20230723_figshare_DIRECT.zip

All data required to reproduce the DIRECT sampling results in the manuscript "Robust Training of Machine Learning Interatomic Potentials with Dimensionality Reduction and Stratified Sampling".
It contains training, validation and test data for M3GNet-MS and M3GNet-DIRECT UPs, training and test data for MTP-AL and MTP-DIRECT for Ti-H, as well as pre-processed M3GNet features of the 1.3 million structures in <em>MPF.2021.2.8.all</em> and the 274,000 M3GNet MD structures of Ti-H.

本数据集包含复现论文《结合降维与分层采样的机器学习原子势鲁棒训练》中DIRECT采样结果所需的全部数据。其涵盖内容包括：用于M3GNet-MS与M3GNet-DIRECT统一势能（Unified Potentials）的训练集、验证集与测试集；用于钛-氢（Ti-H）体系的MTP-AL与MTP-DIRECT的训练集与测试集；以及<em>MPF.2021.2.8.all</em>中130万条结构的预处理M3GNet特征，与钛-氢体系的27.4万个M3GNet分子动力学（Molecular Dynamics，MD）结构。