Microsoft and Ai2 unveil machine learning tools for weather forecasting

The problem the tool addresses affects both traditional physics-based forecasting and newer AI models. Physics-based models simulate the atmosphere on supercomputers, while AI models look for patterns in historical data. Both rely on approximations to fill gaps caused by sparse weather stations or poorly understood atmospheric phenomena. Those approximations can become systematic bias, and the resulting errors compound as forecasts extend further in time, making predictions less precise beyond two weeks.

Microsoft calls its method Probabilistic Bias Correction, or PBC. The tool uses machine learning to identify systematic bias in a model and correct it. Lester Mackey, a senior principal researcher at Microsoft Research New England, told The Verge that one example is correcting for climate change: a weather model trained on historical records may be biased toward colder temperatures than the world is actually experiencing because of greenhouse gas emissions. Unlike physics-based modeling that runs on supercomputers and generative AI that depends on powerful GPUs, PBC can run on CPUs, Mackey said. “I have a lot of hope both for the increased resource efficiency of weather forecasting and the extended reliability of forecasts, especially for extreme weather,” he told The Verge.

Microsoft is joining Google, Nvidia and other companies developing AI tools for weather forecasting. Instead of creating a standalone weather model, its latest research uses machine learning to correct bias in existing models. The researchers found that PBC improved the accuracy of a leading physics-based weather model from the European Centre for Medium-Range Weather Forecasts for 98 percent of precipitation targets and 92 percent of temperature targets. The machine learning framework also doubled the accuracy of ECMWF’s new AI forecasting system in the study’s analysis. Before the paper was published, ECMWF had awarded the authors’ research on PBC in its AI Weather Quest competition.

Matthew Chantry, ECMWF’s strategic lead for machine learning, said in an emailed comment to The Verge that the paper, by researchers at Microsoft, MIT, ECMWF and other institutions, “is a good example of new ideas being developed for the benefit and consideration of the community.” He added that “turning ideas into reality [in] an operational process is a most challenging step.”

The researchers hope PBC’s lightweight design will make AI-enhanced forecasting more accessible to weather agencies with limited computing resources. Genevieve Flaspohler, one of the paper’s authors, said the models are lightweight and easy to train, allowing scientists without huge pools of resources to configure them for specific localities. Flaspohler worked on PBC as an intern at Microsoft with Mackey and is now executive director of the nonprofit Rhiza Research. The group is working with the Kenya Meteorological Service Authority, with funding from the Gates Foundation, to test whether AI can help deliver more accurate early warnings ahead of an expected more extreme rainy season during a record-breaking El Niño. PBC is one of the methods being tested.

In a separate announcement the same day, the team behind the High Resolution Output Ai2 Climate Emulator introduced HiRO-ACEv2. The system can generate global 3 km climate data at about 1/300th the cost of X-SHiELD, a high-resolution global atmospheric model that explicitly simulates storms and other small-scale weather, according to the team’s blog post. HiRO-ACE, introduced earlier in 2026, combines ACE, a coarser 100 km AI model, with HiRO, a downscaling model that sharpens the coarse outputs into 3 km detail.

The biggest change in HiRO-ACEv2 is a new downscaling model, HiROv2, which predicts several high-resolution fields at once instead of only rainfall, allowing it to capture a wider range of extreme weather. The team said HiROv2 matches the target simulation’s high-resolution statistics to within 3 percent, even for events so rare they occur only once in a million data points. It also accurately estimates the total energy of tropical cyclones, including storms it never saw during training, and generates hurricanes with realistic eyes and rain bands, though winds and pressure around the eye remain somewhat more lopsided than in X-SHiELD. It also realistically generates strong winds flowing down mountain slopes.

The team also trained an updated ACE model. Over a 10-year simulation, its hurricanes, average climate patterns and overall distributions of rain, wind and pressure lined up well with X-SHiELD, the blog post said. Putting HiROv2 and ACE together, the team generated a full year of global 3 km data whose statistics closely matched X-SHiELD’s, even though HiRO was working from a simulation ACE ran on its own rather than from real X-SHiELD data. Across a 10-year run, the energy of HiRO-ACEv2’s hurricanes was comparable to X-SHiELD’s over the same period.

Producing global high-resolution data with HiRO-ACEv2 costs roughly 300 times less than running X-SHiELD, the team said. If only one region is needed, savings can be more than a hundred times greater depending on the area. The team is also working on a streamlined version of HiROv2 that could make it more than ten times faster. Limitations remain: because HiRO learned from X-SHiELD, it only works with the version of ACE tuned to match X-SHiELD, so it cannot yet produce high-resolution data for other climates such as a warmer future or the historical record in datasets like ERA5. The team plans to train on new X-SHiELD datasets, including one simulating a world 4°C warmer, and to add outputs such as near-surface temperature and snow and ice for uses including snowpack prediction. A web demo shows pre-generated precipitation, wind and sea-level pressure output from HiROv1 and HiROv2 side by side with X-SHiELD, and the models are available for download with tutorial notebooks.

Editor's Summary

Microsoft published an open-source machine learning tool in Nature that corrects bias in existing weather models and improved ECMWF forecasts for most precipitation and temperature targets, while the Ai2 team released HiRO-ACEv2, a 3 km climate emulator that can simulate rare extreme weather at a fraction of the cost of X-SHiELD. Both efforts aim to make more accurate medium-range forecasts and high-resolution extreme-weather data more accessible, though operational and climate-generalization limits remain.