(Estimated reading time: 8 minutes)
Editor’s note: Recently, the Microsoft Research Center for Scientific Intelligence proposed an innovative generative AI material design tool, MatterGen, which breaks through the limitations of traditional material screening methods. It can efficiently explore a broader material space and directly generate new materials based on application requirements. The model can optimize for specific geometric structures of materials, generating stable new materials across various aspects such as chemical composition and physical properties, opening a new paradigm of generative AI-assisted material design that will have a profound impact on innovations in fields such as batteries and magnets.
Material innovation is one of the key factors driving major technological breakthroughs. The discovery of lithium cobalt oxide in the 1980s laid the foundation for today’s lithium-ion battery technology. Today, lithium-ion batteries power modern smartphones and electric vehicles, impacting the daily lives of billions. Material innovation is also essential for designing more efficient solar cells, low-cost batteries for grid-level energy storage, and adsorbents necessary for carbon dioxide recovery from the atmosphere.
Finding a new material that meets specific needs is like searching for a needle in a haystack. In the past, this task required expensive and time-consuming trial-and-error experiments. In recent years, computational screening of large-scale material databases has accelerated this process, but it still requires screening millions of candidate materials to find a few suitable ones.
Figure 1: Schematic diagram of material design screening and generation methods
In a recent paper published in the top scientific journal Nature, researchers from the Microsoft Research Center for Scientific Intelligence proposed a generative AI tool, MatterGen, that addresses the material discovery problem from different angles. Unlike screening candidate materials, MatterGen generates new materials directly based on the design requirements of application needs, producing materials with desired chemical, mechanical, electronic, or magnetic properties that meet various constraints. MatterGen opens a new paradigm of generative AI-assisted material design, capable of efficiently exploring materials and surpassing known material boundaries.
Paper link:
https://www.nature.com/articles/s41586-025-08628-5
GitHub link:
https://github.com/microsoft/mattergen
Innovative Diffusion Architecture
As a diffusion model based on the three-dimensional geometric structure of materials, MatterGen can generate the desired material structure by adjusting elements, positions, and periodic lattices in random structures, similar to how image diffusion models modify pixel colors of noisy images based on text prompts to generate pictures. This diffusion architecture is specifically designed for material design, aiming to address special property issues such as periodicity and three-dimensional geometry.
The foundational model of MatterGen has achieved leading levels in generating novel, stable, and diverse materials (as shown in Figure 2). The model is trained on 608,000 stable materials from the Materials Project (MP) and Alexandria (Alex) databases. Its performance improvement is attributed to architectural optimization and high-quality, large-scale training data.
Figure 2: Performance comparison of MatterGen with other methods in generating stable, novel, and unique material structures. The training datasets for each method are indicated in parentheses. Purple indicates performance improvement due solely to the MatterGen architecture, while cyan indicates performance improvement due to a larger dataset.
Materials Project
https://next-gen.materialsproject.org/
Alexandria Database
https://alexandria.icams.rub.de/
MatterGen can also be fine-tuned using labeled datasets to generate new materials under any specified conditions. Figure 3 demonstrates examples of generating new materials under constraints of target chemical composition and symmetry, as well as electronic, magnetic, and mechanical property limits.
Figure 3: Schematic of MatterGen. MatterGen can fine-tune the model based on different design requirements, such as specific chemical compositions, crystal symmetries, or material properties.
Superior to Traditional Screening Methods
The main advantage of MatterGen over screening methods is its ability to access the complete space of unknown materials. As shown in Figure 4, MatterGen excels in generating new candidate materials with high bulk moduli (e.g., greater than 400 GPa) that are difficult to compress. In contrast, screened benchmark methods are limited to known materials and tend to saturate.
Figure 4: Performance comparison of MatterGen (cyan) and traditional screening methods (yellow) in finding novel, stable, and unique structures that meet design requirements (e.g., with bulk moduli greater than 400 GPa).
Handling Component Disorder
Component disorder is a common phenomenon, referring to the random exchange of different atoms in their crystallographic positions within synthesized materials. In recent years, the community has continued to explore what constitutes novel materials in the context of computational material design, as widely used algorithms cannot distinguish between two structures that differ only in the arrangement of similar elements.
Figure 5: Schematic of component disorder. Left: A perfect crystal without component disorder, with repeating unit cells (black dashed line). Right: A crystal with component disorder, where each position has a 50% probability of being occupied by yellow or cyan atoms.
By introducing a new structural matching algorithm that considers component disorder, researchers provide a preliminary solution to this problem. This algorithm can evaluate whether two structures can be recognized as ordered approximate structures of the same base components with disorder, providing a new definition of novelty and uniqueness for materials. Currently, this method has been adopted in computational evaluation metrics and publicly released as part of evaluation tools.
GitHub link:
https://github.com/microsoft/mattergen?tab=readme-ov-file#evaluation
In addition to extensive computational evaluations, the capabilities of MatterGen have also been validated through experimental synthesis. The Microsoft Research Center for Scientific Intelligence collaborated with the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, to synthesize a new material, TaCr2O6, whose structure was generated by MatterGen under the condition of a bulk modulus of 200 GPa. The structure of the synthesized material matches that proposed by MatterGen, but there is component disorder between Ta and Cr.
Furthermore, the measured bulk modulus of this material is 169 GPa, which is less than 20% error compared to the design specification of 200 GPa, indicating a strong experimental agreement. If similar results can be generalized to other fields, MatterGen will have a profound impact on the design of materials for batteries, fuel cells, and more.
Figure 6: Experimental validation of the proposed compound TaCr2O6
AI Simulator and Generator Flywheel
MatterGen provides new opportunities for AI-accelerated material design, complementing the AI simulator MatterSim previously proposed by the Microsoft Research Center for Scientific Intelligence. MatterSim follows the fifth paradigm of scientific discovery, significantly accelerating the speed of material property simulations. MatterGen, on the other hand, accelerates the exploration of new material candidates based on property generation. The synergy between the two can accelerate both material simulations and the exploration of new materials, creating a flywheel effect.
To maximize the impact of material design, researchers have now released the source code of MatterGen under the MIT license and made the training and fine-tuning data publicly available for community use and further development.
GitHub link:
https://github.com/microsoft/mattergen
Using generative AI technology, MatterGen opens a new chapter in material design. It explores a broader material space than screening methods and efficiently drives material exploration through instructions. Similar to the impact of generative AI on drug discovery, MatterGen will have far-reaching implications in fields such as batteries, magnets, and fuel cells.
The Microsoft Research Center for Scientific Intelligence will also continue to collaborate with partners to further develop and validate this technology. “At the Johns Hopkins University Applied Physics Laboratory (APL), we are committed to exploring tools that can drive the discovery of new task-driven materials. Therefore, we are particularly interested in the impact of MatterGen on material discovery,” said Christopher Stiles, a computational materials scientist responsible for multiple material discovery projects at APL.
You may also want to see:
