Climate models reveal 73 nonlinear surprises arriving sooner than expected

Climate scientists have long warned that the Earth system does not always respond to warming in smooth, predictable increments. Now, a systematic search through the world’s largest archive of climate model simulations has produced the most comprehensive inventory yet of the moments when the climate lurches. In a study published in Earth System Dynamics, Joran R. Angevaare and Sybren S. Drijfhout of the Royal Netherlands Meteorological Institute catalogued 73 instances of what they call Strong Nonlinear Surprises, or SNS, in future projections from the Coupled Model Intercomparison Project Phase 6, known as CMIP6. These are changes so large and so fast that they cannot be explained by the gradual rise in greenhouse gas forcing alone; they require powerful internal feedbacks within the ocean, sea ice, and atmosphere to take over and drive the system into a new state.

The scale of the analysis is formidable. The researchers examined 13,266 distinct datasets, spanning 54 climate models run under five shared socio-economic pathways, from the ambitious SSP119 scenario to the high-emissions SSP585. Twelve ocean, sea-ice, and atmospheric variables were scrutinised, including sea-surface temperature and salinity, sea-ice concentration, mixed-layer depth, sea-surface height, and the overturning streamfunctions that track the Atlantic Meridional Overturning Circulation, or AMOC. Crucially, the team replaced the expert visual inspection used in an earlier CMIP5 catalogue with a fully automated, reproducible detection algorithm, harnessing modern computing power and machine-learning tools such as the unsupervised clustering method HDBSCAN to isolate candidate regions of at least one million square kilometres.

The detection pipeline works in phases. Monthly model output is first averaged to yearly values and regridded onto a common Gaussian grid, and the historical runs are stitched to the scenario runs while checking that no artificial jump appears at the 2014 seam between them. Four different region-finding approaches then flag promising areas, using thresholds iteratively relaxed from the 99.99th percentile down to the 85th, and comparing variability against preindustrial control runs. Finally, six formal sets of criteria, labelled Ab1, Ab2, and St1 through St4, decide whether a region hosts an abrupt change on decadal timescales or a slower state transition too large to be attributed to the forcing alone. The criteria demand, for example, that a jump exceed 4.5 standard deviations of preindustrial variability, or that sea ice decline by at least 95 percent across an area of five million square kilometres.

The results are grouped into 11 categories, and their composition is telling. Roughly 45 percent of the surprises involve sea-ice cover, 29 percent mixed-layer depth, 19 percent ocean currents, and 7 percent atmospheric systems such as the Intertropical Convergence Zone. The single largest category concerns the transition to large, year-round ice-free areas in the Arctic, detected in 22 models. In the previous CMIP5 catalogue, only five models showed a winter sea-ice collapse, and only under the most extreme scenario, with thresholds between 4.5 and 8.2 degrees Celsius of global warming. In CMIP6, the corresponding warming levels fall to a 68 percent confidence interval of 2.5 to 4.1 degrees, meaning winter Arctic sea-ice disappearance could plausibly begin by the end of this century even under moderate emissions.

The physics behind these sea-ice surprises involves a well-understood web of positive feedbacks. As open water replaces ice, the lower atmosphere moistens and warms, increasing downward longwave radiation and inhibiting freezing. Thinner winter ice then lowers the surface albedo, boosting springtime shortwave absorption, while enhanced ocean heat uptake further erodes the ice pack. The catalogue also documents how sea-ice loss cascades into other variables: in several models, the disappearance of the ice lid unleashes abrupt shifts in sea-surface salinity, near-surface air temperature, and even sea-surface height, as the ocean and atmosphere suddenly exchange heat and freshwater far more vigorously than before.

Perhaps the most consequential finding concerns the North Atlantic. Ten models show a collapse of winter deep convection and mixed-layer depth in the subpolar gyre and Nordic Seas, often beginning around 2020 and in some cases even at warming levels near one degree. In nine models, this convective collapse is followed by a transition of the AMOC to an extremely weak state in which the cell associated with North Atlantic Deep Water formation has essentially vanished, leaving only a shallow, wind-driven residual overturning of about three Sverdrups. The onset of AMOC decline begins between 1980 and 2020 in these models, and the five-Sverdrup boundary is crossed between 2100 and 2140 in most of them. Strikingly, in the MRI-ESM2 model the collapse proceeds even under the low-emissions SSP126 scenario, a novel result suggesting the overturning may become largely independent of the forcing once a tipping point is passed.

