Microsoft Unveils Machine Learning Tool to Enhance Medium-Range Weather Forecasting Accuracy

Image: Illustrative image · Logos of the Nokia Corporation and that of Windows Phone 8 by the Microsoft Corporation. · Public domain · Source

Microsoft developed a machine learning method to improve weather forecasts two to six weeks ahead, correcting biases in existing models to assist planning in agriculture and emergency response.

Microsoft researchers have introduced a novel machine learning technique aimed at enhancing the accuracy of medium-range weather forecasts, addressing a longstanding limitation in traditional and AI-based forecasting methods. Published in the journal Nature, this cost-effective tool—dubbed Probabilistic Bias Correction (PBC)—is designed to improve weather predictions from two to six weeks in advance, a timeframe where forecast reliability typically declines.

Unlike conventional physics-based forecasting models that simulate Earth's atmospheric physics or AI models that predict based on historical weather data patterns, both face challenges from approximations used to fill data gaps. These approximations can introduce systematic biases, which accumulate over longer forecast periods and reduce precision. For example, climate changes such as global warming can cause models relying on historical data to underestimate current temperatures.

Microsoft’s PBC uses machine learning to identify and correct these systematic biases in existing forecasting models. It has been tested on the European Centre for Medium-Range Weather Forecasts' (ECMWF) leading physics-based model, improving forecast accuracy for precipitation in 98% of cases and temperature predictions in 92%. Additionally, PBC doubled the accuracy of ECMWF's new AI forecasting system during study analyses. This research was recognized by ECMWF through its AI Weather Quest competition.

Importantly, PBC is computationally efficient, capable of running on CPUs without the need for power-hungry supercomputers or GPUs, making it accessible to weather agencies with limited resources—particularly in economically developing regions. Researchers highlight its lightweight architecture allows scientists globally to adapt the tool for localized weather prediction improvements.

Genevieve Flaspohler, an author of the study and executive director of Rhiza Research, is currently collaborating with the Kenya Meteorological Service Authority, backed by Gates Foundation funding, to implement AI-driven early warning systems ahead of an anticipated more extreme rainy season linked to El Niño. The PBC method is part of this initiative to deliver more accurate, timely forecasts that can benefit farmers and emergency responders.

While promising, experts caution that operational deployment of novel AI forecasting methods remains complex, but Microsoft’s approach signifies a meaningful step toward more reliable medium-range weather prediction and broader access to advanced forecasting technology.