Researchers at MIT have developed an AI system designed to forecast extreme weather scenarios that have never appeared in a region’s historical record.
The system, developed by mechanical engineering graduate student Kai Chang and Professor Themis Sapsis, can generate maps of statistically plausible extreme events while estimating their intensity, duration, and geographic impact.
The researchers call the method Extreme Event Aware, or η-learning. Their work was published in Nature Communications.
Looking beyond historical disasters
Traditional extreme-weather models often rely heavily on historical observations. If a model is trained on past storms, rainfall, floods, or other disasters, it can estimate the likelihood of similar events occurring again.
The problem is that the most severe events may occur so rarely that there are few examples available for an AI system to learn from.
The MIT researchers want to address that limitation.
For example, planners may have extensive data about previous hurricanes but still need to understand what an event substantially worse than anything previously observed could look like.
The new approach is intended to estimate those plausible but unprecedented scenarios rather than simply reproducing events already present in the data.
Combining statistics with spatial information
The system combines two different types of information.
The first is point statistics, which describe how frequently particular levels of intensity occur. In the case of rainfall, this could involve measuring how often extremely high precipitation levels appear in a dataset.
The second is spatial information, which describes how an event is distributed across a geographic area.
By learning the relationship between these two forms of information, the AI can generate spatial patterns representing extreme events that go beyond the specific examples contained in its training data.
This means the system does not necessarily need to have previously observed the exact event it is attempting to model.
Testing the system with rainfall
The researchers tested their approach using precipitation data from across the continental United States.
They began with 25 years of hourly rainfall observations and converted the information into daily maps. From this dataset, they calculated statistics describing the frequency of different rainfall intensities.
The researchers then deliberately restricted the spatial model’s training data. It was trained using low- and high-resolution rainfall maps from only the first six months of the 25-year dataset, a period that contained few or no examples of the most extreme rainfall levels.
The AI learned how detailed high-resolution rainfall patterns corresponded with lower-resolution information.
The broader statistical data was then used to constrain the extreme scenarios generated by the model.
Generating storms that have never happened
One example illustrates the potential of the technology.
New York City has recorded rainfall reaching approximately 200 millimetres. The MIT system can generate a statistically plausible scenario involving around 300 millimetres of rainfall, despite there being no equivalent event in the historical dataset.
A user could also ask the system to generate scenarios corresponding to a once-in-a-century event for a particular location.
Rather than producing a single prediction, the system can generate numerous possible scenarios. These can differ in their intensity, geographic coverage, and duration.
That could give planners a much broader picture of the risks they may need to prepare for.
Potential applications for infrastructure
The technology could eventually be useful for testing infrastructure against extreme scenarios that have never actually occurred.
For example, a city could use generated weather maps to examine how a seawall might perform during an unusually severe storm.
Power-grid operators could investigate the potential effects of an exceptionally long heatwave, while emergency services could examine how they might respond to a wildfire larger than anything previously recorded in the region.
The key advantage is the ability to explore events outside the boundaries of historical experience.
The technology still has limitations
The approach is not a universal weather simulator that can immediately predict every type of disaster.
According to the researchers, applying the method to a new hazard requires appropriate point statistics and spatial data for that particular hazard.
That means additional work would be necessary before applying the technique to hazards such as floods or wildfires in the same way it was demonstrated with precipitation.
The researchers nevertheless see broader potential as suitable datasets become available.
Preparing for risks that have never happened
Modern infrastructure is often designed around efficiency, leaving limited spare capacity when extreme events occur.
That can make unprecedented disasters particularly disruptive. A sufficiently severe event could affect not just one piece of infrastructure, but supply chains, energy markets, food systems, and other interconnected networks.
By generating statistically plausible scenarios outside the historical record, MIT’s approach could give governments, insurers, utilities, and other organizations another tool for stress-testing their systems.
The broader goal is not to predict exactly which unprecedented disaster will happen next. Instead, it is to provide a way to quantify and visualize extreme scenarios that historical data alone cannot adequately represent.
