Climate scientists can confidently tie global warming to impacts such as sea-level rise and extreme heat. But ask how rising temperatures will affect rainfall and storms, and the answers get a lot shakier. For a long time, researchers chalked the problem up to natural variability in wind patterns—the inherently unpredictable fluctuations of a chaotic atmosphere.
Now, however, a new analysis has found that the problem is not with the climate, it’s with the massive computer models designed to forecast its behavior. “The climate is much more predictable than we previously thought,” says Doug Smith, a climate scientist at the United Kingdom’s Met Office who led the 39-person effort published this week in Nature. But models don’t capture that predictability, which means they are unlikely to correctly predict the long-term changes that are most influenced by large-scale wind patterns: rainfall, drought, flooding, and extreme storms. “Obviously we need to solve it,” Smith says.
The study, which includes authors from several leading modeling centers, casts doubt on many forecasts of regional climate change, which are crucial for policymaking. It also means efforts to attribute specific weather events to global warming, now much in vogue, are rife with errors. “The whole thing is concerning,” says Isla Simpson, an atmospheric dynamicist and modeler at the National Center for Atmospheric Research, who was not involved in the study. “It could mean we’re not getting future climate projections right.”
The study does not cast doubt on forecasts of overall global warming, which is driven by human emissions of greenhouse gases. And it has a hopeful side: If models could be refined to capture the newfound predictability of winds and rains, they could be a boon for farming, flood management, and much else, says Laura Baker, a meteorologist at the University of Reading who was not involved in the study. “If you have reliable seasonal forecasts, that could make a big difference.”
The study stems from efforts at the Met Office to predict changes in the North Atlantic Oscillation (NAO), a large-scale wind pattern driven by the air pressure difference between Iceland and the Azores. The pressure difference reverses every few years, shunting the jet stream north or south; a more northerly jet stream drives warm, wet winters in northern Europe while drying out the continent’s south, and vice versa. In previous attempts to project the pattern decades into the future, a single model might yield opposite forecasts in different runs. The uncertainty seemed “huge and irreducible,” Smith says.
At first, the Met Office model did no better. But when the team ran the same model multiple times, with slightly different initial conditions, to forecast the NAO a season or a year into the future, a weak signal appeared in the ensemble average. Although it did not match the strength of the real NAO, it did match the overall pattern of its gyrations. But on individual model runs, the signal was drowning in noise.
The new work uses an ensemble of 169 model runs to find the same weak but predictable NAO pattern persisting for up to a decade. For each year since 1960, the team forecasted the NAO pattern 2 to 9 years in the future. When compared with weather records, the ensemble results showed the same pattern, ultimately explaining four-fifths of the NAO’s behavior. The massive computational effort suggests changes in the NAO are more predictable than models capture by an order of magnitude, Smith says. It also suggests individual models aren’t properly accounting for the ocean or atmospheric forces shaping the NAO.
The missed predictability appears to be universal. “This is being pursued everywhere,” says Yochanan Kushnir, a climate scientist at Columbia University, whose team reported last week in Scientific Reports that rainfall in the Sahel zone is more predictable than models indicate. In forthcoming work, a group led by Benjamin Kirtman, an atmospheric scientist and model developer at the University of Miami, will flag similar missed predictability in wind patterns above many of the world’s oceans.
Kirtman thinks something fundamental is wrong with the models’ code. For the time being, he says, “You’re probably making pretty profound mistakes in your climate change assessment” by relying on regional forecasts. For example, models predicted that the Horn of Africa, which is heavily influenced by Indian Ocean winds, would get wetter with climate change. But since the early 1990s, rains have plummeted and the region has dried.
The missing predictability also undermines so-called event attribution, which attempts to link extreme weather to climate change by using models to predict how sea surface warming is altering wind patterns. The changes in winds, in turn, affect the odds of extreme weather events, like hurricanes or floods. But the new work suggests “the probabilities they derive will probably not be correct,” Smith says.
What’s not clear yet is why climate models get circulation changes so wrong. One leading hypothesis is that the models fail to capture feedbacks into overall wind patterns from individual weather systems, called eddies. “Part of that eddy spectrum may simply be missing,” Smith says. Models do try to approximate the effects of eddies, but at just kilometers across, they are too small to simulate directly. The problem could also reflect poor rendering of the stratosphere, or of interactions between the ocean and atmosphere. “It’s fascinating,” says Jennifer Kay, a climate scientist at the University of Colorado, Boulder. “But there’s also a lot left unanswered.”
While researchers around the globe hunt down the missing predictability, Smith and his colleagues will take advantage of the weak NAO signal they have in hand. The Met Office and its partners announced this month they will produce temperature and precipitation forecasts looking 5 years ahead, and will use the NAO signal to help calibrate regional climate forecasts for Europe and elsewhere.
But until modelers figure out how to confidently forecast changes in the winds, Smith says, “We can’t take the models at face value.”