MIT Develops AI System to Predict Extreme Weather Events Never Before Recorded

A team of MIT engineers has unveiled an artificial intelligence system capable of generating forecasts for extreme weather events that lack historical precedent in a given region. Kai Chang, a graduate student in mechanical engineering, and Professor Themis Sapsis created the tool, which generates spatial maps depicting statistically-plausible phenomena alongside projections of their probable duration, intensity, and geographic reach.

Forecasting extreme weather events without historical precedent

Sapsis holds the William I. Koch Professorship in Mechanical and Ocean Engineering at MIT. Both researchers work within the MIT Center for Computational Science and Engineering, with Sapsis also serving in a role at the MIT Institute for Data, Systems, and Society. Their methodology, termed Extreme Event Aware or η-learning, appeared in a Nature Communications publication dated 20 August.

Traditional risk assessment frameworks operate on different principles. Insurance companies, municipal authorities, and power system operators frequently seek to understand what a once-per-century weather event might resemble in their specific location. Conventional simulation techniques typically rely on historical datasets that already include extreme occurrences, extracting the conditions that produced them before extrapolating comparable patterns into the future.

Chang identifies a fundamental constraint in this conventional strategy. "These methods assume there are very disastrous events that we have seen in the dataset, and they build a method to either estimate the risk of those events, or they try to predict exactly the events that have happened," he states.

Sapsis illustrates the same constraint using a specific example. "An event like Hurricane Katrina is something that happens every 30 to 40 years," he observes. "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."

Combining point statistics with spatial detail

The system draws upon two distinct categories of information. Point statistics measure how regularly particular intensity thresholds—such as peak rainfall values across a geographic area—manifest within a dataset. Spatial maps illustrate how an occurrence's consequences distribute across different locations within a region.

By understanding the mathematical connection between these two data types, the system can construct spatial configurations representing phenomena that exceed anything present in its training material, without requiring prior instances of those particular extremes.

The researchers validated their approach using precipitation patterns throughout the continental United States. They utilized 25 years of hourly precipitation information, aggregated into daily spatial representations, and determined point statistics indicating how regularly the maximum precipitation on a map reached specific levels throughout the complete dataset.

The temporal scope for training the spatial component remained limited. They trained this portion of the system using matched lower-resolution and higher-resolution representations sourced exclusively from the initial six months of the 25-year period, a timeframe that included minimal or no instances of the most intense rainfall events.

The system identified how patterns in the lower-resolution representations aligned with specifics in the higher-resolution versions, then employed the point statistics derived from the complete record to establish boundaries on how severe the generated patterns could become.

Testing infrastructure against worst-case maps

New York City's documented maximum precipitation stands at 200 millimetres. The technique can create realistic maps depicting a storm delivering 300 millimetres instead, a quantity with no corresponding instance in recorded observations.

Users can direct the trained system to illustrate what a once-per-century weather event might appear as for a particular metropolitan area. The resulting output comprises spatial maps representing statistically-feasible storms at that frequency. Every map includes its own dimensions and coverage zone, with rainfall levels fluctuating throughout the collection. Chang notes that the system can produce substantial quantities of these projections simultaneously.

Such generated maps could enable municipalities to evaluate their coastal defenses against storm surge surpassing anything previously documented. Comparable maps could demonstrate whether electrical infrastructure would remain functional during an extended heat episode, or whether firefighting personnel could manage a wildfire surpassing any prior occurrence.

Limits of the demonstration so far

Extending the technique to additional hazard categories demands pertinent point statistics and spatial information specific to each hazard, as Chang and Sapsis clarify. They reference potential applications once such information becomes obtainable, including depicting catastrophic flooding and wildfires lacking equivalents in the observational past.

Sapsis underscores that contemporary worldwide infrastructure has been engineered for maximum efficiency, providing minimal buffer capacity in the systems it maintains.

"A single extreme event propagates through supply chains, energy markets, and food systems in weeks," he elaborates. "Being able to put a probability on an event that hasn't happened yet is now a question of national and economic resilience."