MIT develops method to map plausible unprecedented extreme events

MIT engineers have developed a machine-learning method that generates plausible maps of unprecedented extreme events without requiring examples of such events in its training data. The approach, called Extreme Event Aware or η-learning, can model scenarios such as a once-in-100-year storm and describe its likely size, intensity, duration, and area of impact.

Kai Chang, an MIT graduate student in mechanical engineering and an affiliate of the MIT Center for Computational Science and Engineering, and Themis Sapsis, the William I. Koch Professor of Mechanical and Ocean Engineering, presented the method in an open-access paper published on August 20 in Nature Communications. The research received support in part from a Vannevar Bush Faculty Fellowship and the U.S. Air Force Office of Scientific Research.

Generating events beyond the historical record

Risk assessments for infrastructure, public policy, and insurance commonly depend on historic extreme events. That creates a basic limitation: rare events are sporadic by definition, while planners often need to examine conditions that exceed anything recorded in the available data.

The MIT method is designed to generate scenarios that are more severe than observed cases while remaining statistically plausible. In the example given by the researchers, if New York City’s highest recorded rainfall is 200 millimeters, the tool can generate possible storms with 300 millimeters of rainfall and map where they might occur, how large an area they could affect, and how intense they could be.

How η-learning combines statistics and maps

The researchers demonstrated the approach using precipitation across the continental United States. They started with 25 years of hourly precipitation maps, pooled them into daily maps, and calculated point statistics for how frequently the maximum rainfall on a map reached particular levels.

They then trained the algorithm on paired low- and high-resolution spatial maps from only the first six months of that record, a period containing few or no examples of the highest rainfall levels. The model learned the relationship between broad, low-resolution weather patterns and more detailed precipitation maps.

Point statistics then constrain the level of rainfall extremes represented in generated maps. This combination lets the method produce plausible spatial patterns for rare events outside the training examples. A user can ask what a storm occurring at a specified frequency might look like, and the trained model can generate thousands of possible realizations with characteristics including coverage, rainfall intensity, size, and duration.

Potential uses beyond precipitation

MIT says the same framework could be used where relevant spatial data and point statistics are available, including for floods and wildfires. The researchers also identify possible applications in robotic navigation and financial markets, where rare outcomes can emerge from interactions among many factors.

For businesses and public agencies, η-learning does not replace a forecast of a particular future event. Its practical value is to provide statistically constrained stress scenarios for evaluating whether infrastructure, operational capacity, and supply-chain assumptions can withstand conditions that have not yet appeared in local records.