Can a city’s seawall survive a blockbuster storm? Will a region’s power grid endure record‑breaking heat? Can a town’s fire crews contain a massive wildfire? To answer these questions, planners first need plausible outlines of how such extremes could unfold—wildfire spread, storm impact area, heat‑wave duration. Traditional risk models rely on past extreme records, yet by definition extremes are rare outliers.
MIT engineers have introduced a machine‑learning approach that generates plausible extreme scenarios without any historical extreme examples, called Extreme Event Aware (η‑learning). The algorithm only requires everyday observations, such as daily weather logs and spatial maps. It first computes point‑level statistics describing how often a given extreme magnitude occurs, then learns the relationship between low‑resolution and high‑resolution spatial fields. By constraining the generated fields with the point statistics, the model can synthesize spatial patterns that are more extreme than anything seen in the training set, for example a once‑in‑100‑years storm with a peak rainfall of 300 mm.
In a demonstration over the continental United States, the team pooled 25 years of hourly precipitation into daily maps, derived the maximum‑rainfall point statistics, and trained the model on paired low‑ and high‑resolution maps from the first six months, which contained few or no extreme cases. After training, a user can ask “What would a 100‑year storm look like in New York City?” and receive dozens of statistically plausible storm maps showing size, coverage area, and intensity.
The framework is not limited to weather; with appropriate point statistics and spatial data it can visualize unprecedented floods, wildfires, or even financial market crashes. As modern systems operate with minimal slack, a single extreme event can cascade through supply chains, energy markets, and food systems. Providing probability estimates for events that have never occurred yet is now a question of national and economic resilience.
Review