Inverse design method of optical devices for generating target near-fields via physics-informed neural networks

We propose a method for inverse design of optical devices to generate target near-fields using physics-informed neural networks (PINNs). A finite-difference Helmholtz equation is embedded in the loss function, so the predicted field satisfies the wave equation and device boundary conditions without extra boundary-loss terms. A data-loss function forces the network output to match a prescribed near-field, enabling design for target distributions. Additional penalty and binarization losses restrict the refractive index to realizable values. The method requires no precomputed training data or problem-specific adjoint gradients. We validate the method accuracy by comparing PINN-predicted fields with those from generalized Lorenz-Mie theory and full-wave simulations. Using photonic nanojets and photonic hooks as benchmarks, we demonstrate the framework's ability to discover non-intuitive structures for arbitrary near-field distributions automatically.