Enabling Inverse Design of Nanoparticle Photophysics via Deep Learning on Heterogeneous Graphs

Deep learning (DL) is commonly employed to accelerate materials design and development. However, to date, few efforts have applied DL to nanomaterials, which lack suitable structural representations and adequate data for model training. We report efforts to overcome these limitations, leveraging DL to optimize the optical properties of heterostructured core-shell upconverting nanoparticles (UCNPs) for applications in e.g. biosensing, super-resolution microscopy, and 3D printing. The nonlinear photophysical properties of UCNPs, which allow for the emission of visible and ultraviolet light from near-infrared excitations, depend on the number of shells, shell thickness, and dopant concentrations. Though kinetic Monte Carlo (kMC) simulations allow for reasonably accurate prediction of UCNP optical properties, high simulation cost has limited the exploration of this vast design space to relatively small, simple systems. Here, we describe the first large-scale dataset of UCNP spectra based on kMC simulations (N > 6000). Training on this dataset, we develop an accurate physics-infused heterogeneous graph neural network and use it to perform inverse design of UCNPs via gradient-based optimization. We identify novel structures with 6.5x higher predicted emission under 800nm illumination than any UCNP in our training set, validated by months-long kMC simulations. Our work reveals new design principles for UCNPs and presents a roadmap for de novo nanomaterial design.