Electromagnetic Inverse Design Based on Physics Informed Neural Network with Generalization Across 1D PEC Geometries

This work introduces a physics-informed neural operator (PINO) that integrates the equivalence principle to enable electromagnetic computation with generalization across 1D PEC geometries and inverse design. Unlike traditional physics-informed neural networks (PINNs), which perform instance-specific regression for a single, fixed geometry, our framework learns a geometry-agnostic operator that maps any structure from a given parametric class (e.g., 1D surface structure) to its electromagnetic responses by encoding Maxwell’s equations as physical constraints. We demonstrate this through 1D meta-surface design, where the operator—trained on procedurally generated geometries—uses an attention-driven architecture to model integral equation correlations inspired by the Method of Moments. The resulting network serves as a universal function approximator for Maxwell’s equations, bridging the gap between PINNs and neural operators, and facilitates inference-driven inverse design.