Scientists Develop AI Tool to Enhance Long-Range Weather Forecasting, Improving Extreme Event Predictions
A team of scientists, including researchers from Harvard College, the University of Toronto, MIT, and Microsoft Research, has developed a new machine-learning tool designed to significantly improve weather forecasts three to four weeks in advance. This tool, named probabilistic bias correction, addresses the long-standing challenge of the 'predictability desert'—the period beyond two weeks where traditional weather models often perform worse than simply guessing the long-term average. The AI tool works by learning from past forecasting errors and adjusting the probabilities of a week being unusually warm, cool, wet, or dry. It compares 20 years of a model's past forecasts with actual weather events, subtracting recurring errors. Additionally, it blends the forecast with recent observed weather and long-term averages. This innovation has doubled the skill of a leading AI weather model and improved the European model, considered a gold standard, in approximately 90% of grid squares. The tool also demonstrate...
