From random patterns to nanophotonic structures: Electromagnetic field prediction without retraining

Abstract

Electromagnetic surrogate solvers can accelerate nanophotonic design, but their broader use requires accurate predictions for unseen geometries and material arrangements. Training on random dielectric patterns enables direct prediction of complex electromagnetic fields in diverse two-dimensional nanophotonic structures without retraining. The network takes spatial permittivity distributions as inputs, with training samples drawn from a dataset comprising 700,140 combinations of 5,001 random freeform dielectric patterns and 140 relative permittivities from 2.1 to 16.0. The trained model predicts fields for aperiodic bar arrays, manually defined patterns, inverse-designed metalenses, inverse-designed binary structures, and two-material freeform patterns excluded from training. Across 35,441 geometry-permittivity samples, a dilated residual network achieves average relative field errors of 1.4%-7.6%. For 6,950 inverse-designed metalens samples, the three most accurate architectures predict transmitted power with median relative errors of 0.7%-0.9%. Accuracy depends on refractive index and structural scale, with greater errors for patterns coarser than those used for training. These results demonstrate transfer to independently constructed geometries and to new spatial arrangements of dielectric values not represented in the training. Random-pattern training therefore supports reuse of electromagnetic surrogates across multiple structural families, providing a basis for sharing training costs among material screening, iterative optical design, and fabrication-tolerance analysis.