Inverse Design of Dual-Layer Chiral Metasurfaces through Tandem Deep Learning Network

This letter introduces a tandem cascade network model using deep learning for inverse design of dual-layer chiral metamaterials. We uniquely employ the chirality coefficient κ, a direct measure of magnetoelectric coupling, as the optimization target, providing a physically grounded approach to inverse design that offers a more comprehensive characterization of chiral responses compared to traditional transmission coefficients. Our model efficiently predicts chiral responses across six fundamental patterns at different rotation angles, demonstrating its generality. Using 22,000 training samples, both forward and inverse networks achieve computation speeds approximately 10⁶ times faster than conventional simulations. Validated experimental results demonstrate high prediction accuracy with MSE values well within 7.5 × 10^-3 for all test samples, and robust generalization across diverse dual-layer patterns.