GARFfield = Genetic Algorithm based Reactive Force Field optimizer. 

GARFfield is a multi-platform, multi-objective parallel hybrid genetic algorithm (GA) / conjugate-gradient (CG) based force field optimization framework.  It enables first-principles based force fields prepared from large quantum mechanical data sets, which are now the norm in predictive molecular dynamics simulations for complex chemical processes, as well as from phenomenological data. The former allow improved accuracy and transferability over a wider range of molecular compositions, interactions and environmental conditions unexplored by experiments.

GARFfield currently supports a range of force field engines, via the LAMMPS Parallel Molecular Dynamics Simulator, including the adiabatic ReaxFF and COMB potentials for modeling reaction processes, the non-adiabatic eFF electron force field with effective core potentials, and Morse potentials (atomistic and coarse-grain).

As opposed to other efforts found in the literature, GARFfield is not limited to a single force field!!. It is intended to handle electronic, atomistic (reactive and non-reactive), and coarse-grain type force field parameter optimization problems, albeit with implicit considerations for each energy/force type of engine in order to improve global optimization efficiency.

GARFfield provides multiple optimization features that are specific to force field training, including:

  • Automatic hooks to the LAMMPS force field engines and automatic force field detection,
  • Multi-objective fitness functions (including charges, relative energies, lattice parameters, geometric parameters, and others)  in a weighted or un-weighted scalar sum,
  • Partial parameter sub-set selection and optimization,
  • Training sets with periodic and finite system models,
  • Geometry model specifications in extended xyz, biograf (.bgf) and LAMMPS native formats,
  • Relative restraints on bond-order based valence interactions, transition state bonds, and electron sizes,
  • Non-deterministic solutions in the Pareto-front using random/fixed weighted and un-weighted training sets,
  • Systematic hill-climbing option from local minima solutions,
  • RMS force fitness for geometric objectives (as opposed to gradient-based energy minimization),
  • Deterministic CG minimization switch option when GA is within parabolic minima wells, and
  • Others (e.g. parallelization of GA string population)

Support for other force field engines and new features will be added through the modular software architecture design.  Current efforts include: template-based force field support (to avoid syntax dependencies from LAMMPS), heuristic sequence training (e.g. valence then non-bond or vice-versa, finite then periodic training cases, etc.), and training set parallelization (in addition to the current GA string population parallelization).

We are releasing the GARFfield code to promote community-development, so if you want to contribute bug fixes, new extensions (e.g. force field parsers and hooks) or new features (e.g. GUI) please register (see instructions below), download and compile your own GARField copy.  We ask that once you integrate, validate and test your contribution you submit it back to us (ajaramil at for review and integration into the published distribution.

To cite GARFfield use:

NOTE: GARFfield is currently available for non-profit academic and research activities. Its use for commercial purposes requires a commercial license that can be requested via email to

If you want to try GARFfield, please read and accept the terms of our LICENSE. If you have registered and been given a personal login/password key pair, you can download GARFfield binaries for:

Or download the C/C++ source code (including all dependencies) and examples from GARFfield.tar.gz and the GARFfield User Manual. If you have not registered, please fill out and submit the form below.




This page was last updated 9/15/2013 4:17:00 PM

Note: I've added an underscore "_" after each e-mail address to avoid automated spam e-mail