Developing Neural Networks for Inverse Design of Chiral Metamaterials

Abstract Chiral metamaterials, renowned for their unique optical properties such as circular dichroism, are pivotal in applications like spectroscopy, sensing, and imaging. However, their inherent asymmetry and complex light‐matter interactions present substantial design challenges. This study harnesses the power of deep neural networks (DNNs) for the inverse design of chiral nanohole arrays (CNAs). A bidirectional neural network (Bi‐DNN) is developed to address the one‐to‐many mapping issue, achieving high prediction accuracy (0.98). Various input‐output configurations are examined, including combining inputs and leveraging different input‐output models (e.g., using left‐handed circularly polarized light spectra to predict right‐handed circularly polarized spectra), enhancing prediction precision while reducing experimental workload. Additionally, the potential of CNAs as high‐performance surface‐enhanced Raman spectroscopy substrates for chiral detection is demonstrated. The Bi‐DNN enabled rapid and accurate design solutions, showing strong agreement with experimental validations. These findings emphasize the transformative role of DNNs in advancing chiral metamaterial design, unlocking efficient and customizable optical materials for next‐generation sensing and imaging technologies.