Efficient models for nanophotonic systems: optimization of compact representations and physics-guided deep learning.
Sofia Ponomareva PhD defense
30.11.26 - 30.11.26
Nanophotonic devices, whose operation relies on light–matter interactions at the nanoscale, offer a high degree of freedom for shaping light propagation. This flexibility, however, comes with a significant computational challenge when modeling large and complex nanostructures. Therefore, numerical methods that provide compact models while retaining essential physical effects are of particular interest. The aim of this thesis is to obtain efficient representations of optical systems through hybrid approaches that combine data-driven methods with known physical laws, leveraging the complementary strengths of both paradigms. The first contribution develops a physics-informed deep learning method to reconstruct a representation of interactions with an unknown nano-optical system from far-field intensity measurements alone. The learned "Neural Green's functions" enable modeling of arbitrary particles interacting with the characterized environment. The second contribution presents an optimization framework to derive compact yet accurate models of nanostructures based on the Global Polarizability Matrix method. By combining neural representations, gradient-based optimization, and a greedy pruning technique, the framework provides representations that are able to capture near-field interactions between neighboring structures while enabling full-scale optimization of large ensembles of scatterers. Finally, to reduce the high computational cost of generating training data for models describing the optical response of large scattering assemblies, this thesis proposes a physics-driven deep learning approach that minimizes the need for expensive simulations by imposing Green's dyadic formalism on neural network predictions.
published on 20.11.26