MatterGen

MatterGen is a generative diffusion model for inorganic materials design from Microsoft Research's AI for Science team, unveiled in January 2024 with a paper in Nature. Given a prompt specifying desired material properties — hardness, conductivity, magnetism, bulk modulus — it generates novel crystal structures (chemical composition plus 3D lattice) from scratch, using a two-stage diffusion process over the crystal's shape, atom types, and coordinates, with property-conditioning networks injected at multiple layers. In a headline demonstration, the team collaborated with researchers at China's Shenzhen Institutes to synthesize TaCr2O6, a material MatterGen had designed, with the real-world material closely matching the AI's predictions. Microsoft released the model and code under an open-source license alongside MatterSim, its companion property-prediction model, via the Azure Quantum Elements platform.

Limitations: MatterGen generates crystals only — no liquids, gases, or amorphous solids like glass — and computationally designed materials still require extensive experimental validation before any industrial use, as the authors and outside experts stress. Training data is available non-commercially, so check the current license terms on Microsoft's repo if you plan commercial R&D. It is a research tool for materials discovery, not a substitute for lab synthesis and testing.