Ultra-fast interpretable machine-learning potentials 

Stephen R. Xie, Matthias Rupp & Richard G. Hennig 

npj Computational Materials 9: 162 (2023)

doi.org/10.1038/s41524-023-01092-7

Published online: 02 September 2023

编辑概述

全原子动力学模拟之光：超快可解释机器学习势 

全原子动力学规模模拟在物理、化学和材料科学是一种不可或缺的定量工具，然而，由于计算效率与预测精度之间的制约，处理大规模系统和长时间模拟仍然具有挑战。该研究旨在开发超快的、可解释的机器学习势，以保持高准确性的同时实现与最快的传统经验势相媲美计算效率。

来自美国弗罗里达大学材料科学与工程系的Richard G. Henning教授领导的研究团队，采用一种创新的方法，通过将在一个立方B样条基础上有效的二体和三体势与正则化线性回归结合，成功地开发了一种超快、可解释的机器学习势。这些势在应用方面足够准确，速度上与最快的传统经验势相当，甚至比最先进的机器学习势快2到4个数量级。这些机器学习势在元素钨中展现了与最先进MLPs接近的准确性，同时与传统经验势的速度相匹配。这一方法在模拟元素钨的性质方面取得了显著的成功，包括弹性常数、声子谱、表面能和熔点等。新开发的机器学习势不仅提高了计算效率，还同时保持了预测准确性，扩展了预测精度和计算成本的帕累托前沿，为大规模原子系统实现长时间尺度的全原子动力学模拟提供了可靠的解决方案。

Editorial Summary

The Prospect of Full Atomic-Scale Dynamics Simulation: Ultra-Fast Interpretable Machine Learning Potentials

Full atomic-scale dynamic simulation is an indispensable quantitative tool in physics, chemistry, and materials science. However, dealing with large systems and long simulation times remains challenging due to the trade-off between computational efficiency and predictive accuracy. This study aims to develop an ultra-fast interpretable machine learning potential, striking a balance between maintaining high accuracy and achieving computational efficiency comparable to the fastest traditional empirical potentials.

A team led by Prof. Richard G. Henning from the Department of Materials Science and Engineering, the University of Florida, USA, employed an innovative approach. They successfully developed ultra-fast interpretable machine learning potentials by combining effective two- and three-body potentials in a cubic B-spline basis with regularized linear regression. These potentials exhibit sufficient accuracy in applications and match the speed of the fastest traditional empirical potentials, surpassing even state-of-the-art machine learning potentials by two to four orders of magnitude. Demonstrating close accuracy to advanced ML potentials, particularly in the case of tungsten, and matching the speed of traditional empirical potentials, this method achieved notable success in simulating tungsten's properties, including elastic constants, phonon spectra, surface energy, and melting point. The newly developed machine learning potentials not only enhance computational efficiency but also maintain predictive accuracy, extending the Pareto frontier of predictive accuracy and computational cost. This provides a reliable solution for achieving accurate all-atom dynamics simulations of large atomistic systems over long-time scales.