Multi-Modal Foundation · microsoft
MatterGen
A diffusion model that generates novel, stable inorganic crystal structures with user-defined target properties.
What MatterGen is
MatterGen is a generative diffusion model designed for inorganic materials discovery, developed by the Materials Design Team at Microsoft Research AI for Science. It addresses the challenge of exploring vast chemical search spaces by proposing new inorganic crystal structures either unconditionally or guided by specific material properties.
The model simultaneously generates atomic types, fractional coordinates, and unit cell lattice vectors. By proposing stable, novel, and unique crystalline candidates, it provides materials scientists and chemists with computationally viable leads for applications across energy, catalysis, and functional materials research.
How it works
MatterGen is a diffusion model with an architecture based on GemNet. It operates on primitive crystal cells by jointly predicting atomic fractional coordinates, element types, and lattice vectors. Unit cell lattices are preprocessed using Niggli reduction followed by polar decomposition to ensure that lattice matrices are symmetric.
Training was restricted to structures containing up to 20 atoms in the unit cell and an energy below 0.1 eV/atom above the reference convex hull. Noble gases, elements with atomic numbers above 84, and the radioactive elements technetium (Tc) and promethium (Pm) were removed prior to training.
Key facts
- Model architecture
- GemNet-based diffusion model
- Parameter count
- 46.8M
- Training data
- DFT-relaxed crystal structures from Materials Project (v2022.10.28) and Alexandria
- Maximum unit cell size
- 20 atoms
- Sampling throughput
- 1,000 structures in approximately 2 hours on a single NVIDIA V100 GPU
What you can use it for
- Unconditional crystal structure generation
- Generates novel, stable, and unique inorganic crystalline candidates directly from random noise without predefined target constraints.
- Chemical system-guided generation
- Proposes valid crystal structures constrained to specific chemical systems, covering well-explored, partially explored, or entirely unexplored elemental combinations.
- Property-targeted materials discovery
- Produces candidate structures conditioned on target physical parameters, such as a bulk modulus of 400 GPa or a magnetic density exceeding 0.2 Å⁻³.
- Custom property fine-tuning
- Allows researchers to fine-tune the base model using several thousand labeled property data points to target arbitrary, user-defined material properties.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Achieves a 38.57% rate of stable, unique, and novel (S.U.N.) materials in unconditional generation, with an average root mean square distance of 0.021 Å to DFT-relaxed local minima.
- Can sample extreme target properties with very sparse training precedents, such as generating bulk modulus materials exceeding 400 GPa when only two such examples existed in the reference set.
- Jointly predicts lattice vectors, atomic positions, and element identities within a unified diffusion framework.
Watch out for
- Cannot generate or process crystal structures with more than 20 atoms inside the unit cell.
- Cannot generate organic crystals or non-crystalline materials.
- Excluded noble gases, radioactive elements (including Tc and Pm), and all elements with atomic numbers above 84 from its chemical space.
- Performance on property-guided tasks degrades for extreme values where training examples are sparse and requires at least several thousand labeled examples when fine-tuning for new properties.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use MatterGen commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
What is meant by a 'S.U.N.' structure?
S.U.N. stands for stable, unique, and novel. In the context of MatterGen, a structure is considered stable if its energy is within 0.1 eV/atom of the reference convex hull, novel if it does not match reference dataset entries using a disordered structure matcher, and unique if it appears only once among the generated batch.
Can MatterGen generate disordered or organic materials?
No. MatterGen is explicitly designed for inorganic crystalline materials and was trained only on primitive crystal structures. It does not support organic crystals or non-crystalline phases.
How much data is required to fine-tune the model for a new material property?
Fine-tuning effectively requires a dataset of at least several thousand labeled structures with sufficient coverage across the target property range.
How to cite it
Zeni, Claudio, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabbé, Shoko Ueda, Roberto Sordillo, Lixin Sun, Jake Smith, Bichlien Nguyen, Hannes Schulz, Sarah Lewis, Chin-Wei Huang, Ziheng Lu, Yichi Zhou, Han Yang, Hongxia Hao, Jielan Li, Chunlei Yang, Wenjie Li, Ryota Tomioka, and Tian Xie. "A generative model for inorganic materials design." Nature (2025). https://doi.org/10.1038/s41586-025-08628-5
Written from MatterGen's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.