Microsoft’s new open-source AI tool aims to make long-range weather forecasts far more accurate

Microsoft researchers say they have developed an open-source machine learning tool that can improve weather forecasts two to six weeks ahead, a range where both traditional physics-based systems and newer AI models typically lose reliability. The company says the method could help farmers, emergency managers and national weather agencies make better decisions with less computing power.

The technique, called Probabilistic Bias Correction, is designed to fix systematic errors in existing forecast models rather than replace them. Microsoft says the approach can run on CPUs instead of expensive supercomputers or high-end AI chips, making it a potentially cheaper path to better long-range forecasting for organizations that lack deep technical infrastructure.

The findings were published in Nature on October 14, 2026, and the software has been released openly so researchers and weather services can test it themselves.

Why Microsoft says the breakthrough matters

The practical importance of forecasting two to six weeks out is hard to overstate. That window sits between daily weather prediction and seasonal climate outlooks, giving decision-makers time to prepare for floods, heat waves, storms, drought conditions or planting delays.

For farmers, better mid-range forecasts can help determine when to irrigate, plant or harvest. For emergency agencies, more reliable predictions can influence how quickly to pre-position supplies, warn communities or plan evacuations. For public meteorological services, the main attraction is simple: if the forecast is more accurate and cheaper to compute, more people can use it.

Microsoft’s researchers argue that the cost advantage may be just as important as the accuracy gains. Many forecasting advances depend on massive computing budgets, which are out of reach for agencies in lower-income countries. By contrast, the new method is intended to be light enough to run on ordinary hardware.

How does Probabilistic Bias Correction work?

Probabilistic Bias Correction, or PBC, does not try to build a brand-new weather model from scratch. Instead, it learns where an existing model tends to drift and then adjusts the output to reduce those errors.

In Microsoft’s telling, the key idea is that forecasts become less reliable over longer periods because small mistakes compound. The model can inherit biases from missing data, imperfect approximations or atmospheric processes that are still too complex to represent fully. PBC uses machine learning to spot those recurring distortions and correct them.

That matters because the problem is not confined to one kind of weather system. Physics-based models simulate the atmosphere using scientific equations, while AI models learn from historical patterns. But once forecasts stretch beyond roughly two weeks, both approaches encounter the same issue: uncertainty grows, and the errors can become systematic.

What kind of bias is Microsoft trying to fix?

Microsoft says the method is especially useful when a forecasting model reflects outdated conditions. One example the researchers gave involves climate change: if a model is trained on historical weather that was generally cooler than today’s environment, it may underpredict current temperatures unless corrected.

That kind of bias can affect not just temperature but rainfall and other variables. In practice, the point is to make the original model more honest about the world it is trying to describe, rather than assuming that yesterday’s climate is still the right baseline.

“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,” Lester Mackey, a senior principal researcher at Microsoft Research New England, said in an interview with The Verge.

What did Microsoft test, and how well did it perform?

Microsoft says the method improved a leading physics-based weather model used by the European Centre for Medium-Range Weather Forecasts, or ECMWF, on most of the forecast targets tested in the study. The company says PBC improved accuracy for 98% of precipitation targets and 92% of temperature targets.

The researchers also say the framework roughly doubled the accuracy of ECMWF’s newer AI forecasting system in the analysis. The study was notable enough that ECMWF had already recognized the work before publication in its AI Weather Quest competition.

Those results matter because ECMWF is among the world’s most respected weather organizations. If a correction layer can improve one of its leading systems, that suggests the method could potentially be adapted to other forecasting workflows as well.

How is this different from other AI weather efforts?

Microsoft’s approach stands out because it is less ambitious in one sense and more practical in another. Rather than building a giant standalone weather model, the company is using machine learning as a corrective layer on top of forecasts that already exist.

That is a meaningful distinction. Large weather systems can require massive compute clusters and specialized hardware. Microsoft says PBC is much lighter to train and deploy, which could make it easier for smaller forecasting teams to adopt and customize for local conditions.

