A comprehensive machine learning strategy for designing high-performance photoanode catalysts

Abstract

Engineering photoelectrodes with cocatalysts is an effective approach to enhance their photocatalytic performance. However, selecting an appropriate cocatalyst for a specific photoelectrode is challenging due to various factors within the system. Machine learning techniques are revolutionizing the catalytic field and are good at addressing multifactorial problems. In this study, several fundamental factors of the photoanode catalytic system and their specific mechanisms were systematically summarized. A comprehensive machine learning process was introduced to guide cocatalyst selection for BiVO₄ photoanodes. A multi-layer perceptron neural network and several tree-based ensemble models were trained to capture intricate relationships among photoanodes, cocatalysts, and electrolytes, enabling the prediction of unstudied cases. The best-performing random forest model accurately captured the essential features of high-performance cases, achieving a generalization accuracy of 96.30% for binary classification. Furthermore, the model's built-in feature importance analysis revealed that the type and preparation method of cocatalysts were the two most significant factors affecting the catalytic system's performance. According to the Shapley additive explanations interpretation, some heuristic rules were provided to propose a class of promising cocatalyst/photoanode catalysts.