Train Models for Your Chemistry

• Train or refine ReaxFF, DFTB/GFN1-xTB, and machine learning potentials
• Fit ReaxFF parameters for a bond scan
• Build a larger ReaxFF parameterization workflow
• Parameterize GFN1-xTB for lithium fluoride
• Fit a DFTB repulsive potential
• Train or fine-tune M3GNet with the ParAMS GUI
• Train MACE with the ParAMS GUI
• Train ML potentials on the fly with Simple Active Learning
• Run parameter sensitivity analysis before optimizing
• Run multiple optimizers in parallel

How ParAMS works

1. Collect reference data: use AMS jobs, trajectories, PES scans, external calculations, or experimental data.
2. Choose the model: select ReaxFF, DFTB/GFN1-xTB, MACE, M3GNet, Lennard-Jones, or another supported parameter interface.
3. Define the target: fit energies, forces, stresses, charges, bond lengths, angles, cell parameters, reaction energies, or custom extracted properties.
4. Optimize parameters: choose the parameters, bounds, weights, optimizer settings, and validation set.
5. Validate and export: inspect errors, correlation plots, validation performance, and use the final model in AMS calculations.

Current ParAMS workflows

ParAMS now covers classical parameter fitting and machine learning potential training in the same AMS environment. Use it for ReaxFF and DFTB parameterization, MACE and M3GNet training, single-point benchmarks, sensitivity analysis, validation sets, and active learning during molecular dynamics.

Recent ParAMS improvements include a simpler single-point benchmark setup, support for importing molecular dynamics simulations with changing atom counts, machine learning potential training, Simple Active Learning, parallel optimizers, recommended ReaxFF parameter constraints, and ASE calculator parameterization.

Related AMS tools

Use ReaxFF when you need reactive molecular dynamics. Use DFTB/GFN1-xTB for fast electronic-structure models. Use machine learning potentials for fast molecular and materials simulations, and Simple Active Learning when new reference data should be generated during MD.

• ParAMS is part of Advanced workflows and tools
• You also need to license the compute engine you want to train: ReaxFF and/or DFTB and/or ML potentials
• For reference data, we recommend to use ADF, BAND, or Quantum ESPRESSO. You may also user other training data, for example from other software or experimental data.
• See also Pricing and Licensing