Interpretable Machine Learning Unveils Hydroxyl/Amino Synergy and Guides Discovery of Optimal MOF Photocatalysts for Hydrogen Evolution

Abstract

Abstract Metal–organic frameworks (MOFs) are premier platforms for photocatalytic hydrogen evolution (PHER), yet navigating their multidimensional parameter space typically relies on inefficient trial-and-error approach. While machine learning (ML) can accelerate discovery, it is often hindered by ″black-box″ predictions that lack mechanistic transparency and experimental validation. Herein, we establish an interpretable ML-to-experimental framework for rational MOF engineering. By training a CatBoost model on a curated database and employing SHapley Additive Explanations (SHAP), we deconstructed the hierarchical influence of ligand motifs on catalytic activity. This revealed the cooperative effect of hydroxyl and amino dual functionalization, which optimizes the electronic landscape through balanced bandgap dynamics and hard–soft acid–base (HSAB) matching. Guided by these insights, we synthesized benzophenanthrene-based mixed-ligand MOFs. The champion catalyst achieved a peak HER rate of 73.7 mmol g–1 h–1─without external photosensitizers or cocatalysts─exhibiting a 4.18% deviation from algorithmic predictions and a 15.8% enhancement over the top of the data set. This work develops a high-performance photocatalytic system and provides a generalizable, interpretable paradigm for data-driven discovery of advanced energy materials.