MIT generated plausible extreme events without extreme-event data

Infrastructure planners, policymakers, and insurers need to know what a once-in-a-century storm, heat wave, wildfire, or flood could look like. Historical records are a poor training set for that question: extreme events are rare, and the worst event a region has seen may still be milder than the one it must withstand next.

MIT engineers Kai Chang and Themis Sapsis built Extreme Event Aware, or η-learning, to generate plausible unprecedented scenarios without examples of those extremes. The algorithm learns relationships between point statistics, such as how often rainfall reaches a level, and spatial maps, such as where an event spreads. The statistics rule out implausible magnitudes while the maps supply realistic geography.

Using 25 years of hourly precipitation maps over the continental United States, the team demonstrated how the method could generate thousands of realizations of a once-every-100-year storm, including events more intense than anything in the training data. A planner could ask what a 300-millimeter storm might look like in New York City and receive maps of its possible size, coverage, and intensity.

The approach can also visualize plausible floods and wildfires and may apply to robotic navigation and financial-market crashes. The work was published in Nature Communications; it is a planning and simulation method, not evidence that any particular future event will occur.

Constrain spatial generation with point statistics

A model trained only on past extremes cannot directly show planners an event more severe than the historical record.

Point statistics provide the rare-event frequency and magnitude constraint without requiring a matching extreme map in the training examples.

Spatial maps preserve the geographic relationships that make a generated storm or wildfire plausible rather than merely large.

Thousands of realizations let planners explore a distribution of credible scenarios instead of betting on one forecast.

If the event you must plan for is absent from history, separate its frequency from its spatial shape and generate scenarios satisfying both constraints instead of waiting for more history.

The method gives communities a way to stress-test seawalls, power grids, firefighting resources, and supply chains against plausible events not yet observed. The article presents it as a research tool that can extend to floods, wildfires, robotics, and markets; it does not report a deployed planning system or a measured reduction in losses.