MIT and Microsoft Researchers Use Machine Learning to Double Weather Forecasts Accuracy

Machine Learning Breakthrough Tackles the Predictability Desert

Researchers at MIT and Microsoft have developed a machine learning tool that significantly improves weather forecasts in the notoriously difficult two to six week timeframe, known as subseasonal forecasting1

1

. The Probabilistic Bias Correction (PBC) framework doubles forecasting accuracy when applied to existing models and placed first in the European Centre for Medium-Range Weather Forecasts (ECMWF) AI Weather Quest competition2

2

. This range has long been called the "predictability desert" because forecasts cannot capitalize on short-term initial conditions or long-term boundary forces like El Niño patterns1

1

.

The research, published in Nature, addresses a critical gap in disaster preparedness. The 2021 Texas winter storms killed at least 246 people partly because energy officials lacked reliable long-range weather prediction to prepare infrastructure1

1

3

3

. "Not only does this do better on time averaged forecasts for this range, which is already very useful for farmers and members of the energy industry trying to match supply with demand, but it also improves predictions of extreme weather events," says MIT Research Scientist Judah Cohen1

1

.

How Probabilistic Bias Correction Works

Instead of creating a standalone AI weather model, the PBC tool uses machine learning to identify and correct systematic bias in existing forecasting models2

2

. Lester Mackey, senior principal researcher at Microsoft Research New England, initiated the concept of using machine learning to reduce forecasting errors through postprocessing1

1

. Traditional physics-based models simulate atmospheric conditions on supercomputers, while AI weather models learn patterns from historical data. Both approaches rely on approximations to fill data gaps, creating systematic bias that compounds over time2

2

.

The PBC framework corrects these biases by learning from a model's historical errors. One example involves adjusting for climate change impacts, where models trained on historical records may bias toward colder temperatures than current conditions warrant2

2

. When tested on ECMWF's leading numerical model, PBC improved accuracy for 98% of precipitation and 92% of temperature targets2

2

. Applied to ECMWF's AI forecasting system, the tool doubled subseasonal accuracy3

3

.

Extreme Weather Warnings See Dramatic Improvement

The most significant finding centers on extreme weather prediction. For forecasts made three weeks ahead, PBC significantly improved the ECMWF numerical model's ability to predict extreme heat, extreme precipitation, and unusually high sea level pressure3

3

. Testing on real-world floods from the last decade showed the tool improved flood prediction skill by 76% to 92%3

3

. For severe floods causing more than 100 deaths or displacing at least 80,000 people, forecasting skill improved by 86% to 150%3

3

.

"Extreme weather has the largest impact on our society and improved forecasts of these events can yield large societal benefits," Cohen told Gizmodo. "Interestingly, previous bias correction schemes actually made forecasts of extreme weather worse, and our model showed clear improvement of extreme weather forecasts"3

3

. This advancement matters particularly as climate change drives more anomalous weather patterns.

Accessible Technology for Resource-Limited Regions

Unlike physics-based modeling requiring supercomputers or generative AI depending on powerful GPUs, the PBC tool runs on standard CPUs2

2

. The framework is open-source, making it accessible even where funding and infrastructure for weather forecasting are limited3

3

. "One of the things that's really exciting to me about models like PBC is how lightweight they are, how easy to train, that scientists without huge pools of resources can work on trying to configure [these] models and make them perform better in specific localities," says co-author Genevieve Flaspohler2

2

.

The Kenya Meteorological Service Authority is already experimenting with the tool to improve monsoon forecasting, supported by Gates Foundation funding through the nonprofit Rhiza Research2

2

3

3

. This testing comes ahead of what's expected to be a more extreme rainy season during a record-breaking El Niño2

2

. "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," Mackey says2

2

.

What to Watch: Operational Implementation and Global Adoption

Matthew Chantry, ECMWF's strategic lead for machine learning, acknowledged the research represents "a good example of new ideas being developed for the benefit and consideration of the community," while noting that "turning ideas into reality [in] an operational process is a most challenging step"2

2

. The transition from research to operational forecasting systems will test whether PBC can deliver consistent improvements across diverse weather conditions and geographic regions.

The tool's model-agnostic design means it can enhance both traditional numerical models and newer AI weather models, potentially accelerating improvements across the forecasting landscape3

3

. For farmers planning planting schedules, energy operators matching supply with demand, and emergency response officials preparing for disasters, reliable long-range weather prediction could transform decision-making in the coming years. The collaboration between MIT, Microsoft, ECMWF, and other institutions signals growing momentum behind integrating machine learning with traditional meteorology to push past the two-week forecasting barrier that MIT Professor Edward Lorenz once theorized as a fundamental limit1

1

.