The AMOC's 51% Problem: Ridge Regression Picks the Pessimists

For a generation of climate assessments, the question "how fast will the Atlantic's great ocean conveyor slow down this century?" came with an answer that felt more like a shrug than a forecast: somewhere between not very much and catastrophically. A new paper in Science Advances borrows a decades-old technique from high-dimensional statistics — ridge regression — and applies it to the Coupled Model Intercomparison Project's AMOC projections. The result is uncomfortable, precise, and roughly twice as bleak as the multimodel mean: a 51 ± 8% slowdown by 2100 under a middle-of-the-road emissions scenario, with 90% probability (Portmann et al., Science Advances, 2026).

The headline number matters less than the move that produced it. Instead of averaging dozens of climate models as if each one were equally trustworthy, the authors weighted each model by how closely it matched nineteen real-world ocean observations, then let a ridge-regularised regression carry the weights forward into the 2100 projection. The models that reproduce the recent South Atlantic ocean best are, it turns out, the ones that have been warning us most loudly.

Why This Matters Now

The Atlantic Meridional Overturning Circulation is the system that sends warm, salty water north toward Greenland, gives up its heat to the atmosphere over northwestern Europe, sinks in the Labrador and Nordic Seas, and returns south along the deep Atlantic as cold bottom water. It is one of the largest heat engines on the planet, currently moving on the order of seventeen sverdrups of water, and every sverdrup of that flow carries roughly 0.05 petawatts of poleward heat — an invariant the paper's authors underline in the body of the text itself (Portmann et al., 2026).

The IPCC's most recent best guess, derived from the CMIP6 multimodel ensemble under SSP2-4.5, was a 2100 weakening of 32% with a margin of error of ±37 percentage points. That is a central estimate of about a third, with an uncertainty window so wide it spans almost nothing happening to the circulation collapsing outright. Policymakers, coastal planners, and insurers have been asked to make real decisions against that range for years. The new paper replaces it with 51 ± 8%: still a forecast, but now one narrow enough to plan around.

A Direct Observation, Not a New Model

One temptation, on reading "51% weakening," is to imagine a new climate model running for a few simulated centuries and spitting out a more pessimistic number. That is not what happened. The ensemble is the same CMIP6 ensemble everyone else uses: thirty-two models, running SSP2-4.5, for the same window. What changed is how the projection is read out of that ensemble.

Portmann, Swingedouw, Khattab and Chavent — three oceanographers and a statistician at Université Bordeaux, with collaborators at CNRS and INRIA — treat the ensemble as a regression problem. Each model's simulated 2091–2100 AMOC strength is the target. The predictors are nineteen observable quantities that the same model also simulates for the historical period: the mean AMOC observed from 2005–2023 (where direct measurements exist), and the 1900–2020 means of sea surface temperature and sea surface salinity across nine distinct Atlantic, Indian and Equatorial Pacific subregions.

The model-by-model mapping from those nineteen historical observables to the 2100 AMOC is fit using ridge-regularised linear regression. Ridge regression is the workhorse of high-dimensional statistics: it fits a linear model while shrinking coefficients toward zero to prevent overfitting when predictors are correlated or when there are more predictors than data points. In climate constraint work, where models share ancestry and their historical fingerprints are heavily entangled, that shrinkage matters. To make sure the fit is not just memorising the ensemble, the authors use leave-one-out cross-validation: train on thirty-one models, predict the thirty-second, rotate, repeat. The technique is picked because, per the paper's own discussion, it delivers the lowest leave-one-out prediction error among the constraint strategies tested.

Plugging the actually observed historical values into the fitted mapping yields the constrained 2100 projection. That constrained estimate is 8.1 ± 1.4 sverdrups of 2091–2100 AMOC, compared with a multimodel-mean estimate of 12.0 ± 6.5 sverdrups from the raw ensemble. Expressed as a percentage reduction from the historical baseline, it is the now-famous 51 ± 8%.

The uncertainty reduction is arguably more consequential than the central estimate. Model uncertainty — the spread across CMIP6 projections — dominates the 2100 AMOC forecast, accounting for about 78% of total variance; scenario uncertainty, meaning the choice of emission pathway, contributes only about 14% (ocean2climate summary of Portmann et al.). Cutting the model-uncertainty component dramatically — roughly by four-fifths — is the real deliverable here. The 51 ± 8% range sits at the IPCC "very likely" threshold of 90% probability, a confidence level the unconstrained ensemble cannot reach.

Where the Models Went Wrong

If the ridge-regression procedure is effectively telling CMIP6 which of its members to trust, a fair follow-up is: what exactly are the low-weighted members getting wrong? The paper decomposes the adjustment into contributions from each observable and points at two culprits.

