Hitchhikers Guide To Training More General Machine Learning Potentials in Heterogeneous Catalysis

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• Machine learning potentials (MLPs) are advanced simulation tools for heterogeneous catalysis.
• Current methods focus on system-specific applications, limiting generalizability.
• Training general MLPs requires diverse, nonequilibrium data, presenting unique challenges.

Purpose of the Study:

• To investigate challenges and pathologies in training general MLPs using diverse datasets.
• To provide practical recommendations for developing transferable MLPs for catalysis.

Main Methods:

• Systematic investigation of training pathologies using the REICO method for data generation.
• Analysis of deviations from standard system-specific MLP training protocols.
• Evaluation of data cleaning, model selection, and error metrics for diverse datasets.

Main Results:

• Identified critical deviations and pathologies when training MLPs on diverse, nonequilibrium data.
• Demonstrated challenges in standard training discipline and evaluation logic for general MLPs.
• Highlighted the need for specialized approaches beyond system-specific workflows.

Conclusions:

• Training general and transferable MLPs requires addressing unique challenges associated with diverse data.
• Recommendations provided for data cleaning, model selection, and error metrics enhance MLP robustness.
• Guidance offered for developing reliable MLPs applicable to a wider range of catalytic systems.