Scaling deep learning for materials discovery

ABSTRACT

The discovery of novel functional materials is essential for technological progress in batteries, solar cells, and computation, but traditionally relies on expensive, trial-and-error experimentation. We present Graph Networks for Materials Exploration (GNoME), a deep learning framework that scales the prediction of stable crystal structures to unprecedented levels. By actively learning from density functional theory (DFT) calculations, GNoME discovered over 2.2 million new stable crystal structures, expanding the known database of stable materials by an order of magnitude. These predictions have already led to the successful physical synthesis of hundreds of novel compounds in independent laboratories worldwide.