Multi-feature deep learning framework for predicting CO adsorption mechanisms at metal oxide interfaces: a...

Area of Science:

• Materials Science
• Computational Chemistry
• Catalysis

Background:

• Predicting CO adsorption mechanisms on metal oxides is crucial for designing efficient catalysts.
• Traditional methods like DFT calculations are computationally expensive and time-consuming.
• Developing rapid screening tools for catalyst design is essential for sustainable catalysis.

Purpose of the Study:

• To present a novel multi-feature deep learning framework for predicting CO adsorption mechanisms at single metal oxide interfaces.
• To develop a computationally efficient alternative to DFT for catalyst screening.
• To provide mechanistic insights into CO adsorption processes.

Main Methods:

• Integration of Transformer architecture with molecular descriptors (structural, electronic, kinetic).
• Utilization of cross-feature attention mechanisms for capturing catalytic process complexity.
• Development of specialized encoders for different descriptor types.

Main Results:

• The framework achieved high accuracy in predicting adsorption energies (MAE < 0.12 eV) and correlation coefficients (> 0.92) across seven metal oxide systems.
• Demonstrated superior performance compared to traditional machine learning methods.
• Ablation studies highlighted the critical role of structural information in the predictions.

Conclusions:

• The multi-feature deep learning framework provides a rapid and accurate method for predicting CO adsorption mechanisms.
• The model successfully captures material-specific mechanisms, coverage-dependent effects, and defect influences.
• This approach lays the foundation for data-driven catalyst design and mechanism elucidation in sustainable catalysis.