Machine Learning Potentials in AMS

Within the Amsterdam Modeling suite several Machine Learning potentials are available such as the popular M3GNET universal potential. 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 thermochemistry calculations and visualize the results with the GUI.

Training or refining your own Machine Learning potentials is possible within the dedicated parametrization environment of AMS (ParAMS). Your own (proprietery) in house ML potentials,can be used via the ASE engine, offering instant access to all AMS driver functionality such as our advanced Molecular Dynamics or potential energy exploration methods.

AMS features the following foundation models, which are applicable for a wide range of molecules/materials:

• ANI-1x, ANI-1ccx, ANI-2x : Suitable for conformer search and thermochemistry of organic molecules
• AIMNet2: Suitable for conformer search and thermochemistry of molecules
• M3GNet-UP-2022: Suitable for general materials science applications

More details about these models can be found in the MLPotential documentation.

If you have trained your own machine learning potential using an external package, you may be able to couple it to the AMS Driver, GUI, and PLAMS using the AMS interface to the Atomic Simulation Environment (ASE).

Fast and accurate Li intercalation potentials in layered cathodes

Using the machine learning potential M3GNET Li intercalation potentials of layered cathode materials are screened, fast and accurate.

Concentration dependent Li migration barriers in LiTiS2

Calculate Li migration barriers inside a crystal structure with the fast and accuratue M3GNET machine learning potential.

Fast and accurate prediction of Kevlar’s mechanical properties

Learn how to use the ANI-2x machine learning potential to calculate the Elastic Tensor and mechanical properties of Kevlar.