Physics-Informed Inverse Design of Optical Coatings using a Differentiable Transfer Matrix Method

We tackle the challenging inverse design of optical coatings using an artificial intelligence (AI) framework for optical thin-film coating design. Our approach is based on a physics-informed autoencoder with a differentiable physics decoder. Unlike data-driven approaches, our model embeds Maxwell’s equations directly through an analytical forward model, enabling end-to-end, gradient-based optimization from target optical properties to physical layer structures, without requiring any prior design examples. We demonstrate our method by designing a complex broadband mirror with a target reflectivity >99\% and a precise group delay dispersion of$-200\,\text{fs}^2$ over the 940–1120\,nm wavelength range. The AI-generated design reaches performance characteristics competitive with state-of-the-art commercial software, demonstrating a powerful and generalizable framework for solving physics-based inverse design problems.