Microsoft Unveils Open-Source MatterGen AI Model That Can Generate Inorganic Material Designs

Microsoft researchers unveiled a new artificial intelligence (AI) model last week that can design new inorganic materials with desired properties. Dubbed MatterGen, the open-source large language model (LLM) is currently available in the public domain. The researchers have also published a paper highlighting that the technology can be used to accelerate the material design of new energy sources, semiconductors, and carbon capture. MatterGen has a diffusion-based architecture that is also the foundation for AI models such as OpenAI’s Dall-E and Stability AI’s Stable Image Ultra.

Microsoft Unveils MatterGen AI Model

Typically, generative AI models are associated with the generation of text, images, audio, and video based on prompts made by human users. However, MatterGen is different from the usual large language models, as detailed in a post last week. It can take requirements from a user and then generate a wide range of inorganic material designs.

Material design currently is a slow and methodical process where scientists use their knowledge and intuition to create better designs. An example of material design innovation is the recent usage of lithium carbide batteries in smartphones that allow a larger capacity while using less space. However, one downside of human-generated designs is that the experimentation and creation process takes a long time.

MatterGen, on the other hand, can generate crystalline structures across the periodic table, and even combine different elements. The model works on an atomic scale and can refine atom types, coordinates, as well as the periodic lattice. It can generate material designs at a high speed while running simulation-based experiments on them to determine each design’s viability, efficiency, and durability.

The researchers stated that a diffusion architecture was used to build this AI model. This type of architecture is generally seen in image and video generation models, as it has a better spatial and geometric understanding of shapes and designs.

In a paper published in the Nature journal, the researchers revealed that the base model was trained on a large dataset of more than 6,00,000 stable inorganic crystal structures. The dataset was compiled from the Materials Project and Alexandria databases. Next, adapter modules were added to enable fine-tuning for specific properties. This will allow scientists to add criteria such as specific chemical composition or magnetic density.

Currently, MatterGen’s source code is available to download and build upon in a GitHub listing. The AI model is available with an MIT license, which allows for both academic and commercial usage. The researchers believe this AI model can significantly benefit the world by accelerating the discovery of new materials for a variety of applications.