Explainable artificial intelligence for materials discovery: application to catalysts for the HER and ORR

作者：Valentín Vassilev-Galindo, Javier LLorca · 发表于：Chemical Science · 年份：2025 · DOI：10.1039/d5sc06442b · 被引用次数：8 · 研究领域：Machine Learning in Materials Science、Catalysis and Oxidation Reactions、Inorganic Chemistry and Materials

calculations and machine learning (ML) has opened the door for both fast and accurate chemical/physical property predictions and for the virtual design of materials. However, these techniques are very often used as a "black box" with the sole objective of obtaining high accuracy with scarce or no special attention on how ML models obtain their predictions. This can be improved by leveraging explainability of ML models, which, at the same time, would increase the chance of ML to offer new insights into the chemistry and physics of materials. Hence, the next generation of ML models in these realms must guarantee explainability by embedding explainable artificial intelligence (XAI) tools into their pipelines. Specifically, ML-assisted materials discovery and design can take great advantage of the use of XAI. Enabling explanations would increase the impact of these approaches by providing not only a set of candidates, but also insights into what makes a given material better than others. With this in mind, using the example of heterogeneous catalysts for hydrogen production and energy generation, here we propose a novel strategy for materials design based on counterfactual explanations. We were able to find materials featuring properties close to the design targets that were later validated with density functional theory calculations. Explanations were devised by comparing original samples, counterfactuals, and discovered candidates. Such explanations allowed us to unveil subtle ...