mlse2020-workshop

Tutorial Workshop: Machine Learning in Materials Science, Dec 13-15, 2020: https://www.eventbrite.com/e/tutorial-workshop-machine-learning-in-materials-science-tickets-128297271593

See also: Strategies for the Construction of Neural-Network Based Machine-Learning Potentials (MLPs), A.M. Miksch*, T. Morawietz, J. Kästner, A. Urban, N. Artrith*, “Strategies for the Construction of Machine-Learning Potentials for Accurate and Efficient Atomic-Scale Simulations”, Mach. Learn.: Sci. Technol. 2 (2021) 031001. DOI: https://doi.org/10.1088/2632-2153/abfd96
Or https://github.com/atomisticnet/MLP-beginners-guide
Contact: Nong Artrith ([email protected])

SUNDAY, DECEMBER 13 – DATA FROM SIMULATION

9:00 AM - 9:30 AM: Overview: Machine Learning for Materials Science: Alexander Urban, Columbia University

9:30 AM - 11:00 AM: Artificial Neural Network Potentials for Materials: Nongnuch Artrith, Columbia University

The link to the ænet Google Colab notebook is: https://colab.research.google.com/drive/1Uz3xulDPMoAEHj1RNPYdq-fAwBwId-aB?usp=sharing

Contact: Nong Artrith ([email protected]) and Alex Urban ([email protected])

To learn more about ænet, sign up to the Google Group so that you don’t miss any announcements (e.g., for new releases) and can reach a wider community with any questions/issues related to ænet. Once subscribed, you can also post by sending emails to [email protected].

Example of construction and application of ANN potentials

This directory contains an example that showcases the Chebyshev descriptor for local atomic environments proposed in Phys. Rev. B 96 (2017) 014112.

Owing for the large amount of data the reference data set and the training set files are not included in the example. Otherwise the example is self-contained. The reference data set used is the TiO₂ data set from Comput. Mater. Sci. 114 (2016) 135-150, which can be downloaded from ann.atomistic.net. Using that data set, the training set file can be generated as described in the first subdirectory 01-generate.

Output files generated at each step are contained in output subdirectories.

01-generate

Evaluation of the Chebyshev descriptor for all structures in the reference data set. The resulting feature vectors are written to a training set file.

• generate.in: Input file for generate.x
• O.fingerprint.stp and Ti.fingerprint.stp: Descriptor definitions for the atomic species O and Ti. For both species relatively small descriptor sizes with a radial expansion order of 16 and an angular order of 4 are used for set001 and set002. Another set (set003) uses different order of expansions: radial 22 and angular 6.

02-train

Training examples using the training set files generated in the previous step (not included because of the file size). The results of three training runs with different neural network sizes are shown in the subdirectories set001, set002, and set003.

• train.in: Input file for train.x
• get-energies: Subdirectory demonstrating how to write out the energies and errors of all training and testing samples (after training has completed).

Training will generate the ANN potential files O.15t-15t.nn and Ti.15t-15t.nn which are also provided in the output subdirectory.

03-predict

Usage of the ANN potentials trained in the previous step for the prediction of the entire reference data set.

• predict.in: Input file for predict.x
• O.15t-15t.nn and Ti.15t-15t.nn: ANN potential files

The output (energies and atomic forces) can be found in the output subdirectory.

04-aenetLib-ann-md

Usage of the trained ANN potentials from set003 with external solftware. Example for the Python (API) with the Atomistic Simulation Environment (ASE). See the ænet Google Colab notebook.

• O.40t-40t.nn and Ti.40t-40t.nn: ANN potential files