June 16, 2016
A modular approach to atomistic machine learning
We have created Amp, a modular package that allows for atomistic machine-learning in both atom-centered and image-centered mode. By allowing the user to specify both the transformation and the machine-learning model independently -- as well as build their own -- this software provides a flexible means to approximate the potential energy surface.
The code integrates with ASE and behaves as a trainable ASE calculator, which gives it the ability to interface with a wide variety of electronic structure calculators from NWChem and QuantumEspresso to Gaussian and VASP.
Our description of Amp was recently published in Computer Physics Communications, where you can find many more details on the approach. Documentation for the code lives on amp.readthedocs.io while the project's git repository is maintained on bitbucket.org/andrewpeterson/amp.