NEP89: universal neuroevolution potential for inorganic and organic materials across 89 elements

Area of Science:

• Computational Materials Science
• Artificial Intelligence in Chemistry
• Atomistic Simulations

Background:

• Machine-learned interatomic potentials offer high accuracy but are often material-specific or computationally expensive.
• Existing models limit the scope and scale of atomistic simulations.

Purpose of the Study:

• Introduce NEP89, a novel foundation model for atomistic simulations.
• Achieve near-empirical-potential speed with high accuracy across a wide range of elements.
• Enable large-scale simulations for diverse inorganic and organic systems.

Main Methods:

• Developed NEP89 using a neuroevolution potential architecture.
• Curated a compact, comprehensive training dataset via descriptor-space subsampling and iterative refinement.
• Benchmarked NEP89 against existing foundation models for accuracy and computational efficiency.

Main Results:

• NEP89 demonstrates competitive accuracy compared to other foundation models.
• Achieved computational efficiency three to four orders of magnitude greater than existing models.
• Successfully applied NEP89 to diverse large-scale simulations, including alloy compression and ion diffusion.

Conclusions:

• NEP89 offers a significant advancement in atomistic simulations, balancing speed and accuracy.
• Its broad applicability and fine-tuning capabilities make it a versatile tool for materials science research.
• Enables previously intractable large-scale simulations across inorganic and organic materials.