MIT's η-Learning Generates Extreme-Weather Scenarios That Never Happened — Yet

Can a city’s seawall survive a blockbuster storm? Will a regional power grid hold under record-breaking heat? Can a town’s firefighting resources contain a wildfire bigger than anything in its records? Communities can only answer these questions if they can first picture how such events would unfold — and that has always been the hard part, because by definition, unprecedented events have no precedent to learn from.

On August 24, 2026, MIT News detailed a machine-learning approach that breaks this circular dependency. Engineers Kai Chang, a mechanical engineering graduate student, and Professor Themis Sapsis developed an algorithm dubbed Extreme Event Aware learning — or η-learning — that generates realistic maps of extreme events which have never appeared in a region’s historical record, yet remain statistically plausible. The underlying paper appeared open-access in Nature Communications on August 20, 2026.

The problem with learning from disasters

Almost every risk model in use today — the kind insurers, city planners, and grid operators rely on — works backward from history. To estimate what a once-in-a-century storm looks like for New York City, simulations must be trained on datasets that already contain once-in-a-century events, so the model can learn the conditions that produced them and extrapolate forward.

Chang argues this creates a fundamental ceiling on what such models can show. “These methods assume there are very disastrous events that we have seen in the dataset, and they build a method to either estimate the risk of those events, or they try to predict exactly the events that have happened,” he says. “We are trying to see: What do unprecedented extreme events look like that are riskier than everything that has happened before and yet are still plausible?”

Sapsis frames the limitation through a concrete example. “An event like Hurricane Katrina is something that happens every 30 to 40 years. What will be the Katrina that happens every 100 years? How bad will it be? That’s exactly what we’re trying to quantify, to help planners prepare for plausible extreme scenarios.”

The stakes of this blind spot are rising. As the climate warms, the historical record increasingly understates tail risk: the past is becoming a weaker guide to the future precisely when planners most need forward-looking estimates.

Two data types, one statistical bridge

The algorithm’s trick is to combine and learn the statistical relationship between two different representations of the same data:

• Point statistics — how often a given intensity level, such as the maximum rainfall across a map, occurs within a full record.
• Spatial maps — how an event’s impact varies across a region at high resolution.

Learning the relationship between the two lets the algorithm construct spatial patterns for events beyond anything in its training data, using the point statistics as a constraint that filters out physically implausible scenarios.

The demonstration is striking in its data efficiency. The researchers applied η-learning to precipitation over the continental United States, starting from 25 years of hourly rainfall data pooled into daily maps. From the full record they computed point statistics describing how often map-scale maximum rainfall reached each level. But the spatial part of the model was trained on paired low- and high-resolution maps from just the first six months of the 25-year record — a window that contained few or no examples of the heaviest rainfall levels.

From that deliberately impoverished training set, the algorithm learned how coarse patterns correspond to fine-grained precipitation detail, then used the full-record statistics to bound how extreme its generated storms could plausibly become.

Prompting the once-in-a-century storm

Once trained, the tool works like a scenario generator with a return-period dial. A user can ask, in effect, “What could a once-in-a-century storm look like in New York City?” The algorithm responds with maps of statistically plausible storms at that frequency, each annotated with its likely size, area of coverage, and rainfall intensity, plus separate estimates of duration.

The headline example: the highest rainfall ever recorded in New York City measures about 200 millimeters. η-learning can generate plausible storm maps producing 300 millimeters — a level with zero match in the observational record, yet consistent with the region’s statistical tail. And it doesn’t produce one canonical answer. “Someone can say, ‘I’m interested in building things to withstand the risk of an event that happens every 100 years,’” Chang says. “What we can do then is produce thousands of possible realizations that will happen with this sort of rare frequency.”

Those thousands of realizations are exactly what stress-testing needs. A city could run the map ensemble against its seawall designs for storm surges beyond anything recorded; a grid operator could check whether the network survives a longer, hotter heatwave than history contains; fire agencies could probe whether containment resources hold against a wildfire larger than any on file.

Beyond the weather map

The method is not weather-specific. As long as relevant point statistics and spatial (or analogous structural) data exist, the same framework can be pointed at other rare-event domains. The team points to extreme floods and wildfires as natural extensions, and to fields outside Earth science entirely: robotic navigation in adversarial conditions, and financial markets.

“Financial market crashes are extreme events that are a complicated combination of things, involving many different sectors,” Chang notes. “What is the interaction that leads to a market crash? That is something that this method could explore.”

Sapsis places the work in a geopolitical-economic frame: “Extreme events have become a strategic concern, not just an environmental one — we’ve optimized global systems for efficiency, and the price of that efficiency is that there’s very little slack left anywhere. A single extreme event propagates through supply chains, energy markets, and food systems in weeks. Being able to put a probability on an event that hasn’t happened yet is now a question of national and economic resilience.”

Honest limits

The current demonstration covers precipitation over the continental US, and applying the method to a new hazard requires point statistics and spatial data specific to that hazard — it is not a turnkey universal risk engine. The generated scenarios are statistically plausible by construction; validating them as physically faithful would ideally involve comparison with physics-based simulations, an area where hybrid physics-plus-ML approaches remain active research. And as with any generative statistical model, the quality of the tail depends on the quality of the record feeding the point statistics.

Still, the core inversion is significant: instead of waiting for a disaster to enter the training data, η-learning manufactures the disaster digitally, in advance, with a probability attached. For infrastructure designed to last a century, that is the difference between preparing for the past and preparing for what has simply not happened yet.

The research was supported in part by a Vannevar Bush Faculty Fellowship and the U.S. Air Force Office of Scientific Research. The paper, “Extreme Event Aware (η-) Learning,” is open-access in Nature Communications, with a preprint also available on arXiv (2510.19161).