MatterGen by Microsoft: status and timeline

Milestones

No announced next step

MatterGen model developed and trained2024Complete.
Nature paper publishedJan 2025Complete.
TaCr2O6 crystal synthesized and validated2025Complete.

Most important updates

Microsoft Research publishes MatterGen in Nature: AI designs crystals to specificationJan 2025

Current obstacles

Synthesis validation at scaleTesting every structure experimentally is costly. High-throughput synthesis is a bottleneck.
Complex property constraintsTargeting multiple properties simultaneously is harder; trade-offs not always clear.

Physics limits

Training data limits explorationCan't reliably generate structures beyond its training chemical space.
Stability prediction is probabilisticGenerated structures are likely stable but not guaranteed. Validation is expensive.

How it works

3 parts

Diffusion for structures

Diffusion in crystal space

Diffuses atomic positions, elements, and lattice parameters, learning to generate realistic crystals.

Property conditioning

Adapter modules for properties

Fine-tuned adapters enable design for bulk modulus, thermal conductivity, band gap, etc.

Validation

Synthesis and measurement

Generated structures are synthesized experimentally and measured to validate properties.

Spec sheet

Spec	MatterGen
Training materials database size	608,000 stable structures^R (reported)
TaCr2O6 bulk modulus vs target	169 GPa (target 200 GPa)^R (reported)
Stable structures vs SOTA	2.9x more than CDVAE^R (reported)

^R reported by the company

Papers & demos

1 item

Jan 2025paper
MatterGen: Property-guided materials design with generative AI (opens microsoft.com)Nature paper demonstrates diffusion models design novel crystals to specification, validated end-to-end.

Update log

1 update

Jan 2025

Major: Microsoft Research publishes MatterGen in Nature: AI designs crystals to specificationPaperResearch

About Microsoft

The team behind MatterGen

Microsoft

AI models, cloud computing, Azure, GitHub Copilot

Microsoft develops software, cloud, and hardware across computing and AI. It operates Azure and GitHub. In June 2026, Microsoft launched frontier AI models in-house, gaining freedom after its OpenAI partnership ended in April 2026.

Microsoft launched seven frontier AI models June 2026, trained in-house from scratch with zero competitor distillation.
GitHub Copilot reached 50M users by July 2026; 1 in 3 PRs now involve an AI agent.
Maia 200 (3nm, 216GB HBM3e) ships to Azure US Central, claims 3x FP4 vs Trainium3.

microsoft.com (opens microsoft.com)Company profile →

Founded: 197551 yrs
Headquarters: United States
Status: Public
Valuation: $3.5Tmarket capOct 2026
Works in: AIFrontier modelsMaterialsComputing
Coverage: 5 programs · 6 updateslatest 9d agochecked 10 Oct
Partners: OpenAI (opens openai.com)GitHub (opens github.com)
Suppliers: NVIDIA (opens nvidia.com)TSMC (opens tsmc.com)
People: Satya Nadella (opens en.wikipedia.org)CEOKevin Scott (opens en.wikipedia.org)Chief Technology Officer

Spec

MatterGen

Training materials database size

608,000 stable structures^R (reported)

TaCr2O6 bulk modulus vs target

169 GPa (target 200 GPa)^R (reported)

Stable structures vs SOTA

2.9x more than CDVAE^R (reported)