Physics-aware graph neural networks for optically interacting plasmonic nanoparticle assemblies

Artificial intelligence (AI)-based surrogate models have accelerated the design of nanophotonics and nano-optics, though the high cost of generating large, high-fidelity training datasets continues to constrain their development and deployment. Here, we introduce a drastically accelerated graph neural network (GNN) framework that employs multi-fidelity hybrid datasets constructed from the coupled-dipole approximation (CDA) as a fast analytical model and high-accuracy numerical simulations. The GNN is constructed with a physics-aware architecture to investigate the optical responses of nanostructures composed of strongly interacting plasmonic nanoparticles. The proposed polarizability-augmented GNN is consistent with the dipole approximation (DA), representing nanoparticles as nodes and their optical interactions as physics-aware edge features. Through the established efficiency and performance for the nano-dimer and trimer cases, the framework's capacity to integrate theoretical and semi-analytical models within a multi-fidelity, physics-aware data strategy is highlighted. Consequently, the framework can be readily extended to other nanophotonic systems and to designs based on more intricate assemblies of interacting plasmonic nanoparticles.