The new approach could help governments, cities and infrastructure operators prepare for disasters that have never occurred in their region, including storms, floods and other weather events that could be more severe than anything previously recorded.

The system was developed by MIT mechanical engineering graduate student Kai Chang and Professor Themis Sapsis.

Their method, known as Extreme Event Aware or η-learning, is designed to overcome a major challenge facing conventional extreme-weather modelling.

Traditional risk models often rely on historical observations to estimate how frequently extreme events occur and what they might look like.

But the most dangerous disasters can be extremely rare. This means a region may have never experienced an event severe enough to properly prepare for the possibility of one occurring in the future.

The MIT researchers use a different approach. Their system combines statistical information about how often certain levels of extreme weather occur with maps showing how weather conditions are distributed across a region.

The researchers tested the method using 25 years of hourly rainfall data covering the continental United States.

They trained part of the model using only a relatively short portion of the dataset that contained few or no examples of the most extreme rainfall events.

Despite this limitation, the system was able to generate detailed maps of rainfall events beyond those contained in the training examples.

For example, the researchers demonstrated how the method could produce a plausible scenario involving 300 millimetres of rainfall in New York City, even though the highest recorded rainfall level used for comparison was 200 millimetres.

The potential applications extend beyond rainfall. According to the researchers, similar approaches could eventually help planners examine the possible effects of severe floods, wildfires, heatwaves and other hazards for which there are limited historical records.

For city authorities, the technology could provide a way to test whether infrastructure such as seawalls, electricity networks and emergency-response systems could withstand events more extreme than those previously experienced.

However, the researchers stress that the system does not predict that a specific extreme event will definitely happen.

Instead, it generates statistically plausible scenarios that can help decision-makers understand risks that may otherwise be difficult to visualise.