MIT's η-Learning Generates Worst-Case Weather Scenarios Without Ever Seeing One

Can a city’s seawall survive a blockbuster storm? Will a regional power grid hold through record-breaking heat? Can a town’s firefighting resources contain a major wildfire? Communities can only answer these questions if they can first picture how such events would unfold — how far the wildfire spreads, how much area the storm covers, how long the heat wave lasts. The problem is that extreme events are, by definition, outliers: rare, sporadic, and poorly represented in the historical record. Most risk-assessment methods therefore extrapolate from past disasters to imagine future worst cases, which means they are anchored to the very events that have already happened.

A team of MIT engineers has now broken that anchor. In a paper published in Nature Communications on August 20, 2026, graduate student Kai Chang and Professor Themis Sapsis introduced a machine-learning framework called Extreme Event Aware (η-) learning, which generates statistically plausible, unprecedented extreme scenarios without requiring any extreme events in its training data. MIT News covered the work on August 24, and the timing could hardly be better: as AI weather models proliferate, their shared weakness at rare, unprecedented events has become one of the most consequential open problems in applied machine learning.

The problem with learning from disasters

Modern data-driven risk modeling has a built-in blind spot. As the authors write in their abstract, existing methods “often require multiple extremes in the training data or sampling process, leading to accurate predictions in quiescent regimes but high epistemic uncertainty in extreme-event regions.” In plain terms: a model trained on decades of ordinary weather is excellent at ordinary weather and unreliable exactly where the stakes are highest.

Consider the concrete example the MIT team offers. If the most extreme rainfall ever recorded in New York City is 200 millimeters, what kind of storm would produce 300 millimeters? No such event exists in the dataset, yet it remains physically plausible — and a city planner deciding on drainage capacity or seawall height desperately needs to know where such a storm would hit, how large an area it would cover, and how intense it would be.

“An event like Hurricane Katrina is something that happens every 30 to 40 years,” Sapsis explains. “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.”

How η-learning works

The method’s core trick is statistical regularization. Instead of demanding examples of disasters, the algorithm learns from a dataset such as a region’s daily weather records and maps — records that may contain few or no extreme deviations at all. It combines two types of information: point statistics (probabilities describing how often a measurable quantity, like maximum rainfall, reaches a given level) and spatial maps (the geographic patterns of events).

During training, the model is constrained to remain consistent with the prescribed observable statistics, even in regions of the input space where no data exists. Theoretical results based on optimal transport provide rigorous justification and establish key optimality properties. The result is a model that fits observed data faithfully while remaining statistically honest about the tails — enabling it to generate unprecedented events that are extreme yet plausible, rather than extreme and absurd.

The demonstration is instructive. The researchers applied the method to generate maps of extreme precipitation over the continental United States. They started with 25 years of hourly precipitation maps pooled into daily maps, computed point statistics over the full record, but trained the algorithm on paired low- and high-resolution spatial maps from just the first six months of the record — a slice containing few or no examples of the most extreme rainfall. The algorithm learned how coarse patterns correspond to detailed precipitation maps, then used the point statistics to constrain the extremes. From this, it could generate plausible spatial patterns for events more extreme than anything in its training data: the possible locations, sizes, and intensities of a once-in-a-century rainfall event peaking at 300 millimeters.

A user can then prompt the trained model with a question like “What could a once-in-a-century storm look like in New York City?” and receive thousands of statistically consistent realizations — maps showing the storm’s size, coverage area, and rainfall intensity. “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.”

Why this matters beyond weather

Weather is the flagship application, but the framework is deliberately general. The team notes it can be applied to robotic navigation and financial markets wherever relevant point statistics and spatial data exist — extreme floods and wildfires are natural next targets.

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

Sapsis frames the stakes in strategic rather than environmental terms. “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,” he says. “Being able to put a probability on an event that hasn’t happened yet is now a question of national and economic resilience.”

Context: AI weather models’ known weakness

The work lands amid an active debate about AI’s role in forecasting. Physics-based numerical models have historically outperformed AI systems for record-breaking weather, and a University of Chicago study published in late 2025 found that AI weather models “do not extrapolate to gray swan events that they have not seen during training, despite thousands of years” of synthetic training data. That finding — neural networks failing at precisely the unprecedented events η-learning targets — explains why the MIT approach is structurally different: rather than hoping a network extrapolates gracefully, it hard-codes statistical honesty about the tails into the training objective itself.

The research was supported in part by a Vannevar Bush Faculty Fellowship and the U.S. Air Force Office of Scientific Research. The paper is open access, and Chang has released the code on GitHub (repository: eta), lowering the barrier for insurers, municipal planners, and infrastructure operators to experiment with the method.

Outlook

η-learning will not replace ensemble forecasting, and its outputs are scenario generators, not deterministic predictions — the value lies in stress-testing infrastructure against a statistically defensible worst case rather than forecasting a specific storm on a specific day. But for the growing list of institutions whose exposure is defined by events that have never happened yet — seawall designers, grid operators, supply-chain planners, reinsurers — a method that can credibly imagine the unprecedented fills a genuine methodological gap. As extreme events shift from environmental curiosity to economic and national-security concern, expect statistical regularization of the unknown to become a standard tool in the resilience toolkit.