Machine Learning Potentials & Force Field
Machine learning interface, ML potential backends, classical force fields
Machine learning potentials can provide very accurate descriptions of chemical systems at the computational cost of a force field. In combination with the AMS driver, machine learning potentials allow you to run accurate Molecular Dynamics simulations or to explore potential energy surfaces for e.g. frequency calculations or transition state searches.
AMS2020 features the pre-parameterized models ANI-1ccx and ANI-2x for direct usage but also allows you to develop and employ your own machine learning potentials using one of the available backends: PiNN, SchNetPack, sGDML, and TorchANI.
In AMS2023 we added the M3GNet-UP-2022 universal graph neural network potential, with which you can study the geometry or thermodynamic stability of almost any material. You can also run molecular dynamics, or use other tasks through our AMS driver.
Other machine learning potentials could be employed through our AMS and PLAMS interface Atomic Simulation Environment (ASE).
Video: Machine Learning Potentials in AMS2020
SCM’s expert Matti Hellström demonstrates the new machine learning potential engine in the Amsterdam Modeling Suite.
Classical Force Fields
Since AMS2022, the Machine Learning Potentials module also includes the classical force fields:
- GFN-FF: all-round force field for most elements in the periodic table
- GAFF: General Amber Force Field, with automatic atom typing option
- APPLE&P: Polarizable force field for electrolytes and ionic liquids. APPLE&P parameters are sold separately, see list with (classes of) molecules which can be simulated.