MIT researchers Kai Chang and Themis Sapsis developed a machine-learning method called Extreme Event Aware (η-learning) that generates plausible worst-case scenarios for extreme events, such as storms, heat waves, or wildfires, without needing historical data on similar extreme events. The algorithm learns from point statistics and paired low- and high-resolution spatial maps, then generates plausible spatial patterns for events more extreme than anything in the training data. Demonstrated on continental U.S. precipitation records, the method could predict what a once-in-a-century storm might look like for a city, including its size, intensity, and area of impact, and could extend to floods, wildfires, robotic navigation, and financial market crashes. The work is detailed in a paper published in Nature Communications.
Questions this post answers
How can you predict what a once-in-a-century storm would look like without historical data on such storms?
A machine-learning method called Extreme Event Aware (η-learning), developed at MIT, combines point statistics on rainfall extremes with paired low- and high-resolution spatial maps learned from as little as six months of ordinary weather data. It uses these statistics to constrain plausible spatial patterns, generating maps of storms more extreme than anything in the training data, including size, intensity, and area of impact. daily.dev surfaces research like this for engineers exploring ML approaches to climate risk modeling.