Abstract
In this work, we introduce a combined DFT and machine learning approach to obtain insights into the chemical design of metal-organic framework (MOF) photocatalysts for hydrogen (HER) and oxygen (OER) evolution reactions. To train our machine learning models, we evaluated a dataset of 314 MOFs using a dedicated DFT workflow that computes a set of five descriptors for both closed and open shell MOFs. Our dataset is composed of a diverse selection of the QMOF database and experimentally reported MOF photocatalysts. In addition, to ensure a balanced dataset, we designed a set of MOFs (CDP-MOF) inspired by insights obtained regarding different types of photocatalytic materials. Our machine-learning approach allowed us to screen the entire QMOF and CDP-MOF databases for promising candidates. Our analysis of the chemical design space shows that we have many materials with a suitable spatial overlap of electron and hole, band gap, band-edge alignment to HER, and charge-carrier effective masses. However, we have identified in the QMOF database only a very small percentage of materials that also have the right band-edge alignment to OER. With the CDP-MOF database, we successfully targeted building blocks that potentially have the correct OER band alignment, and indeed obtained a larger percentage of materials that obey these criteria. Among those, a few motifs stood out, such as Au-pyrazolate, Ti clusters and rod-shaped metal nodes, and a particular MOF designed with the Mn(4)Ca cluster, which mimics the OER center in the photosystem II of photosynthesis.