npj Computational Materials

A universal strategy for the creation of machine learning-based atomistic force fields（创建基于机器学习的原子力场通用方案)
Tran Doan Huan, Rohit Batra, James Chapman, Sridevi Krishnan, Lihua Chen & Rampi Ramprasad
npj Computational Materials 3:37 (2017)
doi:10.1038/s41524-017-0042-y
Published online:18 September 2017
Abstract| Full Text | PDF OPEN

摘要:新兴的机器学习（ML）技术为研究各种物理和化学问题提供了强大而新颖的工具。本研究概述了创建基于ML的原子力场的通用方案，可用于执行高保真分子动力学模拟。这个方案包括：（1）准备一个足够低噪的原子环境和各种原子力的大参考数据集（如，使用密度泛函理论或更高层次的方法）；（2）利用结构特征的通用化分类来再现原子环境；（3 ）从参考数据中优选多种、非冗余的训练数据集；（4）提出各种学习方法以便从原子配置角度直接（并且快速）地预测各种原子力。依据原子力，再依据沿反应坐标或沿分子动力学轨迹的适当积分，就可获得精确的势能。基于这一方案，本研究为六种元素块体（包括Al、Cu、Ti、W、Si和C）创建了ML力场，并展示了它们都能达到的化学精确度（chemical accuracy）。所提出的方案是一般性和通用性的，有可能为任何材料生成ML力场，任何材料都可使用相同的统一工作流程，几乎无需人为修正。另外，本方案可通过逐步添加新的训练数据来代表前所未遇的原子环境，从而可以系统地改进力场。　　

Abstract:Emerging machine learning (ML)-based approaches provide powerful and novel tools to study a variety of physical and chemical problems. In this contribution, we outline a universal strategy to create ML-based atomistic force fields, which can be used to perform high-fidelity molecular dynamics simulations. This scheme involves (1) preparing a big reference dataset of atomic environments and forces with sufficiently low noise, e.g., using density functional theory or higher-level methods, (2) utilizing a generalizable class of structural fingerprints for representing atomic environments, (3) optimally selecting diverse and non-redundant training datasets from the reference data, and (4) proposing various learning approaches to predict atomic forces directly (and rapidly) from atomic configurations. From the atomistic forces, accurate potential energies can then be obtained by appropriate integration along a reaction coordinate or along a molecular dynamics trajectory. Based on this strategy, we have created model ML force fields for six elemental bulk solids, including Al, Cu, Ti, W, Si, and C, and show that all of them can reach chemical accuracy. The proposed procedure is general and universal, in that it can potentially be used to generate ML force fields for any material using the same unified workflow with little human intervention. Moreover, the force fields can be systematically improved by adding new training data progressively to represent atomic environments not encountered previously.

Editorial Summary

Machine learning: molecular dynamics simulations feel the force（机器学习：分子动力学模拟感受原子力)

该研究提出了一个用于计算原子力场的基于机器学习的方案，可以扩大分子动力学模拟范围。分子动力学模拟是探索原子尺度化学和物理过程随时间演变的有力工具。为了执行这种模拟，需要定义初始原子配置，并为每个时间步长输入原子力。虽然计算原子力场可用基于量子力学的各种方法，但这些方法在长时程上应用于大型系统很是不便。来自美国康涅狄格大学的Rampi Ramprasad领衔的研究团队，在本文中提出了一个通用策略：基于机器学习的方法生成高精度原子力场，为纳米级几纳秒系统的分子动力学模拟提供了一条有效途径。

A machine learning-based strategy for calculating atomic force fields could expand the range of molecular dynamics simulations. Molecular dynamics simulations are a powerful tool for exploring how atomic-scale chemical and physical processes evolve over time. To perform such simulations, an initial atomic configuration needs to be defined, and atomic forces are input for each time step. Whilst quantum mechanics-based methods for calculating the fields are available, these approaches cannot easily be applied to large systems over long timescales. A team of researchers from the University of Connecticut, led by Rampi Ramprasad, now present a general and universal strategy for using machine learning-based methods to generate highly accurate atomic force fields that may provide a pathway for performing efficient molecular dynamics simulations on nanometer-sized systems over several nanoseconds.