10 scientific breakthroughs from Microsoft researchers
As AI becomes a bigger part of everyday life, scientists are finding exciting new ways of harnessing its transformative power to tackle some of society’s biggest challenges.
From designing new materials to mapping flood risks through clouds, Microsoft researchers are using AI to solve problems more quickly and effectively than ever before.
With sustainability and accessibility in mind, they are also overcoming challenges in surprising new ways — like using seaweed to lower cement carbon emissions and building an energy-efficient computer that uses smartphone camera sensors and light.
In 2025, Microsoft published numerous research papers in peer-reviewed journals sharing their findings with others to build upon. Here are 10 examples that show how AI and other technologies are accelerating innovation in banking, healthcare, life sciences and energy — and charting a path for much-needed breakthroughs.
Majorana 1: The world’s first quantum processor powered by topological qubits
Imagine self-healing materials that repair cracks in bridges or airplane parts, catalysts that can break down pollutants into valuable byproducts — or breakthroughs that boost soil fertility to increase yields or promote sustainable growth of foods in harsh climates.
Research published in Nature earlier this year detailed how Microsoft researchers were able to create exotic quantum properties that led to a new type of quantum chip called the Majorana 1. The chip is powered by a new type of quantum architecture that is expected to realize quantum computers capable of solving meaningful, industrial-scale problems that today’s computers cannot — within years, instead of decades.
The chip leverages the world’s first topoconductor, a breakthrough type of material that can observe and control Majorana particles to produce more reliable and scalable qubits, the building blocks for quantum computers. While engineering work is still ahead, many difficult scientific and engineering challenges have now been met.
BioEmu-1: Faster protein stability predictions could lead to more effective medicines
Proteins make up the functional building blocks of life and are central to drug discovery and biotechnology. While there has been extraordinary progress in recent years toward better understanding protein structures using AI, many of these methods offer only a snapshot of a highly flexible molecule or require simulation times of years or even decades.
Enter Biomolecular Emulator-1 (BioEmu-1), a generative deep-learning model that provides scientists with a glimpse into the rich world of different structures each protein can adopt. This is significant because a deeper understanding of proteins could enable the design of more effective drugs, as many medications work by influencing protein structures to boost their function or prevent them from causing harm.
As explained in the journal Science, BioEmu-1 can generate thousands of protein structures per hour on a single graphics processing unit (GPU) at a fraction of the computational cost of traditional simulations. Based on this, BioEmu-1 can predict functionally relevant structure changes of proteins at unprecedented speed and predict protein stability, an important factor when designing proteins for therapeutic purposes.
MatterGen and MatterSim: AI-powered breakthroughs in materials discovery
Materials innovation drives technological progress — from batteries and fuel cells to magnets — and is essential for creating future energy breakthroughs. But identifying the next new material has long relied on costly and time-consuming experiments. Even computer-powered screening requires evaluating millions of options.
MatterGen is a generative AI tool that skips screening and instead seeks to produce novel materials based on prompts that outline design requirements for specific applications, as explained in the journal Nature. Much like how an AI image generator turns blurry pictures into clear ones with a prompt, it starts with a random 3D structure and gradually adjusts atoms, elements and repeating patterns to create a realistic material with defined chemical, mechanical, electronic or magnetic properties. Trained on over 600,000 examples, MatterGen achieves the state of the art in generating inorganic materials across the periodic table. MatterGen can also work with MatterSim, an AI-powered tool that rapidly simulates material properties. Together, they can create a feedback loop that accelerates both simulation and exploration.
RAD-DINO: X-ray data meets AI technology
In healthcare, faster access to information can save lives.
Findings published in Nature Machine Intelligence shows that generative AI foundation models may be able to give clinicians more accurate information and improve patient care.
A collaboration between Microsoft Research and Mayo Clinic is focused on building multimodal foundation models that integrate text and X-ray images. The project pairs Microsoft Research’s AI technology with Mayo Clinic’s X-ray data to help doctors get better and more comprehensive medical data so they can analyze radiology results in less time.
The technology, called RAD-DINO, is named for its focus on radiology and a specific computer learning method. It works by identifying anatomical matches between the chest X-rays of different subjects, while indicating similarities through the proportional brightness of the heatmap, which is a visual overlay that uses color to show areas of interest or importance on an X-ray, CT or MRI and other types of images.
Aurora: Advanced atmospheric and weather forecasting
Microsoft’s Aurora AI foundation model leverages the latest advances in AI to more accurately predict not just the weather, but a wide range of environmental events.
Developed by Microsoft Research, Aurora forecasts this range of atmospheric events with greater precision and speed and at much lower computational cost when compared to traditional numerical forecasting and previous AI approaches. What sets Aurora apart is its versatility. It can be specialized through finetuning to go beyond what is considered traditional weather forecasting, such as predicting air pollution, ocean waves and tropical cyclones.
