Rosenberger David, Smith Justin S, Garcia Angel E
Los Alamos National Laboratory, Theoretical Division, Chemistry and Physics of Materials Group, Los Alamos, 87545 New Mexico, United States.
Los Alamos National Laboratory, Theoretical Division, Center for Nonlinear Studies, Los Alamos, 87545 New Mexico, United States.
J Phys Chem B. 2021 Apr 15;125(14):3598-3612. doi: 10.1021/acs.jpcb.0c10401. Epub 2021 Apr 2.
The replacement of classical force fields (FFs) with novel neural-network-based frameworks is an emergent topic in molecular dynamics (MD) simulations. In contrast to classical FFs, which have proven their capability to provide insights into complex soft matter systems at an atomistic resolution, the machine learning (ML) potentials have yet to demonstrate their applicability for soft materials. However, the underlying philosophy, which is learning the energy of an atom in its surrounding chemical environment, makes this approach a promising tool. In particular for the exploration of novel chemical compounds, which have not been considered in the original parametrization of classical FFs. In this article, we study the performance of the ANI-2x ML model and compare the results with those of two classical FFs, namely, CHARMM27 and the GROMOS96 43a1 FF. We explore the performance of these FFs for bulk water and two model peptides, trialanine and a 9-mer of the α-aminoisobutyric acid, in vacuum and water. The results for water describe a highly ordered water structure, with a structure similar to those using molecular dynamics simulations. The energy landscape of the peptides described by Ramachandran maps show secondary structure basins similar to those of the classical FFs but differ in the position and relative stability of the basins. Details of the sampled structures show a divergent performance of the different models, which can be related either to the short-ranged nature of the ML potentials or to shortcomings of the underlying data set used for training. These findings highlight the current state of the applicability of ANI-2x ML potential for MD simulations of soft matter systems. Simultaneously, they provide insights for future improvements of current ML potentials.
用新型基于神经网络的框架取代经典力场（FFs）是分子动力学（MD）模拟中一个新兴的话题。与已证明能够在原子分辨率下深入了解复杂软物质系统的经典力场不同，机器学习（ML）势尚未证明其对软材料的适用性。然而，其基本原理，即在周围化学环境中学习原子的能量，使这种方法成为一个有前途的工具。特别是对于探索新型化合物，而这些化合物在经典力场的原始参数化中并未被考虑。在本文中，我们研究了ANI-2x ML模型的性能，并将结果与两种经典力场，即CHARMM27和GROMOS96 43a1 FF的结果进行比较。我们在真空和水中探索了这些力场对体相水以及两种模型肽（三丙氨酸和α-氨基异丁酸的九聚体）的性能。水的结果描述了一种高度有序的水结构，其结构与使用分子动力学模拟得到的结构相似。由拉马钱德兰图描述的肽的能量景观显示出与经典力场相似的二级结构盆地，但盆地的位置和相对稳定性有所不同。采样结构的细节显示了不同模型的不同性能，这可能与ML势的短程性质或用于训练的基础数据集的缺点有关。这些发现突出了ANI-2x ML势在软物质系统MD模拟中的适用性现状。同时，它们为当前ML势的未来改进提供了见解。
