Distinguishing the optical properties of plasmonics and dielectric nanostructures via machine learning

We demonstrate how machine learning, coupled with dimensionality reduction techniques, allows us to understand and classify the optical response of nanophotonic components with characteristic plasmonic and Mie-dielectric signatures. To reduce the dimensional space of our analytically-calculated spectra of all metallic and dielectric materials, we use principal component analysis (PCA) and t-Distributed Stochastic Neighbor Embeddings (tSNE) as a linear and nonlinear dimensionality reduction technique respectively. Hereafter, by coupling this lower-dimensional space with an XGBoost-based multi-output regressor, we map optical spectra to the linear and nonlinear reduced maps. As experimental validation, we use a 3D electromagnetic full-field simulation solver to generate optical spectra to show that our machine learning technique can identify properties of experimentally-computed nanostructures such as size and intrinsic material properties and which dielectric environment they are in. Our methodology highlights an effective path to discover novel materials for optical sensing.