A team of researchers from Carnegie Mellon University and Los Alamos National Laboratory have developed a model that can simulate reactive processes in various organic materials and conditions using machine learning.
“It’s a tool that can be used to study more reactions in this field,” said Shuhao Zhang, a graduate student in the Department of Chemistry at Carnegie Mellon University. “We can provide a full simulation of reaction mechanisms.”
Zhang is the lead author of the paper describing the creation and results of this new machine learning model. The paper is titled “Exploring the Frontiers of Chemistry with the Potential of General Reactive Machine Learning” and is published in Nature Chemistry.
Although researchers have simulated reactions before, previous methods had many problems. Reactive force field models are relatively common, but they often need training for specific types of reactions. Traditional models, which simulate chemical reactions based on underlying physical principles, are known to require a lot of computational power and time, which can be used for any material and molecule that can be implemented using supercomputers.
This new general machine learning interatomic potential (ANI-1xnr) can perform simulations for any material containing the elements carbon, hydrogen, nitrogen and oxygen, and requires significantly less computational power and time than traditional quantum mechanics models. According to chemistry professor Olexandr Isayev from Carnegie Mellon, head of the laboratory where the model was developed, this achievement is based on advances in machine learning.
“Machine learning, with its regression algorithms, is emerging as a powerful tool to construct diverse forms of transferable atomistic potentials,” noted Isayev. “The overarching goal of this project is to develop a machine learning method capable of predicting reaction energetics and rates for chemical processes with high accuracy, yet at a significantly reduced computational cost.”
“We have demonstrated that these machine learning models can be trained to high levels of quantum mechanics theory, enabling them to predict energies and forces with quantum mechanics-level accuracy while achieving a speed increase of up to 6–7 orders of magnitude. This represents a new paradigm in reactive simulations.”
ANI-1xnr was subjected to various chemical problems by researchers, including comparisons of biofuel additives and monitoring methane combustion. They even replicated the Miller experiment, a renowned chemical demonstration illustrating the origin of life on Earth. Through this experiment, ANI-1xnr exhibited precise results in condensed phase systems.
Zhang suggested that with further training, the model could potentially simulate biochemical processes such as enzymatic reactions.
“We discovered its potential utility in simulating biochemical processes like enzymatic reactions,” Zhang mentioned. “While it wasn’t initially designed for such applications, with modifications, it could serve this purpose.”
In the future, the team aims to refine ANI-1xnr, expanding its compatibility with additional elements and chemical domains, and enhancing its capacity to handle larger-scale reactions. This advancement could find utility in various fields where the innovation of new chemical reactions holds significance, such as drug discovery.
Contributors to this study included Małgorzata Z. Makoś, Ryan B. Jadrich, Elfi Kraka, Kipton Barros, Benjamin T. Nebgen, Sergei Tretiak, Nicholas Lubbers, Richard A. Messerly, and Justin S. Smith.
More information: Exploring the Frontiers of Chemistry with a General Reactive Machine Learning Potential, Nature Chemistry (2024). DOI: 10.1038/s41557-023-01427-3
