Chemical hardness-driven interpretable machine learning approach for rapid search of photocatalysts

Abstract Strategies combining high-throughput (HT) and machine learning (ML) to accelerate the discovery of promising new materials have garnered immense attention in recent years. The knowledge guiding principles is usually scarce such studies, essentially due ‘black-box’ nature ML models. Therefore, we devised an intuitive method interpreting opaque models through SHapley Additive exPlanations (SHAP) values coupling them with HT approach for finding efficient 2D water-splitting photocatalysts. We developed a database 3099 consisting metals connected six ligands octahedral geometry, termed as 2DO (octahedral materials) database. were constructed using combination composition chemical hardness-based features gain insights into thermodynamic overall stabilities. Most importantly, it distinguished target properties isocompositional differing bond connectivities by advantages both elemental structural features. interpretable regression, classification, data analysis lead hypothesis that highly stable follow HSAB principle. most further screened based on suitable band gaps within visible region alignments respect standard redox potentials GW method, resulting 21 potential candidates. Moreover, HfSe 2 ZrSe found high solar-to-hydrogen efficiencies reaching their theoretical limits. proposed methodology will enable scientists engineers formulate predictive models, which be accurate, physically interpretable, transferable, computationally tractable.