Towards universal neural network potential for material discovery applicable to arbitrary combination of 45 elements

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Atomistic simulations

Background:

• Computational material discovery aims to explore vast chemical spaces.
• Neural network potentials (NNPs) are effective for atomistic simulations but often material-specific.
• Existing NNPs lack universality, limiting their application in broad material discovery.

Purpose of the Study:

• To develop a universal neural network potential (NNP) applicable to any combination of 45 elements.
• To overcome the limitations of narrow-target NNPs in computational material discovery.
• To provide a versatile tool for accelerating the discovery of new materials.

Main Methods:

• Development of a universal NNP named PreFerred Potential (PFP).
• Creation of diverse datasets including virtual structures to achieve universality.
• Application and validation of PFP across various material systems and phenomena.

Main Results:

• PFP demonstrates capability in simulating diverse systems, including lithium diffusion, molecular adsorption, alloy transitions, and catalyst discovery.
• The universal nature of PFP allows for handling any combination of 45 elements.
• The developed datasets are crucial for achieving the broad applicability of PFP.

Conclusions:

• The PreFerred Potential (PFP) offers a universal solution for NNP-based atomistic simulations.
• PFP significantly enhances the scope and efficiency of computational material discovery.
• This technology represents a powerful and highly useful tool for researchers in materials science.