The larger is the South Atlantic surface salinity bias, which accounts for roughly 47% of the total correction from the multimodel mean to the constrained estimate (Above The Norm News summary). CMIP6 models generally simulate a South Atlantic that is too fresh — less salty than observations show. That matters because salt-advection is one of the feedbacks thought to govern AMOC stability: salty water moving northward is denser, sinks more readily, and helps sustain the overturning. A model that runs too fresh in the South Atlantic underestimates that feedback, which translates into an overturning that looks more stable under greenhouse forcing than the real ocean probably is.

The second largest contributor, about 36% of the correction, is a North Atlantic surface temperature bias: models that simulate the subpolar gyre as too cold tend to project an overturning that responds less to warming than the real ocean likely will. Together these two fingerprints — too fresh in the south, too cold in the north — account for a large majority of the gap between the unconstrained 32 ± 37% and the constrained 51 ± 8%.

The framing matters. The paper is not saying that CMIP6 is wrong in some global sense; CMIP6 remains the evidentiary backbone of climate science. It is saying that the spread inside CMIP6 is not random noise around a correct mean. It is structured, and it is structured along axes — South Atlantic salinity and North Atlantic temperature — where we have enough observations to tell the members apart. When we let the data do the sorting, the higher-weakening members are the ones that survive. As the lead author, Valentin Portmann, summarised it to CleanTechnica: "We found that the AMOC is going to decline more than expected compared to the average of all climate models" (CleanTechnica, 2026-04-16).

The Prior-Art Ladder

The Portmann paper does not arrive in a vacuum. It is the latest in a ladder of AMOC results that have, over roughly three years, tightened the community's view of circulation risk.

Ditlevsen & Ditlevsen (2023) used sea-surface-temperature fingerprints of the North Atlantic to argue for an early-warning signal of an impending AMOC tipping point, publishing in Nature Communications. That paper was controversial — its tipping date range was broad — but it reset the conversation about whether CMIP6's relative optimism was evidentially load-bearing.

Van Westen et al. (2024), also in Science Advances, complemented this with a physics-based early-warning analysis of AMOC stability in the CESM model, arguing that the circulation sits on a slow trajectory toward its tipping point even under moderate forcing (DOI 10.1126/sciadv.adk1189).

Bellomo et al. (2024) examined the state-dependence of the AMOC's weakening impacts in Geophysical Research Letters, emphasising that the consequences of a given sverdrup decline scale nonlinearly with baseline state — another reason the magnitude of the 2100 weakening matters beyond its headline number.

Portmann et al. contribute the observational-constraint lens to this ladder. Where Ditlevsen offered a fingerprint-based warning and van Westen a physics-based one, the new work is a statistical-constraint argument: inside the CMIP6 ensemble itself, the members that reproduce the ocean we actually live in are the pessimists.

What This Paper Is Not

Three disclaimers deserve to be made up front, because the coverage around the study has been inconsistent on them.

• It is not a prediction of collapse by a specific date. A 51 ± 8% weakening by 2100 is a magnitude claim, not a tipping-date claim. Potsdam's Stefan Rahmstorf, reacting to the paper in CleanTechnica, called it an "important and very concerning result" and argued separately that his own reading of the risk landscape leaves him more worried about a mid-century tipping point. That gloss is his, not the paper's.
• It is not an engineered tipping-point fingerprint. Ditlevsen and van Westen offered warning signals aimed at the collapse phase transition. Portmann et al. offer a statistical weighting of CMIP6's slow-weakening projections. The two approaches are complementary, not substitutes.
• It is not Greenland-inclusive. CMIP6's SSP2-4.5 runs do not fully represent accelerating Greenland meltwater discharge, and the paper's constrained projection inherits that blind spot. Multiple outlets have noted that including Greenland's freshwater flux plausibly pushes the projection further toward the high-weakening tail.

What the Rewrite Means Downstream

A 51% Atlantic overturning, if realised, is not a European problem, an African problem, or an American problem. It is all three, weighted differently.

For Europe and the subpolar North Atlantic, the loss of about half a petawatt of northward heat transport means less warm water reaching the eastern boundary of the Atlantic. The UK, Ireland, coastal Scandinavia and Iceland would see the background warming from greenhouse forcing partially offset — and in some winter months possibly overpowered — by reduced ocean heat delivery. A colder North Atlantic in a warmer world is the classic AMOC-weakening signal, and it does not show up uniformly; it shows up as regime shifts in storm tracks, marine heat-wave distributions and winter severity.

For the Sahel and tropical Atlantic, the story is different and arguably more humanitarian. AMOC weakening drives an equatorward shift of the Intertropical Convergence Zone — the band of rising air and heavy rainfall that tracks the thermal equator. A southward ITCZ shift dries the Sahel, a region home to hundreds of millions of people whose agriculture is already climate-stressed. CleanTechnica's coverage notes that roughly 400 million people live across Sahelian Africa; while any single headcount for a multi-country region carries its own uncertainty, the order of magnitude underlines why the Atlantic's northward heat transport is a global-south story as well as a European one.

