Generating scenarios for extreme events, without extreme data

What if a once-in-a-century storm hit a coastal city? How much destruction could it cause, and what preparations might be needed? In a recent paper published in Nature Communications, MIT engineers have unveiled a novel method to generate plausible extreme weather events, even without historical examples to draw upon. The technique, dubbed "η-learning," relies on machine learning algorithms trained on a region's weather data and maps, rather than solely on past extreme events.

The researchers feed the algorithm statistical information about the relationships between various data points, such as daily weather records and geographical maps. By learning these patterns, the algorithm can then produce plausible scenarios of extreme events, even those that have never been observed before. The team tested the method on generating maps of future extreme precipitation events across the United States, using 25 years of hourly precipitation data.

From this data, they calculated the likelihood of different rainfall levels and trained the algorithm using only the first six months of data, which contained few examples of extreme rainfall. Through this process, the algorithm learned how low-resolution maps could be used to infer higher-resolution precipitation patterns, even those exceeding the most extreme levels seen in the training data.

This new approach could help city planners, policymakers, and insurance companies better prepare for the worst-case scenarios, even when past data is limited or scarce. By enabling the creation of maps depicting the potential size, intensity, and duration of extreme events, the algorithm empowers decision-makers to allocate resources more effectively and design infrastructure capable of withstanding the most extreme conditions.