Data-driven design of porphyrin photocatalysts using deep transfer learning from quantum chemical databases

Metalloporphyrins and Porphyrins (MpPs) have received increasing attention as potential photocatalysts for hydrogen and oxygen evolution and the reduction of carbon dioxide (CO2). Recent studies have found that both HOMO/LUMO energy levels and energy gaps are important factors controlling the performance of MpPs photocatalysts. Since pure quantum chemistry methods consume a large amount of computational resources and cannot learn directly from previously accumulated data, in this study, we proposed a deep transfer learning approach to rapidly predict the HOMO/LUMO energy levels and energy gaps based on two quantum chemistry databases. To complement the open-source Porphyrin-based Dyes Database (PBDD), which contains porphyrin structures designed as photosensitizers, we curated a new database, the Metalloporphyrins and Porphyrins Database (MpPD), in which MpPs were specifically designed as potential photocatalysts. We also adapted two state-of-the-art chemical deep learning models to accommodate more complex systems of MpPs. The PorphyBERT model and PorphyDMPNN model, both pretrained with PBDD and fine-tuned with MpPD, performed satisfactorily in predicting HOMO/LUMO level and energy gap. We concluded the paper by recommending 14 MpPs as potential photocatalysts for CO2 reduction, whose structures have similar value in out-of-sample model prediction and DFT calculations.