanced Nonreciprocity by Machine Learning

Abstract

Nonreciprocity plays a fundamental role in governing direction‐dependent asymmetric wave propagation. Previous approaches to nonreciprocity involve ferrite‐based devices with bulky systems. Herein, the controlled synthesis of a space–time modulation (STM) metamaterial for enhanced nonreciprocity using machine learning (ML) is investigated. The design of STM metamaterial poses great challenges due to the nonlinear nature of time‐periodic Floquet harmonics, which are inefficiently handled in traditional methods such as numerical optimization. To deal with the challenge, an ML approach is proposed that learns from the accumulated training data using the guided objective function and generates high‐quality designs by leveraging the learned features. This approach first trains a residual neural network (ResNet) to learn the nonlinear relationships between the STM parameters and nonreciprocal responses. The trained ResNet achieves a high testing accuracy, with 96.7% of the 9000 instances having a mean square error less than 0.6 × 10−4. For the synthesis of STM metamaterial, a customized Wasserstein generative adversarial network (WGAN) is proposed, which leverages the discovered nonlinearity and synthesizes large‐scale datasets using small computational costs. The histogram obtained using 90 000 data samples shows that WGAN generates designs with an average normalized nonreciprocity of 0.83, four times higher than the measured data.

Keywords

• generative adversarial network
• machine learning
• metamaterials
• nonreciprocity
• space–time modulation