On 17th January 2025, Satya Nadella, Microsoft's Chairman and CEO, took to X to announce a research paper published in Nature that caught our attention.
He wrote, “Today in @Nature: Our MatterGen model represents a paradigm shift in materials design, applying generative AI to create new compounds with specific properties with unprecedented precision.”
This was the announcement of one of the newest tools in material discovery: Microsoft’s MatterGen, an AI-driven model that can generate new materials with specific desired properties. In a paper published in Nature, the team from Microsoft Research AI for Science, UK, and the Chinese Academy of Science, Shenzhen, China, demonstrated MatterGen’s uncanny ability to predict structures for new compounds with the desired properties.
Traditional methods of finding new materials are a bit like searching for a needle in a haystack. Scientists have had to rely on either blind luck or tedious experiments, trying countless combinations and hoping that one of them results in a stable material with the desired traits. But thanks to advances in artificial intelligence and machine learning, this process has largely been handled by computers.
Today, several AI tools and models are employed in material discovery. Some of the more notable ones are:
- Materials Project (MP): While not purely an AI tool, this database provides a wealth of information about various materials, which can be used in conjunction with AI algorithms to predict properties and stability. Mp was also used to train MatterGen
- Cedar Database: Focused on thermodynamic properties, Cedar aids AI models in understanding the stability and potential reactions of materials.
- Open Quantum Materials Database (OQMD): This database holds a vast set of calculated materials properties that AI tools use for training models, enabling predictive analytics in material science.
- Atomate: An open-source software that uses workflows to automate calculations based on existing materials databases and machine learning models.
- DeepChem: This platform provides machine learning tools for scientific discovery, including materials science, where it helps in predicting material properties from data.
- SchNet: A neural network for predicting quantum-mechanical properties, capable of operating directly on molecular systems to estimate a variety of material properties.
What is MatterGen?
MatterGen is a generative model created by researchers to design and discover novel inorganic materials. The core of MatterGen lies in a technology called diffusion modelling. If you've ever seen AI art generators that can create images from text descriptions, you'll have some idea of what diffusion models do. They start with randomness and, step by step, build up a coherent structure guided by the input “constraints” (or rules). In the context of MatterGen, this means beginning with a random arrangement of atoms and methodically adjusting them to form a stable material that fits the desired criteria.
MatterGen has been trained on over 600,000 stable materials from vast databases like the Materials Project and others. With this training, it can recognize patterns and characteristics that contribute to stability and uniqueness. Moreover, its predictions continue to improve as it generates more materials, ensuring they're not only novel but also feasible to produce.
MatterGen employs a unique diffusion model tailored to handle the 3D geometry and periodicity of materials. This allows for better precision when generating structures that meet the desired constraints. Compared to other generative models, MatterGen materials are found to be more than twice as likely to be novel and stable and more than ten times closer to the local energy minimum, suggesting better viability for practical use. The model is also versatile in fulfilling multiple constraints simultaneously, including chemistry, symmetry, mechanical, electronic, and magnetic properties, offering a broader application scope.
One of the materials suggested by MatterGen was synthesized by the researchers as a real-world validation of its capabilities. The properties of the synthesised material were found to be remarkably close to the intended targets. This successful experiment showcases its potential to bring rapid breakthroughs in materials science.
While existing AI tools in material discovery are excellent at leveraging and analyzing existing data, MatterGen represents a leap forward by not only analyzing but also creating new materials tailored to specified requirements. This shift from a reactive to a proactive approach to material discovery heralds new opportunities in the field.
MatterGen's application can be tailored to more specific domains in materials science, such as developing biodegradable materials or those with low environmental impacts. By leveraging the power of AI, it accelerates the discovery of materials that can spearhead new technological advancements. From energy storage to carbon capture, the innovations MatterGen facilitates today could shape a brighter and more sustainable tomorrow.
This research news was partly generated using artificial intelligence and edited by an editor at Research Matters.