AI-driven materials discovery of novel solar cell materials

AI-driven materials discovery has become a new paradigm for condensed matter physics. In this work, we optimize an atomistic line graph neural network (ALIGNN) model for rapidly predicting the dielectric function of potential optoelectronic materials. The graph neural network, which is trained on approximately 7000 dielectric functions from the Joint Automated Repository for Various Integrated Systems-Density Functional Theory (JARVIS-DFT) database, can accurately reproduce spectral features, allowing it to effectively characterize derived properties, including the solar efficiency. This success is encouraging evidence for the general application of advanced graph neural networks to the prediction of spectral properties. We are thus able to confidently employ this model to analyze over 400,000 3D DFT materials in the Alexandria materials database, and identified that the perovskite class of materials tends to have a higher proportion of high-efficiency solar cell materials [1].