language:
- en
license: cc-by-nc-sa-4.0
size_categories:
- 10K<n<100K
pretty_name: Catalys_mxenes
task_categories:
- graph-ml
tags:
- MLIP
- 2DMxenes
This repository contains the dataset and models presented in the paper Benchmark Dataset for Catalysis on 2D MXenes.
The dataset encompasses a comprehensive collection of density functional theory (DFT) calculations for Ti$_2$CT$_y$ MXene configurations and molecular systems. It includes 50,000 calculations for training, 10,000 for testing, and an additional 1,000 larger systems to evaluate how well models generalise. These quantities (atomic forces and formation energies) are used to train and validate machine learning interatomic potential (MLIP) models.
Repository Structure
Datasets are stored in the datasets/ directory and are provided in both .h5 and .xyz formats. The directory also includes the 1000 larger systems dataset. For .h5 datasets (used with EquiformerV2), the training and validation split is handled automatically during loading. The remaining models use the .xyz format.
All model implementations, training code, evaluation code, and checkpoints are located in the models/ directory. Each model subdirectory contains:
runs/— stores trained model checkpoints.scripts/— contains training and evaluation shell scripts.
All scripts are intended to be launched from the repository root directory. For example, run a model evaluation using:
./models/<model_name>/scripts/eval_<model_name>.sh
Additional seetings, like other models, can be changed inside these shell scripts
Installation
We provide a conda/mamba environment file that should work for all four models.
Create the environment with:
mamba env create -f catalyst_env.yml
Activate the environment using:
mamba activate catalyst
If some packages fail to install due to dependency conflicts, install the missing packages manually using pip or mamba.