Deep Learning for the Inverse Design of Non-Hermitian Photonic Structures

Deep learning is becoming a powerful tool for designing photonic systems with complex behaviors. In this work, we use neural networks for the inverse design of non-Hermitian photonic structures, where optical gain and loss are used to control light in new ways. The model learns how to connect desired optical responses—such as transmission or reflection spectra—to the structural parameters that produce them. This allows fast and accurate design of systems showing key non-Hermitian effects, including exceptional points and unidirectional transparency. We demonstrate the approach on multilayer and resonator structures and show that it can explore large design spaces much faster than traditional methods. Our results show how deep learning can speed up the discovery and understanding of next-generation photonic devices that rely on non-Hermitian physics