Release notes

Development version

(Significant changes since the last release should be itemized here.)

A community page has been added with resources such as the new mailing list and issue tracker.

The default convergence parameters have been changed to energy-only training; force-training can be added by the user via the loss function.

Convergence plots show maximum residuals as well as root mean-squared error.


Release date: February 24, 2017

The code has been significantly restructured since the previous version, in order to increase the modularity; much of the code structure has been changed since v0.4. Specific changes below:

  • A parallelization scheme allowing for fast message passing with ZeroMQ.
  • A simpler database format based on files, which optionally can be compressed to save diskspace.
  • Incorporation of an experimental neural network model based on google’s TensorFlow package. Requires TensorFlow version 0.11.0.
  • Incorporation of an experimental bootstrap module for uncertainty analysis.

Permanently available at


Release date: February 29, 2016

Corresponds to the publication of Khorshidi, A; Peterson*, AA. Amp: a modular approach to machine learning in atomistic simulations. Computer Physics Communications 207:310-324, 2016.

Permanently available at


Release date: July 13, 2015

First release under the new name “Amp” (Atomistic Machine-Learning Package/Potentials).

Permanently available at


Release date: July 13, 2015

Last version under the name “Neural: Machine-learning for Atomistics”. Future versions are named “Amp”.

Available as the v0.2 tag in


Release date: November 12, 2014

(Package name: Neural: Machine-Learning for Atomistics)

Permanently available at

First public bitbucket release: September, 2014.