Machine learning interatomic potentials – active learning techniques

Everything connected to the active learning tutorial for the CNPEM-MPG meeting 07/2025.

Overview

This tutorial illustrates the basic concepts of committee-based active learning using MACE potentials to represent the potential energy surfaces in simple, illustrative systems.

You will find two tutorials in the notebook folder:

• 1-zundel.ipynb: apply Query by Committee to the zundel cation from labeled data
• 2-eigen.ipynb: apply Query by Committee to the eigen cation from unlabeled data

Resources

• Website of the Workshop: https://pages.cnpem.br/ilum-maxplanck-meeting/
• Webpage of our research group (SAbIA): https://www.mpsd.mpg.de/research/groups/sabia
• GitHub repository of this tutorial: https://github.com/sabia-group/AL-tutorial
• MACE repository: https://github.com/ACEsuit/mace
• ase repository: https://gitlab.com/ase/ase

Installation

System Dependencies

This tutorial requires the system package python3-tk for GUI support (used by ase and related tools).
Please install it before proceeding:

sudo apt-get install python3-tk

Option 1: Using pip (recommended for simplicity)

Create and activate a virtual environment and install the package and its dependencies through the pyproject.toml file:

mkdir -p ~/venv
python -m venv ~/venv/alt
source ~/venv/alt/bin/activate  
# On Windows use: %USERPROFILE%\venv\alt\Scripts\activate
python -m pip install -e .

Option 2: Using conda (alternative approach)

Create a conda environment using the provided environment.yaml file and activate the alt environment:

conda env create -f environment.yml
conda activate alt

Verify Installation

Check that the main packages that we need are installed:

python tests/check.py

References

These tutorials are based on the following work:

• Schran C., Brezina K., Marsalek O. JCP, 153, 104105, 2020