MatPES
收藏资源简介:
MatPES数据集是一个基础性的潜在能量面数据集,由劳伦斯伯克利国家实验室等机构开发,旨在为材料科学研究提供高质量的训练数据。该数据集包含了从281百万个分子动力学快照中精心采样的约504,811个结构,覆盖了16亿个原子环境。数据集通过使用预训练的M3GNet UMLIP进行采样,并结合2DIRECT采样方法,确保了数据质量。MatPES数据集的应用领域主要是材料科学,用于训练更可靠、通用且高效的UMLIP,以支持大规模的材料发现和设计。
The MatPES dataset is a foundational potential energy surface dataset developed by institutions including Lawrence Berkeley National Laboratory, aiming to provide high-quality training data for materials science research. It contains approximately 504,811 structures carefully sampled from 281 million molecular dynamics snapshots, covering 1.6 billion atomic environments. The dataset is sampled using the pre-trained M3GNet UMLIP combined with the 2DIRECT sampling method to ensure data quality. The MatPES dataset is primarily applied in materials science, and is used to train more reliable, generalizable and efficient UMLIPs to support large-scale material discovery and design.