For the US East Coast, AMOC weakening raises local sea level beyond the global mean. The overturning's Gulf Stream component produces a dynamic sea-surface slope; weaken it, and the slope relaxes, allowing sea level along the eastern seaboard to rise more than the global-average figure implies. This is an additive risk on top of thermal expansion and ice-sheet contributions, and its magnitude scales with the magnitude of the slowdown.

The paper's authors close their own discussion by noting, with characteristic restraint, that a substantial weakening of this size would carry adaptation implications around the Atlantic and in teleconnected regions (Portmann et al., 2026). Translating that into policy language: the AMOC is no longer an outlier scenario one hedges against; it is a central-case planning constraint under moderate emissions.

A Methodological Note Worth Pausing On

There is an interesting second-order story here about climate methodology. Observational constraint — using real-world data to downweight bad models — is not new. It has been applied to equilibrium climate sensitivity, Arctic sea ice, tropical precipitation, and more. What is unusual is the form of the constraint: nineteen predictors fed into a ridge-regularised regression, with leave-one-out validation, treated as a high-dimensional statistical problem rather than a hand-picked one-line fingerprint.

In climate science that kind of multi-observable, regularised approach has been rare, partly because of domain preference for physically interpretable constraints and partly because the number of independent ensemble members (tens, not thousands) makes classical machine-learning tooling feel oversized. The Portmann paper's contribution is partly to show the technique works on an ensemble this size — that its cross-validation error beats simpler single-observable constraints — and partly to demonstrate that the technique picks up physically interpretable signals anyway. Salinity and temperature biases, both well-known problem areas in CMIP6, emerge as the dominant drivers of the weighting. The machine-learning surface layer, in other words, is pointing at the physics underneath, not replacing it.

If this style of constraint travels well beyond AMOC, the next five years of IPCC-relevant assessments may routinely look like this: a ridge or lasso regression over historical observables, leave-one-out validated, delivering narrower 2100 ranges on the quantities policymakers actually have to plan around.

What This Does Not Tell Us — Yet

Five honest open questions the paper does not claim to answer, ordered roughly by priority for downstream planners.

1. Whether the 51% projection is itself still optimistic. With Greenland freshwater underrepresented in SSP2-4.5 runs, the constrained central estimate is plausibly a floor rather than a middle. The paper is transparent about this boundary condition; downstream coverage has sometimes elided it.
2. When the weakening is most rapid. A 51% decline by 2100 can occur through a slow linear drift or a compressed nonlinear slump. The constraint method projects an endpoint, not a trajectory. Decadal planners — insurers, port authorities, fisheries managers — need the trajectory.
3. Whether the South Atlantic salinity bias is fixable. If CMIP7 narrows the fresh bias, the unconstrained multimodel mean should move closer to the 51% figure on its own. That is a testable prediction of the Portmann framework.
4. How the constraint behaves across non-AMOC metrics. Ridge regression applied to ocean observables might reweight models for AMOC while leaving other diagnostics — sea-ice, ENSO, regional precipitation — unaffected or even mis-affected. The generality of the weighting is an open empirical question.
5. Where the tipping threshold sits, if there is one. The Portmann paper is silent on whether 51% weakening puts the circulation dangerously close to a bifurcation. Physics-based work (van Westen, Ditlevsen) has views; the constraint paper does not. Magnitude is now well-constrained; stability is not.

Key Takeaways

• A new Science Advances paper uses ridge-regularised regression, trained on nineteen ocean observables and leave-one-out cross-validated, to re-weight CMIP6's thirty-two AMOC projections. The constrained 2100 estimate under SSP2-4.5 is a weakening of 51 ± 8%, with 90% probability.
• The central finding narrows uncertainty from the IPCC's ±37 percentage points to ±8 — roughly a four-fifths reduction in model-spread uncertainty, at a confidence level IPCC labels "very likely."
• The correction is driven largely by two biases well known inside CMIP6: a fresh bias in the South Atlantic (47% of the adjustment) and a cold bias in the North Atlantic (36%). The methodologically novel regression ends up pointing at physics the community already had reason to distrust.
• Prior work by Ditlevsen, van Westen and others approached the AMOC from tipping-point fingerprint and physics-based early-warning angles. Portmann et al. add an observational-constraint angle inside the existing ensemble — a complementary tool, not a replacement.
• For planners, the relevant shift is from "central estimate around one-third, uncertainty swamps everything" to "central estimate around one-half, uncertainty tight enough to plan against." Europe, the Sahel and the US East Coast inherit the downstream impacts differently, but they all inherit them.