Revolutionizing Material Discovery with Generative AI: Introducing MatterGen
Brief news summary
Materials innovation is key to technological progress, as shown by the development of lithium-ion batteries. Traditionally, discovering new materials involves a slow, costly trial-and-error process. Although computational screening can assess millions of materials, it remains time-intensive. Enter MatterGen, an AI tool presented in a Nature paper, aiming to revolutionize this process. Using generative techniques, MatterGen designs materials based on specific criteria and is trained on over 600,000 examples from sources like the Materials Project. It employs a diffusion model that focuses on 3D structures, enabling the exploration of new material spaces and customization of materials with desired properties, surpassing traditional methods. MatterGen addresses challenges such as compositional disorder and has achieved experimental success at the Shenzhen Institutes of Advanced Technology. Together with MatterSim, an AI simulation tool, it greatly accelerates material discovery and simulation. As an open-source platform, MatterGen encourages community collaboration for ongoing improvements. Like AI’s impact on drug discovery, MatterGen could lead to breakthroughs in material design, especially for batteries and fuel cells. It is supported by entities such as Johns Hopkins University Applied Physics Laboratory and is part of Microsoft Research AI for Science’s initiatives.Materials innovation is crucial for technological breakthroughs, as demonstrated by the discovery of lithium cobalt oxide, which underpins current lithium-ion batteries that power mobile phones and electric cars. Materials innovation is also essential for efficient solar cells, economical batteries for energy storage, and CO2 recycling adsorbents. Traditionally, finding new materials involves costly trial-and-error, but computational screening has sped up this process by evaluating extensive materials databases. MatterGen, detailed in a Nature paper, presents a novel approach to materials discovery using generative AI. Instead of screening materials, MatterGen directly creates them based on specific application requirements, making it possible to design materials with various desired properties. This generative AI tool supports efficient exploration beyond well-known materials. MatterGen uses a diffusion model operating on material 3D geometries, generating structures by adjusting the positions and elements in a random setup. It's trained with data from 608, 000 stable materials and can be fine-tuned with labeled datasets to generate novel materials tailored to chemistry, symmetry, and various properties. Unlike traditional screening, MatterGen accesses unexplored materials and continues to generate novel candidates with specific traits. To address compositional disorder—where atoms swap sites within a material—MatterGen introduces a new structure-matching algorithm.
This algorithm redefines novelty by assessing whether structures represent variations of the same compositionally disordered template. Experimental validation involved synthesizing a new material, TaCr2O6, which showed results closely aligning with MatterGen's predictions. MatterGen complements the AI emulator MatterSim, forming a "flywheel" that accelerates both simulation and exploration of materials, potentially enhancing applications in batteries, magnets, and fuel cells. The MatterGen model, source code, and data are publicly released under the MIT license. Looking ahead, continued work with collaborators, such as at the Johns Hopkins University Applied Physics Laboratory, aims to realize MatterGen's full potential. This project emerged from teamwork at Microsoft Research AI for Science, involving a diverse group of researchers.
Watch video about
Revolutionizing Material Discovery with Generative AI: Introducing MatterGen
Try our premium solution and start getting clients — at no cost to you
