aMACEing toolkit

Foundational machine learned force field models are trained on structures from large databases and aim to enable simulations of a wide range of compounds without further modification. However, their accuracy can often be significantly improved by fine tuning the models using a small amount of compound specific reference data.

During our own fine tuning efforts, we experienced the fragmentation of current software ecosystems for machine learned force fields first hand. This motivated the development of our Python based aMACEing toolkit [2], which provides tools for the semi automatic fine tuning of foundational machine learned force fields. The toolkit supports the generation of training data, the fine tuning process itself, and the execution of machine learned force field based molecular dynamics simulations across multiple software packages.

The aMACEing toolkit for the semi automatic fine tuning of several foundational machine learned force fields is available at
https://amaceing-toolkit.readthedocs.io/