AI meets materials discovery: The vision behind MatterGen and MatterSim

MatterGen and MatterSim are cutting edge tools reshaping how we design and innovate advanced materials. Explore the journey from concept to creation behind these AI-powered technologies.

In 2018, Tian Xie was immersed in his doctoral studies at MIT, exploring the complex world of materials science and engineering. Midway through his PhD, a question began to take root—both in his mind and in conversations with peers: Can you build a model that takes constraints and criteria as input and generates a viable material as output?

Xie’s answer was a confident yes. But he couldn’t predict how swiftly this vision would materialize—or just how transformative its applications would prove to be. 

After a two-year postdoctoral stint at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), Xie joined Microsoft Research’s newly minted AI for Science initiative in 2022. What began as a big bet under renowned scientist Chris Bishop—a bold experiment to harness AI for tackling humanity’s most pressing challenges, from sustainability to drug discovery—has since expanded into a global endeavor. In just two-and-a-half years, the team has grown into a far-reaching group spanning five time zones, united by a mission to redefine the boundaries of innovation.

MatterGen and MatterSim

Two of the transformative tools that play a central role in Microsoft’s work on AI for science are MatterGen and MatterSim. In the world of materials discovery, each plays a distinct yet complementary role in reshaping how researchers design and validate new materials. 

MatterGen is the idea generator, the visionary in the partnership. It crafts detailed concepts of molecular structures by using advanced algorithms to predict potential materials with unique properties, grounded in scientific principles and computational precision. “MatterGen generates thousands of candidates with user-defined constraints to propose new materials that meet specific needs,” Xie said. “This represents a paradigm shift in how materials are designed.” 

This generative approach is a radical departure from traditional methods of screening existing materials. It replaces the meticulous observation and precise assembly required when fitting puzzle pieces from a box with a tool that designs entirely new puzzles customized to defined parameters. Once MatterGen has proposed its possibilities, MatterSim steps in as the gatekeeper, the realist to MatterGen’s visionary. MatterSim applies rigorous computational analysis to predict which of those imagined materials are stable and viable, like a sieve filtering out what’s physically possible from what’s merely theoretical. Together, these tools accelerate a process that once relied on years of trial and error in the lab. This tandem functionality enables researchers not only to explore the vast, uncharted territories of material possibilities but also allows them to do so with a newfound efficiency and confidence. 

“From an industrial perspective, the potential here is enormous. Human civilization has always depended on material innovations. If we can use generative AI to make materials design more efficient, it could accelerate progress in industries like energy, healthcare, and beyond.”

—Tian Xie, Principal Research Manager, Microsoft Research AI for Science

---

---

Building bridges between research and real-world innovation

Materials are, in many ways, the unsung heroes of human progress. From the steel girders forming the backbone of modern cities to silicon chips powering smartphones, advances in materials science have propelled technological innovation for centuries. Every leap in civilization—from the Bronze Age to the Space Age—has been defined by the human ability to discover, manipulate, and deploy materials. MRI machines, for example, rely on superconductors, which were only made possible through advances in materials science. 

Yet the process of developing new materials has traditionally been a slog. Despite their pivotal role, identifying and refining new materials has often required years of painstaking attempts, with researchers relying heavily on intuition, experience, and luck. This approach can cost millions, if not billions, of dollars, with no guarantee of success. 

In industries like energy and healthcare, where the right material can revolutionize markets, the stakes are higher than ever. Developing new efficient battery material could unlock the potential for more sustainable energy storage, while advances in superconductors could lead to groundbreaking improvements in medical imaging or quantum computing. But these outcomes hinge on solving an enduring challenge: identifying and testing viable materials at a speed and scale that match the urgency of global demands. By leveraging artificial intelligence, the AI for Science team has built tools that promise to reshape materials discovery entirely. 

“One of the fundamental ideas driving our approach is that the more computational power you invest in these tools, the more insights and discoveries you can generate,” said AI for Science Principal Researcher Ziheng Lu. “This is a crucial mindset shift for the field.” 

Lu began his work at Microsoft Research Asia working on sustainability initiatives. Like Xie, Lu did not realize how quickly AI for Science would evolve. 

“We thought it might be possible for machine-learning-based methods to replace 80–90% of quantum-mechanical calculations within one or two years,” he says. “But progress happened so quickly that the tools began outperforming our initial expectations.” 

Lu’s work focuses on enabling in silico characterization of material properties, integrating factors like temperature and pressure to make predictions more realistic. This ambition has led to breakthroughs that could fundamentally alter subfields of materials science. One notable example is their work on the limits of heat transfer in matter, a question that has eluded scientists for 125 years.

A glimpse behind the curtain

For the researchers at AI for Science, the typical workday unfolds as a blend of intellectual exploration and strategic planning. Some moments are spent huddled over digital whiteboards or sketching ideas on scrap paper, brainstorming breakthroughs. Other times, the focus shifts to reviewing fresh results and debating whether the current approach needs recalibrating to meet the broader goals of the initiative. 

“We hold a team meeting once a week for MatterGen and MatterSim,” said Daniel Zügner, a senior researcher with AI for Science. “Below that level, smaller groups from both teams collaborate on applying the models in new settings and transferring insights between them.” 

“Although MatterGen and MatterSim are often described as separate, they work together in an extraordinarily interconnected way, having complementary properties that create a whole which is far more than the sum of its parts.”

—Daniel Zügner, Senior Researcher, Microsoft Research AI for Science

Zügner’s journey to AI for Science began in 2022, fresh from completing his PhD at the Technical University of Munich, where he specialized in machine learning and graph neural networks. Zügner has contributed to the development of MatterGen by helping to build its machine-learning architecture, optimizing it for property-guided materials design, and ensuring its scalability in collaboration with materials scientists. Reflecting on his early days at the organization, he recalled an emphatic message from Chris Bishop: “The first couple of years would be all about building a world-class organization for science. Only then could we hope to achieve the kind of transformative, large-scale impact we envisioned.” 

Central to this vision is the connection of groundbreaking scientific research with real-world applications. A key collaborator is Microsoft’s Azure Quantum team, with the Azure Quantum Elements platform, using high performance computing, artificial intelligence, and quantum computing to solve complex problems in chemistry and materials science. This partnership allows AI for Science researchers to stay closely aligned with industry needs, ensuring their work tackles challenges that matter most to our customers. It also means that the latest discoveries don’t stay stuck in the lab. Instead, they’re quickly integrated into the company’s broader offerings, helping businesses across industries—from healthcare to energy—apply cutting-edge science to accelerate scientific discovery and solve big problems faster than ever before. 

---

---

A transformative future

Recent milestones reached by MatterGen and MatterSim researchers are a testament to their potential to redefine how complex problems are approached across disciplines. The open release of MatterSim has gained remarkable traction with tens of thousands of downloads, signaling broad interest in the tools the initiative is creating. A recent publication in Nature (opens in new tab) showcases in detail MatterGen’s innovative approach to materials discovery.  

The impact of AI for Science extends beyond materials science to climate prediction and biomedicine. The atmospheric foundation model, Aurora, has already drawn significant interest in the climate research community. Another recent paper in Nature highlights AI_²BMD’s advances (opens in new tab) in biomolecular dynamics. These breakthroughs demonstrate how artificial intelligence can accelerate scientific understanding in fields that have traditionally required time-consuming experimentation. 

Yet AI for Science is not about chasing flashy milestones or producing publications for the sake of metrics. 

“Our focus is on driving science in a meaningful way,” Zügner said. “The team isn’t preoccupied with publishing papers for the sake of it. We’re deeply committed to research that can have a positive, real-world impact, and this is just the beginning.”