Discovering the Origin of Catalyst Performance and Degradation of Electrochemical CO₂ Reduction through Interpretable Machine Learning

Catalyst degradation is a significant challenge for the commercialization of the electrochemical reduction of CO₂, as it decreases activity and selectivity. However, the high experimental cost of catalyst characterization hinders the generation of sufficient and valuable information regarding catalyst degradation. Recently, machine learning (ML) models have exhibited high potential to replace costly processes, but their low interpretability makes their application challenging. Herein, we introduce an interpretable ML framework that accurately projects the catalyst status using simple linear sweep voltammetry (LSV) within subseconds while providing insights into the origin of catalyst degradation. A convolutional neural network trained on experimentally collected 5196 LSV results achieved superior performance in total current and Faradaic efficiency predictions. The ML framework demonstrates an impressive accuracy of mean absolute error below 0.5% in predicting the Faradaic efficiency of various products, irrespective of the operating conditions and catalyst types. The prediction mechanism learnt by the model was interpreted via explainable artificial intelligence (XAI), and critical degradation factors were identified. We performed catalyst surface analyses at milestone points to verify the XAI interpretation and demonstrate the reliability of the proposed framework. This approach can potentially be applied to a wide range of electrochemistry involving catalytic process, battery degradation, and chemical process monitoring, suggesting that it offers a viable means of rapidly and reliably monitoring performance.