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 of new guiding principles is usually scarce in such studies, essentially due to the ‘black-box’ nature of the ML models. Therefore, we devised an intuitive method of interpreting such opaque ML models through SHapley Additive exPlanations (SHAP) values and coupling them with the HT approach for finding efficient 2D water-splitting photocatalysts. We developed a new database of 3099 2D materials consisting of metals connected to six ligands in an octahedral geometry, termed as 2DO (octahedral 2D materials) database. The ML models were constructed using a combination of composition and chemical hardness-based features to gain insights into the thermodynamic and overall stabilities. Most importantly, it distinguished the target properties of the isocompositional 2DO materials differing in bond connectivities by combining the advantages of both elemental and structural features. The interpretable ML regression, classification, and data analysis lead to a new hypothesis that the highly stable 2DO materials follow the HSAB principle. The most stable 2DO materials were further screened based on suitable band gaps within the visible region and band alignments with respect to standard redox potentials using the GW method, resulting in 21 potential candidates. Moreover, HfSe2 and ZrSe2 were found to have high solar-to-hydrogen efficiencies reaching their theoretical limits. The proposed methodology will enable materials scientists and engineers to formulate predictive models, which will be accurate, physically interpretable, transferable, and computationally tractable.