MatterGen

MatterGen is a diffusion-based generative model for inorganic materials that directly designs novel crystal structures conditioned on desired properties. It iteratively refines atomic positions, element types, and lattice parameters from random initialization, jointly predicting all three under crystallographic constraints such as periodicity. Trained on roughly 608,000 stable materials from the Materials Project and Alexandria databases, it supports property-guided generation across bulk modulus, band gap, chemical system, and magnetic density. The work has been experimentally validated: researchers synthesized a novel material (TaCr₂O₆) generated by MatterGen targeting a 200 GPa bulk modulus, measuring 169 GPa.

MatterGen shifts materials discovery from exhaustive screening toward intelligent generation conditioned on application requirements, accessing the much larger continuous space of hypothetical stable structures. The model was published in Nature and is designed to work in concert with MatterSim — the simulator validates and ranks the generator’s proposals — forming a closed loop that accelerates both exploration and evaluation. Target domains include batteries, fuel cells, catalysts, magnets, and carbon-capture materials, where discovery has historically required years of experimental iteration. Source code, training data, and fine-tuning recipes are publicly available under the MIT license.