Deep-Learning-Assisted Inverse Design of Chiral Metasurfaces With Dual-Band Circular Dichroism

Abstract

Chiral metasurfaces exhibit distinct electromagnetic responses to left- and right-handed circularly polarized light, offering significant potential for applications in optical sensing, chiral recognition, and circular dichroism (CD) manipulation. However, conventional design methods for chiral metasurfaces typically rely on exhaustive parameter sweeps and empirical optimization, resulting in low design efficiency and long optimization cycles. To overcome these limitations, this work introduces a deep-learning-assisted inverse design framework for chiral metasurfaces with dual-band CD. A multilayer perceptron (MLP) neural network is employed to establish the nonlinear mapping between structural parameters and CD spectra, enabling rapid and accurate prediction of optical responses. Based on the trained forward prediction model, an inverse design framework is further constructed to automatically retrieve the corresponding structural parameters according to target CD responses. Through extensive training and validation, the proposed model achieves an average coefficient of determination (R2) of 0.98, a mean absolute error (MAE) of 0.026 and a root mean square error (RMSE) of 0.035, outperforming conventional machine-learning methods including support vector machine (SVM), random forest (RF), and k-nearest neighbor (KNN). The optimized metasurface realizes dual-band CD responses with peak values of 0.99 and −0.94. This deep-learning-driven architecture establishes an efficient, reliable paradigm for accelerating the design of high-performance and versatile chiral metaphotonic devices.