colabfit/ANI-1xBB

ANI-1xBB是一个包含约1310万个非平衡构象的小有机分子（仅含H、C、N、O元素；最多7个重原子；最多23个原子总数）的数据集，旨在支持反应性机器学习原子间势能的训练。单点量子化学性质在三个电子温度（T_el = 0、1000和5000 K）下使用B97-3c复合DFT在ORCA 4.2.1中通过有限温度DFT（费米展宽）计算。所有几何结构均被视为闭壳层（电荷=0，多重度=1）；T_el = 5000 K下的费米展宽近似模拟了键解离过程中闭壳层和开壳层状态的叠加，是该数据集相关出版物中模型训练使用的主要标记方案。本数据集包含T_el = 5000 K（b973c_etemp5000）的能量和力；T_el = 0 K和1000 K的数据可在原始源文件中获取。配置集代表：约束几何优化步骤（snap_source=opt，约9%的数据）和固定距离NVT MD快照（snap_source=md，约91%的数据）。ANI-1xBB is a dataset of approximately 13.1 million nonequilibrium conformers of small organic molecules (H, C, N, O only; up to 7 heavy atoms; up to 23 atoms total), designed to support the training of reactive machine learning interatomic potentials. Single-point quantum chemistry properties were computed at three electronic temperatures (T_el = 0, 1000, and 5000 K) using B97-3c composite DFT in ORCA 4.2.1 via finite-temperature DFT (Fermi smearing). All geometries were treated as closed-shell (charge = 0, mult = 1); Fermi smearing at T_el = 5000 K approximates the superposition of closed- and open-shell states during bond dissociation and is the primary labeling scheme used for model training in the associated publication. This dataset contains the T_el = 5000 K (b973c_etemp5000) energies and forces; data at T_el = 0 K and 1000 K are available in the original source files. Configuration sets represent: constrained geometry optimization steps (snap_source=opt, ~9% of data) and fixed-distance NVT MD snapshots (snap_source=md, ~91% of data).