Machine learning aided design of high performance copper-based sulfide photocathodes

Abstract

Copper-based sulfide photocathodes have shown impressive performance in solar water splitting applications due to their narrow bandgaps, high absorption coefficients, and good carrier transport properties. Several factors, such as composition, thickness, and doping, have a direct influence on the onset potential, photocurrent density, and solar-to-hydrogen efficiency. Screening for the optimal combination in the presence of multiple variables is undoubtedly a challenging task. However, constructing a comprehensive database, developing photocathode models, and utilizing machine learning to derive the best results clearly save a significant amount of experimental effort. This approach efficiently reduces the experimental workload, streamlines the process, and expedites the development of high-performance materials for photoelectrochemical water splitting applications. Here, we introduce a comprehensive machine learning process to guide the preparation of copper-based sulfide photocathodes. The random forest model was selected to train and capture the complex relationship between different layers of copper-based sulfide photocathodes and electrolytes to predict unstudied conditions, and the accuracy of the test set reached 96.7%. Through SHAP interpretability analysis, we provide heuristic rules to deepen the understanding of the influence of different factors on the performance of the catalytic system. We also developed a prediction platform to share our prediction models.