mattergen

Why we included this project

MatterGen lets you generate new inorganic crystal structures rather than just predicting how known compounds behave. It is a diffusion-based generative model, and the repo ships pre-trained checkpoints, so you can start producing candidates without training anything yourself. The base model generates unconditionally, while fine-tuned variants steer output toward properties like chemical composition, symmetry, density, or formation energy. The repo also includes generation, evaluation, and training code, and researchers with their own labeled data can fine-tune the model to hit the specific constraints they care about. Backed by a peer-reviewed paper, it is a practical starting point for screening candidate materials or working on inverse design.