MIT engineers have developed a machine-learning algorithm that generates plausible worst-case scenarios without needing past disaster data in its training set, MIT reported.
The method, detailed in an open-access paper published on Aug. 20 in the journal Nature Communications, is called Extreme Event Aware, or “η-learning.” Developed by graduate student Kai Chang and Themis Sapsis, the William I. Koch Professor of Mechanical and Ocean Engineering at MIT, the tool models unprecedented occurrences that traditional risk-assessment frameworks struggle to anticipate.
Precipitation mapping
Conventional computer simulations used by insurers and municipal planners rely on historical records of extreme events to simulate future conditions. When historical datasets lack record-breaking storms or heat waves, those models often fail to capture unprecedented extremes. The MIT algorithm circumvents this requirement by combining point statistics—which calculate how often a metric reaches a given threshold—with spatial maps.
To test the system, the researchers evaluated precipitation patterns across the continental United States. They took 25 years of hourly rainfall maps, compiled them into daily records, and extracted point statistics describing rainfall maximums across the entire span. They then trained the algorithm on paired low- and high-resolution spatial maps drawn from only the first six months of the record, a timeframe containing few or no extreme rainfall events.
By learning how low-resolution patterns relate to detailed high-resolution data and applying the point statistics as constraints, the model generated detailed maps of plausible, unobserved storms. For example, if a region's historical maximum rainfall was 200 millimeters, the model could project the size, coverage area, and intensity of a hypothetical once-in-a-century 300-millimeter storm.
Broader applications
Beyond meteorology, the researchers indicated the framework can evaluate other rare, complex occurrences such as wildfires, floods, robotic navigation failures, and financial market crashes.
The research received partial funding from the U.S. Air Force Office of Scientific Research and a Vannevar Bush Faculty Fellowship.
