New AI Tool Significantly Improves Weeks 3-4 Weather Forecasts, Bridging 'Predictability Desert'
A new machine learning method has substantially improved weather forecasts for the three-to-four-week period, a timeframe previously considered a 'predictability desert' by meteorologists. Published in Nature, this research indicates that the AI tool, developed by a team including researchers from Microsoft, MIT, and other institutions, doubles the skill of a leading AI forecasting system and enhances Europe's flagship physics-based model across most of the globe. The method, called probabilistic bias correction, acts as an add-on to existing models by learning from approximately 20 years of past forecast errors and blending them with recent observations. While the gains are modest in absolute terms, they are significant for decisions that are costly to reverse, such as those made by farmers, utility companies, and emergency managers. The tool does not require a supercomputer and its code is publicly available.