MatterGen
Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,092 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,092 words
Add missing citations, update stale details, or suggest a clearer explanation.
MatterGen is a generative artificial intelligence model from Microsoft Research for designing novel inorganic crystalline materials. Built by the company's AI for Science division, it is a diffusion model that generates three-dimensional crystal structures by jointly producing the atom types, the atomic coordinates, and the periodic lattice, and it can be fine-tuned to steer generation toward target properties such as a desired magnetic density, electronic band gap, mechanical bulk modulus, chemical system, or crystal symmetry. Microsoft first released a preprint describing MatterGen in December 2023, published a peer-reviewed account in the journal Nature on 16 January 2025, and open-sourced the code and trained model weights under an MIT license. The work is part of a broader push to apply generative AI to materials science, and it is frequently compared with Google DeepMind's GNoME effort, though the two systems take different approaches. MatterGen has drawn both substantial interest and pointed independent criticism over how many of its outputs are genuinely new.
The traditional computational route to a new material is screening: enumerate or sample many candidate structures, compute their properties, and filter for the ones that meet a requirement. MatterGen inverts that workflow. Rather than searching a fixed list, it directly generates candidate crystals conditioned on the properties an application needs, an approach often called inverse design. Microsoft positions this as a shift from screening toward generation, arguing that a generative model keeps producing useful candidates after a finite screening library has been exhausted. In one comparison reported by Microsoft, MatterGen continued to generate stable structures with a bulk modulus above 400 gigapascals (more than 100 of them) where a screening baseline produced roughly 40 before saturating.
MatterGen was developed by Microsoft Research's AI for Science group, with Tian Xie, a senior researcher and project lead, among the authors; the Nature paper lists a large team including Claudio Zeni, Robert Pinsler, Daniel Zuegner, Andrew Fowler, Matthew Horton, and Ryota Tomioka. The model is available for research use, including through Microsoft's Azure AI Foundry Labs, in addition to the open-source release on GitHub and Hugging Face.
| Fact | Detail |
|---|---|
| Developer | Microsoft Research, AI for Science division |
| Type | Diffusion-based generative model for inorganic crystals |
| Preprint | arXiv 2312.03687, December 2023 |
| Peer-reviewed paper | Nature, 16 January 2025; "A generative model for inorganic materials design"; DOI 10.1038/s41586-025-08628-5 |
| Training data | About 608,000 stable structures (Alex-MP-20, drawn from the Materials Project and Alexandria databases) |
| Reported novelty/stability | More than twice as likely to be new and stable versus prior generative models |
| Reported structure quality | More than ten times closer to the local energy minimum; average RMSD of about 0.021 angstrom to the DFT-relaxed structure |
| Conditioning targets | Chemistry, symmetry, magnetic density, electronic band gap, mechanical bulk modulus, supply-chain risk |
| Experimental test | TaCr2O6 synthesized with collaborators at SIAT; bulk modulus target 200 GPa, measured about 169 GPa |
| Companion model | MatterSim (deep-learning atomistic simulator) |
| License | MIT (code and weights) |
MatterGen adapts the diffusion-model recipe that powers image generators such as Stable Diffusion to the periodic, symmetric world of crystals. A crystal is represented by three coupled objects: the discrete atom types A, the fractional atomic coordinates X, and the lattice L that defines the repeating unit cell. The forward process corrupts all three toward a simple random distribution, and a learned score network is trained to reverse that corruption, denoising a random configuration step by step into an ordered, stable structure.
The design choices reflect the physics. Because crystal coordinates are periodic, MatterGen diffuses the fractional coordinates using a wrapped normal distribution, so positions wrap around the unit cell and the model respects translational symmetry. Atom types are discrete, so the model uses a discrete diffusion scheme (the D3PM framework) in which a Markov chain randomly swaps element identities according to a transition matrix and the network learns to recover the correct elements. The lattice is corrupted toward a distribution of physically plausible cells rather than arbitrary noise. The underlying network is an equivariant graph neural network, meaning its predictions transform consistently under rotations and translations of the crystal, which matters because a material's properties do not depend on how it is oriented in space. The base model is pre-trained on a large dataset of stable structures.
Property control is added through fine-tuning rather than retraining from scratch. Microsoft inserts lightweight adapter modules into the score network and fine-tunes them on labeled data so the model can condition generation on a target property. This is what lets MatterGen aim for a particular bulk modulus, band gap, magnetic density, chemical system, or space group, and even combine constraints. The Nature work demonstrates multi-property design by requesting structures that have both a high magnetic density and a composition with low supply-chain risk, as scored by a Herfindahl-Hirschman index over the constituent elements.
The peer-reviewed paper, "A generative model for inorganic materials design," appeared in Nature on 16 January 2025. It reports that MatterGen was trained on roughly 608,000 stable materials assembled from the Materials Project and the Alexandria database (the Alex-MP-20 dataset, restricted to cells of at most 20 atoms and within 0.1 eV per atom of the convex hull). Microsoft evaluates outputs on the fraction that are stable, unique, and new (the SUN metric, where stability is judged against the convex hull and novelty against the training and reference sets) and on how close generated structures sit to their relaxed ground state.
