On an August afternoon in 2005, Hurricane Katrina barreled toward the Gulf Coast. Less than three weeks later, New Orleans was underwater, more than 1,800 people were dead, and the United States confronted the limits of its infrastructure. Katrina was, by any measure, an extreme event — but statistically, it was not the worst imaginable. Estimates place Katrina as roughly a 1-in-30 or 1-in-40-year storm for that region. What, then, would a 1-in-100-year storm look like? What would a 1-in-500-year event do to a city’s seawall, a region’s power grid, or a town’s fire-fighting resources?
These are not idle questions. They are the central challenge of modern climate risk assessment. And until now, the tools available to answer them have been fundamentally limited by a paradox: to predict an extreme event, most models require data on previous extreme events. But extreme events, by definition, are rare. They are outliers in the historical record. A method that can only extrapolate from what has already happened is poorly equipped to anticipate what has never happened — but plausibly could.
Researchers at the Massachusetts Institute of Technology have now developed a machine-learning tool that breaks this constraint. Called Extreme Event Aware, or η-learning, the algorithm generates plausible worst-case scenarios for extreme weather events — including hurricanes, heat waves, floods, and wildfires — without requiring any past examples of those events in its training data. The method, detailed in an open-access paper published August 20 in Nature Communications, represents a fundamental shift in how communities, insurers, and policymakers can prepare for the unprecedented.
The Problem with Predicting the Unprecedented
To understand why η-learning matters, it helps to understand why conventional risk assessment falls short. When a city wants to know what a 1-in-100-year storm might look like, it typically turns to computer simulations. These models are trained on historical data — decades of rainfall records, temperature logs, wind measurements. They learn the patterns that precede extreme events and then project those patterns forward.
The catch is that the training data must include examples of the very events the model is trying to predict. To simulate a 1-in-100-year storm, the model needs to have seen at least one storm of that magnitude. But for most regions, such an event may never have been recorded. The most extreme rainfall ever measured in New York City might be 200 millimeters in a single day. A storm producing 300 millimeters has never happened — at least, not in the instrumental record. Yet such a storm is physically plausible, and city planners need to know where it would hit, how big an area it would cover, how intense the rainfall would be, and how long it would last.
“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,” explains Kai Chang, an MIT graduate student in mechanical engineering and affiliate of the MIT Center for Computational Science and Engineering. “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?”
How η-Learning Works: Statistics Without Past Examples
The core innovation of η-learning is that it decouples the generation of extreme-event scenarios from the need for historical extreme-event data. Instead, the algorithm learns from two separate but complementary types of information: point statistics and spatial maps.
Point statistics describe, for a given location, how often a certain level of rainfall (or temperature, or wind speed) occurs. For example, a point statistic might tell you that rainfall of 300 millimeters in a single day over a specific grid cell has a 1-in-100-year probability. These statistics are typically well-characterized even for rare events, because they are derived from long-term observational records and probabilistic distributions, not from specific storm events.
Spatial maps show the patterns of weather across a region — where rain falls, how intense it is, how the storm is shaped. These maps are abundant in the historical record, but they rarely contain examples of the most extreme events. A 25-year dataset of daily precipitation maps over the continental United States, for instance, might include thousands of storms, but only a handful of truly extreme ones.
The η-learning algorithm is trained on paired low-resolution and high-resolution spatial maps from just the first six months of such a dataset — a period that contains few or no examples of extreme rainfall levels. From these, it learns how broad, coarse weather patterns correspond to detailed, high-resolution precipitation maps. Then, it uses the point statistics to constrain the extremes: it tells the algorithm, in effect, “These are the probabilities for rainfall at each intensity level. Now generate spatial patterns that respect these probabilities, even if the patterns themselves have never been observed.”
The result is a tool that can generate thousands of plausible realizations of a 1-in-100-year storm for any given location, complete with maps showing the storm’s likely size, its area of coverage, and the spatial distribution of rainfall intensity. A user can prompt the trained algorithm with a question such as, “What could a once-in-a-century storm look like in New York City?” and receive statistically robust visualizations of storms that are unprecedented but physically possible.
Beyond Weather: Financial Markets and Robotic Navigation
While the immediate application of η-learning is in climate risk assessment, the researchers emphasize that the method is domain-agnostic. Any field that deals with extreme, rare events in high-dimensional spatial or temporal data could benefit. The team specifically points to financial markets and robotic navigation as promising areas.
“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.”
In finance, a crash is a rare event. Historical data may contain only a handful of examples, leaving models poorly equipped to anticipate the next one. η-learning could, in principle, generate plausible crash scenarios based on the statistical properties of market behavior and the spatial patterns of sector interactions, without needing to have observed a crash of that magnitude before.
Similarly, in robotics, an autonomous vehicle might need to anticipate rare but catastrophic scenarios — a pedestrian stepping out from behind a blind corner, or a sudden loss of traction on an icy patch — that were not present in its training data. η-learning could generate these adversarial scenarios for testing and validation, making autonomous systems more robust to the unexpected.
A Practical Tool for Infrastructure Planning
The most immediate impact of η-learning will likely be in infrastructure planning and climate adaptation. Cities and regions around the world are investing heavily in hardening their defenses against extreme weather — building higher seawalls, upgrading power grids, expanding firefighting capacity. But these investments are only as good as the risk assessments that guide them.
Themis Sapsis, the William I. Koch Professor of Mechanical and Ocean Engineering at MIT and a core member of the Center for Computational Science and Engineering, puts it bluntly: “We want to predict maps of these worst-case scenarios. There is no method that does this efficiently to predict events that happen rarely.”
With η-learning, a planner could generate maps of a 1-in-100-year flood event for a coastal city, showing not just the expected flood depth but the spatial extent — which neighborhoods are most exposed, which roads are likely to be cut off, which critical infrastructure sits in the danger zone. The same tool could be used to model extreme heat waves, informing decisions about grid capacity and cooling center placement, or extreme wildfires, helping fire departments pre-position resources and plan evacuation routes.
The method’s efficiency is also a major advantage. Because it does not require training on rare events — which are inherently scarce — the algorithm can be trained on just a few months of relatively ordinary data, combined with well-established point statistics. This makes it practical for regions that lack decades of high-quality extreme-event records, including many developing countries where climate risk is highest and data is thinnest.
Strategic and Economic Stakes
As climate change accelerates, the frequency and intensity of extreme weather events are increasing. What was once a 1-in-100-year storm may become a 1-in-50-year event, or even a 1-in-10-year event. The past is no longer a reliable guide to the future. Tools like η-learning, which can generate plausible scenarios for events that have never happened, are becoming not just useful but essential.
Sapsis frames the stakes in strategic 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. Being able to put a probability on an event that hasn’t happened yet is now a question of national and economic resilience.”
This is not hyperbole. The 2021 winter storm that knocked out power across Texas, causing hundreds of deaths and billions of dollars in damage, was an extreme event for which the state’s infrastructure was not designed. The 2022 floods in Pakistan, which submerged a third of the country, were unprecedented in scale. The 2023 wildfires in Canada, which sent smoke plumes across the entire eastern United States, were beyond anything in the historical record. In each case, the event was not just extreme — it was outside the envelope of what planners had considered possible.
η-learning offers a way to expand that envelope. By generating plausible worst-case scenarios that are statistically grounded but not historically constrained, it gives planners a more complete picture of the risks they face. It does not predict the future — no tool can do that. But it does something almost as valuable: it maps the space of what is possible, so that when the unprecedented arrives, we are not caught entirely unprepared.
The Technical Foundation: A Step-by-Step Explanation
For readers who want a deeper understanding of how η-learning achieves its results, the technical mechanism is worth examining more closely.
What Are Point Statistics and Spatial Maps?
Point statistics are single-location probability distributions. For a given grid cell on a map, the point statistic tells you, “Over many years, what fraction of days had rainfall between X and Y millimeters?” These statistics can be reliably estimated even for rare events, because they are derived from long-term records and are well-described by extreme-value theory, a branch of statistics that models the tails of probability distributions.
Spatial maps are full-field data — images or grids showing weather across a region at a given time. They contain rich information about spatial correlations: where rain is likely to fall relative to the storm center, how intensity tapers off at the edges, how storm shape relates to underlying topography. But because extreme events are rare, the spatial maps in any dataset are overwhelmingly dominated by ordinary, non-extreme weather.
How Does the Algorithm Combine Them?
η-learning trains on paired low- and high-resolution spatial maps from a short period (the researchers used six months of data). From these pairs, the algorithm learns a mapping: given a coarse, low-resolution description of a weather pattern, what does the corresponding high-resolution map look like? This is essentially a super-resolution task, similar to what AI image upscaling does.
Then, the algorithm incorporates the point statistics as a constraint. It generates candidate high-resolution maps and checks whether their pixel-level intensity distributions match the point statistics for each grid cell. If a generated map shows rainfall of 300 millimeters in a cell where the point statistic says that level occurs once every 100 years, the map is kept as a plausible 1-in-100-year event. If the map shows patterns that violate the point statistics — for example, rainfall that is implausibly high everywhere — it is rejected.
By iterating this process, η-learning generates an ensemble of maps that are statistically plausible, spatially coherent, and maximally extreme within the bounds set by the point statistics. The result is a set of scenarios that are not just extrapolations of past events but genuinely novel configurations that respect the physical and statistical constraints of the system.
What Makes This Different from Generative AI?
One might ask: Could a generative adversarial network (GAN) or a diffusion model do the same thing? The key difference is that those models learn from the distribution of data in the training set. If the training set contains no extreme events, the model will not generate them — it has no way to know what they look like. η-learning, by contrast, uses point statistics as an external constraint that pushes the generation toward extremes that are outside the training distribution. It is a form of what the researchers call “extreme-aware” learning, where the algorithm is explicitly designed to explore the tails of the probability distribution rather than the bulk.
Future Outlook: From Weather to Financial Crises
The research, supported by a Vannevar Bush Faculty Fellowship and the U.S. Air Force Office of Scientific Research, is still in its early stages. The team has demonstrated the method on extreme precipitation events over the continental United States, producing maps that show plausible spatial patterns for 1-in-100-year rainfall events. But the framework is general.
“As long as relevant point statistics and spatial data are available, the method could be applied to visualize other unprecedented events such as extreme floods and wildfires,” the researchers write. The same approach could be extended to extreme heat, drought, windstorms, and even compound events — for example, a heat wave combined with a drought, or a hurricane followed by heavy rainfall, where the interaction of two hazards creates a risk greater than the sum of its parts.
Beyond environmental applications, the method could be used to model rare but catastrophic failures in engineered systems — such as power grid cascades, supply chain disruptions, or financial contagions — where the goal is to understand what a plausible worst-case scenario looks like, even if it has never been observed.
The broader implication is profound. For centuries, humanity has looked to the past to understand the future. We build flood walls based on the worst flood in living memory. We set insurance premiums based on historical loss data. We design power grids for the highest demand ever recorded. But in a world of rapid climate change and tightly coupled global systems, the past is no longer a sufficient guide. The unprecedented is becoming the new normal. Tools like η-learning — which can map the plausible but never-seen — are not just academic curiosities. They are becoming operational necessities for a resilient future.
