TSGen

Training Set Generator (TSGen) constructs datasets and coordinates adaptive ReaxFF parameterization. It extracts molecular configurations from reactive trajectories, combines them with seeds, screens the candidate space and selects DFT batches using SOAP novelty and Gaussian-process uncertainty. Its outer loop returns fitted force fields to reactive MD and assesses the resulting expansion of the reference domain.

Installation

Use Python 3.10 or newer in a dedicated environment. Run from the repository root.

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
tsgen --help

Quickstart

Run from the repository root after installation.

tsgen explore run --config examples/quickstart.yaml

The example uses mechanistic seeds and downsampled ReaxFF-MD trajectories for n-heptane, methylcyclohexane and toluene. Example inputs and outputs

Run Explore	labeling.gaussian.mode	loop.mode	Behavior
Automatic on Linux	local	auto	Run Gaussian in the current environment and advance through the acquisition rounds.
Automatic on a cluster	slurm	auto	Submit Gaussian jobs with sbatch, poll for results and advance through the acquisition rounds.
Prepare inputs only	manual	step	Write Gaussian inputs. Supply the logs and run tsgen explore resume to collect them and advance.

The supplied example uses manual and step. See the configuration for local Linux or Slurm to run Gaussian automatically.

Reproducing the manuscript's numerical results requires the original study trajectories, split manifests, DFT labels and fitting inputs. This repository provides the workflow code and downsampled Quickstart inputs.

Relative-energy model

For a sampled configuration i of species g, the GP target in kcal/mol is

y_i = 627.5094740631 * (E_SP(i) - E_Opt(g))

Both electronic energies are calculated at M06-2X/6-311G(d,p) in Hartree. Each sampled geometry supplies a single-point energy, and all configurations of the same species share a fixed reference energy from geometry optimization.

Workflows

Command group	Inputs and outputs
explore	Seed XYZ files and reactive trajectories → screened pool, fixed test manifest, DFT acquisition history, labels and fitting representatives.
trainset	Optimized structures and Gaussian relaxed scans → paired ReaxFF geo and trainset.in.
sensitivity	Force field, fitting targets and a ReaxFF evaluator → LHS responses, HSIC ranking and selected parameter bounds.
outerloop	Fitting datasets, successive force fields and MD results → assessed candidates, expanded reference archive and next fitting dataset.
benchmark	Saved screened pool, fixed test manifest and explicit DFT oracle → matched-budget TSGen, random and FPS metrics.
diagnostics	Production run artifacts → geometry checks, sensitivity diagnostics and Slurm concurrency adjustments.

Gaussian 16 supplies the DFT calculations. ReaxFF evaluation, external INDEEDopt optimization and AMS MD require their respective installations and input files. See outerloop.md for the engine handoff contracts.

See the study protocol for research inputs, independent evaluation and matched-budget comparisons.

Citation and license

Use CITATION.cff to cite this software. TSGen is distributed under the MIT License.

Development

python -m pip install -e '.[test]'
python -m pytest

src/tsgen/ contains the installable workflows. docs/ describes data contracts and execution. tests/ contains automated regression checks for file handling, numerical methods, job execution and workflow recovery.