MIT engineers have created an AI model that predicts statistically possible extreme weather events without needing historical data from past disasters.

MIT mechanical engineering graduate student Kai Chang and Professor Themis Sapsis have developed an AI tool capable of forecasting extreme weather events that have not appeared in historical records. Published in Nature Communications on August 20, the method—known as Extreme Event Aware or η-learning—generates maps of statistically plausible scenarios, including estimates for event duration, intensity, and affected areas.

Overcoming Data Limitations

Traditional risk models rely on datasets containing past extreme events to project future patterns, which limits their ability to predict unprecedented disasters. The new AI approach bypasses this by learning the statistical relationship between point statistics and spatial maps. This allows the algorithm to construct patterns for extreme events that exceed the intensity of anything previously recorded.

Testing and Potential Applications

The researchers tested the model using 25 years of hourly rainfall data from the continental U.S. By training the spatial component on only six months of data and applying broader point statistics, the tool successfully generated maps of storms more intense than those in the historical record. For example, the model can simulate a 300-millimeter rainfall event for New York City, even though the highest recorded rainfall there is 200 millimeters. These simulations could help city planners and grid operators test infrastructure resilience against scenarios like larger wildfires or more severe storm surges.