Introduction

The Arctic has been experiencing pervasive changes in recent decades across its marine, coastal and terrestrial environments1, portrayed by the severe and rapid surface warming and sea-ice loss observed in the central Arctic2,3,4. Arctic warming is progressing at a pace much faster than in the rest of the planet, a phenomenon which has become widely known as Arctic amplification of climate change5. Recently a ‘new’ phenomenon has emerged in the European Arctic –the area encompassing the Fram Strait and Barents Sea–, where seawater properties have evolved from a typical condition of a stratified water column with a layer of cold and fresh Arctic waters confining warm and salty Atlantic waters at intermediate depths6, to a more homogenous water column composed by warmer and saltier waters resembling Atlantic properties7,8. This “Atlantification” of the European Arctic has been especially observed in the Barents Sea7, the West Spitsbergen Shelf and fjords9,10 and the area north of Svalbard11 and has far-reaching influences on sea-ice formation and vertical mixing, hence on the local circulation, productivity and ecosystems12,13,14.

Atlantification revolves around the following mechanisms. In the open-ocean and yearly ice-free areas, Atlantification has been explained by strengthened advective and diffusive heat transport convergence14,15 modulated by large-scale wind patterns16. In the marginal ice zones, it stems from weakened surface ocean heat losses due to atmospheric warming15, and reduced ocean stratification by weakened sea-ice advection13. Lateral transport of sea ice from the main Arctic basin toward the Barents Sea provides the lateral freshwater input necessary to keep the marginal ice zone in a cold and stratified state, dampening the upward heat and salt fluxes from the underlying Atlantic water layer17. In ice-covered areas, ocean heat advection and air-sea fluxes act together to drive a gradual warming of the upper ocean15. As a result of strengthened Atlantic water inflows, the area north of Svalbard and the Barents Sea is transitioning from an ice-covered to an ice-free region for longer periods during the year18,19,20, which can initiate further reinforcement of the warming through the sea ice-albedo feedback. Mechanisms of Atlantification thus intertwine with or may represent a regional expression of Arctic amplification mechanisms involving cryosphere and ocean5,21,22,23.

Characterization of the early development of Atlantification is limited by structured oceanographic observations and ocean reanalyses covering only the past few decades24. Available paleoceanographic reconstructions and paleoclimate model simulations describing the initial phases of the Arctic response to global warming consistently depict the exceptional character and an early onset of modern surface ocean warming in the European Arctic via enhanced poleward heat-conducive conditions towards the Nordic Seas25,26 and the Arctic21. However, studies disagree on magnitude and timing of European Arctic warming27,28,29, and only one study29 explicitly invokes Atlantification to explain oceanographic changes reconstructed in the eastern Fram Strait. It remains therefore unclear whether reconstructed and simulated changes are truly descriptive of an expansion of the Atlantic domain in the European Arctic, rather than the regional expression of pan-Arctic changes largely independent from Atlantification.

An ongoing acceleration of Atlantification30 and suggestions that it may continue in the future18,31 contribute to increasing scientific interest in this phenomenon. Atlantification, as pan-Arctic warming, is critically connected with changes in the stratification and vertical density distribution of the upper Arctic Ocean and changes in sea ice, which are known crucial factors controlling Arctic state and variability6,32,33,34. Both also express the natural exposure of the Arctic environment to Atlantic forcing21. This phenomenological superposition calls for a careful framing of Atlantification toward an unambiguous fingerprinting of its historical and potential future impacts. This need is further stressed by the lack of quantitative estimates for the current magnitude of Atlantification.

Hence, this study provides a coherent historical and future perspective on the Atlantification by using ORAS5 ocean reanalyses covering the past six decades and a multi-model ensemble of climate simulations extending from the preindustrial millennium to the end of the twenty-first century (see Methods). A clustering algorithm applied on upper-ocean thermohaline properties is used to divide the European Arctic into two domains –Atlantic and Arctic– and build an Atlantification index, ATLi, that quantifies changes in the areal extent of the Atlantic domain in the European Arctic (see Methods). Accordingly, an increase in ATLi values describes Atlantification, whereas a decrease describes Arctification, that is, the Arctic domain gaining area at the expense of the Atlantic domain. ATLi accounts for the concomitant relative state of the Atlantic and Arctic oceans, thus accounting for their vast historical and projected changes, and allows surpassing limitations imposed by the severe model biases in ocean temperature and salinity known in the region35,36. Our results show that modern Atlantification is rooted in late preindustrial times yet projected to stop after the 2050s due to an oceanic damping mechanism triggered by an increased density gradient between Atlantic and Arctic waters in the European Arctic.

