Integrating Machine Learning and High‐Throughput Calculations for the Rational Design of Photocatalytic 2D‐COFs on Overall Water Splitting

Abstract

Two‐dimensional covalent organic frameworks (2D‐COFs) with tunable structures offer a promising platform for photocatalytic overall water splitting (OWS) under visible‐light. However, it is challenging to quickly find photocatalytic OWS materials from the massive amounts of 2D‐COFs and then achieve precise synthesis. In this work, we constructed 11 934 hcb‐type 2D‐COFs by assembling 28 building blocks and 9 linkages. The machine learning (ML) and high‐throughput computation (HTC) were integrated to predict feasible photocatalytic OWS hcb‐type 2D‐COFs. Through training 10 initial algorithms and optimizing the hyperparameters of top 4 algorithms in terms of performance, the ETR model for predicting the band‐edge levels with R 2 of 0.97 and 0.99 was developed, and the KNR and RFR models were built for predicting Δ G *H and Δ G *OH with R 2 of 0.99 and 0.83, respectively. 2581 2D‑COFs (21.63% of dataset) are screened to be potential structures for visible‑light‑driven water splitting. After applying the optimal ML models on all assembled 2D‐COFs, a list of high‐frequency building blocks and linkages is suggested as a set of suitable candidates for the first time. Then TBTZ_FBN0_Imine (TBTZ‐FBN0‐COF) was assembled and experimentally synthesized. Its OWS activity verified our ML‐HTC paradigm, which provides the researchers recommendation to construct photocatalytic OWS 2D‐COFs.