OMatG/MP-20-DNG
Updated
AI & ML interests
None defined yet.
Recent Activity
Organization Card
Open Materials Generation (OMatG)
About
OMatG is a generative model for crystal structure prediction and de novo generation of inorganic crystals.
This repository hosts our model checkpoints and benchmark datasets.
Models
Each of our models have been trained with a variety of hyperparameters as, e.g., various positional stochastic interpolants for the fractional coordinates in the unit-cell representation. Checkpoints for different positional stochastic interpolants are included in subdirectories within each model repository.
The tables below indicate the recommended checkpoints for each model, as well as the suggested use case.
Try our models live at OMatGenerate.
Crystal Structure Prediction (CSP)
| model | best checkpoints | match rate (%) | RMSE | notes |
|---|---|---|---|---|
| Alex-MP-20-CSP | Trig SDE Gamma | 72.50 | 0.1261 | Predict inorganic crystal structures of compositions with up to 20 atoms per unit cell. Largest training set; recommended over MP-20-CSP. |
| MP-20-CSP | Linear ODE | 69.83 | 0.0741 | Predict inorganic crystal structures of compositions with up to 20 atoms per unit cell. |
| MPTS-52-CSP | Linear ODE | 27.38 | 0.1970 | Predict inorganic crystal structures of compositions with up to 52 atoms per unit cell. Should only be used if strictly necessary; use Alex-MP-20-CSP models if possible. |
| perov-5-CSP | VPSBD ODE | 83.06 | 0.3753 | Predict perovskite structures with exactly 5 atoms per unit cell. Should only be used if strictly necessary. |
De Novo Generation (DNG)
| model | best checkpoints | S.U.N rate (%) | RMSD | notes |
|---|---|---|---|---|
| MP-20-DNG | Linear SDE Gamma | 22.48 | 0.6357 | Generate de novo crystal structures with up to 20 atoms per unit cell. |
Citation
Please cite our paper on OpenReview if using our models or datasets.
Links
OMatG on GitHub: See this repository for installation, training and usage instructions.
KIM Initiative: Knowledgebase of Interatomic Models. Tools and resources for researchers in materials science and chemistry.
Fermat-ML on GitHub: Foundational Representation of Materials. Machine learning foundation model for materials and chemistry discovery.
OMatGenerate: Try our models live at OMatGenerate, hosted on New York University's High Speed Research Network.