Materials innovation is one of the key drivers of major technological advances. The discovery of lithium cobalt oxide in the 1980s laid the foundation for today’s lithium-ion battery technology. It now powers modern mobile phones and electric cars, impacting the daily lives of billions of people. Materials innovations are also needed to design more efficient solar cells, cheaper batteries for grid-level energy storage, and sorbents to recycle CO2 from the atmosphere.
Finding new materials suitable for a given application is like finding a needle in a haystack. Historically, this work has been done through expensive and time-consuming experimental trial and error. More recently, computer screening of large materials databases has allowed researchers to speed up this process. Nevertheless, millions of candidates still need to be screened to find a small number of materials with desirable properties.
In a paper published today in Nature, (Opens in new tab)Now we’re sharing MatterGen, a generative AI tool that approaches materials discovery from a different angle. Instead of screening candidates, new materials are generated directly given the prompts of an application’s design requirements. Materials can be produced with desired chemical, mechanical, electronic, and magnetic properties, as well as various combinations of constraints. MatterGen enables a new paradigm of generative AI-assisted materials design that enables efficient exploration of materials beyond a limited set of known materials.
Novel diffusion architecture
MatterGen is a diffusion model that operates on the 3D geometry of materials. In the same way that image diffusion models generate images from text prompts by changing the color of pixels from noisy images, MatterGen was proposed by adjusting positions, elements, and periodic lattices from random structures. Generate the structure. Diffuse architectures are specifically designed to handle special materials such as periodicity and 3D geometry.
MatterGen’s base model delivers state-of-the-art performance in generating novel, stable, and diverse materials (Figure 3). Trained with 608,000 stable materials from the Materials Project (Opens in new tab) (MP) and Alexandria (Opens in new tab) (Alex) Database. The performance improvement can be attributed to both architectural advances and the quality and size of the training data.
MatterGen can be fine-tuned using labeled datasets to generate new materials given any desired conditions. We provide an example of generating novel materials by considering constraints on target chemistry and symmetry, as well as electronic, magnetic, and mechanical properties (Figure 2).
excellent screening
The main advantage of MatterGen over screening is the ability to access the entire space of unknown materials. Figure 4 shows that MatterGen continues to generate new candidate materials with high bulk modulus, for example above 400 GPa, that are difficult to compress. In contrast, screening baselines reach saturation because they run out of known candidates.
Spotlight: Event Series
microsoft research forum
Join us for an ongoing exchange of ideas about general AI-era research. Watch the first four episodes on demand.
Dealing with compositional disorders
Compositional disorder (Figure 5) is a commonly observed phenomenon in which different atoms can randomly exchange crystallographic positions within a synthesized material. recently (Opens in new tab)The community has been exploring what it means for a material to be novel in the context of computer-designed materials. This is because widely adopted algorithms cannot distinguish between pairs of structures that differ only by permutation of similar elements in each site.
We provide the first solution to this problem by introducing a new structure matching algorithm that takes compositional perturbations into account. This algorithm evaluates whether pairs of structures can be identified as ordered approximations of the same underlying compositionally disordered structure. This provides a new definition of novelty and originality, which we adopt in our calculated metrics. The algorithm is also published (Opens in new tab) As part of the evaluation package.
Laboratory validation
In addition to extensive computational evaluation, we validated MatterGen’s functionality through experimental synthesis. Cooperation with a team led by Professor Li Wenjie of Shenzhen Institute of Advanced Technology (Opens in new tab) In collaboration with the Chinese Academy of Sciences (SIAT), we synthesized a new material, TaCr2O6. The structure was generated by MatterGen after tuning the model with a bulk modulus value of 200 GPa. The structure of the synthesized material is consistent with the structure proposed by MatterGen, but note the compositional disorder between Ta and Cr. Additionally, we experimentally measured a bulk modulus of 169 GPa versus 200 GPa specified as the design specification. The relative error is less than 20%, which is very close from an experimental point of view. If similar results can be applied to other areas, it will have a major impact on the design of batteries, fuel cells, etc.
AI emulator and generator flywheel
MatterGen provides new opportunities for AI-accelerated materials design and complements the AI emulator MatterSim. MatterSim follows the fifth paradigm of scientific discovery and significantly accelerates the speed of material property simulation. MatterGen accelerates the search for new material candidates through property-based generation. MatterGen and MatterSim can work together as a flywheel to speed up both simulation and exploration of new materials.
Make MatterGen available
We believe the best way to make an impact in Material Design is to make your models publicly available. We will publish the MatterGen source code (Opens in new tab) Used with training and fine-tuning data under the MIT license. We welcome the community to use and build on our model.
Looking to the future
MatterGen represents a new paradigm in materials design enabled by generative AI technology. Investigate a significantly wider range of materials than screening-based methods. You can also be more efficient by using prompts to guide your material exploration. Similar to the impact generative AI has had on drug discovery (Opens in new tab)It will have a major impact on how materials are designed in a wide range of areas, including batteries, magnets, and fuel cells.
We plan to continue to further develop and validate our technology in collaboration with external collaborators. “At the Johns Hopkins Applied Physics Laboratory (APL), we are dedicated to exploring tools that have the potential to advance the discovery of new mission-enabling materials. We’re interested in understanding the impact,” said Christopher Stiles, a computational materials scientist who leads multiple materials discovery efforts at APL.
understand
This achievement is the result of a highly collaborative team effort at Microsoft Research AI for Science. All authors include Claudio Zeni, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabbé, Shoko Ueda, Roberto Sordillo, Lixin Sun, Jake Smith, Biclien Nguyen, and Hannes Schulz. Included. Sarah Lewis, Qingwei Huang, Zhiheng Lu, Yiyi Zhou, Han Yang, Hongxia Hao, Jieran Li, Chunlei Yang, Wenjie Li, Ryuta Tomioka, and Tian Xie.
