A Machine Learning and Explainable AI Framework for Adsorption Energy Prediction via Effective Representation of the Adsorbate Chemical Environment

Adsorption energies, which capture the interactions between adsorbates and solid surfaces, are central to heterogeneous catalysis. Machine learning (ML) offers a powerful approach for rapidly and accurately predicting adsorption energies from computational data, thereby accelerating catalyst screening. The effectiveness of ML models depends on features that accurately represent the chemical environments of atoms, incorporating both geometric and electronic properties that influence adsorbate-surface interactions. In this study, we present an ML framework that leverages advanced electronic structure descriptors via Gaussian Multipole (GMP) featurization. GMP approximates electron density using Gaussian basis functions, providing a novel representation of elemental identity. Combined with robust geometric features, our model predicts CO and H binding energies (∆ECO* and ∆EH*) on multimetallic alloys, achieving mean absolute errors (MAEs) of 0.07 eV for ∆ECO* and 0.06 eV for ∆EH*. To interpret the model's predictions, we applied SHAP (SHapley Additive exPlanations), a post hoc explainable AI (XAI) method. The analysis revealed that GMP features associated with adsorbates and their first-nearest neighbours (FNNs) played the most significant role in determining binding energies, while features from second-nearest neighbours (SNNs) had minimal influence. Additionally, broader elemental properties-such as boiling point, group number, and atomic number-were found to be more predictive of adsorption behavior than conventional features like electronegativity. Clustering and t-SNE analyses showed that similar FNN environments yield consistent binding energies, supporting the model's ability to generalize. Overall, this work demonstrates that integrating electronic structure features with explainable AI improves both predictive accuracy and interpretability, offering a powerful strategy for accelerated catalyst screening and rational catalyst design.