Official repository for the paper "Higher Rank Irreducible Cartesian Tensors for Equivariant Message Passing". It is built upon the ALEBREW repository and implements irreducible Cartesian tensors and their products.
Please consider citing us if you find the code and paper useful:
@misc{zaverkin2024higherrankirreduciblecartesiantensors,
title={Higher-Rank Irreducible Cartesian Tensors for Equivariant Message Passing},
author={Viktor Zaverkin and Francesco Alesiani and Takashi Maruyama and Federico Errica and Henrik Christiansen and Makoto Takamoto and Nicolas Weber and Mathias Niepert},
year={2024},
eprint={2405.14253},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2405.14253},
}
This repository implements:
- Irreducible Cartesian tensors up to a rank of three;
- Irreducible Cartesian tensor products (currently, only even tensor products);
- MACE-like architecture based on irreducible Cartesian tensors and their products.
This source code has a non-commercial license; see LICENSE.txt for more details.
An environment with PyTorch (>=2.3.1) and ASE (==3.22.1) installed. Also, some other dependencies may be necessary; see the ictp-cuda.yml file.
First, clone this repository into a directory of your choice git clone https://github.com/nec-research/ictp.git <dest_dir>. Then, move to <dest_dir> and install the required packages into a conda environment using, e.g., conda env create -f ictp-cuda.yml. Finally, set your PYTHONPATH environment variable to export PYTHONPATH=<dest_dir>:$PYTHONPATH.
We provide example scripts for training ICTP models for molecular (examples/run_training_DHA.py) and material (examples/run_training_HEA.py) systems. For the DHA molecule, first, download the corresponding data set by running wget http://www.quantum-machine.org/gdml/repo/static/md22_DHA.zip and unzip it with, e.g., unzip md22_DHA.zip. Then, store the md22_DHA.xyz file in the datasets/md22 subfolder and run python run_training_DHA.py to train your first ICTP model. The HEA data set can be downloaded from DaRUS. Please refer to examples/using_ictp.ipynb for more details on training ICTP models and using them in, e.g., molecular dynamics simulations.
In the experiments subfolder, we provide scripts to reproduce all results from the paper, along with data preparation scripts in the datasets subfolder. For the experiments with the original MACE source code, we used the commit 88d49f9ed6925dec07d1777043a36e1fe4872ff3.