Synthesizable materials discovery via interpretable, physics-informed machine learning models

Abstract

The critical roles of computations and machine learning in accelerating materials discovery have become increasingly recognized, particularly in predicting and interpreting the synthesizability and functionality of new materials. Here, we develop a synthesizable materials discovery scheme using interpretable, physics-informed models. Our approach is based on an integration of high-throughput computations that capture the essence of materials properties, including the impact of point defects, and explainable machine learning models. These models provide quantitative predictions and interpretations of the materials’ synthesizability and functionality based on structural and chemical descriptors in a vast compositional space. Applying this scheme to proton-conducting cubic inorganic electrolytes for fuel cells, two unconventional materials for proton-conducting electrolytes, Pb-doped Bi12SiO20 and Sr-doped Bi4Ge3O12, were discovered in the first two synthesis trials. This scheme effectively bridges the existing gap between computational predictions, human interpretations, and experimental feasibility in machine learning, offering new insights into materials discovery and accelerating the development of new functional materials.