Aurora learns how to generate forecasts through training on general weather patterns from over one million hours of data. And it generates forecasts in seconds, compared to traditional systems that require hours on large supercomputers to generate comparable predictions. Aurora’s early results, published in Nature, have stirred particular interest in seeing how it can be adapted to better predict rain, enhance crop logistics and protect energy grids. Microsoft continues to advance Aurora as an open-source platform, deepening research partnerships through a Microsoft AI for Good grant and investing in community weather stations.
FCDD: Improving early breast cancer screening with AI
Breast cancer is the most common cancer among women worldwide. And while early screening can save lives, it often leads to high rates of false positives, significantly increased anxiety for patients and unnecessary biopsies. The problem is especially acute for women who have dense breast tissue, a condition that increases the risk of breast cancer and makes it harder to detect abnormalities through traditional imaging methods like mammograms.
But a new AI model called FCDD (Fully Convolutional Data Description) aims to improve early detection by generating MRI heatmaps that locate suspected tumors with a very high degree of accuracy, outperforming other AI models. Developed through a collaboration between Microsoft’s AI for Good Lab, the University of Washington and Fred Hutchinson Cancer Center, the findings were published in Radiology, and the model has since been made open source. While AI won’t replace radiologists, it can give them better tools for evaluating difficult cases or reducing their workload.
Seaweed-infused cement could cut concrete’s carbon footprint
The modern world is built with concrete. And cement, the key component in concrete, is everywhere. It’s the second most-used material on Earth after water — and one of the biggest contributors to greenhouse gas emissions.
Now, researchers at the University of Washington and Microsoft have developed a new type of low-carbon concrete made from seaweed that is designed to cut emissions without sacrificing performance.
While most cement emissions come from the fossil fuels used to heat raw materials during production, seaweed is a carbon sink. It pulls carbon out of the air and stores it while it grows. The teams findings, published in Matter, showed that dried, powdered seaweed mixed with cement had a 21% lower global warming potential, or GWP, a metric used to compare how much heat gases trap compared to carbon dioxide. Thanks to custom machine learning models, the team developed this new formulation in just 28 days, compared to a more typical five years of trial and error.
Mapping floods from space — even when clouds get in the way
Floods cause extensive global damage annually. While satellite observations have proven invaluable for flood detection and tracking, comprehensive global flood datasets spanning extended time periods remain scarce, complicating disaster preparedness.
But a deep learning flood detection model from the Microsoft AI for Good Lab leverages the cloud-penetrating capabilities of a powerful Earth observation satellite using radar imagery, enabling researchers to map areas impacted by floods even through cloud cover and in the dark of night.
As explained in Nature Communications, the model allowed researchers to analyze specialized data to build a global map showing where floods have happened over a 10-year period, providing a reliable picture of flood-prone areas. That long-term view gives policymakers greater insight into flood trends, so communities can better prepare. The researchers’ long-term analysis suggests global floods may be increasing, though more research is needed. The team’s predictions and code are publicly available so researchers and responders worldwide can improve flood monitoring and disaster response.
Analog optical computer: Accelerating AI and optimization with light
Microsoft has developed an analog optical computer (AOC) that uses light instead of conventional digital electronics to efficiently tackle complex optimization problems and accelerate AI inference, the process of running a trained AI model to generate outputs, without further training. Optimization problems aim to find the best solution from among nearly endless possibilities.
The findings, published in the journal Nature, show the potential of harnessing light to enable key calculations, potentially at a fraction of the energy and significantly higher speed than the GPUs commonly used today. It was built using existing and scalable technologies like micro-LED lights to be more affordable and easier to manufacture with existing supply chains. The prototype successfully solved two types of optimization problems in banking and healthcare — finding the most efficient way to settle complex banking transactions and slashing the time it takes to do MRI scans.
Managing the risk behind the promise of AI in biology
Advances in AI are opening extraordinary frontiers in biology. Yet these same technologies also introduce biosecurity risks and may lower barriers to designing harmful toxins or pathogens. This “dual-use” potential, where the same knowledge can be harnessed for good or misuse to cause harm, poses a critical dilemma for modern science.
A Microsoft-led paper published in Science describes a two-year confidential project begun in late 2023. Microsoft researchers recognized that the work itself — detailing methods and failure modes — could be exploited by malicious actors if published openly. To guide decisions about what to share, they held a multi-stakeholder deliberation involving government agencies, international biosecurity organizations and policy experts.
The authors also devised a tiered-access system for data and methods, implemented in partnership with the International Biosecurity and Biosafety Initiative for Science (IBBIS). To their knowledge, this is the first time a leading scientific journal has formally endorsed a tiered-access approach to manage an information hazard.
Images by Microsoft and John Brecher for Microsoft (Majorana 1), Jonathan Banks for Microsoft (MatterGen, MatterSim), Frank Ramspott / Getty Images (typhoon image), Mark Stone for University of Washington (seaweed cement) and Chris Welsch for Microsoft (analog optical computer).