Predicting the Outcomes of Material Syntheses with Deep Learning

A key bottleneck for material discovery is synthesis. While significant advances have been made in computational design, synthesis pathways are still often determined through trial and error. In this work, we develop a method that predicts the major product of solid-state reactions. The main advance presented here construction fixed-length, learned representations Precursors represented as nodes on “reaction graph”, message-passing operations between used to embody interactions precursors reaction mixture. We show deep learning framework not only outperforms baseline methods but also more reliably assesses uncertainty its predictions. Moreover, our approach establishes quantitative metric inorganic similarity, allowing user explain model predictions retrieve relevant literature sources.