MIT engineers Kai Chang and Themis Sapsis have developed a machine-learning method called η-learning that forecasts extreme weather events without relying on historical disaster data. Published in Nature Communications, the system combines point statistics and spatial mapping to generate realistic scenarios for unprecedented storms, such as a hypothetical 300-millimeter rainfall event in New York City. This tool aims to help planners test infrastructure resilience and protect highly optimized, vulnerable global supply chains.
MIT η-learning algorithm development
- ▪MIT researchers Kai Chang and Themis Sapsis developed a machine-learning method named Extreme Event Aware, or η-learning, to forecast extreme weather without training on historical disaster data.
- ▪The details of the η-learning method were published in the journal Nature Communications on August 20, 2026.
Forecasting unprecedented extreme weather
- ▪The η-learning system produces maps showing the likely size, intensity, and duration of extreme weather events, alongside estimates of the areas they might affect.
- ▪The η-learning algorithm generates maps of statistically-possible extreme weather events that have not previously appeared in a region's historical record.
Point statistics methodology
- ▪Researchers tested the η-learning method on continental United States precipitation by pooling 25 years of hourly rainfall data into daily maps to compute point statistics.
- ▪The η-learning algorithm utilizes point statistics, which capture how often a given intensity level occurs within a dataset, to constrain how extreme the generated patterns can become.
Spatial mapping methodology
- ▪Researchers trained the spatial model of the η-learning algorithm using paired low-resolution and high-resolution maps drawn from only the first six months of a 25-year record.
- ▪The η-learning algorithm utilizes spatial maps to learn how an event's impact varies across a region, connecting broad patterns in low-resolution maps to high-resolution details.
Infrastructure resilience testing applications
- ▪The η-learning method can generate plausible maps of a storm producing 300 millimeters of rainfall in New York City, where the highest recorded rainfall is 200 millimeters.
- ▪The generated maps from the η-learning algorithm can help cities test seawalls against unprecedented storm surges, evaluate power grids during heatwaves, or assess wildfire containment resources.
Supply chain vulnerability concerns
- ▪MIT Professor Themis Sapsis warned that a single extreme event can propagate through supply chains, energy markets, and food systems within weeks.
- ▪MIT Professor Themis Sapsis stated that global infrastructure has been optimized for efficiency, leaving little slack or capacity to absorb extreme events.
Story comments
Loading comments…