Results

Modern Atlantification

The clustering algorithm applied to ORAS5 yields the current Atlantic domain to mainly encompass the southern side of the European Arctic, the West Spitsbergen area, the area close to the Barents Sea Opening and the southern Barents Sea (Fig. 1a,c,e,g, red lines), where the Norwegian Atlantic and West Spitsbergen currents flow (supplementary Figure S1). The position of the Atlantic-Arctic boundary is robust against different considerations of the vertical structure of the water column (see Methods). The multi-model ensemble captures the gross spatial features of the modern (1981–2010) Atlantic-Arctic boundary, especially in its eastern branch (Fig. 1a,c,e,g, blue shading). There, the S-shaped profile in MPI-ESM1-2-LR and IPSL-CM6A-LR reflects a realistic representation of the flow of the Norwegian Atlantic and Spitsbergen Polar currents, whereas the more linear profile in MRI-ESM-2-0 and MIROC-ES2L suggests a too weak circulation in the area. Models diverge more regarding the typical position of the Atlantic-Arctic boundary in the Fram Strait, reflecting uncertainties in the simulation of the West Spitsbergen current.

Fig. 1
figure 1

Atlantic-Arctic boundary position in reanalysis and climate simulations. (a,c,e,g) Expected Atlantic-Arctic boundary position in the modern (1981–2010) period for ORAS5 ocean reanalyses (red dots) and climate simulations (shading) from MPI-ESM1-2-LR (a), MRI-ESM-2-0 (c), MIROC-ES2L (e) and IPSL-CM6A-LR (g). (b,d,f,h) Change in the expected Atlantic-Arctic boundary position between modern and future (2071–2100, ssp585 scenario) periods simulated by MPI-ESM1-2-LR (b), MRI-ESM-2-0 (d), MIROC-ES2L (f) and IPSL-CM6A-LR (h). Red (blue) colors indicate areas where the front position is more (less) probable in the future compared to the modern period. Filled circles and crosses represent the position of the average sea-ice edge in March in the modern and future period, respectively.

Full size image

Mean state biases in the Atlantic-Arctic boundary are reflected in a mean state bias in the areal extent of the Atlantic domain in the European Arctic that is ubiquitous across models (Fig. 2a). Models overestimate ATLi from reanalysis especially in the Fram Strait, whereas distributions overlap in the Barents Sea region -most relevant for Atlantification- within the 1.5 interquartile range, except for IPSL-CM6A-LR (supplementary Figure S2). Notwithstanding mean state biases, simulated ATLi anomalies encompass reanalyses during the observational period (Fig. 3b), illustrating model-data consistency in the temporal evolution of Atlantification. ATLi trends (period 1958–2014) consistently quantifies the progress of Atlantification in ORAS5 (0.7–1.4%/decade) and in the ensemble mean (0.9–1.2%/decade), supporting that the forced component is predominant since the second half of the twentieth century.

Fig. 2
figure 2

Boxplots of single-model ensemble and reanalysis distributions of ATLi (a) and PAT (b) in various experiments. Shown are median (thick black line), interquartile range (IQR, boxes), 1.5*IQR (whiskers) and outliers (circles). Models are vertically aligned in panels a and b.

Full size image
Fig. 3
figure 3

Evolution and trends of Atlantification and Pan-Arctic warming in the historical and future scenario (ssp585) periods. Trend analysis of ATLi (a) and PAT (c) ensemble means. Trend coefficients with p-value > 0.05 are masked out with black dots. The black cone encompasses trends estimated for the historical period alone. Time series of ATLi (b) and PAT (d) for all considered models and realizations, as well as the multi-model ensemble mean (black line). The green line in panel (b) illustrates the ATLi time series from ORAS5. Data are anomalies referred to the 1981–2010 average.

Full size image

As expressed by pan-Arctic temperature, PAT (see Methods), the Arctic Ocean warms consistently across models and realizations during the historical period (Figs. 2, 3d). Ocean thermal inertia dampens high-frequency variability while highlighting multidecadal-to-centennial trends in PAT, which remain below 0.05 °C/decade in the ensemble mean during the historical period. Coherent decadal trends in ATLi and PAT indicate that Atlantification events historically occur during phases of pan-Arctic warming, while Arctification events occur during phases of pan-Arctic cooling (Fig. 3a,c). Multivariate regression of upper-ocean (0–200 m depth) temperature and salinity on ATLi and PAT supports that Atlantification and pan-Arctic warming constructively superposed in driving historical upper-ocean warming and salinification of the European Arctic, with comparable relative impacts (Fig. 4a,c, spatial fingerprints in supplementary Figures S3 and S4). The historical period thus features a tight coupling between local dynamics in the European Arctic and pan-Arctic properties.

