Overview

OMatGenerate is a web-based platform for generating crystal structures using state-of-the-art machine learning models. Built on the OMatG (Open Materials Generation) framework, it provides two generation modes for materials discovery, detailed below.

OMatG and OMatGenerate are part of Fermat-ML, a project of the KIM Initiative. Explore related projects on the KIM Initiative homepage.

Supported by the U.S. National Science Foundation: Award #2311632.

Also supported by resources from High Speed Research Network of New York University.

Generation Modes

Crystal Structure Prediction (CSP)

Generate crystal structures for specific chemical compositions. Provide a formula like "GaTe" or "Li2O", and the model predicts stable crystal structures with realistic atomic positions and lattice parameters.

  • Input: Chemical composition (e.g., "Ga4Te4", "SiO2")
  • Output: Multiple candidate crystal structures
  • Use case: Predicting polymorphs and structure candidates for known materials

De Novo Generation (DNG)

Generate novel crystal structures without fixing the chemical composition. Specify only the number of atoms, and the model creates diverse structures with varying element combinations.

  • Input: Number of atoms (2-100)
  • Output: Novel crystal structures with varied compositions
  • Use case: Materials discovery and exploring chemical space

Models

CSP mode uses the Trig-SDE-Gamma stochastic interpolant model trained on OMatG's custom splits of the Alex-MP-20 dataset.

DNG mode uses Linear-SDE-Gamma stochastic interpolant models trained on the Materials Project (MP-20) dataset, containing 45,229 structures.

Model weights are available on OMatG's HuggingFace page:

Training Data: Alex-MP-20 (CSP); Materials Project (MP-20) (DNG)
Architecture: Trig-SDE-Gamma Interpolant (CSP); Linear-SDE-Gamma Interpolant (DNG)
Framework: PyTorch + PyTorch Lightning

Features

  • Generate crystal structures using CSP and DNG modes
  • Relax structures using CHGNet
  • Download structures in CIF and XYZ formats
  • View and download OVITO-rendered structure images
  • View interactive 3D visualization of generated structures

How to Use

  1. Choose a mode: Select CSP for specific compositions or DNG for exploration
  2. Enter parameters:
    • CSP: Type chemical formulas (e.g., "GaTe, Li2O")
    • DNG: Choose number of atoms with the slider
  3. Set quantity: Choose how many structures to generate (max 100 total)
  4. Generate: Click and wait for results
  5. Explore: View 2D and 3D structures
  6. Relax: Optionally, select structures and click (under the Download Options tab) to generate relaxed structures using CHGNet
  7. Download: Download your generated or relaxed structures in CIF or XYZ format, optionally including PNG images

Technical Details

ML Framework: PyTorch 2.8, PyTorch Lightning 2.5
Visualization: OVITO (static PNG images); 3DMol (interactive 3D viewer)

Citation

OMatG is provided under the MIT license.

If you use OMatGenerate or OMatG in your research, please cite:

@inproceedings{
    hoellmer2025,
    title={Open Materials Generation with Stochastic Interpolants},
    author={Philipp H{\"o}llmer and Thomas Egg and Maya Martirossyan and Eric
    Fuemmeler and Zeren Shui and Amit Gupta and Pawan Prakash and Adrian
    Roitberg and Mingjie Liu and George Karypis and Mark Transtrum and Richard
    Hennig and Ellad B. Tadmor and Stefano Martiniani},
    booktitle={Forty-second International Conference on Machine Learning},
    year={2025},
    url={https://openreview.net/forum?id=gHGrzxFujU},
    archivePrefix={arXiv},
    eprint={2502.02582},
    primaryClass={cs.LG},
}

Links