Materials Discovery through Interpretation

Machine learning has become a pivotal tool in materials discovery, but conventional data-driven models often behave as “black boxes” limited by the scope of their training data. In this perspective, we outline the current status of machine learning-driven materials discovery and argue that incorporating scientific interpretation into the machine learning loop, an approach we term Materials Discovery through Interpretation (MDI), improves the efficacy of materials design. Instead of relying solely on statistical correlations, the MDI approach iteratively interweaves domain knowledge and insights from experimental and/or computational results for phase formability and materials functionality into the predictive process. This interpretive strategy guides the selection of new candidate materials and refines models in light of physicochemical understanding, enabling discoveries that pure algorithms might miss. Applied to proton-conducting oxides, this approach identified several compounds exhibiting its conductivity over 0.01 Scm⁻¹ at 300°C, a key threshold for practical application as a fuel cell electrolyte, and provided clear rationales for material selection. The MDI approach offers a flexible and interpretable pathway to accelerate materials innovation beyond the confines of existing data.