Abstract

This thesis investigates Spherical Fourier Neural Operators (SFNOs) as surrogate models for electromagnetic scattering on clusters of spherical scattering objects. Traditional numerical solvers like finite difference time domain (FDTD) and finite element method (FEM) are computationally expensive, creating bottlenecks in optimization and inverse design scenarios. The architecture of our SFNO model respects SO(3)-equivariance, mapping 3-dimensional distributions of scattering objects as 250 concentric shell maps to complex scattered electric fields on spherical observation surfaces. Training data consists of 30000 samples generated using T-matrix methods, with clusters of five non-overlapping dielectric spherical scattering objects (n=1.5) at lambda=1000 nm. Two different model variants were tested: model one trains on normalized inputs/outputs while model two trains on raw data. Both models were trained with multiple different architectural configurations and while neither delivered good results, the analysis revealed possible limitations, which consist of architectural design, model capacity or loss function design.