LASP软件包与相关体系神经网络势函数

（1）LASP程序(large-scale atomistic simulation with neural network potential)是一个大规模原子模拟软件，软件包整合了刘智攀课题组过去十年内开发的两个主要计算模拟方法，即(1)基于随机势能面行走方法(SSW)的全局势能面搜索方法，和(2)全局神经网络势能面(G-NN)以进行快速的势能面计算。除此之外，LASP还提供了多种外部程序接口，比如密度泛函(DFT)计算程序VASP、分子动力学软件LAMMPS等。利用高维神经网络势能面对DFT数据高精度的拟合，LASP实现了相对DFT计算103~104倍的加速，并得以利用SSW进行大规模高效的采样模拟，搜索物理、化学、材料工作者所关心的全局最稳定结构和过渡态结构，或进行第一性原理精度的长时间、大体系的分子动力学模拟。（2）本课题采用全局势能面采样法构建神经网络势函数的训练集。对于每一个NN势函数，所需的结构数据集的采样主要通过LASP程序使用SSW算法产生，结构通过VASP程序进行第一性原理计算得到能量、力和应力，形成的结构数据集使用LASP训练NN势函数。具体实践上，上述的采样、标记、训练过程需进行多轮迭代，并最终产生NN势函数。通过采用同时训练能量、力和应力张量的网络训练方法，我们使得神经网络势能面具备了结构弛豫和全局采样的能力。此外，我们还设计了一系列新型的结构描述符作为神经网络的输入，以提高神经网络势能面的精度。

(1) The LASP (large-scale atomistic simulation with neural network potential) program is a large-scale atomic simulation software package that integrates two main computational simulation methods developed by the Liu Zhipan Research Group over the past decade: (1) the global potential energy surface (PES) search method based on the stochastic surface walking (SSW) algorithm, and (2) the global neural network potential (G-NN) for rapid PES calculations. In addition, LASP provides interfaces to multiple external programs, such as the density functional theory (DFT) calculation code VASP and molecular dynamics software LAMMPS. Leveraging the high-precision fitting of DFT data by high-dimensional neural network potentials, LASP achieves a speedup of 10³ to 10⁴ times relative to standalone DFT calculations. This enables large-scale and efficient sampling simulations using SSW, allowing researchers in physics, chemistry, and materials science to search for globally stable structures and transition states of interest, or to perform long-time, large-system molecular dynamics simulations at first-principles accuracy.
(2) This study constructs the training set for neural network potential functions using global PES sampling methods. For each NN potential function, the sampling of the required structural dataset is primarily generated by the LASP program using the SSW algorithm. The energies, forces, and stresses of these structures are calculated via first-principles simulations with the VASP code, and the resulting structural dataset is used to train the NN potential functions with LASP. In practical implementation, the aforementioned sampling, labeling, and training processes require multiple iterative rounds to ultimately produce the NN potential functions. By adopting a network training approach that simultaneously fits energies, forces, and stress tensors, we endow the neural network PES with the capabilities of structural relaxation and global sampling. Additionally, we designed a series of novel structural descriptors as inputs for the neural network, to improve the accuracy of the neural network PES.

搜集汇总

背景与挑战

背景概述

该数据集包含LASP软件包及其相关的神经网络势函数，用于大规模原子模拟。LASP整合了随机势能面行走和全局神经网络势能面方法，能够高效搜索化学结构的全局稳定态和过渡态，并进行高精度分子动力学模拟，相比传统密度泛函计算加速数千倍。数据集通过迭代采样和训练生成，支持机器学习在催化化学等领域的应用，数据量为294.41MB，包含142个文件。