Microsoft recently released a research blog titled "MatterGen: A new paradigm of materials design with generative AI." I found time to read it today.
MatterGen was recently published in *Nature*, where the Microsoft team shared this material discovery tool based on generative AI. Instead of traditionally screening candidate materials, MatterGen directly generates new materials according to application-specific design requirements.
MatterGen can generate materials with specific chemical, mechanical, electrical, or magnetic properties based on input design requirements, and can even meet multiple different constraints simultaneously.
MatterGen: A Diffusion Model for Generating Novel Stable Materials
MatterGen is a tool based on diffusion models specifically designed to handle the three-dimensional geometric structures of materials. Similar to how image diffusion models generate images by modifying pixel colors in noisy images, MatterGen generates desired material structures by adjusting positions, elements, and periodic lattices in random structures.
This diffusion architecture has been optimized specifically for the characteristics of materials, effectively handling their periodicity and three-dimensional geometry. MatterGen can be fine-tuned according to different design requirements, generating materials that meet specific chemical compositions, crystal symmetries, or material properties. Through this innovative design, MatterGen can more accurately generate material structures that meet design requirements.
Performance Demonstration and Optimization of MatterGen
The base model of MatterGen has achieved industry-leading levels in generating novel, stable, and diverse materials. The performance improvement not only benefits from architectural innovation but also originates from the quality and scale of the training data.
Comparison of MatterGen's performance with other methods in generating stable, unique, and novel material structures. The training dataset for each method is noted in parentheses. The purple bar indicates the performance improvement brought by the MatterGen architecture itself, while the teal bar indicates the performance improvement achieved through a larger training dataset.
A key advantage of MatterGen over traditional screening methods lies in its ability to explore the entire unknown material space. As shown in the figure below, MatterGen can continuously generate more innovative candidate materials that meet high bulk modulus requirements (greater than 400 GPa), which are typically difficult to compress. In contrast, traditional screening methods gradually saturate in performance due to reliance on known candidate materials, making it difficult to discover new materials.
Comparison of the performance of MatterGen (teal) and traditional screening methods (yellow) in finding stable, novel, and unique material structures that meet design requirements. The design requirement is that the material's bulk modulus must exceed 400 GPa.
Handling Composition Disorder
Composition disorder is a common phenomenon in synthesized materials, where different atoms can randomly exchange their crystallographic positions. Recently, the academic community has begun discussing what constitutes "novelty" in computationally designed materials. Traditional algorithms often fail to distinguish between structural pairs whose only difference lies in the arrangement order of similar elements in lattice positions.
To address this issue, MatterGen proposes a new structural matching algorithm that specifically considers composition disorder. This algorithm determines whether a pair of structures can be identified as ordered approximations of the same composition-disordered structure. This provides a new definition of novelty and uniqueness, which has been incorporated into our computational evaluation metrics.
The figure above illustrates composition disorder. Left: A perfect crystal without composition disorder, showing repeated unit cells (black dashed lines). Right: A crystal with composition disorder, where each position has a 50% probability of being occupied by either a yellow or teal atom.
Laboratory Verification
In addition to extensive computational evaluations, the team experimentally validated MatterGen's capabilities through synthesis.
Experimental validation of the compound TaCr₂O₆ proposed by MatterGen. The image shows a scientist operating an experimental bench in the lab, using tweezers to hold a small sample.
Its structure was generated by MatterGen under a condition of 200 GPa bulk modulus. The synthesized material's structure matches the one proposed by MatterGen, with the only difference being compositional disorder between Ta and Cr. Additionally, the experimentally measured bulk modulus was 169 GPa, which has a relative error of less than 20% compared to the design specification of 200 GPa, making it very close from an experimental perspective. If such results can be generalized to other fields, it will have far-reaching implications for the design of battery, fuel cell, and other materials.
AI Simulator and Generator Flywheel
MatterSim is a deep learning model used for accurate and efficient material simulation and property prediction across a wide range of elements, temperatures, and pressures. MatterSim follows the fifth paradigm of scientific discovery, significantly accelerating the simulation speed of material properties. Meanwhile, MatterGen accelerates the exploration of new material candidates through property-guided generation. MatterGen and MatterSim can work together to form a flywheel effect, simultaneously accelerating the simulation and exploration of new materials.
Five Paradigms of Scientific Discovery
First Paradigm: Empirical Observation Second Paradigm: Theoretical Models Third Paradigm: Numerical Computation Fourth Paradigm: Data-Intensive Scientific Discovery Fifth Paradigm: Machine Learning and Artificial Intelligence (AI-Driven Scientific Discovery)
Looking Ahead
MatterGen represents a new paradigm of material design driven by generative AI technology, exploring a material space far beyond what screening-based methods can achieve. Moreover, by guiding material exploration, MatterGen demonstrates superior efficiency. Similar to the impact of generative AI on drug discovery, MatterGen will profoundly influence how we design materials, spanning a wide range of applications including batteries, magnets, fuel cells, and more.