Graph Neural Networks for Accelerated Optoelectronic Materials Prediction

We employ an atomistic line graph neural network (ALIGNN) model for analyzing photovoltaic materials. The model converts an input crystal structure directly to the real and imaginary components of its dielectric function, and was trained on approximately 7000 dielectric functions from the JARVIS-DFT database whose optoelectronic properties were computed with the meta-GGA exchange-correlation functional. The trained model is able to replicate reference dielectric function spectra with sufficient accuracy to serve as a drop-in replacement for materials workflows that involve finding the dielectric function as an intermediate step. We demonstrate this at a large scale by analyzing over four hundred thousand insulating 3D DFT structures from the Alexandria database collection to search for promising new solar energy materials. Our analysis highlights perovskites, and more specifically vanadium-based perovskites, as promising avenues for novel solar materials design[1].