The sequencing matters as much as the outcome. By comparing cumulative distributions of the warming level at which each change is maximal, the authors show that the mixed-layer collapse precedes the AMOC shutdown by roughly 0.3 to 0.5 degrees Celsius of global warming, equivalent to about twenty to thirty years. This is a clear example of a tipping point cascade: freshening and reduced heat loss stratify the subpolar North Atlantic, deep mixing fails, and the resulting buoyancy changes then starve the overturning of the dense water it needs. Notably, the freshening is driven by a weakening AMOC transporting less salty water northward, because meltwater from the Greenland Ice Sheet is not included in CMIP6 simulations, meaning the real-world freshwater forcing could be even stronger.

Not every surprise in the catalogue carries equal weight, and the authors are candid about model limitations. An abrupt shift in the North Atlantic Current pathway in the GISS-E2-1-H model appears tied to a known cold bias east of Newfoundland. A reorganisation of the Intertropical Convergence Zone in four models stems from the notorious double-ITCZ bias, though the underlying cloud-convection feedbacks exist in nature. A projected 50 percent intensification of the Antarctic Circumpolar Current is likely spurious, since coarse-resolution models cannot capture eddy saturation, which buffers the current against strengthening winds. By contrast, abrupt sea-ice loss in the Southern Ocean linked to the onset of deep convection, and the eastward expansion of the Ross Gyre, are corroborated by high-resolution regional modelling and observations of the Weddell Polynya mechanism.

To characterise timing, the team built probability density functions for the global warming level at which each surprise’s change is maximal, applying flexible 20-to-50-year windows to account for methodological uncertainty. These distributions show the frequency of surprises rising steeply below two degrees of warming and remaining high between two and four degrees. Importantly, the authors stress that these are conditional distributions describing when changes peak, not probabilities that a given event will occur, a deliberate choice to avoid politicised threshold debates. They also note that if a tipping point underlies a surprise, it is crossed before the temperature at which the transition is fastest.

The overall message is a sharpened warning. Compared with the 2015 CMIP5 catalogue, which identified 37 abrupt shifts, the new analysis finds nearly twice as many, occurring earlier and at lower warming levels, particularly for Arctic sea ice, North Atlantic convection, and the AMOC. Part of the increase reflects more extended simulations running to 2300 and improved detection methods, but the authors argue that genuinely higher sensitivity in CMIP6 also plays a role, and that observational constraints and the absence of Greenland meltwater both point toward underestimation rather than exaggeration of the risk. Because the criteria remain somewhat subjective and individual surprises can be stochastic, appearing in some ensemble members but not others, any single case demands caution. Yet the pattern across 54 models is difficult to dismiss: the newest generation of climate models suggests that strong nonlinear surprises, from an ice-free Arctic year-round to a shutdown of the Atlantic overturning, may arrive sooner, and at lower levels of warming, than the scientific community had previously assessed.

Subject of Research: Abrupt climate changes and state transitions in ocean, sea-ice, and atmospheric variables in CMIP6 future scenarios

Article Title: Catalogue of strong nonlinear surprises in ocean, sea-ice, and atmospheric variables in CMIP6

Article References: Angevaare, J. R., & Drijfhout, S. S. (2026). Catalogue of strong nonlinear surprises in ocean, sea-ice, and atmospheric variables in CMIP6. Earth System Dynamics, 17(4), 1081-1115. https://doi.org/10.5194/esd-17-1081-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1081-2026

Keywords: CMIP6, climate tipping points, Arctic sea ice, AMOC, mixed-layer collapse, abrupt climate change, North Atlantic, deep convection, climate models, nonlinear feedbacks, sea-ice albedo, Earth System Dynamics