Fig. 4
figure 4

Dependency of upper-ocean properties in the European Arctic on Atlantification and pan-Arctic warming. Box plots (line: median, box: interquartile range—IQR, whiskers: 1.5*IQR, + : outliers) are for local (grid-point) regression coefficients in the European Arctic domain: (a,b) upper-ocean potential temperature and (c,d) upper-ocean salinity on ATLi and PAT calculated for the (a,c) historical and (b,d) ssp585 experiments. Statistical significance, tested by means of a Mann–Whitney U-test, is reported in the stars above the plot as follows: *0.01 < p ≤ 0.05; **0.001 < p ≤ 0.01; ***10−4 < p ≤ 0.001; ****p ≤ 10−4.

Full size image

Consistently across simulations and reanalyses, the Atlantic-Arctic boundary approximates the winter sea-ice edge but remains more locally confined than the latter through the year (Fig. 1b,d,f,h and supplementary Figures S5 and S6) maintained by buoyancy fluxes. The Atlantic-Arctic boundary position is therefore not a mere expression of sea-ice dynamics, although the latter contribute to explain the diversity of simulated behaviors.

Projected future evolution of Atlantification

The ensemble coherence extends beyond the observational period, as the models yield a strongly consistent temporal evolution of the Atlantification across models and realizations over the whole historical and ssp585/ssp245 experiments, independent of the future scenario considered. Following the initial stationary phase with lack of interdecadal trends from 1850 to the first quarter of the twentieth century, the ensemble-mean evolution of ATLi exhibits a progressive acceleration culminating around the mid-twenty-first century (Fig. 3b, ssp245 in supplementary Figure S7). The Atlantic maximizes its areal gain around 2060, followed by a stationary phase until 2100 characterized by a northward displacement of the eastern branch of the Atlantic-Arctic boundary (Fig. 1b,d,f,h). This sustained Atlantification occurs across all models, realizations and scenarios (Fig. 2a), thus representing a forced response under global warming. Under ssp585 conditions, MIROC-ES2L and MRI-ESM-2–0 yield similar peak ATLi values (0.45–0.50) but achieved through a small and homogeneous progression along the modern position in MIROC-ES2L, and a stronger progression along the northern branch of the Atlantic flow in MRI-ESM-2–0 (Fig. 1d,f). In IPSL-CM6A-LR the Atlantic domain exhibits a geographically incoherent progression yielding an areal gain of only about 5% (Fig. 1h). MPI-ESM1-2-LR yields the strongest future Atlantification (ATLi exceeding 0.50), associated with an advancement of the Atlantic similar to MRI-ESM-2-0 but extending further north (Fig. 1b). The simulated Atlantic-Arctic boundary evolution is even more diverse across models in the Fram Strait, further contributing to the ensemble spread of projected future Atlantification.

The strong projected (ssp585) sea-ice loss results in the European Arctic facing absence of sea ice in late summer in the second half of the twenty-first century and a substantial retreat of the sea-ice edge in winter at the end of the century, when in the Barents Sea it is confined to above around 80°N, that is, much north of the projected Atlantic-Arctic boundary position (Fig. 1b,d,f,h and supplementary Figure S8). This spatial detachment suggests that in a warmer world different mechanisms govern Arctic sea-ice dynamics and the Atlantic-Arctic boundary. The strong sea-ice melting can be interpreted within the framework of pan-Arctic warming. As expressed by PAT, Arctic warming accelerates around 2010 up to exceeding 0.1 °C/decade by 2030 and 0.2 °C/decade by the end of the twenty-first century. Despite the diversity of biases in the modern PAT climatology, the ensemble agrees on a future pan-Arctic warming of about 1 °C under ssp585 and 0.7 °C under ssp245 (Fig. 2b), suggesting that the destabilizing feedbacks that determine Arctic amplification are consistently still operative at the end of the twenty-first century under both scenarios. The rate of Atlantification increases since 2010 as pan-Arctic warming accelerates, until Atlantification detaches from pan-Arctic warming around 2060 (Fig. 3b,d). Pan-Arctic warming becomes predominant in driving an overall upper-ocean warming and freshening of the European Arctic (Fig. 4b,d), while Atlantification remains a significant contributor to temperature changes only in the eastern and north-eastern region (supplementary Figures S9 and S10). Therefore, Atlantification/Arctification aligns with pan-Arctic warming/cooling in the historical period, but both phenomena decouple under sustained global warming.

