Machine Learning with Interpretable Local Descriptors for Hydrogen Evolution Reaction on Borophene: Prediction and Physical Insights

The local environment of an adsorption site is crucial in determining the adsorption energy. Understanding these local structural characteristics is thus key to optimizing adsorption properties, and this can be enhanced through machine learning (ML) models. In this study, we developed a compact set of five local descriptors (A, B, C, α, and β) that efficiently capture the structural features of the nearest neighbors (NNs) of around 18 different adsorption sites across six different borophene phases under varying external in-plane strain. Using an expanded data set of 131 data points and a fine-tuned ML model, we successfully mapped the descriptors to the hydrogen adsorption free energy (ΔG_H), achieving a low cross-validation MAE of 0.05 eV and a high R² of 0.95. Interpretation with SHapley Additive ExPlanations (SHAP) and radar charts revealed that the structure and coordination of the second nearest neighbors (B and β) play a dominant role in determining ΔG_H. Intermediate values of B and β are identified as optimal for HER, whereas extreme descriptor values lead to anomalous ΔG_H, which appear to be associated with outlier-like behavior in unrelaxed vacancy formation energy and adsorption charge density distortion cost. The insights from this study enhance the understanding of adsorption in borophene and demonstrate the effectiveness of compact local descriptors for adsorption studies.