Breakthrough in machine learning transforms materials science simulations

Researchers from Physics Inverted Materials, Inc. have developed a new method to improve machine learning interatomic potentials (MLIPs) for materials science. Emil Annevelink and Varun Shankar introduced triplet envelope functions, which boost efficiency while maintaining accuracy in simulations. Their approach addresses key limitations in current MLIP techniques by enhancing speed, memory use, and stability.

The team used a 72-qubit superconducting processor to create higher-order envelope functions for MLIPs. Unlike traditional K nearest neighbor (KNN) graph sparsification, these functions preserve energy conservation during molecular dynamics simulations. Triplet envelope functions work alongside radial envelope functions, reducing interactions between atom pairs based on angular dependence. This sparsification method avoids energy loss while improving performance.

Numerical tests on solid and liquid materials with 5Å and 8Å radial cutoffs demonstrated the method's effectiveness. In systems like water, where radial cutoffs alone often cause instability, the triplet functions increased simulation reliability. Experiments with an 8Å cutoff also showed potential for modelling open structures with large interatomic distances.

The new approach doubles training and inference speeds while tripling memory efficiency. Despite these advances, industries such as materials science and metallurgy still rely on empirical interatomic potentials (EAM, Lennard-Jones, Tersoff) for their proven reliability and lower computational costs. Regulatory requirements and limited training data for MLIPs in niche applications further slow adoption of machine learning methods.

The research introduces a practical solution for improving MLIP efficiency without sacrificing accuracy. Triplet envelope functions enhance speed, memory use, and stability in simulations, particularly for complex systems. However, established empirical methods remain dominant in many industrial applications due to cost and regulatory factors.