On those measures the paper reports that MatterGen produces structures that are more than twice as likely to be new and stable as those from earlier generative baselines, and more than ten times closer to the local energy minimum, with an average root-mean-square displacement of about 0.021 angstrom between a generated structure and its density functional theory (DFT) relaxed counterpart. DFT, a first-principles quantum-mechanical method, is used throughout as the reference for energy and stability.
The headline validation is an experimental synthesis. Working with Professor Li Wenjie's group at the Shenzhen Institutes of Advanced Technology (SIAT) of the Chinese Academy of Sciences, the team fine-tuned MatterGen to target a high bulk modulus, generated thousands of candidates, and filtered them by uniqueness, novelty, energetic and phonon stability, and composition. A short list was taken into the lab, and one candidate with the nominal formula TaCr2O6, conditioned on a target bulk modulus of 200 gigapascals, was successfully synthesized (in a disordered form). The measured bulk modulus was about 169 gigapascals, which Microsoft describes as within roughly 20 percent of the design target. The data release accompanying the paper includes the crystallographic files plus X-ray diffraction, X-ray photoelectron spectroscopy, and nanoindentation measurements of the sample.
MatterGen is designed to work alongside MatterSim, a separate deep-learning model from the same Microsoft group that acts as a fast atomistic simulator and property predictor. Described in a 2024 preprint, MatterSim is a universal machine-learning force field trained by active learning on large-scale first-principles calculations; it covers a wide range of elements and conditions, spanning temperatures from 0 to about 5,000 kelvin and pressures up to about 1,000 gigapascals, and Microsoft reports up to a tenfold accuracy improvement over prior universal force fields under demanding conditions. It uses M3GNet and Graphormer architectures as backbones and was trained on roughly 17 million structures.
The intended pairing is a loop: MatterGen proposes candidate structures, and MatterSim cheaply screens and relaxes them and estimates stability (including phonon stability), narrowing the field before expensive DFT or laboratory work. Microsoft describes the two models as a flywheel for materials discovery and has folded them into its wider AI for Science and Azure efforts, including making MatterGen available to researchers through Azure AI Foundry Labs. Both MatterGen and MatterSim were released as open source under permissive licenses, with the stated aim of letting outside groups reproduce and build on the work.
MatterGen attracted considerable attention on release as one of the most visible examples of generative AI applied to physical-science design, and the open-source release was welcomed by researchers who could test it directly. The experimental synthesis was singled out as a notable step, since most generative-materials claims stop at computation.
That same flagship result became the focus of independent criticism. In a paper titled "Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset," Mikkel Juelsholt of Humboldt University of Berlin and the Max Planck Institute for Coal Research (Max-Planck-Institut fuer Kohlenforschung) re-examined the synthesized TaCr2O6. The work appeared as a ChemRxiv preprint in January 2026 and was published in the journal Materials Horizons, first online on 20 April 2026. Juelsholt argues, on the basis of crystallographic analysis and Rietveld refinement of powder X-ray diffraction data, that the material is not new: it is the same disordered phase as Ta1/2Cr1/2O2, first reported in 1971 and already present in MatterGen's training data (cited as ICSD collection code 9516). The broader point is that MatterGen, like other high-throughput pipelines, struggles with disordered compounds in which several elements share a crystallographic site, and that the system "cannot distinguish between its training data and the compounds it predicts." The critique also flags refinement problems in the original synthesis report, such as atomic displacement parameters pinned at zero.
This sits within a wider debate about how to measure success in AI-driven materials discovery. Similar novelty-and-usefulness critiques were leveled at GNoME: materials scientists Anthony Cheetham and Ram Seshadri of the University of California, Santa Barbara argued that the GNoME haul consisted of crystalline inorganic compounds that should be described as such rather than as "materials" in general, and questioned how many were actually useful. The recurring concern across both projects is that headline counts of stable or novel structures can overstate practical impact, because thermodynamic stability on paper does not guarantee synthesizability, and apparent novelty can dissolve once disorder, known phases, and database coverage are accounted for. These remain open methodological questions rather than settled verdicts.
MatterGen is significant less for any single discovered compound than for what it demonstrates about method. It showed that a deep learning diffusion model can generate crystal structures conditioned directly on target properties, that the same approach can carry through a full pipeline from generation to filtering with a learned simulator to an actual laboratory synthesis, and that releasing the code and weights openly lets the community scrutinize the claims. The skeptical follow-up work is itself part of that significance: it sharpened the field's standards for what counts as a genuinely novel, useful material and underscored that AI proposals still require careful human and experimental verification. Together with GNoME and a growing set of AI for science tools, MatterGen marks an early, contested step toward inverse design of inorganic materials, in which a researcher specifies the properties they want and a model proposes structures to make.