# Release notes¶

## Development version¶

(Significant changes since the last release are itemized here.)

• Added the Nearsighted Force Training approach.
• Added image and feature selection methods.
• Added offsets in G2 fingerprints; that is, G2 symmetry functions with shifted Gaussian centers can be used.
• Fast force calls are now supported, via third-party codes. See the fast force calls portion of the documentation.
• A documented bootstrap module, complete with examples of use, is included for uncertainty predictions.
• Improved interprocess communication which should reduce network traffic for parallel jobs.
• Amp is now part of the Debian archives! This means it should soon be available in package managers for linux releases such as Ubuntu.
• The convergence plots (via amp.analysis and amp-plotconvergence) now handle multiple training attempts from a single log file.
• The image hashing routine, used to uniquely identify images, has been updated to correctly handle permutations in very large atomic systems. (Note this means that images hashed with a prior version of Amp will have a different unique identifier, so you should not mix databases of fingerprints.)
• Added Kernel Ridge Regression to Amp.
• Incorporation of Behler’s G5 angular symmetry function.
• Neural network training scripts are now re-submittable; that is, if a job times out it can be re-submitted (unmodified) and will pick up from the last checkpoint.

## 0.6.1¶

Release date: July 19, 2018

• Installation via pip is now possible.

## 0.6¶

Release date: July 31, 2017

• Python 3 compatibility. Following the release of python3-compatible ASE, we decided to jump on the wagon ourselves. The code should still work fine in python 2.7. (The exception is the tensorflow module, which still only lives inside python 2, unfortunately.)
• 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. This makes convergence easier for new users.
• Convergence plots show maximum residuals as well as root mean-squared error.
• Parameters to make the Gaussian feature vectors are now output to the log file.
• The helper function make_symmetry_functions() has been added to more easily customize Gaussian fingerprint parameters.

Permanently available at https://doi.org/10.5281/zenodo.836788

## 0.5¶

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 https://doi.org/10.5281/zenodo.322427

## 0.4¶

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. http://dx.doi.org/10.1016/j.cpc.2016.05.010

Permanently available at https://doi.org/10.5281/zenodo.46737

## 0.3¶

Release date: July 13, 2015

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

Permanently available at https://doi.org/10.5281/zenodo.20636

## 0.2¶

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 https://bitbucket.org/andrewpeterson/neural/commits/tag/v0.2

## 0.1¶

Release date: November 12, 2014

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

Permanently available at https://doi.org/10.5281/zenodo.12665.

## Alpha version milestones¶

First public code (bitbucket): September 1, 2014.

First project commit: May 5, 2014.