Machine Learning Interatomic Potentials for Reactive Hydrogen Dynamics at Metal Surfaces Based on Iterative...

Area of Science:

• Surface science
• Computational chemistry
• Materials science

Background:

• Molecular hydrogen's surface reactions are vital for energy storage and fuel cells.
• Accurate theoretical prediction of these reactions is computationally demanding.
• Machine learning potentials offer promise but require robust data generation and uncertainty assessment.

Purpose of the Study:

• To develop and apply an ensemble learning approach with uncertainty quantification for gas-surface dynamics.
• To investigate the performance of SchNet and PaiNN models for hydrogen scattering on copper.
• To assess the impact of model uncertainty on predicting reaction probabilities.

Main Methods:

• Adaptive training data generation using ensemble learning.
• Full uncertainty quantification (UQ) for reaction probabilities.
• Application and comparison of SchNet and PaiNN message-passing neural networks.

Main Results:

• Ensemble-based UQ identified limitations in SchNet's invariant feature representation for gas-surface dynamics.
• Iterative refinement of training data improved model reliability.
• The study provides a framework for robust machine learning in surface reaction dynamics.

Conclusions:

• Ensemble learning with UQ is crucial for reliable machine learning in surface chemistry.
• PaiNN shows better performance than SchNet for this specific gas-surface dynamics problem.
• Further development is needed for feature representations in machine learning potentials for surface reactions.