How to validate machine-learned interatomic potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Machine learning (ML) potentials are increasingly used for large-scale atomistic simulations.
• These potentials achieve near-quantum-mechanical accuracy by learning from reference data.
• Validation is crucial, especially for physically agnostic ML potentials.

Purpose of the Study:

• To review the principles of ML potentials for atomic-scale material modeling.
• To discuss best practices for validating ML potentials.
• To provide recommendations for the research community.

Main Methods:

• Review of ML potential principles.
• Discussion of numerical error metrics for validation.
• Exploration of physically guided validation strategies.

Main Results:

• ML potentials enable efficient and accurate atomistic simulations.
• Careful validation is essential for reliable application of ML potentials.
• Established best practices enhance the trustworthiness of ML potential predictions.

Conclusions:

• This review provides a framework for validating ML potentials in materials modeling.
• Recommendations are offered to ensure the robust application of ML potentials.
• The findings aim to support researchers using ML potentials for materials discovery and design.