Natural roots of Atlantification

Preindustrial experiments under transient (past1000/past2k) and unperturbed (piControl) conditions allow framing modern and projected Atlantification, and its relation with pan-Arctic warming, in the context of natural climate variability. All models agree that Atlantification is already distinguishable from natural variability: ATLi values for the modern period are consistently significantly higher (pranksum < 10–5) than transient late preindustrial (1851–1880) and unperturbed preindustrial values (Fig. 2a). However, modern Atlantification does not fully exceed the range of natural variability, as early-21st-century ATLi values remain occasionally below the 95th percentile of piControl values depicting intrinsic variability (supplementary Figure S11). This is expected to change under continued global warming, where future distributions for both ssp585 and ssp245 only overlap with intrinsic variability above its 95th percentile, except for IPSL-CM6A-LR. Accordingly, the exceptional character of modern Atlantification will be fully revealed in the next decades. Further pointing to the exceptional contribution of external forcing to modern Atlantification, non-smoothed ensemble-mean values since 2007 exceed previous maxima during the whole preindustrial millennium and ensemble-mean trends in the historical period exceed preindustrial trends on all time scales (Fig. 5a,b).

Fig. 5
figure 5

Atlantification and Pan-Arctic warming in the climatic context of the last millennium. Trend analysis of ATLi (a) and PAT (c) ensemble means. Trend coefficients with p-value < 0.05 are masked out with a black dot. The black line cone encompasses trends estimated for the historical period alone. Time series of ATLi (b) and PAT (d) for all considered models, as well as ensemble mean (black line), for the historical and parent past1000/past2k experiments. All series are anomalies with respect to the period 850–2014, smoothed with a 15-year moving average. For IPSL-CM6A-LR, spurious trends of 0.002 1/century for ATLi and 0.04 °C/century for PAT are preliminarily removed from the data.

Full size image

In both individual simulations and the ensemble mean, preindustrial variability includes recurrent potent Atlantification and Arctification events, with duration and intensity comparable to modern Atlantification, revealing an almost symmetrical natural pattern of variability of the Atlantic-Arctic boundary around its natural climatological position (Fig. 5a,b and supplementary Figure S11). In all models, the ratio of ATLi standard deviations calculated for the past1000/past2k and piControl experiments only slightly exceeds the value of one (MPI-ESM1-2-LR: 1.12; MRI-ESM-2-0: 1.04; MIROC-ES-2L: 1.07; IPSL-CM6A-LR: 1.24). Arctification and Atlantification therefore pertain to the intrinsic features of climate variability, and this intrinsic component remains predominant for transient natural variability of the Atlantic-Arctic boundary. External forcing thus appears to primarily exert its role by setting the phase of variability generated by intrinsic dynamics.

A major Arctification episode is coherently simulated across models in the early nineteenth century (Fig. 5b), arguably as a forced response to the 1809-Tambora volcanic cluster37. The beginning of modern Atlantification is rooted in the first half of the nineteenth century, if it also encompasses the recovery from an earlier Arctification episode. Spatial patterns further support a natural origin of modern Atlantification: the ensemble yields a coherent progression of the Atlantic-Arctic boundary across the Barents Sea from late preindustrial to modern times, whereas models largely disagree on changes in the Fram Strait (supplementary Figure S12), similar to what is obtained for projected future changes (Fig. 1b,d,f,h).

Consistent with indications from the historical experiment, the preindustrial period features a recurrent coexistence between Arctic warming/cooling episodes and Atlantification/Arctification throughout the past1000/past2k experiment (Fig. 5c,d). Global warming scenarios therefore project a disruption of a coherent natural behavior coupling Atlantification/Arctification and pan-Arctic warming/cooling that lasted on a millennial time scale.

Dynamics of limited future Atlantification

Consistently across ensemble members, upper-ocean warming accelerates in the European Arctic during the late twentieth century while salinity decreases starting from the early twenty-first century, both changes contributing to a decrease in upper-ocean density (Fig. 6a,b,c). This general tendency masks a differential upper-ocean warming in the Atlantic and Arctic domains: except for IPSL-CM6A-LR, the average seawater temperature gradient (Atlantic minus Arctic) increases substantially from the 1970s until peaking in the 2050s, indicating that the Atlantic domain warms at a faster rate compared to the Arctic domain (Fig. 6d). The co-occurrence of this peak with sea-ice free late summers in the European Arctic (supplementary Figure S8) points toward a potential central role of sea-ice dynamics. For instance, sea-ice albedo and ocean latent heat loss by the ocean due to sea-ice melting contribute to slower warming of the Arctic domain. From the 2050s onwards, when sea ice retreats north of the European Arctic (supplementary Figure S8), the Atlantic-Arctic temperature gradient weakens in all models, indicating that the Arctic domain warms faster than the Atlantic domain. The Atlantic-Arctic gradient in salinity slightly increases in the historical experiment, but trajectories diverge in ssp585, yielding near-zero ensemble-mean projected future trends (Fig. 6e). Consistently across ensemble members, the Atlantic-Arctic gradient in density remains stable until around 2020 and then becomes more negative –it strengthens– until around 2050, when it stabilizes again (Fig. 6f).

