Learning and Predicting the Structure-Property Relationship of Plasmonic Gold Nanoparticles and Assemblies via Dual Variational Autoencoder

Plasmonic colloids or assemblies attract tremendous interests across research communities due to their exceptional light-matter interaction at the nanoscale. Data science approach including machine learning provides novel avenue for nanomaterials characterization and discovery. Herein, we employ our dual variational autoencoder (dual-VAE) model to predict structures or photonic response of plasmonic gold nanomaterials with great precision. Upon training both optical dark-field scattering spectra data and electron microscopy images data in two separate VAEs and establishing correlations between the two, we can enable predictions of geometries of the nanostructures based solely on their optical scattering patterns. Owing to this correlative VAE training, we achieve structural probe under the optical diffraction limit while avoiding perturbation due to electron beam irradiation. Specifically, we first predict numbers of particles in a gold assembly given their scattering response intensifying and red-shifting with increasing numbers of particle in each cluster. Secondly, by exploiting the optical asymmetry of gold nanorods and collecting linearly polarized scattering spectra, we predict the orientation angle of individual nanorod. This work delivers insight on how data science approach can be used to advance characterization of plasmonic nanomaterials and explore new structure-property relationships.