Physics-informed deep learning for 3D modeling of light diffraction from optical metasurfaces
Vlad Medvedev, Andreas Erdmann, Andreas Roßkopf
Abstract
We propose an alternative data-free deep learning method using a physics-informed neural network (PINN) to enable more efficient computation of light diffraction from 3D optical metasurfaces, modeling of corresponding polarization effects, and wavefront manipulation. Our model learns only from the governing physics represented by vector Maxwell's equations, Floquet-Bloch boundary conditions, and perfectly matched layers (PML). PINN accurately simulates near-field and far-field responses, and the impact of polarization, meta-atom geometry, and illumination settings on the transmitted light. Once trained, the PINN-based electromagnetic field (EMF) solver simulates light scattering response for multiple inputs within a single inference pass of several milliseconds. This approach offers a significant speed-up compared to traditional numerical solvers, along with improved accuracy and data independence over data-driven networks.