An MIT AI imagines storms the historical record can't show

Every hurricane risk model on the market learns from storms that have already happened, which means by construction it cannot show a worse one. The question Themis Sapsis wants his new method to answer is what the Katrina that happens every 100 years looks like, so that planners can size stormwater systems, insurance policies, and grid hardening for the storm the historical record has not yet shown.

MIT engineers Kai Chang and Themis Sapsis published a method in Nature Communications on 20 August that aims at that gap. Their approach, called Extreme Event Aware learning, or η-learning, learns the statistical link between how often an intensity occurs and where its impact lands, then draws spatial maps for events beyond anything in its training data. Sapsis holds the William I. Koch Professorship in Mechanical and Ocean Engineering and an appointment with the MIT Institute for Data, Systems, and Society; both authors work at the MIT Center for Computational Science and Engineering.

The mechanism works by splitting the problem into two pieces and learning the relationship between them. The first is point statistics, the numbers describing how often a given rainfall intensity occurs at a place. The second is spatial maps of how an event's impact varies across a region. Once the algorithm knows how the two relate, it can build patterns for storms that are statistically plausible but absent from the historical record, without needing a prior example of the extreme itself.

"What we're really interested in is to understand the relation between the statistics of the intensity at a single point, and the spatial field, what the impact is as a function of space," Sapsis said in the MIT News release.

Each output map carries three separate estimates: how long the event is likely to last, how intense it will be, and what area it might affect. Together they form a planning artifact rather than a forecast, a picture of a worst-reasonable-case storm that has not been observed but could plausibly occur.

Sapsis frames the use case through Hurricane Katrina. In his telling, Katrina was a roughly 30-to-40-year event, which is bad enough on its own. The 100-year version is the storm the planning community actually needs to prepare for, the one whose rainfall footprint and surge reach have not been recorded but are within the statistical envelope the method tries to draw.

Insurers price tail risk against the worst event a model can produce. City planners size stormwater systems to handle the storms they assume are possible. Grid operators harden substations against the footprints they expect. All three run into the same wall: a training set that by definition cannot contain the very event they need to plan for.

The published work validates the approach on precipitation across the continental United States. That is enough to demonstrate the mechanism but it is not the same as an operational forecast. The paper is a method paper, not a deployed product, and replacing the existing risk models in insurance, planning, and grid operations is a question of regulatory acceptance, integration cost, and validation against new data that the published work has not yet taken on.

The next step the team flags is extending the framework to wind and storm surge, where the historical record is shorter and the tail matters more. Until that work lands, the method is a way to widen the planning envelope, not a replacement for the risk models already in use.