In the realm of energy storage and nuclear technologies, liquid metals play a pivotal role. However, our quantitative understanding of their thermophysical properties is still somewhat limited. A team of researchers from the University of Cambridge, led by Alex Tai, Jason Ogbebor, and Rodrigo Freitas, has been exploring ways to bridge this knowledge gap using machine learning. Their recent work, published in the journal Nature Communications, offers a promising approach to improve the accuracy and stability of simulations involving liquid metals.
The researchers focused on a method called machine learning interatomic potentials (MLPs), which can provide the accuracy of ab initio molecular dynamics (AIMD) at a much lower computational cost. However, applying MLPs to liquids has been challenging due to inadequate sampling of atomic configurations in training datasets, leading to unphysical force predictions and unstable trajectories.
To address this issue, the team introduced a physically motivated dataset-engineering strategy. Instead of relying solely on AIMD configurations, they synthetically constructed liquid-like training data. This method leverages the known icosahedral short-range order of metallic liquids and their twelvefold, near-close-packed local coordination. By systematically perturbing crystalline references, the researchers generated “synthetic-liquid” structures that better represent the atomic configurations in liquid metals.
The MLPs trained on these engineered datasets showed significant improvements. They closed the sampling gaps that previously led to unphysical predictions, remained numerically stable across different temperatures, and accurately reproduced experimental data for liquid densities, diffusivities, and melting temperatures of multiple elemental metals.
This research provides a practical route to predictive modeling of liquid-phase thermophysical behavior beyond the limits of direct AIMD. For the energy sector, this could mean more accurate simulations and better understanding of liquid metals in energy storage systems and nuclear technologies. The improved stability and accuracy of these simulations can aid in the design and optimization of these systems, potentially leading to more efficient and safer energy solutions.
The researchers’ work highlights the potential of machine learning in advancing our understanding of complex systems like liquid metals. As the energy industry continues to evolve, such computational tools will be invaluable in driving innovation and improving existing technologies.
Source: Nature Communications
This article is based on research available at arXiv.