Fig. 6
figure 6

Simulated evolution of oceanic conditions in the European Arctic in the historical and future scenario (ssp585) periods. Anomaly of the average potential temperature (a), salinity (b) and potential density (c) in the European Arctic. Anomaly of the differences in potential temperature (d), salinity (e) and potential density (f) between the Atlantic and the Arctic domain (Atlantic minus Arctic). Anomalies are calculated by subtracting the average value in the 1981–2010 period from the time series. Color code: MPI-ESM1-2-LR (orange), MRI-ESM-2-0 (red), MIROC-ES2L (blue), IPSL-CM6A-LR (purple), ensemble mean (black).

Full size image

Therefore, the density decrease of Arctic waters is consistently slower than Atlantic waters in the ensemble, despite model specificities in the relative contributions of salinity and temperature changes. The spatial patterns of end-of-21st-century upper-ocean density anomalies further reveal ensemble consistent features and model specificities (Fig. 7a,b,c,d, shading). All models agree on seawater density decrease near eastern Svalbard, where the Spitsbergen Polar Current typically flows southward. Models disagree instead on the density decrease in the Central Bank and surrounding areas, which traces back on the diversity of simulated changes in the Atlantic-Arctic boundary. MPI-ESM1-2-LR and MIROC-ES2L yield the most consistent anomalous density gradients across the Atlantic-Arctic boundary in the Barents Sea, seen as alignment between the boundary position and the anomalous isopycnals. In IPSL-CM6A-LR Atlantic-Arctic gradients are weakest in the Central Bank, whereas they even have an opposite sign, i.e., negative Atlantic-Arctic anomalous gradient, in the southeastern basin.

Fig. 7
figure 7

Projected ocean circulation changes contrasting Atlantification. Upper-ocean density (color maps) and baroclinic pressure gradient (arrows) differences between the averages calculated over the future (2071-2100, ssp585 experiment) and modern (1981–2010) periods for (a) MPI-ESM1-2-LR, (b) MRI-ESM-2-0, (c) MIROC-ES2L and (d) IPSL-CM6A-LR. For each model, the spatial average of the density difference is removed before mapping, amounting to (a) −0.52 kg/m3, (b) −1.32 kg/m3, (c) −0.78 kg/m3 and (d) −0.71 kg/m3, hence an upper-ocean density decrease prevails over the European Arctic. Thick black lines correspond to the Atlantic-Arctic boundary position calculated for the future period.

Full size image

These changes reflect altered ocean–atmosphere interactions in the region: the anomalous density pattern superposes in the Barents Sea on winter wind-stress-curl anomalies, especially in MPI-ESM1-2-LR and MRI-ESM2-0 (supplementary Figure S13); changes in the winter net surface freshwater fluxes also contribute to the density anomaly, consistently following across models a strengthening of their historical climatological pattern of precipitation minus evaporation (supplementary Figure S14). The retreat of the sea ice edge further contributes to enhanced surface ocean turbulent heat losses in the Arctic portion of the Barents Sea (supplementary Figure S15). The overall picture is that local atmosphere–ocean interactions and local wind forcing substantially contribute to the density gradient between Atlantic and Arctic waters in the European Arctic, where the effect is mediated by sea ice changes.

The density anomalies are associated with a baroclinic pressure gradient force (Fig. 7 a,b,c,d, arrows). Besides model specificities –e.g., the southeastern basin in IPSL-CM6A-LR as mentioned above–, the ensemble consistently yields anomalous baroclinic pressure gradient force vectors in the area near the Atlantic-Arctic boundary that point south-westward, that is towards the Atlantic domain, which describes a countercurrent that contrasts Atlantification. Models lose coherence in strength and direction of the baroclinic pressure gradient force away from the Atlantic-Arctic boundary, which supports the interpretation that the dynamics described above pertain to a regional oceanic mechanism.

Large-scale oceanic precursors of the density anomalies are unclear. The ensemble yields a consistent 21st-century progressive weakening of the Atlantic meridional overturning circulation in its historical climatological core at tropical northern latitudes, but also stationary conditions or even temporary strengthening in its shallow polar component until around 2050 (supplementary Figure S16). In MPI-ESM1-2-LR these changes are reflected in latitudinal divergent trends in the meridional ocean heat transport driven by the overturning circulation, with reduction at tropical/mid latitudes and increase at subpolar/polar latitudes (supplementary Figure S17). These changes dominate the total meridional ocean heat transport and are associated with a shrinking and north-eastward intensification of the subpolar gyre with respect to its historical climatology (supplementary Figure S18a). The subpolar gyre also strengthens in MRI-ESM-2-0, but with different magnitude and spatial pattern compared to MPI-ESM1-2-LR, whereas IPSL-CM6A-LR yields a slowdown of the subpolar gyre (supplementary Figure S18b,c).

