Elemental Reactivity Maps for Materials Discovery

Abstract

When searching for novel inorganic materials, limiting the combination of constituent elements can greatly improve the search efficiency. In this study, we used machine learning to predict elemental combinations with high reactivity for materials discovery. The essential issue for such prediction is the uncertainty of whether the unreported combinations are non-reactive or not just investigated, though the reactive combinations can be easily collected as positive datasets from the materials databases. To construct the negative datasets, we developed a process to select reliable non-reactive combinations by evaluating the similarity between unreported and reactive combinations. The machine learning models were trained by both datasets and the prediction results were visualized by two-dimensional heatmaps: elemental reactivity maps to identify elemental combinations with high reactivity but no reported stable compounds. The maps predicted high reactivity (i.e., synthesizability) for the Co_Al_Ge ternary system, and two novel ternary compounds were synthesized: Co4Ge3.19Al0.81 and Co2Al1.24Ge1.76.

14 pages, 5 figures

References in corpus (5)

• Data-driven materials science: status, challenges and perspectives
• Learning from positive and unlabeled data: a survey
• Machine learning modeling of superconducting critical temperature
• Predicting materials properties without crystal structure: Deep representation learning from stoichiometry
• Compositional descriptor-based recommender system accelerating the materials discovery