Accelerating discovery through integration: a DFT validated machine learning framework for screening MOF photocatalysts

Abstract

The discovery of Metal–Organic Framework (MOF) photocatalysts for CO₂ reduction is hindered by the computational cost of quantum chemical screenings. To overcome this barrier, we introduce a Machine Learning (ML)-accelerated workflow that integrates the speed of ML with the accuracy of Density Functional Theory (DFT). While a DFT-based screening of over 20 000 MOFs identified 105 promising candidates in nearly a month, a ML-driven approach using the Molecular Graph Transformer (MGT) required only 4.5 hours. Here, we present a quantitative assessment of ML performance compared with hybrid DFT for MOF electronic screening, showing that prediction errors are related to the chemistry of the MOFs. We therefore derive an error-aware ML candidate selection strategy that raises DFT candidate recovery from 20% to 70% while keeping a sensible selection set. Building on this, we propose a practical ML to DFT screening workflow in which ML serves as a fast pre-filter to define a small subset for hybrid DFT evaluation, enabling efficient discovery of promising MOFs.