colabfit/AIMNet2

AIMNet2(2025)是用于AIMNet2（第二代原子在分子网络）神经网络原子间势的扩展训练数据集，旨在改进模型对非共价相互作用（NCIs）的描述，包括氢键、π-π堆积、色散、σ-孔、离子和静电接触。该数据集涵盖由最多14种非金属元素（H、B、C、N、O、F、Si、P、S、Cl、As、Se、Br、I）组成的中性和带电闭壳层分子系统，每个系统最多包含193个原子。结构来源于三个互补来源：(a) SPICE v2.0.1中的分子几何结构（溶剂化系统、氨基酸-配体对、水簇）和CREMP数据集（大环肽）；(b) 通过正常模式采样和元动力学引导的几何探索从PubChem采样的中小型中性和带电分子；(c) 从剑桥结构数据库（CSD）单体组装的二聚体几何结构（最多14种支持元素，少于200个原子），并使用AIMNet2-wB97M-D3(2023)进行预优化以消除空间冲突，同时保持构型多样性。所有量子化学计算均使用ORCA 6.0.1，在限制性Kohn-Sham（RKS）形式下使用复合B97-3c DFT泛函。通过TightSCF和SlowConv强制SCF收敛；全程应用RIJCOSX积分加速和DEFGRID2积分网格。AIMNet2(2025)从AIMNet2(2023)权重初始化，并在此数据集上持续预训练，无需权重冻结或正则化，使用能量（权重1.0）、力（权重0.2）和Hirshfeld部分电荷（权重0.5）的多任务损失函数。AIMNet2(2025) is the extended training dataset for the AIMNet2 (second generation atoms-in-molecules network) neural network interatomic potential, curated to improve the models description of noncovalent interactions (NCIs) including hydrogen bonding, pi-pi stacking, dispersion, sigma-hole, ionic, and electrostatic contacts. The dataset covers neutral and charged closed-shell molecular systems composed of up to 14 non-metal elements (H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I) with up to 193 atoms per system. Structures were drawn from three complementary sources: (a) molecular geometries from SPICE v2.0.1 (solvated systems, amino acid-ligand pairs, water clusters) and the CREMP dataset (macrocyclic peptides); (b) small neutral and charged molecules from PubChem sampled via normal mode sampling and metadynamics-guided geometry exploration; (c) dimer geometries assembled from Cambridge Structural Database (CSD) monomers (up to 14 supported elements, fewer than 200 atoms) and pre-optimized with AIMNet2-wB97M-D3(2023) to remove steric clashes while preserving configurational diversity. All quantum chemical calculations used ORCA 6.0.1 with the composite B97-3c DFT functional under restricted Kohn-Sham (RKS) formalism. SCF convergence was enforced with TightSCF and SlowConv; RIJCOSX integral acceleration and DEFGRID2 integration grid were applied throughout. AIMNet2(2025) was initialized from AIMNet2(2023) weights and continually pretrained on this dataset without weight freezing or regularization, using a multi-task loss over energy (w=1.0), forces (w=0.2), and Hirshfeld partial charges (w=0.5).