Novel machine learning potential developed to understand materials deformation

-

-

Temperature phase diagram of tin comparing the results from the current potential (lines) to the experimental data available (points). 

-

To design materials for extreme applications, understanding and predicting phase transitions and their influence on material properties under high pressures and temperatures is key. Atomic-scale modeling is a useful tool for assessing these behaviors; however, its accuracy depends on the precision of interatomic potentials — that is, information regarding the interaction of the atoms that make up the material.

In Physical Review Materials, a team of Los Alamos researchers and their university colleagues present a hybrid potential, one that combines the speed of conventional embedded atom method with the accuracy of machine learning. The result is a one-of-a-kind potential that accurately reproduces the pressure-temperature phase diagram of tin, making the hybrid potential the only interatomic potential that can quantitatively replicate phases transitions in this material under high pressure.

The researchers used a neural-network-based approach to train the machine learning potential using density functional theory calculations. Tin was selected given its complex phase system and its structural, electrical and thermodynamic properties, which have interesting current and potential technological applications in areas such as lithium-ion batteries, solar photovoltaics, hydrogen generation and more.

The researchers noted that this potential demonstrates the necessity of adding an underlying physics model to the machine learning potentials to create a stable interatomic potential. The inclusion of a physics model represents a difference from the general approach, which does not use subject matter expert judgement or underlying materials science to train potentials. Using the new, hybrid approach not only ensures a more accurate potential but also reduces the training data-set size.

The researchers’ method is generally applicable to other elements and multicomponent systems that require a high level of accuracy and can confidently be used to gain insights into deformation mechanisms of complex materials.

Funding and mission

This work was funded by the Advanced Technology Development and Mitigation (ATDM) project within the Advanced Simulation and Computing (ASC) Program.

Reference

“Hybrid interatomic potential for Sn,” Physical Review Materials, 7, 043601 (2023); DOI: 10.1103/PhysRevMaterials.7.043601. Authors: Mashroor S. Nitol, Michael I. Baskes, Khanh Dang, Saryu J. Fensin (Los Alamos National Laboratory); Doyl. E Dickel, Christopher D. Barrett (Mississippi State University).

Technical contact: Saryu Fensin (MPA-CINT)