AI weather models cannot forecast what they never saw. A hybrid method aims at that gap.

An international team in the United States and France has published a hybrid forecasting method, AI+RES, in Physical Review Letters, aimed at the one thing machine-learning weather models reliably get wrong: events that were not in their training data ^[1][8]. The work was co-led by members of the Climate Extremes Theory and Data Group at the University of Chicago, run by associate professor of geophysical sciences Pedram Hassanzadeh ^[1].

The trade-off being attacked is familiar to anyone who has costed out a compute budget. Physics-based supercomputer models can produce genuine extremes, but need large amounts of time and energy ^[2]. AI models are fast and good at day-to-day forecasting, and then fall over on outliers ^[3]. Hassanzadeh puts it bluntly: AI weather and climate models are "one of the great achievements of AI in science, but they're not magical - they fail on gray swans, the rarest and most extreme events," while detailed physics-based models capture extremes at prohibitive cost ^[4].

The cost problem is a sampling problem. To estimate the odds that Chicago hits 90F (32C) in July you do not need many simulations, because it is not uncommon; to estimate the odds of 105F (41C) you need far more runs before the extreme shows up at all ^[5]. Rare event sampling (RES) reduces that bill by scoring conditions so the climate model spends its cycles only on the promising ones ^[6]. Its known weakness is duration: RES does not work well for short events such as a weeklong heat wave, as opposed to a whole season that runs unusually hot ^[7]. AI+RES replaces the scoring step with an AI prediction of which conditions lead to shorter, rapidly developing extremes ^[8]. Run iteratively, according to co-first author Alexander Wikner, a Schmidt AI in Science Postdoctoral Fellow in Hassanzadeh's group, you end up with a set of simulations that do capture the rare event, and the more of them you have the better the probability estimate ^[9].

The benchmark is the part worth noting. The team ran 50,000 simulations with a traditional climate model to predict heat waves over parts of France and the U.S. Midwest, and reports that AI+RES produced nearly identical results with one-hundredth as many simulations ^[10] - on the order of 500 runs ^[11].

The stakes are not abstract. The 2003 European heat wave was associated with roughly 70,000 deaths, and Russia's 2010 event with 56,000 ^[12]. This past June, nearly half of the United States, about 180 million people, saw dangerous temperatures ^[13].

Two caveats sit inside the paper's own framing. This was a proof of concept run on a model that does not account for climate change, and the researchers say they hope to test it under different warming scenarios ^[14]. And Wikner notes the method could be used to generate rare-event data sets to train better AI models, which would in turn speed up the method ^[15] - a loop that is attractive and also circular, since the AI doing the scoring would then be trained on synthetic extremes.

What to watch: whether the hundredfold reduction survives a model with climate forcing in it, whether AI+RES holds up on extremes other than heat waves, and whether anyone reproduces the France and Midwest result independently. Until then this is one benchmark on one variable, published by the group that designed the method.