MIT engineers have developed an artificial intelligence tool that can predict regional extreme weather events without referring to data on disasters recorded in the past. Developed by Kai Chang and Prof. Themis Sapsis, the method, called Extreme Event Aware or η-learning (eta-learning), was published in Nature Communications on August 20. The model can generate maps and probability values for events that do not appear in a region’s historical records but are statistically likely to occur.
Limitations of Traditional Risk Models
Insurance companies, urban planners and grid operators often want to know what a once-in-a-century storm in a particular region would look like. Existing simulations generally use datasets that already contain extreme events. This limits the scenarios that models can generate. The researchers say that existing methods rely on disasters previously observed in datasets but are inadequate for measuring possible extreme scenarios that have never occurred.
Two Types of Data Combined
The algorithm developed uses two main types of data together. Point statistics show how frequently a particular intensity level, such as the maximum rainfall recorded on a map, appears in the dataset. Spatial maps, meanwhile, reflect how the impact of an event varies across different parts of a region. By learning the statistical connection between these two types of data, the system can generate spatial patterns for extreme events that are not represented in the training data.
25-Year Record and 6-Month Training Trial
Researchers tested the model on precipitation records from the continental United States. In the study, 25 years of hourly precipitation data were converted into daily maps, and point statistics were extracted. The training scope used for the spatial model was kept quite limited. This part of the algorithm was trained using pairs of low- and high-resolution maps obtained from only the first 6 months of the 25-year record. Although this 6-month period contained no examples of heavy precipitation, the system learned how low-resolution patterns corresponded to high-resolution details. It then used the point statistics to generate extreme scenarios.
300-Millimeter Precipitation Scenarios and Infrastructure Trials
Although the highest amount of precipitation ever measured in New York City was 200 millimeters, the new method can produce statistically plausible maps for a 300-millimeter storm that does not appear in observational records. These maps enable cities to test their coastal protection walls against record-level storm surges, assess the resilience of power grids to prolonged heat waves, and measure the adequacy of firefighting resources in the face of large fires.
Global Systems and Resilience
Adapting the model to a new type of hazard requires point statistics and spatial data related to the relevant field. The methodology can be used to study not only meteorological events but also robotic navigation and sudden crashes in financial markets.
Researchers state that global infrastructure systems are designed with efficiency as the guiding principle, and that a single extreme event can trigger cascading effects across supply chains and energy markets within weeks. They say that being able to assign a probability to an event that has not yet occurred is critically important for economic and national resilience.