AnonymouScientist/MaterialsSaddles

MaterialsSaddles 是一个用于固态和表面化学的高通量收敛过渡态库，包含34,135,597个完全收敛的过渡态，这些过渡态是通过对公共材料和催化数据集（如LeMat-Bulk、Open Catalyst 2020、Open Catalyst 2022和Materials Project电池结构）进行大规模并行鞍点搜索计算得到的。数据集使用SaddleMill包和Meta的uma-s-1p2机器学习原子间势作为计算器。每个过渡态事件由三个连续条目组成：反应物最小值、过渡态（一阶鞍点）和产物最小值，端点收敛到0.02 eV/Å（最大力），鞍点收敛到0.05 eV/Å。每个鞍点行还存储其本征模式——一个(N, 3)的每原子位移场，指示鞍点不稳定的方向。数据以ASE-LMDB格式存储，包含丰富的元数据，如task_name、ms_id、eigenmode、curvature等，并提供了按子集分层的训练/验证/测试分割。该数据集旨在用于训练生成模型以预测过渡态、生成DFT标签以改善机器学习原子间势的势垒预测、以及热启动DFT鞍点搜索。

MaterialsSaddles is a high-throughput library of converged transition states for solid-state and surface chemistry, containing 34,135,597 fully converged transition states computed by massively-parallel saddle searches on public materials and catalysis datasets (e.g., LeMat-Bulk, Open Catalyst 2020, Open Catalyst 2022, and Materials Project battery structures). The dataset uses the SaddleMill package and Metas uma-s-1p2 machine-learning interatomic potential as the calculator. Each transition-state event consists of three consecutive entries: reactant minimum, transition state (first-order saddle), and product minimum, with endpoints converged to 0.02 eV/Å (max force) and saddles to 0.05 eV/Å. Each saddle row also stores its eigenmode—an (N, 3) per-atom displacement field giving the direction along which the saddle is unstable. Data is stored in ASE-LMDB format with rich metadata such as task_name, ms_id, eigenmode, curvature, etc., and includes a stratified train/val/test split by subset. The dataset is intended for training generative models for transition-state prediction, generating DFT labels to combat MLIP barrier softening, and warm-starting DFT saddle searches.