“A generative model for inorganic materials design,” by Claudio Zeni, Tian Xie and colleagues at Microsoft Research’s AI for Science lab, was published in Nature in January 2025. It introduced MatterGen, which inverts the usual approach to materials discovery: instead of screening a list of candidate crystals for desired properties, it generates new crystal structures directly from a target specification.
MatterGen is a diffusion model adapted for materials science. Starting from a random arrangement, it iteratively adjusts the atomic positions, chemical elements and periodic lattice to produce a stable structure, and it can be steered toward specified chemistry, mechanical, electronic or magnetic properties, or combinations of them. The model was trained on more than 600,000 stable materials. To show the predictions were real rather than only computational, the team synthesized one AI-generated compound, TaCr2O6, in the laboratory and measured a property close to the target.
Generative design points to a different relationship between scientists and search: rather than asking a model to rank what already exists, researchers can ask it to invent candidates meeting a need, such as a magnet without rare-earth elements. MatterGen, following GNoME, marked the move from screening to generation in computational materials discovery.