Researchers improve weather forecasts looking two to six weeks out (news.mit.edu)

Researchers from MIT and other institutions have developed a groundbreaking machine-learning model that significantly improves weather forecasting accuracy for the subseasonal range of two to six weeks. This advancement addresses a critical gap in weather predictions, following incidents like the catastrophic 2021 Texas storm, where timely forecasts could have mitigated disaster impacts. The new model corrects errors in existing forecasting systems, doubling their predictive accuracy and outperforming traditional climatologic averages for the first time. This achievement highlights the potential for AI in enhancing long-term weather forecasting, crucial for sectors like agriculture and energy. The model employs a novel post-processing approach that adaptively learns from historical errors, allowing for more reliable predictions of extreme weather events, which can have profound societal implications. As the technique can be applied to various forecasting models and is open-source, it is poised for widespread adoption. The potential applications are vast, from improving agricultural strategies during critical weather events to aiding energy management and emergency preparedness. The research, which demonstrated its effectiveness in prestigious forecasting competitions, signals a transformative step in how meteorological predictions could evolve, offering hope for better preparedness against climate-related challenges.