Microsoft develops machine learning tool for extended weather forecasting

Microsoft researchers have developed a new machine learning tool designed to improve the accuracy of weather forecasts in the two-to-six-week range. Findings published in the journal Nature indicate that the tool can address the reliability issues that typically plague both physics-based and AI models beyond a two-week window.

Rather than building a standalone model, the technology uses machine learning to correct systematic biases in existing forecasting models through postprocessing. This method compares inaccuracies against historical data to minimize errors. A key advantage of this approach is its resource efficiency; while traditional physics-based modeling requires supercomputers and generative AI often requires powerful GPUs, this tool can run on standard CPUs.

This advancement targets the ‘predictability desert,’ a period of subseasonal forecasting where variable atmospheric forces like El Niño make predictions difficult. The researchers submitted versions of their model to the European Centre for Medium-Range Weather Forecasts (ECMWF) for testing. The open-source software aims to provide resource-strapped weather agencies with a more accessible way to enhance long-range forecasting for extreme weather, agriculture, and emergency response.