Ocean Circulation That Heats Europe Is on Track to Lose Half Its Strength by 2100

The number that’s been in the reports, the assessments, the headlines for years now is 32. A 32 percent weakening of the Atlantic’s great overturning current by the end of this century, give or take. Uncertain, yes, but that was the multimodel consensus, the product of dozens of the world’s most sophisticated climate simulations all pointing, more or less, in the same direction. Thirty-two percent. Manageable, perhaps. Alarming, certainly, but within the range of things we might adapt to.

The new number is 51.

That shift, from 32 to 51, is the central finding of a study published this week in Science Advances. It comes from a team led by Valentin Portmann and colleagues who’ve spent considerable effort trying to answer a question that climate scientists have quietly fretted over for years: if the models disagree so wildly about what the Atlantic Meridional Overturning Circulation will do, which ones should we trust? Their answer, reached through a statistical approach borrowed from economics and health research rather than climate science, is unsettling. The models that best match real-world observations point to a substantially more dramatic slowdown than the average suggests. And the uncertainty, rather than being something to hide behind, has been quietly pointing us in the wrong direction.

The Current That Heats Half a Continent

The AMOC is, in the most literal sense, what makes northern Europe habitable. It works a bit like a conveyor belt, pulling warm salty water northward through the upper ocean, releasing that heat into the atmosphere somewhere over the North Atlantic, then watching the now-cold, dense water sink and return southward along the ocean floor. Without it, the British Isles would look rather more like Newfoundland in January. Ireland’s famous greenness would be rather less famous. London’s temperature, already bracing, would drop by several degrees.

That the current is slowing is not in dispute. The RAPID monitoring array, a series of sensors strung across the Atlantic at 26 degrees north, has been measuring it continuously since 2004, and the trend is visible in the data. What’s been in dispute is how much slower it will get, and how fast. This is where climate models diverge in ways that, frankly, you might not expect from the outside. The spread of projections across the major modelling centres is enormous, from just a few percent weakening to more than 70%, depending on the model. Ask a room of climate physicists to predict the AMOC’s state in 2100 and you’d expect them to at least roughly agree. They don’t. Mostly because of what goes on in the South Atlantic.

Here’s the problem, as Portmann’s team diagnose it. The CMIP6 models, the current generation of climate simulations used for IPCC assessments, tend to get the salt content of the South Atlantic surface waters systematically wrong. They simulate it as too fresh, too diluted. And that matters because the saltiness of the ocean in this region is a crucial control on the AMOC’s stability; it’s what drives a feedback loop that either sustains or erodes the current’s strength. Get the salt wrong and you get the future wrong. The models, in effect, are running the AMOC on better conditions than actually exist, making it more stable than it is, producing projections that are, on average, too optimistic.

An Unusual Tool for an Unusual Problem

Portmann and colleagues tried something relatively novel. Rather than simply averaging all the model projections (the standard approach, called a multimodel mean) or throwing their hands up at the disagreement, they used what are called observational constraint methods: techniques that take actual measurements of the current climate and use them to score the models, effectively asking which simulations best reproduce what we can directly observe. Better-performing models, in theory, should also do a better job of predicting the future.

They tested four different approaches. The winner, in terms of predictive accuracy assessed through a leave-one-out cross-validation, was ridge-regularized linear regression, a method drawn from fields accustomed to dealing with lots of correlated variables simultaneously. It’s rarely used in climate science; it’s the kind of thing an epidemiologist or economist might reach for. Applied here, using 19 observable variables spanning sea surface temperatures and salinities across nine ocean regions, it reduced the model uncertainty by 79 percent. The constrained estimate it produced: a 51 percent weakening of the AMOC by 2100 under a moderate emissions scenario, with an uncertainty range of just plus or minus 8 percent.

That 8 percent error bar, compared to the original 37 percent, is arguably as significant as the shift in the central estimate. It says something about how much of the disagreement between models was, in principle, resolvable using existing observations. The information was there. The question was whether anyone would look for it in the right way.

