Prediction of adsorption energies of C_mH_nO_p (m ≤ 2, n ≤ 6, p ≤ 2) on transition metals and alloys with machine learning methods.

This article focuses on developing and evaluating machine learning (ML) models to predict adsorption energies of 57 chemical species involved in the hydrogenation of CO₂ to ethanol on seven pure face-centered cubic (FCC) metals and twelve binary FCC alloys. Using surface, adsorbate, and adsorption site features derived from quantum chemical calculations, the study compares several ML methods—including sure independence screening and sparsifying operator (SISSO), multilayer perceptron regression (MLPR), random forest regression (RFR), kernel ridge regression (KRR), support vector regression (SVR), eXtreme Gradient Boosting (XGBoost), and various ensemble models—to identify the most accurate predictors. The results show that ensemble ML models, particularly the combination of KRR, MLPR, and XGBoost, outperform individual models, achieving a mean absolute error of 0.03 eV and a maximum error of 0.17 eV, comparable to density functional theory (DFT) calculation errors. Feature importance analysis highlights the condensed local softness of the adsorbate's bonding atom as the key descriptor correlating linearly with adsorption energies, especially for C₁ species, suggesting its significance in catalyst screening for CO₂ hydrogenation.