(685f) Deciphering the Roots of Catalyst Degradation in Electrochemical CO2 Reduction Via Interpretable AI

The catalyst degradation poses a substantial obstacle for the commercialization of electrochemical CO₂ reduction reactions, as it leads to diminished activity and selectivity. Nevertheless, the considerable difficulties and expesive experimental cost of catalyst assessment make the electrochemical catalyst degradation research underpeformed. Machine learning has demonstrated its potential to supplant costly procedures in recent years, but the limited interpretability of these models complicates their implementation. In this study, we present an explainable machine learning system capable of accurately predicting catalyst conditions using linear sweep voltammetry in a sub-seconds, while also offering insights into degradation mechanisms. A comprehensive database consisting of 5236 linear sweep voltammetry experiments conducted under diverse conditions was also provided. A convolutional neural network, trained on this dataset, shows its capability of predicting total current and faradaic efficiency. The model's acquired physical and chemical rationale is elucidated through explainable artificial intelligence interpretation, which identifies crucial degradation factors. We subsequently conduct surface analysis experiments to corroborate the explainable artificial intelligence interpretation and validate the dependability of the suggested framework.