Microsoft researchers have open-sourced a machine learning tool designed to correct systematic biases in existing weather forecasting models, according to a paper published in the journal Nature. Developed alongside collaborators from MIT and the European Centre for Medium-Range Weather Forecasts (ECMWF), the framework—called Probabilistic Bias Correction (PBC)—aims to enhance prediction accuracy for two to six weeks in advance.

Unlike traditional physics-based models or heavy generative AI frameworks that require supercomputers or dedicated GPUs, PBC operates on standard CPUs. Tests showed the tool improved the accuracy of ECMWF’s leading physics model across 98 percent of precipitation and 92 percent of temperature targets while doubling the accuracy of ECMWF’s AI forecasting system.

Why it matters

  • Enables resource-constrained agencies to run high-accuracy extended weather forecasts on standard CPU infrastructure.

  • Demonstrates a cost-effective, open-source approach to post-processing and bias correction in complex physical forecasting systems.

Source: theverge.com