Machine Learning Powers Up Atomic-Level Modeling

The capacity to model atomic-level system behavior offers a potent tool for a variety of applications, including materials research and drug development. Researchers at Los Alamos National Laboratory have led a team in developing machine-learning interatomic potentials that predict forces and molecular energies acting on atoms. These potentials allow for faster and less expensive simulations than those made possible by current computational methods.

Machine learning potentials increasingly offer an effective alternative to computationally expensive simulations that try to represent complex physical systems on the atomic scale. A general reactive machine learning interatomic potential, applicable to a broad range of reactive chemistry without the need for refitting, will greatly benefit chemistry and materials science.

Benjamin Nebgen, Chemical Physicist and Study Co-Author, Los Alamos National Laboratory

The study was published in the journal Nature Chemistry.

Bridging the Gap in Effective Simulations

Traditionally, physics-based computational models such as quantum mechanics or classical force fields are used to construct efficient molecular dynamics simulations in chemistry. Quantum mechanical models are very costly to compute even though they are precise and widely applicable. In contrast, classical force fields are only applicable to a small range of systems and have relatively low accuracy despite being computationally efficient.

A long-standing gap in chemistry’s speed, accuracy, and generality is closed by the team's revolutionary machine learning model, ANI-1xnr. Machine learning is an artificial intelligence application in which computer programs are trained to “learn.”

ANI-1xnr is the first reactive machine-learning interatomic potential versatile enough to compete with physics-based computational models for large-scale reactive atomistic simulations. It can be applied to a wide range of chemical systems. Reactive molecular dynamics simulations over a broad range of chemical systems containing carbon, hydrogen, nitrogen, and oxygen elements were carried out using an automated workflow used to develop ANI-1xnr.

ANI-1xnr demonstrated the ability to investigate a wide variety of systems, including prebiotic chemistry, combustion, and carbon phase transitions. The simulations were verified by the team by comparing experimental results with traditional computational methods.

A Transformational Interatomic Potential

ANI-1xnr does not require domain expertise or refitting for every new use case, enabling scientists from a diverse range of domains to study unknown chemistry. The general applicability of ANI-1xnr is transformational, representing a significant step toward replacing the long-standing modeling techniques for studying reactive chemistry at scale.

Richard Messerly, Computational Scientist, and Study Co-Corresponding Author, Los Alamos National Laboratory

The research community now has public access to both the team’s data set and the ANI-1xnr code. 

The DOE Office of Science, Basic Energy Sciences, Chemical Sciences, Geosciences, Biosciences Division, and the Los Alamos Laboratory Directed Research and Development program supported this study. The Center for Integrated Nanotechnologies, a DOE Office of Science user facility, and the Center for Nonlinear Studies were used for some of the work at Los Alamos. The Los Alamos Institutional Computing Program provided resources for this study.

Journal Reference:

Zhang, S., et al. (2024) Exploring the frontiers of condensed-phase chemistry with a general reactive machine learning potential. Nature Chemistry. doi.org/10.1038/s41557-023-01427-3

Source: https://www.lanl.gov/