Preparing for extreme events often starts with a difficult question: what should a community plan for when the worst case has not happened yet? MIT engineers have developed a machine-learning tool designed for that problem, generating plausible scenarios for rare events even when the historical record does not contain examples as severe as the scenario being tested.
The method can produce maps of potential extreme events, including details such as duration, intensity, and area of impact. Its central promise is practical: planners could examine what a rare storm, flood, wildfire, or heat-related threat might look like before a comparable event appears in the data.
Why rare risks are hard to model
Extreme events are difficult to anticipate because they sit at the edge of the record. A city may want to know whether a seawall could withstand a major storm, whether a power grid could hold up under record-breaking heat, or whether firefighting resources could contain a large wildfire. But those questions depend on knowing how the event could unfold.
The challenge is that traditional risk methods often rely on past extremes to estimate future extremes. That creates a gap: the event of interest may be worse than anything previously recorded, yet still plausible enough that planners need to account for it.
MIT graduate student Kai Chang, in mechanical engineering and affiliated with the MIT Center for Computational Science and Engineering, describes the goal this way: "We are trying to model extreme, unprecedented events that no one has seen before, that are not in the dataset."
Themis Sapsis, the William I. Koch Professor of Mechanical and Ocean Engineering at MIT, frames the planning question around events that are rarer than familiar disasters. "An event like Hurricane Katrina is something that happens every 30 to 40 years," he says. "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 the new method works
The team calls the approach Extreme Event Aware, or "η-learning." It is a machine-learning algorithm that learns from available datasets such as daily weather records and maps. Those records may include past extremes such as record-setting heat or rain, but the method does not require those examples in order to generate more severe future scenarios.
Instead of simply reproducing disasters that already appear in the record, the algorithm uses statistics to rule out implausible possibilities. It then generates extreme events that are statistically plausible for a region and frequency, such as an event expected once every 100 years.
The output is not just a single number. The method can project what the event might look like across space and time, including:
- the likely size of the event;
- its intensity;
- its duration;
- the area it could affect;
- the spatial pattern shown on a map.
That mapping function matters because infrastructure decisions are often spatial decisions. A rainfall total alone does not tell a city where flooding pressure may concentrate, how broad the affected area could be, or what type of reinforcement might be most relevant.
The rainfall test case
To demonstrate the method, the researchers applied it to future extreme precipitation over the continental United States. They began with 25 years of hourly precipitation maps, which they pooled into daily maps. From that record, they computed point statistics describing how often maximum rainfall across a map reached a given level.
They then trained the algorithm on paired low- and high-resolution spatial maps from only the first six months of the record. According to the source, that training slice contained few or no examples of the most extreme rainfall levels.
From this limited training data, the algorithm learned how low-resolution patterns correspond to detailed, high-resolution precipitation maps. It then used point statistics to constrain the extremes shown in those maps. That combination allowed the method to generate plausible spatial patterns for events more extreme than those represented in the training data.
One example in the source concerns New York City rainfall. If the most extreme recorded rainfall measurement is 200 millimeters, planners may still want to understand what kind of storm could produce 300 millimeters. Such an event has not appeared in the record described, but the method is designed to explore how it could look if it remained statistically plausible.
A user could ask, "What could a once-in-a-century storm look like in New York City?" The trained algorithm would then generate maps of storms likely to occur with that rare frequency, showing characteristics such as coverage area, storm size, and rainfall intensity.
What planners could do with plausible scenarios
The method is aimed at people and institutions that have to plan for rare events before they arrive. The source names planners, policymakers, and insurance companies as groups that already ask questions such as what a once-every-100-year storm might look like for New York City.
Chang explains the use case in terms of building against a rare risk. "Someone can say, ‘I’m interested in building things to withstand the risk of an event that happens every 100 years,’" he says. "What we can do then is produce thousands of possible realizations that will happen with this sort of rare frequency."
That matters because worst-case planning is not only about estimating whether an event can happen. It is also about seeing many ways an event could happen. Thousands of plausible realizations can show different paths to the same level of risk, helping decision-makers think beyond a single imagined storm track, rainfall pattern, or area of impact.
The team says the approach could also be applied to other unprecedented events, including extreme floods and wildfires, as long as relevant point statistics and spatial data are available. Beyond weather, the source says the method could apply to robotic navigation and financial markets.
Chang points to market crashes as another type of extreme event. "Financial market crashes are extreme events that are a complicated combination of things, involving many different sectors," he says. "What is the interaction that leads to a market crash? That is something that this method could explore."
Why the stakes extend beyond weather
Sapsis argues that rare-event modeling has become broader than environmental planning. "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.
His conclusion is direct: "Being able to put a probability on an event that hasn’t happened yet is now a question of national and economic resilience."
Sapsis and Chang detail the method in an open-access paper that appeared on Aug. 20 in the journal Nature Communications. The work does not remove uncertainty from rare-event planning. It offers a way to examine plausible futures when the past is too limited to provide a direct example.