MIT engineers have built an AI system that doesn’t just predict the weather — it imagines disasters that have never occurred in the historical record. The method, called Extreme Event Aware or η-learning, comes from mechanical engineering graduate student Kai Chang and Professor Themis Sapsis, and was published in Nature Communications on August 20.
The core problem with existing risk models is baked into their design. They train on datasets that already contain extreme events, learning the conditions that produced past disasters before projecting similar patterns forward. Chang puts it bluntly: “These methods assume there are very disastrous events that we have seen in the dataset.” That means insurers, city planners, and grid operators are always fighting the last war. Sapsis frames the limitation through Hurricane Katrina, which hits every 30 to 40 years. “What will be the Katrina that happens every 100 years?” he asks. “How bad will it be?”
The algorithm’s trick is combining two data types that previous approaches kept separate. Point statistics capture how often a given intensity level — like maximum rainfall across a map — appears in a dataset. Spatial maps show how an event’s impact varies across a region. By learning the statistical relationship between the two, the algorithm can generate spatial patterns for events beyond anything in its training data. The researchers tested it on 25 years of hourly rainfall data across the continental US, but trained the spatial model on just the first six months — a period with few or no examples of the heaviest rainfall levels. The algorithm learned how low-resolution patterns corresponded to high-resolution detail, then used the full record’s point statistics to constrain how extreme the generated patterns could become.
Here’s what that means in practice: the highest rainfall ever recorded in New York City is 200 millimetres. The tool can generate plausible maps of a storm dropping 300 millimetres — a level with no match in the observational record. Users can prompt the system to show a once-in-a-century storm for a named city, and it produces multiple maps with varying intensity, duration, and area of coverage. Chang says the algorithm can generate large volumes of these scenarios at once.
This matters because infrastructure has been optimized for efficiency, not resilience. Sapsis notes that global systems have little slack: “A single extreme event propagates through supply chains, energy markets, and food systems in weeks.” The ability to put a probability on an event that hasn’t happened yet is, in his words, “now a question of national and economic resilience.” The approach echoes a broader shift in AI risk modeling — moving from extrapolating past data to generating plausible futures — but the researchers are honest about current limits. Applying the method to floods or wildfires requires relevant point statistics and spatial data for each specific hazard, which doesn’t always exist yet.
