Machine learning-assisted screening and optimization of LDH-based hydrogen production photocatalysts

Abstract

To address the large compositional design space, pronounced experimental heterogeneity, and the difficulty of rapidly revealing structure–property–performance relationships for LDH-based hydrogen-evolution photocatalysts through conventional trial-and-error approaches, a literature-derived structured database was established, covering material composition, preparation conditions, structural parameters, and reaction conditions. A total of 1,800 samples and 23 initial descriptors were collected, and 18 effective features were retained after mRMR ranking and Pearson correlation analysis. Four algorithms—gradient boosting regression (GBR), support vector regression (SVR), back-propagation artificial neural network (BPANN), and multiple linear regression (MLR)—were further compared for the prediction of band gap, specific surface area, crystallite size, and hydrogen-evolution rate. The results show that GBR was the optimal model for band gap and crystallite size prediction, with R values of 0.7436 and 0.8368, respectively, whereas SVR was the best model for specific surface area prediction, achieving R = 0.8879. For hydrogen-evolution prediction, the optimized GBR model improved R from 0.9036 to 0.9142 and reduced RMSE from 528.4 to 462.8. The relationship analysis indicates that superior hydrogen-evolution activity is usually associated with a moderate band gap, relatively large specific surface area, and small-to-moderate crystallite size. Virtual screening further identified composite/heterostructure systems such as WS2/MgAl-LDH, 15 wt% Cu2O/CoAl-LDH, Cu3P/NiP2/NiAl-LDH, and 20 wt% NiAl-LDH/Mn0.5Cd0.5S as promising candidates. This work provides a data-driven route for the rapid screening and optimized design of LDH-based photocatalysts.