Machine learning-accelerated discovery of covalent organic frameworks for hydrogen peroxide photosynthesis

Abstract Covalent organic frameworks (COFs) are promising photocatalysts for hydrogen peroxide (H 2 O 2 ) production, yet their rational design remains challenging. Although machine learning has advanced the prediction of properties of porous materials, its application to COF-based photocatalysis faces two major challenges: the representation of multilevel structural features and the limited availability of training datasets. Here we present a comprehensive computational framework, termed ‘information co-evolution’, that accelerates the discovery of efficient COF structures for H 2 O 2 photosynthesis. This framework integrates two pathways: to mitigate data limitations, we introduce data augmentation techniques and ensemble modelling; concurrently, to address the structural encoding challenge, we introduce a cross-level feature fusion strategy that integrates these fragment descriptors with mechanism-driven physical descriptors. These strategies collectively reduced the validation root mean square error from 4.70 to 3.31. Among over 10,000 candidates, our framework can successfully identify high-performance COFs for H 2 O 2 photosynthesis, for example, COF-343 achieves a H 2 O 2 photosynthetic rate of 12,978.7 μmol h −1 g −1 . The model interpretation further unveiled critical structural motifs, offering information for the rational design of COF photocatalysts beyond traditional trial-and-error methods.