That strategy also aligns with a broader trend in applied AI: rather than replacing proven tools, machine learning is increasingly being used to patch weaknesses, reduce error and improve the output of existing scientific systems.

Why does the CPU requirement matter?

Because it lowers the barrier to entry. If a forecasting tool demands the same kind of infrastructure as a frontier AI model, then only a handful of institutions can use it. If it runs on standard processors, more national weather agencies, universities and nonprofits can test it without major capital investment.

For countries where forecasting budgets are constrained, that difference could shape how quickly the technology spreads. It may also make it easier to run local versions tuned for regional weather patterns rather than relying on generic global models.

Who is involved in the research?

The paper brings together researchers from Microsoft, MIT, ECMWF and other institutions. That mix is important because weather forecasting sits at the intersection of science, software engineering and public service. A useful tool in this field usually needs both technical rigor and operational credibility.

Matthew Chantry, ECMWF’s strategic lead for machine learning, said the paper represents a constructive contribution to the field, while also noting that the hard part is turning research into an operational forecasting process.

Chantry described the work as a strong example of ideas being developed for the wider community, but he added that moving those ideas into real-world operations remains the most difficult step.

That caution is worth emphasizing. Academic or experimental success does not automatically translate into reliable use by meteorological agencies, which must maintain forecasts around the clock and prove their outputs are dependable under pressure.

What does this mean for weather agencies in developing countries?

It could make advanced forecast correction more accessible. The amount of compute and infrastructure required for traditional forecasting has long limited adoption in places with smaller technology budgets. Microsoft’s researchers believe PBC could help close that gap.

Genevieve Flaspohler, one of the paper’s authors, said she is excited by how lightweight the models are and how easy they are to train for specific regions. That flexibility, she argued, could allow local scientists to improve forecasts without access to large resource pools.

Flaspohler now serves as executive director of Rhiza Research, a nonprofit working with the Kenya Meteorological Service Authority and the Gates Foundation to explore whether AI can improve early warnings. The group is testing PBC ahead of what is expected to be a more severe rainy season during a record-breaking El Niño pattern.

Why this could matter for extreme weather warnings

Extreme weather is where forecast errors can become most costly. A slight shift in the predicted timing or intensity of heavy rain, heat or wind can change how communities prepare, how governments allocate resources and how many lives are at risk.

Microsoft says that longer-range reliability is one of the biggest benefits of PBC. If the correction system can hold up under severe-weather conditions, it could improve not just convenience but public safety.

That said, the company’s own framing remains careful. The research points to better performance in a controlled study, but the real test will be whether these gains persist across different geographies, seasons and operational forecast systems.

How does this fit into the wider AI weather race?

Microsoft is not alone. Google, Nvidia and other technology companies have been investing in AI systems for weather forecasting, looking for ways to speed up predictions or improve their quality.

What makes Microsoft’s work notable is that it pushes toward a more affordable model of AI adoption. Instead of pursuing the biggest possible model, the company is proposing a lighter tool that can sit on top of existing systems and improve them.

That may prove especially valuable as the weather sector becomes more interested in hybrid approaches: physics-based modeling, machine learning, and corrections layered together rather than one method replacing the others outright.

Timeline of the research and release

The following table summarizes the key milestones described by Microsoft and its collaborators.

What happens next?

The next stage is operational proof. Researchers now need to show that the tool works consistently outside the lab, in the hands of real weather services with real forecasting deadlines.

If those tests go well, PBC could become a practical add-on for agencies looking to improve medium-range forecasts without upgrading to massive compute systems. It may also inspire similar bias-correction tools for other scientific prediction problems where data gaps and systematic errors are hard to avoid.

For now, Microsoft’s release marks an important step in the evolution of AI weather tools: not a flashy attempt to replace meteorology, but a pragmatic effort to make forecasting better, cheaper and more widely available.

That combination of accessibility and accuracy is what makes the research stand out. In a field where every additional day of useful warning can matter, even a modest improvement could have outsized value.

And because the software is open source, the wider scientific community will now be able to test Microsoft’s claims, challenge them and, if they hold up, build on them.