We introduce a physics-informed neural network (PINN) based strategy for the inverse design of nanophotonic metasurfaces that directly incorporates fabrication uncertainties while providing reliable uncertainty estimation. In this approach, Maxwell’s equations are enforced through the loss function, fabrication variations are modeled using Monte Carlo perturbations, and Bayesian PINN ensembles are employed for calibrated predictive intervals. This framework enables the discovery of metasurface geometries that maintain high optical performance while substantially enhancing fabrication yield compared to conventional adjoint and supervised learning techniques. The method is validated on focusing metalens structures at λ = 632.8 nm, where it delivers a 23% improvement in median yield and reduces data requirements by a factor of 25 relative to standard methods. Our design achieves 87.8% focusing efficiency with a 91.3% fabrication yield under ±8 nm geometric variations, marking a significant step toward robust and scalable photonic device design.
Sneha Latha P (Fri,) studied this question.