Machine-Learning-Driven Photocurrent Prediction in Multielement-Doped Hematite Photoelectrodes

Abstract

Reliably predicting heterogeneous photocatalyst efficiency using machine learning (ML) remains challenging, because of variations in synthesis protocols and evaluation methods. To address this, we developed ML models to predict the photocurrent density, which is an indicator of the hydrogen generation rate in photoelectrochemical (PEC) water splitting. For multielement-doped hematite photoelectrodes, which have demonstrated potential for enhancing solar energy conversion, we acquired over 2000 data files uniformly and systematically from approximately 100 samples. A high predictive accuracy with a coefficient of determination (R2) of 0.817 and a root mean squared error (RMSE) of 0.105 mA cm-2 was achieved using elemental features and Raman spectra as explanatory variables. By analyzing this ML model, we identified both semiconductor properties and device characteristics as key factors for achieving high performance, particularly valence-related information, electron delocalization, and particle density on the substrate. Finally, we extrapolated the photocurrent densities for 79 800 virtual hematite samples doped with two or three elements using our high-precision ML model, revealing elements that could potentially enhance or reduce the PEC performance, while also predicting the performance of masked samples accurately. These findings establish a robust framework for data-driven materials discovery and rational dopant selection, thereby facilitating the optimization of hematite-based photoelectrodes.