One to Rule Them All: A Universal Interatomic Potential Learning across Quantum Chemical Levels

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning interatomic potentials are increasingly used, but datasets often vary in quantum chemical (QC) levels.
• A scalable framework for learning across diverse chemical spaces and QC levels is needed.

Purpose of the Study:

• To develop a universal framework for simultaneous learning across arbitrary QC levels.
• To present an all-in-one approach as an alternative to transfer learning for interatomic potentials.

Main Methods:

• Proposed an all-in-one machine learning strategy for interatomic potentials.
• Created OMNI-P1, a universal interatomic potential trained on multiple QC levels.
• Utilized OMNI-P1's prediction capabilities for generating correction terms in Δ-learning models.

Main Results:

• OMNI-P1 demonstrates generalization comparable to GFN2-xTB and DFT methods but with significantly higher speed.
• The approach enables simultaneous learning and prediction across various QC levels.
• Generated the Ω-ωB97X-D4 method by correcting DFT ωB97X-D4 using OMNI-P1, achieving superior accuracy.

Conclusions:

• The all-in-one strategy provides a more general and user-friendly alternative for interatomic potential development.
• OMNI-P1 is the first universal interatomic potential capable of learning and predicting at multiple QC levels.
• This work facilitates accurate and efficient molecular modeling across different computational chemistry standards.