SEAMM Normal Mode Sampling Plug-in

A SEAMM plug-in for Wigner/thermal normal-mode sampling of the Hessian to generate displaced structures (e.g. for MLFF training sets)

• Free software: BSD-3-Clause

• Documentation: https://molssi-seamm.github.io/normal_mode_sampling_step/index.html

• Code: https://github.com/molssi-seamm/normal_mode_sampling_step

History

2026.10.3.1 – The finite-difference Hessian as separate calculations, on this machine or a cluster

• When the job’s calculations go to a cluster queue and the program can run separate calculations (ORCA, MOPAC), the Hessian is the finite difference of the gradients, run there as 6N calculations; nothing is started on the job’s own machine, where the program may not be installed.

• Otherwise the analytic Hessian is used when the program’s MDI engine has one, and ORCA’s model chemistry now says so without ORCA being started to ask. Failing that, ORCA’s finite-difference calculations run several at a time on this machine; MOPAC, xTB and MLFFs use the warm MDI engine as before.

• Rerunning the job reuses finished finite-difference calculations.

• If the program is not installed on this machine, the separate calculations are used and the output says so; other failures to start the engine stop the step with the reason rather than quietly switching method.

• Requires seamm-exec 2026.10.3 or later.

2026.10.3 – Bugfix: the Hessian used the program’s method name and default basis

• The MDI engine for the Hessian was launched with the model chemistry’s method name alone, so ORCA ran def2-SVP whatever basis was chosen, and a functional whose name the Model Chemistry step had to alter was not recognized. It now gets the program’s own keyword and the chosen basis (with model_chemistry_step 2026.10.3).

• The shared CI now runs on uv: devtools/conda-envs/test_env.yaml is removed, so requirements.txt is the one dependency list.

2026.7.15 – Initial release

• Generates an ensemble of displaced structures by normal-mode sampling of a molecule’s Hessian, for building machine-learned-force-field training sets and similar uses.

• Amplitudes follow the quantum (Wigner) distribution by default, so stiff modes such as O-H stretches get the real zero-point spread that classical 300 K sampling misses; classical-thermal and ground-state (0 K) distributions are also available.

• The Hessian is obtained from the Model Chemistry defined earlier in the flowchart, over MDI: the analytic Hessian when the engine provides one, otherwise a finite-difference of the forces over the resident engine.

• A temperature, per-mode selection, amplitude cap, and harmonic-energy outlier rejection are all controllable, and the random seed can be fixed for reproducible ensembles.

• Off-minimum geometries are allowed with a warning, so a transition state is sampled along its real modes while its imaginary reaction coordinate is left alone.

• Each generated structure records its predicted harmonic energy and the per-mode normal-coordinate displacement as properties, for downstream filtering and analysis.