Response to [Reassessing Machine Learning Techniques for Electrocatalyst Design: A Call for Robust Methodologies]

ABSTRACT This article is a response to the comment “Reassessing Machine Learning Techniques for Electrocatalyst Design: A Call for Robust Methodologies”. First, we clarify that the artificial neural network–SHapley Additive exPlanation (ANN–SHAP) method mentioned in the comment originates from the original work of Ding et al., which we only briefly summarized. In that study, nine different machine learning models were employed to predict the performance of proton exchange membrane fuel cells, among which the ANN model performed best. SHAP, together with multiple interpretability techniques (PDP, Tree‐based Rule, EIX, etc.), was used to cross‐validate feature importance, which was further compared with the results from manual feature selection, PCA, and t‐distributed stochastic neighbor embedding, and complemented by experimental validation to reduce the risk of bias amplification. We agree with the commenter that model interpretability should be approached with caution, as the absence of a definitive “ground truth” for feature importance remains a current challenge. However, benchmarking SHAP explanations against domain knowledge or validating them using synthetic datasets can help reduce the risk of misinterpretation. Regarding the unsupervised methods suggested in the comment (FA and HVGS), we consider them to have exploratory value for certain data structures, but caution is needed when applying them to experimental systems involving nonlinearity or high noise.