Atomic Adsorption Energies Prediction on Bimetallic Transition Metal Surfaces Using an Interpretable Machine...

Area of Science:

• Materials Science
• Computational Chemistry
• Surface Science

Background:

• Predicting adsorbate adsorption energies on transition metal (TM) surfaces is crucial for catalyst design.
• Density Functional Theory (DFT) calculations are computationally expensive for large-scale screening.
• Machine Learning (ML) offers a promising avenue for accelerating these predictions.

Purpose of the Study:

• To identify key features and property trends that predict adsorption energies of carbon, hydrogen, and oxygen on TM surfaces.
• To develop and interpret accurate ML models for adsorption energy prediction.
• To establish an interpretable ML-DFT approach for TM surface studies.

Main Methods:

• Generated datasets from 26 monometallic and 400 bimetallic fcc(111) TM surfaces using DFT calculations.
• Extracted fourteen elemental, electronic, and structural properties.
• Employed feature selection and applied Random Forest Regression (RFR), Gaussian Process Regression (GPR), and Artificial Neural Network (ANN) models.
• Utilized model-agnostic interpretation methods like Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP).

Main Results:

• RFR and GPR models achieved the highest prediction accuracies across all datasets.
• Interpretation methods identified key features and directional trends consistent with established TM structure-property-performance relationships (e.g., d-band model, Friedel model).
• Higher-fold adsorption sites were found to be significant predictors.

Conclusions:

• The interpretable ML-DFT approach effectively predicts atomic adsorption energies on TMs.
• This methodology enhances model explainability, providing insights into adsorption mechanisms.
• The approach is applicable to TMs and their derivatives for accelerated materials discovery in catalysis.