Synthesis of covalent organic frameworks for photocatalytic hydrogen peroxide production guided by large language models

The photosynthetic production of hydrogen peroxide (H2O2) from water and oxygen presents a sustainable alternative to the energy-intensive anthraquinone process. Covalent organic frameworks (COFs) have emerged as promising photocatalysts for H2O2 generation. However, most existing COF photocatalysts yield H2O2 at concentrations too low for practical applications, largely due to ongoing challenges in simultaneously optimizing photocatalytic activity and structural stability. Here, we introduce a large language model-driven design strategy for the targeted synthesis of high-performance COF photocatalysts. By analyzing a curated corpus of 355 peer-reviewed articles on COF-based photocatalysis with a language model-driven knowledge extraction pipeline, we extract and structure over 11,000 chemical relationships related to building block identity, linkage robustness, and H2O2 yield. Guided by this artificial intelligence-derived knowledge base, we identify 4,4′,4″-(1,3,5-triazine-2,4,6-triyl)trianiline and benzo[1,2-b:3,4-b′:5,6-b″]trithiophene-2,5,8-tricarbaldehyde as optimal building blocks and thiazole as the preferred linkage motif for constructing a robust, photocatalytic COF. The resulting Thz-COF achieve a high H2O2 concentration of 82.3 mM (~0.28 wt%) in aqueous solution (without using sacrificial agents), with a solar-to-chemical energy conversion efficiency of 1.39%. The development of covalent organic framework (COF) photocatalysts for high-concentration H2O2 production remains challenging. Here, the authors report a large language model–driven design strategy to construct robust and active COF photocatalysts capable of producing high concentrations of H2O2.