Discussion

Our quantitative assessment of modern Atlantification aligns with recent indications regarding its recent acceleration30 and possible late-preindustrial root29. Further, it reveals that modern Atlantification has a predominant forced component but will be hampered by an oceanic mechanisms triggered by the repercussions of sustained global warming in the Barents Sea region, where the Atlantic-Arctic density gradient strengthens progressively up to trigger an Atlantic-ward current that opposes Atlantification. This contrasts with previous indications that Atlantification will continue in the future18,31, which we explain as previous assessments not fully disentangling Atlantification from the local effects of pan-Arctic warming and neglecting the dynamical coevolution of Atlantic and Arctic seawater properties. Improved constraints on the future evolution of Atlantification can benefit from addressing the following sources of uncertainty.

First, simulated historical evolutions and projected changes are affected by climate model biases influencing the hydrography of the Arctic and Atlantic sub-polar regions35,36,38,39. Model consistency therefore partly reflects shared biases, which remains an enduring issue. Besides recent use of similar ensembles37,40, we stress the diversity of model behaviors in our ensemble: model coherency regarding the projected future dampening of Atlantification comes with model diversity regarding onset, magnitude and duration of the modern Atlantification, relative roles of seawater temperature and salinity for density changes, and different strengths of pan-Arctic warming seen in the diversity of projected future trajectories of Arctic seawater temperatures (Fig. 2c). Ultimately, this model diversity reflects critical differences in equilibrium climate sensitivities. IPSL-CM6A-LR, with the strongest climate sensitivity41, pan-Arctic warming and sea-ice loss in the Barents Sea (supplementary Figure S8), yields the strongest damping of Atlantification. This aligns with our general conclusion that Atlantification is damped under sustained global warming and that it is the relative states of Atlantic and Arctic oceans that matter for the underlying mechanism.

We stressed the potential role of air-sea interactions mediated by sea-ice melting for the increased Atlantic-Arctic density gradient and the consequent baroclinic adjustment in the Barents Sea, based on the consistent ensemble co-evolution between density gradient, surface freshwater and turbulent heat fluxes and sea-ice cover anomalies, and on basic physical arguments such as albedo and latent heat of melting effects. Differential lateral heat advection18,21, partly triggered or at least sustained by changes in the large-scale North Atlantic ocean circulation, may be also part of the mechanism: the strengths of the gyre and overturning circulations can affect Atlantification by regulating the heat transport towards the Arctic21 while responding to Atlantification through, e.g., changes in local mechanisms of dense water formation42. The overturning and gyre responses diagnosed here for MPI-ESM1-2-LR are consistent with the model oceanic fingerprint of a warming climate43 and contribute to enhanced poleward heat transport at high latitudes. However, whereas the multi-model ensemble agrees on the general overturning response, substantial discrepancies affect the subpolar gyre evolution. Also, the role of atmospheric heat transports remains unknown. This complexity and model diversity prevents pinpointing a remote origin of our proposed mechanism, further stressing the need to better understand the intertwining of the local ocean circulation, winds and sea-ice dynamics18 with the large-scale coupled gyre and overturning circulations44 to improve constraints on the future evolution of Atlantification.

Remaining open questions concern the impacts of Atlantification beyond the European Arctic considered here12. Fingerprinting impacts in northern basins, such as the Laptev Sea or the Amundsen Basin, may be critical to fully unravel the future of Atlantification. Then, understanding the natural functioning of Arctification and Atlantification expressed in transient and unperturbed preindustrial simulations bridges with the scientific challenge of robustly reconstructing Atlantification before the observational period using proxy data. We remark the broad symmetry between Arctification and Atlantification in the evolution of our Atlantification index in preindustrial times, but more details are needed to support the suggestion that both phenomena pertain to the same underlying mechanism.

In conclusion, this study sheds light on the strong historical relevance of the Atlantification in terms of both magnitude and duration of its effects in the European Arctic, and reveals how its future development may be short lived under continued warming of the planet. As demonstrated here for Atlantification, this reminds us that emerging or changing Earth system phenomena may be an impactful yet temporary manifestation of competing dynamics and feedbacks during certain phases of the transient climate response. This, in turn, requires a robust assessment not only of expected future but also remote past evolutions.

Methods

Data

Seawater temperature and salinity are retrieved from the Ocean Reanalysis System 5 (ORAS5) dataset45, which provides global ocean and sea-ice reanalysis monthly-mean data prepared by the European Centre for Medium-Range Weather Forecasts OCEAN5 ocean analysis-reanalysis system46. ORAS5 performs among the best ocean reanalyses regarding Arctic thermohaline properties, especially sea-surface salinity and its seasonality24,47,48. ORAS5 data covering the period 1957–2014 are used to characterize modern Atlantification and define an observational Atlantification index (ATLi), defined below, that quantitatively describes the evolution of the phenomenon in recent decades. The data also serves as an observational reference for model evaluation. The ORAS5 grid (approximate horizontal resolution of 12 km in the study domain) is preliminarily interpolated on a regular grid with nominal 0.5° resolution to align with the typical model resolution.

