MIT AI forecasts extreme weather without historical data
ainews ·
MIT engineers have developed an AI tool that can forecast extreme weather events even when no historical data exists for such disasters in a given region. Created by graduate student Kai Chang and Professor Themis Sapsis, the method, called Extreme Event Aware or η-learning, was published in Nature Communications on August 20. The tool generates maps of statistically possible but unprecedented events, along with estimates of their duration, intensity, and affected area. It works by learning relationships between point statistics and spatial map data, then applying those relationships to create patterns beyond what was in the training set. The researchers tested it on U.S. precipitation data, using 25 years of rainfall records but training the spatial model on only the first six months. Their approach can produce plausible scenarios such as a storm dumping 300 millimeters of rain on New York City, surpassing the recorded maximum of 200 millimeters. These maps could help planners test infrastructure against worst-case scenarios like extreme storm surges, heatwaves, or wildfires. Applying the method to new hazards would require relevant point statistics and spatial data for each specific event type.
ainews