Data for Machine-Learned Leftmost Hessian Eigenvectors for Robust Transition State Finding

This repository contains the datasets and structural geometries supporting the research on transition state (TS) optimization guided by machine-learned leftmost Hessian eigenvectors (LMHE).1. LMHE Predictor Training Data (LMHE_dataset.tar.xz)This archive contains the training, validation, and test datasets used to develop the E(3)-equivariant LMHE predictors, provided in HDF5 (.h5) format. The data is derived from the augmented T1x dataset. Each .h5 file contains the following keys:pos: Cartesian coordinates of the atomic configurations.z: Atomic numbers.vec: Target leftmost Hessian eigenvectors, which locally approximate the reaction coordinate.slices: Index markers denoting the starting positions of individual molecules within the arrays.2. Transition State Optimization Benchmarks (opt-tests.tar.xz)This archive contains the structural data and optimization trajectories used to evaluate the LMHE optimizer on reactions from the Sella benchmark set. It includes three subdirectories:molecules_kinbotprod_renamed: Ground-truth reactant and product geometries (.xyz format) and their corresponding topological connectivity (.bond format), generated via KinBot.ts-guess: Initial, unoptimized transition state guess geometries (.xyz format) utilized as starting points for the benchmark optimizations.outputs: Results of the TS optimizations. This includes the final converged TS geometries, Intrinsic Reaction Coordinate (IRC) extrapolated reactant and product geometries (.xyz format), full optimization trajectories (.traj format), and detailed optimization log files.3. Code AvailabilityThe associated Python scripts and workflow implementations utilizing this data can be accessed at: https://github.com/THGLab/LMHE-TSopt.