OpenCSP
收藏资源简介:
OpenCSP数据集是一个专门为晶体结构预测任务设计的开放源数据集,适用于从常压到高压的广泛压力条件。该数据集由约152万种DFT标记的配置组成,这些配置是通过随机高压采样并利用不确定性引导的并发学习策略迭代优化得到的。数据集涵盖了化学元素从氢到铋,原子数从5到32个/单元,并具有广泛的应力分布,这对于在极端压力条件下进行准确的压力和焓预测至关重要。OpenCSP数据集与一系列在数据集上训练的潜在能量模型一起提供,这些模型在高压相预测方面达到了与在更大数据集上训练的模型相当或更好的性能。
The OpenCSP dataset is an open-source dataset specifically designed for crystal structure prediction tasks, applicable to a wide range of pressure conditions from ambient to high pressure. This dataset comprises approximately 1.52 million DFT-labeled configurations, which were generated via random high-pressure sampling and iteratively optimized using an uncertainty-guided concurrent learning strategy. The dataset covers chemical elements ranging from hydrogen to bismuth, with 5 to 32 atoms per unit cell, and features a broad distribution of stress states, which is critical for accurate pressure and enthalpy prediction under extreme pressure conditions. The OpenCSP dataset is distributed alongside a suite of potential energy models trained on the dataset, and these models achieve performance comparable to or better than models trained on larger datasets for high-pressure phase prediction.

- 1通过中国科学计算物理研究所, 中国吉林省吉林大学物理学院 · 2025年



