On 16 January 2025 Microsoft Research published MatterGen in Nature, marking a shift in AI-driven materials discovery from screening to generation. Where earlier systems such as GNoME predicted which candidate structures would be stable from a fixed list, MatterGen generates entirely new crystal structures directly from a target specification.
MatterGen is a diffusion model purpose-built for materials. Beginning from a random configuration, it iteratively adjusts atomic positions, elements and the crystal lattice to arrive at a stable structure, and it can be guided toward requested chemistry or toward specific mechanical, electronic or magnetic properties. Trained on more than 600,000 stable materials, the team did not stop at computational claims: they synthesized one AI-designed compound, TaCr2O6, in the laboratory and measured a property close to the requested target, providing experimental confirmation that the generated designs can be real.
The milestone matters because the ability to ask for a material with desired properties, rather than searching for one among known compounds, could accelerate the hunt for better batteries, magnets and other technologies. Together with GNoME, it signaled that generative AI was becoming a practical tool in the physical sciences, not just in text and images.