microsoft/mattergen

What it solves

MatterGen addresses the challenge of designing new inorganic materials with specific desired properties. Instead of relying on trial-and-error or limited search spaces, it allows researchers to generate new crystal structures across the periodic table that are tailored to meet specific property constraints.

How it works

MatterGen is a generative diffusion model. It can be used in an unconditional mode to sample new materials from a base model trained on large datasets (like MP-20). It can also be fine-tuned using adapters to become a property-conditioned model. This allows users to specify target properties—such as magnetic density, band gap, or bulk modulus—and steer the generation process toward materials that exhibit those characteristics. The project also integrates with MatterSim, a machine-learning force field, to relax generated structures and evaluate their stability, novelty, and uniqueness.

Who it’s for

This tool is designed for materials scientists, chemists, and computational researchers who need to discover and design inorganic crystals with specific physical or chemical properties.

Highlights

• Property-Conditioned Generation: Ability to generate materials based on one or multiple target properties (e.g., chemical system and energy above hull).
• Flexible Fine-tuning: Supports fine-tuning base models to steer generation toward specific constraints.
• Integrated Evaluation: Includes tools to assess generated materials for stability, novelty, and uniqueness using machine-learning force fields.
• Crystal Structure Prediction: Supports a specialized mode for generating structures based on a specific chemical formula.