microsoft/mattergen

Official implementation of MatterGen -- a generative model for inorganic materials design across the periodic table that can be fine-tuned to steer the generation towards a wide range of property constraints.

What it solves

MatterGen is designed to automate the discovery and design of inorganic materials. Traditionally, finding new materials with specific desired properties (like magnetic density or band gap) is a slow, trial-and-error process. MatterGen allows researchers to generate new crystal structures that are tailored to meet specific property constraints across the periodic table.

How it works

MatterGen is a generative diffusion model. It can be used in two primary ways:

  1. Unconditional Generation: The model samples from a pre-trained base model to create new materials without specific constraints.
  2. Property-Conditioned Generation: By fine-tuning the base model, it can be steered to generate materials that match target properties, such as magnetic density, bulk modulus, or specific chemical systems. It uses classifier-free diffusion guidance to balance the adherence to target properties with the diversity and realism of the generated samples.

Who it’s for

This tool is primarily for materials scientists, chemists, and computational physicists who need to design new inorganic crystals with targeted physical or chemical properties.

Highlights

  • Multi-Property Conditioning: Ability to generate materials based on one or more target properties simultaneously (e.g., combining chemical system and energy above hull).
  • Pre-trained Models: Provides several checkpoints for various properties, including band gap, magnetic density, and space group.
  • Integrated Evaluation: Includes tools to assess the stability, novelty, and uniqueness of generated structures using machine-learning force fields like MatterSim.
  • Flexible Training: Supports training from scratch on datasets like MP-20 and fine-tuning via adapters for specific properties.