In this study, we identified features with the largest contributions and property trends in predicting the adsorption energies of carbon, hydrogen, and oxygen adsorbates on transition metal (TM) surfaces by performing Density Functional Theory (DFT)-based calculations and Machine Learning (ML) regression models. From 26 monometallic and 400 bimetallic fcc(111) TM surfaces obtained from Catalysis-hub.org, three datasets consisting of fourteen elemental, electronic, and structural properties were generated using DFT calculations, site calculations, and online databases. The number of features was reduced using feature selection and then finely-tuned random forest regression (RFR), gaussian process regression (GPR), and artificial neural network (ANN) algorithms were implemented for adsorption energy prediction. Finally, model-agnostic interpretation methods such as permutation feature importance (PFI) and shapely additive explanations (SHAP) provided rankings of feature contributions and directional trends. For all datasets, RFR and GPR demonstrated the highest prediction accuracies. In addition, interpretation methods demonstrated that the largest contributing features and directional trends in the regression models were consistent with structure-property-performance relationships of TMs like the d-band model, the Friedel model, and higher-fold adsorption sites. Overall, this interpretable ML-DFT approach can be applied to TMs and their derivatives for atomic adsorption energy prediction and model explainability.
Read full abstract