LMHE-TSopt: Molecular Geometry and Optimization Trajectories for Transition State Finding

Data from the THGLab supports research on machine-learned leftmost Hessian eigenvectors for transition state optimization. The repository contains training data for E(3)-equivariant predictors and benchmark results from the Sella set. It includes atomic coordinates, Hessian eigenvectors, and full optimization trajectories for evaluating TS-finding algorithms.

Use Cases

• Training E(3)-equivariant neural networks based on atomic coordinates and target eigenvectors.
• Benchmarking transition state optimization algorithms using the Sella reaction set.
• Analyzing optimization trajectories and Intrinsic Reaction Coordinate (IRC) paths from detailed log files.
• Validating initial TS-guess geometries against converged transition state structures.

Strengths

• Includes full optimization trajectories (.traj) and detailed logs, providing transparency into the optimization process.
• Training data is structured for machine learning with specific keys (pos, z, vec, slices) in HDF5 format.
• Dataset is openly licensed under CC-BY-4.0, facilitating reuse and distribution.
• Cross-platform presence on figshare indicates established sharing and versioning.

Limitations

• Row/record counts are unknown, making the dataset's scale difficult to assess.
• Metadata is sparse; specific molecular counts, reaction types, and data generation parameters are not provided.
• The dataset is large (~18 GB), which may require significant storage and computational resources for full use.

Provenance

Source

THGLab, author Guanchen Wu.

Collection Method

Data derived from the augmented T1x dataset; benchmark geometries generated via KinBot from the Sella set.

Time Range

null

Freshness

Last updated on 2026-03-24.

Geography

null