Scientists have developed a new way to estimate the risk of extremely rare weather events, including heat waves that might occur only once in 1,000 years.
The method combines artificial intelligence with traditional physics-based climate models. Researchers say it could provide reliable estimates of extreme weather risks while requiring only a fraction of the computing power normally needed.
The international study, involving researchers in the United States and France, was published in Physical Review Letters and co-led by scientists at the University of Chicago.
Modern weather forecasting has improved enormously, but predicting exceptionally rare events remains difficult.
Traditional weather and climate models use the laws of physics to simulate how the atmosphere behaves.
They can reproduce extreme conditions, but estimating the probability of a very rare event may require running huge numbers of simulations.
For example, if scientists want to estimate the chance of a fairly common hot summer day, they may encounter one after running only a modest number of simulations.
But if they want to study a devastating heat wave that occurs perhaps once in centuries, they may need thousands or tens of thousands of simulations before seeing enough examples to calculate the risk accurately.
That requires enormous amounts of computing time and energy.
AI weather models offer another option. They can make everyday forecasts extremely quickly, but they have an important weakness: they learn from existing data. If an exceptionally rare event was missing from their training data, AI may struggle to predict it accurately.
The researchers’ solution was to combine the strengths of both approaches.
Their method, called AI+RES, uses AI to help traditional climate models identify the conditions most likely to develop into extreme events.
It builds on a statistical technique known as rare event sampling. Instead of spending computing power on thousands of ordinary weather situations, this technique concentrates simulations on conditions that appear most likely to produce the rare event scientists want to study.
AI makes that selection process more effective, particularly for extreme events that develop over relatively short periods, such as a weeklong heat wave.
To test the approach, the researchers examined extreme heat over parts of France and the U.S. Midwest.
They first ran 50,000 simulations using a traditional climate model. They then tested their AI-assisted method.
Remarkably, AI+RES produced nearly identical estimates while using only about one-hundredth as many simulations.
That could dramatically reduce the computing resources required to calculate the likelihood of rare but dangerous weather.
The researchers focused on heat waves because extreme heat is one of the world’s deadliest weather hazards. Major European heat waves have caused tens of thousands of deaths, and rising global temperatures are increasing concerns about future extremes.
The initial experiment didn’t include the effects of climate change. The team now plans to test the approach under different warming scenarios to see how the probability of extreme events could change.
The same method could also be applied to other disasters, including tropical cyclones and extreme rainfall.
Eventually, the researchers hope their system could provide governments and communities with better estimates of the risks they face from exceptionally severe weather.
By combining AI’s speed with the reliability of physics, scientists may finally have a practical way to prepare for weather events so rare that history provides few clues about when—or how badly—they might strike.
