MIT engineers have a new way to plan for weather disasters that have no precedent.
The AI tool, called Extreme Event Aware, generates maps of events a region has never recorded but that remain statistically possible. Each map carries estimates of duration, intensity, and the area at risk.
Graduate student Kai Chang and professor Themis Sapsis built the system, which appears in Nature Communications. Its name also carries a Greek letter, eta, giving it the shorthand eta-learning.
Conventional risk models train on past disasters, so they can only replay versions of what already happened. Sapsis uses Hurricane Katrina to show the gap. That storm recurs every 30 to 40 years, he notes, so planners still lack a picture of the once-a-century version.
The algorithm connects two kinds of statistics: how often a given intensity occurs, and how an event’s impact spreads across a region. From that relationship it builds spatial patterns for extremes missing from its training data.
Tests ran on US precipitation, using 25 years of hourly rainfall. Insurers, city planners, and grid operators are the target users.
