Inorganic Synthesis Recommendation by Machine Learning the Similarity of Materials from Scientific Literature

Synthesis remains a bottleneck in the discovery of novel inorganic materials. To synthesize the desired compound, various parameters need to be determined, such as precursors, operations, and conditions. Due to the lack of predictive tools, a widely used approach is to manually refer to precedent synthesis recipes for similar materials. However, deciding which materials are similar is often driven by intuition and is limited by the personal experience of the experiment designer. Individuals are typically biased by their experiences working in a specific chemical space, hindering their ability to quickly design syntheses for new chemistries. This work proposes a data-driven framework to quantify the similarity of target materials and recommend synthesis parameters for those targets based on an extensive synthesis database text-mined from scientific literature. We find that vectorized material representations of synthesis parameters can be learned by carefully designing an encoding neural network. Similar to how language models pre-train word representations by predicting context for each word, this neural network trains material representations by predicting synthesis parameters for each target material. We demonstrate that the cosine similarity calculated from the material representations is an effective metric for proposing and ranking potential precursor sets based on analogy to previously studied target materials. Using 35,307 solid-state synthesis recipes from the literature as a knowledge base for precursor recommendation, the observed precursor sets are among the top 10 proposed ones for 72.8% of 2,796 test target materials, providing strong support for our approach. This quantitative recommendation framework can serve as a predictive tool to ease the process of synthesis design and constitutes an important step toward the autonomous synthesis of inorganic materials.