High-temperature multi-element 2021 (HME21) dataset

HME21 is the atomic structure dataset aimed for the neural network potential development. It was created in the development of PFP, a universal neural network potential for material discovery [1].
It contains multiple elements in a single structure and was sampled through a high-temperature molecular dynamics simulation. There are a total of 37 elements in the HME21 dataset, i.e., H, Li, C, N, O, F, Na, Mg, Al, Si, P, S, Cl, K, Ca, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Mo, Ru, Rh, Pd, Ag, In, Sn, Ba, Ir, Pt, Au, and Pb.
They are calculated by Spin-polarized DFT calculations using PBE exchange-correlation functional implemented in VASP [2] version 5.4.4. All structures are under periodic boundary conditions.
For the details of DFT calculation conditions and structure sampling method, please see the reference [1].
Please cite the reference [1] if you use this dataset.
Files
HME21 consists of three files with extxyz format:

train.xyz: 19956 structures
valid.xyz: 2498 structures
test.xyz 2495 structures

The structures were randomly split into training, validation, and test sub-datasets at a ratio of 8:1:1. They are used as training, validation, and test dataset for the benchmark of neural network potentials [1].
The target values are energy and atomic forces. The energy is shifted such that the energy of a single atom located in a vacuum becomes zero. The length is in angstroms (10^−10 m), and the energy is in electronvolts (eV). For supplementary, vasp_shift_energies.json which corresponds to the reference energy of single atom for each element is also included.

HME21为一款旨在促进神经网络势发展的原子结构数据集。该数据集在通用神经网络势PFP的开发过程中被创建，PFP旨在材料发现领域实现广泛应用[1]。数据集中包含多种元素的单个结构，并通过高温分子动力学模拟进行采样。HME21数据集总计包含37种元素，即H、Li、C、N、O、F、Na、Mg、Al、Si、P、S、Cl、K、Ca、Sc、Ti、V、Cr、Mn、Fe、Co、Ni、Cu、Zn、Mo、Ru、Rh、Pd、Ag、In、Sn、Ba、Ir、Pt、Au和Pb。这些元素的计算基于使用VASP [2]版本5.4.4中实现的PBE交换关联泛函进行的自旋极化DFT计算。所有结构均在周期性边界条件下进行。有关DFT计算条件及结构采样方法的详细信息，请参阅参考文献[1]。若使用此数据集，请引用参考文献[1]。

HME21包含三个以extxyz格式存储的文件：

train.xyz：包含19956个结构
valid.xyz：包含2498个结构
test.xyz：包含2495个结构

这些结构以8:1:1的比例随机划分为训练、验证和测试子数据集，用作神经网络势基准的培训、验证和测试数据集[1]。目标值为能量和原子力。能量经过调整，使得位于真空中的单个原子的能量为零。长度单位为埃（10^-10米），能量单位为电子伏特（eV）。此外，还包括vasp_shift_energies.文件，该文件对应于每个元素的单一原子参考能量。