Active Learning‐Driven Discovery of Donor‐Acceptor Covalent Triazine Frameworks for High‐Performance Photocatalysts

Abstract Donor‐acceptor (D‐A) structure enables precise tuning of the electronic and optical properties of materials, enabling widely applicable in organic semiconductors and photocatalysts. However, the vast diversity of donor and acceptor units and their combinations pose considerable challenges to experimental development. Here, this study presents a screening strategy that integrates an active learning (AL)‐based multi‐model framework with experimental synthesis validation to discover high‐performance D‐A covalent triazine frameworks (CTFs) photocatalysts. This framework combines an AL model, trained on experimental data of reported D‐A‐CTFs, with a graph neural networks model that establishes the relationship between molecular structure and electronic properties. Meanwhile, expert chemical knowledge is incorporated into this multi‐model framework to improve the synthesizability and stability, resulting in 113 identified candidates from a database of 21807 structures. Experimental validation confirms that 9 out of 10 newly synthesized D‐A‐CTFs exhibit the predicted photocatalytic performances. Notably, CTF‐[1,1′‐Biphenyl]‐4,4′‐dicarbaldehyde achieved a record hydrogen evolution rate of 33.29 mmol g −1 h −1 for CTF‐based bulk photocatalysts. Further feature engineering analysis reveals that carbon and nitrogen charges critically determine the photocatalytic performance, offering an optimization strategy for D‐A‐CTFs design. This study paves a promising way to accelerate the discovery of effective D‐A structured materials.