This study leverages climate model data sourced from the Coupled Model Intercomparison Project Phase 6, CMIP649, and the Paleoclimate Model Intercomparison Project Phase 4, PMIP450, initiatives. Four-dimensional fields of seawater potential temperature (CMIP6 output variable name: thetao) and seawater salinity (so) are retrieved and preprocessed by first interpolating the data at selected depths (from 10 to 190 m depth, every 10 m) and then vertically averaging to yield three-dimensional fields, which are finally averaged in time at a yearly frequency. The chosen thickness of the upper-ocean layer encompasses depths where Arctic and Atlantic waters are known to interact in the region13,14. We tested other thicknesses in the range 10–300 m depth without obtaining significant changes in the results (not shown). Sea-ice area fraction (sic) data are used to examine the Arctic sea-ice edge.

The main analysis focuses on the 1850–2100 period covered by the historical and shared socio-economic pathways (ssp) scenario experiments, including ssp585, which yields a sustained warming signal throughout the twenty-first century, and the milder ssp245 currently considered to provide for more realistic projected climates51. In addition, past Atlantification events are analyzed through available PMIP4 past1000 and past2k simulations, including their continuation into the historical period described by their corresponding historical simulations.

Four models provide output for the past1000/past2k experiment and therefore contribute to our multi-model ensemble: the CMIP6 version of the Institut Pierre-Simon Laplace coupled atmosphere–ocean general circulation model (IPSL-CM6A-LR)52, the Model for Interdisciplinary Research on Climate, Earth System version2 for Long-term simulations (MIROC-ES2L)53, the Max-Planck-Institute Earth System Model version 1.2 in its low-resolution version (MPI-ESM1-2-LR)54, and the Meteorological Research Institute Earth System Model Version 2.0 (MRI-ESM-2-0)55. Details regarding the models, experiments, and realizations employed in this study are outlined in supplementary Table S1. The models contributing to the past1000/past2k ensemble provide a coherent description of natural multidecadal oceanic37 and atmospheric40 variability in the North Atlantic sector. CMIP6 models are known to weakly represent the Atlantic water layer, which is typically found to be too deep and thick in simulations38,39.

In CMIP6 models, seasonal biases in Arctic sea-ice climatology trace back to biases in various processes, including atmospheric and oceanic circulation35 and sea-ice growth and melt fluxes36. Models in our ensemble consistently show a warm bias over the southern Norwegian Sea, largest in MPI-ESM1-2-LR, but they disagree on magnitude and sign of biases in the eastern part of the European Arctic, where, for instance, MRI-ESM-2-0 shows vanishing errors and MIROC-ES2L a cold bias over extensive areas (supplementary Figure S19). There is also a tendency of the Atlantic domain to be slightly saltier in models than in reanalyses, but the largest salinity biases are in the northern margins of the Barents Sea (supplementary Figure S20). Temperature-salinity profiles further depict model biases in the climatological vertical structure of the upper-ocean column in Atlantic and Arctic waters in the European Arctic (supplementary Figure S21a,b). The ensemble profiles encompass reanalyses for the modern period, typically within ranges of ± 1 °C for temperature and ± 0.2 PSU for salinity for subsurface and deeper layers for both Arctic and Atlantic waters. Especially in layers approaching 300 m of depth, however, no model overlaps with reanalyses meaning that systematic errors in individual models tend to cancel out in the ensemble mean but no individual model represents the vertical structure of Arctic and Atlantic water columns. MIROC-ES2L stands out for its low salinity waters in the Atlantic and around 200 m of depth in the Arctic, where the other models are too salty; IPSL-CM6A-LR stands out for its overly warm and fresh surface waters in the Arctic. The vertical structure of thermohaline gradients between Atlantic and Arctic waters is better represented with ensemble largely overlapping with reanalyses (supplementary Figure S21c), meaning that individual models better represent Atlantic-Arctic gradients than Atlantic and Arctic properties. We exploit this feature to calculate an index for Atlantification.

