MIT AI Breaks Barrier: Predicts Extreme Weather Without Past Data

MIT engineers have developed an innovative AI tool capable of forecasting extreme weather events without relying on historical disaster data. This groundbreaking method, named Extreme Event Aware or η-learning, was created by mechanical engineering graduate student Kai Chang and Professor Themis Sapsis, who holds the William I. Koch Professorship in Mechanical and Ocean Engineering at MIT. The tool generates maps of statistically-possible events that have no precedent in a region’s historical record, providing estimates of their likely duration, intensity, and potential area of effect.

Traditional risk models, often utilized by insurers, city planners, and grid operators, typically aim to predict events like a 'once-in-a-century storm' based on datasets already containing extreme occurrences. These models learn the conditions that produced past disasters and project similar patterns forward. However, Chang highlights a significant limitation in this approach: it assumes that all disastrous events have already been observed. Sapsis further illustrates this by questioning, “An event like Hurricane Katrina is something that happens every 30 to 40 years. 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.”

The η-learning algorithm addresses this gap by combining two distinct types of data. First, point statistics capture the frequency of a given intensity level, such as the maximum rainfall recorded across a map. Second, spatial maps illustrate how an event’s impact varies across a specific region. By learning the statistical relationship between these two data types, the algorithm can construct spatial patterns for events that exceed anything present in its training data, effectively forecasting without direct prior examples of such extreme occurrences. This method was detailed in a paper published in Nature Communications on August 20.

The researchers rigorously tested their approach using 25 years of hourly rainfall data pooled into daily maps for the continental United States. They computed point statistics to describe the frequency of maximum rainfall levels across the entire record. Crucially, the spatial model was trained using a narrow window: paired low-resolution and high-resolution maps from only the first six months of the 25-year record, a period with few or no examples of the heaviest rainfall. The algorithm learned the correspondence between patterns in low-resolution and high-resolution maps, then applied the comprehensive point statistics to constrain the potential extremity of the generated patterns.

The practical applications of this tool are significant. For instance, while the highest rainfall ever recorded in New York City is 200 millimeters, the method can generate plausible maps of a storm producing 300 millimeters—a level unprecedented in observational history. Users can prompt the algorithm to visualize a statistically-plausible once-in-a-century storm for a specific city, yielding maps that include varying sizes, areas of coverage, and rainfall intensities. Chang notes that the algorithm can rapidly generate a large volume of these scenarios. Such maps could be instrumental for cities to stress-test infrastructure like seawalls against previously unrecorded storm surges, assess power grid resilience during extended heatwaves, or evaluate firefighting resources against wildfires larger than any on record.

While the demonstration has been impactful, applying the method to new hazards necessitates corresponding point statistics and spatial data for that specific hazard. With the availability of such data, Chang and Sapsis envision extending the tool to visualize severe floods and wildfires that lack historical equivalents. Sapsis underscores the broader importance of this research, stating that global infrastructure, optimized for efficiency, often lacks resilience. “A single extreme event propagates through supply chains, energy markets, and food systems in weeks,” he explains. “Being able to put a probability on an event that hasn’t happened yet is now a question of national and economic resilience.”