Microsoft researchers have introduced a new machine learning tool called Probabilistic Bias Correction (PBC) designed to improve the accuracy of weather forecasts for the two-to-six-week range. The findings were published in the journal Nature, and the software has been released as an open-source project.
Unlike traditional physics-based models that require supercomputers or generative AI systems that rely on powerful GPUs, the PBC tool is designed to run on standard CPUs. This approach aims to provide a more cost-effective and accessible solution for weather agencies with limited computational resources.
The method works by identifying and correcting systematic biases in existing weather models, such as those caused by climate change or historical data gaps. In testing, PBC improved the accuracy of the European Centre for Medium-Range Weather Forecasts (ECMWF) physics-based model for 98 percent of precipitation and 92 percent of temperature targets.
The framework also doubled the accuracy of the ECMWF’s AI forecasting system during the study. The research was conducted by teams from Microsoft, MIT, and the ECMWF, and the project previously received an award in the ECMWF AI Weather Quest competition.
The tool is currently being evaluated for practical applications, including efforts by the nonprofit Rhiza Research to provide early weather warnings in Kenya. The project is supported by funding from the Gates Foundation and focuses on preparing for extreme weather events.