Methods

Cluster analysis to determine the Atlantic-Arctic boundary position

First, CMIP6/PMIP4 and ORAS5 native grids are geographically subsetted in the European Arctic area that we define as comprising the eastern Fram Strait and the Barents Sea (70–85°N latitude and 0–60°E longitude). For each of the gridded temperature and salinity datasets, two anomaly fields are calculated, defined as local deviations from two reference points, one in the Atlantic domain and one in the Arctic domain. For each year, each anomaly field therefore measures the distances of the local seawater properties from the concomitant state of either the Atlantic or the Arctic ocean. The four anomalous fields are fed into a K-means cluster analysis with two groups: the Atlantic domain and the Arctic domain. We therefore expect two water masses to be always present in the region. Since seawater properties of both domains undergo large spatial and temporal variations, the algorithm does not impose any a priori conditions on the two clusters regarding ranges of temperature and salinity, and spatial extent (see supplementary Figure S21d–f for projected temperature-salinity profiles for Arctic and Atlantic waters, and associated gradients). The two reference points used for calculating anomalous fields consistently fall within the Atlantic and Arctic. This outcome is drawn from an initial cluster analysis of absolute potential temperature and salinity fields conducted over all considered years.

The Atlantic-Arctic boundary position is the area encompassing Atlantic grid cells with at least one Arctic grid cell among their four neighbour grid cells. The climatological position of the Atlantic-Arctic boundary in the late preindustrial, modern and future periods is determined for each model in terms of the local ensemble-mean frequency of occurrence of the boundary within the European Arctic. For each grid cell in the domain, this is calculated by summing up the total annual occurrences of the front in the grid cell, across all realizations, divided by the product of realizations and years in the considered period. Expected changes in the climatological front position between different periods are calculated as the difference between their associated ensemble-mean frequencies of occurrence.

In its main application, the clustering neglects an explicit consideration of the vertical structure of water masses, a known important aspect of Atlantification33. However, in the region and depths relevant for Atlantification, the Atlantic domain is typically characterized by a nearly homogeneous column whereas the Arctic domain exhibits a significant vertical gradient in both temperature and salinity. Hovmoller diagrams of vertical profiles confirm that, for the reanalyses and simulations employed here, vertically-averaged upper-ocean temperature and salinity fields retain the significantly different three-dimensional information for the two domains (supplementary Figures S22 and S23 for MPI-ESM1-2-LR, results are similar for the other models). An extended cluster analysis that explicitly accounts for the vertical structure of the water column (0–200 m of depth) and encompassing various hypotheses about stratification confirms the robustness of our approach for both, reanalyses and climate simulations (supplementary Figures S24 and S25).

Indices

The Atlantification index, ATLi, is defined to overcome limitations by model biases. It is based on the cluster analysis described above, where the Atlantic and Arctic domains in the European Arctic are identified based on the relative similarity of local upper-ocean temperature and salinity with concomitant properties of Atlantic and Arctic water masses. ATLi is defined as the area fraction occupied by the Atlantic domain in the European Arctic. ATLi is computed for each of the considered years, realizations and experiments. The robustness of the index is supported by the climatological position of the Atlantic-Arctic boundary in ORAS5 (Fig. 1) largely agreeing with qualitative descriptions of the Arctic oceanic front position from existing literature13,15,18,56, notwithstanding margins of improvements in current reanalyses regarding Arctic circulation and associated water mass distributions24.

Pan-Arctic temperature (PAT) is used to quantify warming of the Arctic Ocean. PAT is calculated by volumetrically averaging potential temperature across all depths for oceanic grid points north of 70°N, where the latitudinal extent of the Arctic Ocean is chosen to include its smaller fringe seas ranges and the European Arctic.

The sea-ice edge is calculated as follows. For each experiment, the sea-ice areal fraction data are averaged across all considered realizations, and the presence of sea ice is then defined for those grid points where the sea-ice areal fraction exceeds 15%. For each year in the experiment, the sea-ice edge encompasses those grid cells featuring sea ice and having at least one ice-free grid cell among their four closest neighbouring cells.

Meridional ocean heat transports are calculated for MPI-ESM1-2-LR following21,43 as Σz Σx v T cp ρ dx dz, where v is the meridional velocity component, T is temperature, cp is specific heat capacity at constant pressure, ρ is density, and dz and dx represent, respectively, the integrals along depths and longitudes. The overturning (gyre) transport is associated with the zonal-mean (residual) component of the total meridional ocean heat transports.

Statistical analyses

A trend analysis is performed by calculating linear trends at all time steps using different time windows of width ranging from decadal (11 years) to bicentennial (199 years). The reference year for the local trend is the central year in the window and local trends affected by border effects are excluded.

Finally, we use a multivariate linear regression (Ordinary Least-Squares, OLS) model to disentangle the relative contributions of Atlantification and pan-Arctic warming to changes in upper-ocean temperature and salinity in the European Arctic. The multivariate linear regression is performed separately on temperature and salinity data for each realization in the historical and ssp585 experiments (full period) and for each grid point, using the standardized ATLi and PAT as predictors. The regression coefficients for ATLi e and PAT are then averaged across realizations for each model and experiment.