A universal graph deep learning interatomic potential for the periodic table

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Interatomic potentials (IAPs) are crucial for atomistic simulations but existing models lack general applicability.
• Current IAPs are often limited to specific chemistries or lack the required accuracy for broad use.

Purpose of the Study:

• To develop a universal interatomic potential for materials science applications.
• To create a machine learning-based IAP capable of handling diverse chemical spaces and predicting material properties accurately.

Main Methods:

• Developed M3GNet, a graph neural network-based interatomic potential incorporating three-body interactions.
• Trained M3GNet on a large dataset of structural relaxations from the Materials Project.
• Applied M3GNet for screening hypothetical crystal structures and predicting material stability.

Main Results:

• M3GNet demonstrated broad applicability in structural relaxation, dynamic simulations, and property prediction.
• Screening of 31 million hypothetical structures identified 1.8 million potentially stable materials using M3GNet energies.
• Density functional theory (DFT) calculations verified the stability of 1,578 out of the top 2,000 lowest-energy materials.

Conclusions:

• M3GNet provides a universal and accurate interatomic potential for diverse materials.
• Machine learning accelerates the discovery of new, stable, and potentially synthesizable materials.
• This approach offers a pathway to discovering materials with exceptional properties.