If New York’s worst recorded rainfall is 200 mm, what could a plausible 300-mm storm look like? MIT built an AI system to map extreme events that have never happened before

The method, called Extreme Event Aware, or “η-learning”, does not need previous extreme events to create these scenarios. (Representational AP photo)

What would a once-in-a-century storm look like if it was stronger than anything recorded before? Researchers at the Massachusetts Institute of Technology (MIT) have developed a machine-learning method that can generate realistic scenarios for extreme events that have not happened before.According to Massachusetts Institute of Technology, the new system can generate maps showing the likely size, intensity and duration of extreme weather events, even when similar events are missing from the data used to train the system. This can help weather experts issue early warnings and help disaster relief teams prepare.The new method is called Extreme Event Aware, or ‘η-learning’. It learns statistical patterns from available data and uses them to rule out unrealistic possibilities while generating events that could still plausibly happen.The research was published in an open-access paper in Nature Communications.

AI models unseen storms

Extreme weather events are difficult to study because they happen rarely. Existing methods of predictions often rely on records of past extreme events. To estimate the risk from a once-in-a-century storm, therefore, computer simulations may be trained using data that already contains major storms.MIT researchers instead wondered what could an extreme event look like if it was more severe than anything previously recorded, but still possible.The maps generated by AI show where an extreme storm could occur, how large an area it could cover and how intense the rainfall could be. The system learns from two kinds of information- point statistics and spatial maps. Point statistics describe how often rainfall reaches different levels across a region. On the other hand, Spatial maps show how rainfall is distributed across an area.To test the method, scientists used 25 years of hourly precipitation maps covering the continental United States. They combined the hourly data into daily maps and calculated statistics showing how often the maximum rainfall reached different levels.They then trained the algorithm using paired low- and high-resolution maps from only the first six months of the record. That period contained few or no examples of the most extreme rainfall levels.From this information, the algorithm learned how broad patterns in low-resolution maps were connected to detailed, high-resolution precipitation maps. It then used the rainfall statistics to keep the generated extreme events within statistically plausible limits.This allowed the researchers to create possible rainfall patterns that were more extreme than those contained in the training data.

Thousands of possible scenarios

A user can ask the trained system a question such as, “What could a once-in-a-century storm look like in New York City?” The algorithm can then produce maps of statistically plausible storms that could occur at that frequency. The maps will include details about the storm’s size, coverage and rainfall intensity.According to Kai Chang, an MIT graduate student in mechanical engineering and affiliate of the MIT Center for Computational Science and Engineering, the system could also be useful to people designing infrastructure to withstand rare events.“Someone can say, ‘I’m interested in building things to withstand the risk of an event that happens every 100 years. What we can do then is produce thousands of possible realizations that will happen with this sort of rare frequency.”

Uses beyond rainfall

Researchers say the approach could also be used to study other extreme events. Where suitable point statistics and spatial data are available, it could be applied to events such as extreme floods and wildfires.The method may also have applications beyond weather. MIT says the same approach could be used in areas such as robotic navigation and financial markets.“Financial market crashes are extreme events that are a complicated combination of things, involving many different sectors,” Chang said. “What is the interaction that leads to a market crash? That is something that this method could explore,” he added.According to Themis Sapsis, the William I Koch Professor of Mechanical and Ocean Engineering at MIT and an affiliate of the MIT Institute for Data, Systems, and Society, preparing for extreme events has become important beyond environmental planning because modern systems have little spare capacity.“Extreme events have become a strategic concern, not just an environmental one. We have optimized global systems for efficiency, and the price of that efficiency is that there’s very little slack left anywhere. A single extreme event propagates through supply chains, energy markets, and food systems in weeks,” Sapsis said.