What 51 Percent Actually Means

The IPCC, in its most recent assessment, defined a “substantial weakening” of the AMOC as any slowdown exceeding 50 percent. Which means this constrained estimate lands right at that threshold, or just over it. In the constrained picture, the AMOC’s volume transport at the end of the century comes out at roughly 8 sverdrups (the unit used to measure ocean current strength), compared to about 16 or 17 today. That’s not a collapse, technically. But it’s very far from the relatively moderate slowdown that has sometimes featured in public discussions of climate risk.

The consequences ripple outward. A weakened AMOC is strongly connected to a southward shift in the intertropical convergence zone, the band of rainfall that circles the tropics. That shift, in turn, dries out the Sahel, the belt of land stretching across Africa from Senegal to Ethiopia, where agriculture and food security already hang by a thread in many areas. A 51 percent weakening rather than a 32 percent one is not a difference of degree. It’s a difference in the planning horizon, the severity of adaptation strategies, the kind of infrastructure you’d need to build or reroute or abandon.

Under higher emissions, the picture is worse still. The constrained estimate for the most aggressive scenario (SSP5-8.5, roughly “business as usual”) puts the weakening at nearly 58 percent. The paper also raises a possibility that researchers have flagged elsewhere, and that the model biases now identified lend new support to: the real AMOC may already be closer to a tipping point than the consensus figures imply. Climate models, by systematically overestimating the salinity of the South Atlantic and underestimating the temperature of the North Atlantic, have been building simulations in which the current is more robust than it probably is. The corrections are not small. Eighty-three percent of the adjustment in the constrained estimate comes from just two groups of variables, the South Atlantic salinity and the North Atlantic temperature, both biased in directions that made the AMOC look safer than the observations suggest.

Portmann’s team are careful about what they’re not claiming. The observational constraint methods they’ve used rest on assumptions: that models capture the right relationships between variables even if they get individual numbers wrong, that the ensemble of models spans the plausible range of outcomes, that observational noise is small compared to model spread. These are reasonable assumptions but not guaranteed ones. Missing from the analysis, by necessity, is Greenland melt, which is not consistently represented across the CMIP6 models but which could push the AMOC further and faster than any of these projections suggest.

What the study does, perhaps most valuably, is reframe the uncertainty. The problem was never that we didn’t know what the AMOC would do. The problem was that we were averaging over models with systematic biases and calling the result a best estimate. Strip out the bias, and the picture sharpens into something considerably less reassuring.

Frequently Asked Questions

What is the AMOC and why does it matter?

The Atlantic Meridional Overturning Circulation is a system of ocean currents that transports warm water northward through the upper Atlantic and cold water southward along the ocean floor. It functions as a major heat pump for the North Atlantic and northwestern Europe, and its strength influences rainfall patterns across much of the tropics and Sahel. A substantial slowdown would reshape climate across multiple continents.

Why did existing climate models underestimate the slowdown?

The study found two main biases: the models make the South Atlantic’s surface water too fresh (too low in salinity), which weakens a key feedback that destabilises the AMOC under warming. They also simulate the North Atlantic as slightly too cold, which makes the current appear more stable than observations suggest. Both biases push the models toward projecting a milder slowdown than the constrained analysis indicates.

What does a 51 percent weakening actually look like?

It means the volume of water transported by the AMOC would roughly halve relative to pre-industrial levels by the end of the century. The IPCC classifies anything above 50 percent as a “substantial weakening.” Effects would likely include a significant southward shift in tropical rainfall patterns, severe drying in parts of the Sahel, altered storm tracks, and noticeable cooling of northwestern Europe relative to what it would otherwise experience under warming.

Is this a collapse scenario?

Not technically. A collapse would mean the circulation shuts down entirely. A 51 percent weakening is, by the authors’ description, “substantial,” but short of that threshold. Whether the current could continue weakening toward collapse beyond 2100, or in response to factors not fully captured in current models such as Greenland ice loss, remains an open and genuinely urgent question.

What should be done differently in climate modelling as a result?

The authors argue that modelling groups should focus on correcting biases in South Atlantic salinity and North Atlantic temperature, since these are driving most of the divergence between models. They also suggest that climate scientists should use more sophisticated constraint methods, particularly ridge-regularised regression with multiple observed variables, rather than relying on simple model averages when the spread between simulations is this large.