MIT AI Forecasts Extreme Weather Without Historical Data

Engineers at the Massachusetts Institute of Technology have built a novel artificial intelligence tool capable of forecasting extreme weather events without needing to train on historical disaster datasets from the past.

Developed by mechanical engineering graduate student Kai Chang and Professor Themis Sapsis, the system produces maps of events that have never appeared in a region's historical records but remain statistically possible, while also estimating duration, intensity, and affected areas.

The researchers detailed their method in a paper published in Nature Communications on August 20. The algorithm operates using two types of data: point statistics capturing intensity frequency, and spatial maps showing regional impact variations.

Stock photo for illustration only, not from the actual event

To test the approach, the team used 25 years of hourly rainfall data across the continental US, training the spatial portion of the algorithm using only the first six months of records which contained little to no heavy rainfall data, and applying long-term point statistics to constrain the outputs.

For instance, while the highest recorded rainfall in New York City is 200 millimetres, the method generated plausible maps for a 300-millimetre storm with no match in observational records. Users can prompt the algorithm to visualize once-in-a-century storms for specific cities.

"These methods assume there are very disastrous events that we have seen in the dataset, and they build a method to either estimate the risk of those events, or they try to predict exactly the events that have happened."

Kai Chang

These generated scenarios can help cities test seawalls against unprecedented storm surges, evaluate power grid resilience during extended heatwaves, or measure firefighting resources against larger wildfires than any on file.

Source: AI News