Artificial intelligence can now produce credible scenarios of extreme rainfall, even when no event that violent exists in its training data.
Preparing a city for a rare disaster poses a simple problem: examples are precisely lacking when the risk becomes most serious. By definition, exceptional rainfall appears rarely in historical records. Yet many models learn primarily from what has already happened.
Lightning during a storm near Mudanya in Turkey, on June 16.
Credit: Uğur İkizler
Researchers from the Massachusetts Institute of Technology have developed a method to overcome this limitation. Called Extreme Event Aware Learning, it does not seek to predict the next storm. Its goal is different: to imagine many very rare but statistically plausible events.
To test it, the team used 25 years of hourly rainfall data covering the continental United States. These data were compiled into daily maps. However, the part of the model responsible for learning the detailed shape of rainfall used only the first six months. This short period contained few, if any, of the most extreme episodes. This was done to test its predictions against reality.
How can it produce heavier rainfall than it has seen? The idea is to separate two kinds of information. On one hand, the system learns what rainfall maps look like: their shape, extent and spatial organization. On the other, statistics indicate how often certain intensities can occur.
These statistics act as safeguards. The AI cannot simply invent a spectacular map at random. It must propose configurations compatible with the requested level of rarity.
A planner could thus ask what rainfall likely to occur approximately once a century would look like. The model then generates many possible scenarios, with different affected areas and intensities. The researchers give the example of an event reaching 300 mm of rain where the observed record would be 200 mm.
This difference is important for risk management. A city could test a stormwater drainage network, a levee or other infrastructure against several severe scenarios. It would no longer be limited to disasters already present in historical records.
The researchers believe that a similar principle could be used for other risks, such as floods and wildfires. Very different fields, such as robotic navigation or financial market prices, could also be explored.