Shifts from surface density compensation to projected warming, freshening and stronger stratification in the subpolar North Atlantic

Abstract

The hydrography and stratification of the subpolar North Atlantic is highly variable, with convection activating and deactivating across parts of the Labrador and Irminger seas. Likely consequential for the Atlantic Meridional Overturning Circulation (AMOC), this variability is examined in an eddy-rich ocean model hindcast spanning 1958–2021 and in 1950–2050 simulations with four climate models, spanning differences in ocean resolution (eddy-rich or eddy-permitting), code and implementation. Stratification of the Labrador and Irminger seas is quantified with the Potential Energy Anomaly (PEA) in the upper 1000 m of the water column. Monthly PEA anomalies are evaluated alongside corresponding anomalies of sea surface temperature, salinity, and density. For 30-year windows, moving correlations between PEA and surface properties are obtained over the 100-year simulations to characterize the evolving relationships. As climate change progresses, stratification in three of the four models is increasingly associated with variable surface salinity, in both regions. Lagrangian analyses of surface flow pathways in the decades preceding 1990 and 2040 are undertaken for one of the models in which surface salinity grows in influence. The subpolar presence of low-salinity Arctic waters and high-salinity subtropical Atlantic waters are found to increase and decrease respectively by 2040. Furthermore, in three of the four models, surface density compensation associated with correlation of surface temperature and salinity anomalies is progressively replaced by combined surface warming and freshening, lowering surface density, and strengthening stratification. The extent of these model-dependent changes and processes are of consequence for the projected fate of the AMOC by the mid twenty-first century.

1 Introduction

The Atlantic Meridional Overturning Circulation (AMOC) is an integral part of the global ocean circulation (Lee et al. 2019), with extensive influences on regional climate (Zhang et al. 2019). Based on a range of proxies, it has been claimed that the AMOC is currently in a historically weakened state (Caesar et al. 2021), although more recent reconstructions and more direct measurements indicate a high level of uncertainty in changes since 1980 (Jackson et al. 2022). Indeed, one 30-year reconstruction showed no decline at 26ºN (Worthington et al. 2021), while combined use of hydrographic sections and an inverse model indicate stable AMOC strength since the 1990s (Fu et al. 2020). Notwithstanding this considerable uncertainty, several early warning signals indicate that AMOC is approaching rapid transition to a weaker state (Boers 2021) and that this is likely in the coming decades (Ditlevsen and Ditlevsen 2023).

Key to AMOC variability and potential collapse is thermohaline forcing across the subpolar North Atlantic, specifically the Labrador and Irminger seas. It has been asserted that the Irminger Sea is the centre of action for subpolar AMOC variability (Chafik et al. 2022). Despite this current focus, it has also been emphasised that periodic Labrador Sea convection plays an important role in subsequent dense water transformation in the Irminger Sea (Böning et al. 2023) and modulates the AMOC on longer multidecadal timescales (Yeager et al. 2021). Associated with variability of Labrador Sea convection on interannual to decadal timescales is variable formation of Labrador Sea Water (LSW), recorded in modal properties (temperature, salinity) and vertical stratification, accurately measured since the 1960s (Lazier 1973, 1980; Dickson et al. 1996).

Following establishment in 1990 of repeat hydrography across the Central Labrador Sea, Yashayaev (2007) developed a 1960–2005 synthesis of Labrador Sea hydrography, to reveal that LSW transitioned from warm-salty modes during the mid-1960s and early 1970s, to cold-fresh modes between the late 1980s and the 1990s, subsequently warming and freshening by the mid-2000s. Subsequent studies reported extensive interannual variability into the latter 2010s (Yashayaev and Loder 2009, 2016, 2017). This wealth of literature reflects how, for several decades, LSW as the key driver of AMOC variability was established as a leading paradigm.

Attention more recently focussed on episodic dense water formation in the Irminger Sea, east of Greenland and south of Iceland. Since 2002, moorings and Argo floats have revealed extensive interannual variability of mixed layer depth, in the range 400–1400 m (de Jong et al. 2012; de Jong and de Steur 2016). Further attention has been on surface and sub-surface re-stratification processes, examined in reanalysis data over 1993–2019 (Sterl and de Jong 2022); at the surface, re-stratification is dominated by net heat gain and freshwater input, including from excess Greenland melt in recent years; in the sub-surface layer, by contrast, stratification is controlled by lateral advection of warm and saline water from the Irminger Current. Bringing the Irminger Sea fully into focus, recent observations from the Overturning in the Subpolar North Atlantic Program (OSNAP) array indicate that AMOC variability is most sensitive to processes in the eastern subpolar North Atlantic (Lozier et al. 2019).

Convective mixing and formation of dense water in the Labrador and Irminger seas are interconnected via the subpolar gyre circulation and may be associated with the same patterns of atmospheric forcing. Episodic dense water formation is associated with severe winter conditions, associated with storminess and strong surface heat loss. In this context, the renewal of dense modes of LSW has long been associated with positive phases of the basin-scale North Atlantic Oscillation, or NAO (Dickson et al. 1996; Yashayaev and Loder 2017). The wider subpolar gyre, including the Irminger Sea, may thus episodically shift between warm and cold states, on approximately decadal timescales. The most recent shift to a cold state began in winter 2013/14, when the subpolar gyre began to cool through intense air-sea interaction and ocean heat loss (Grist et al. 2016). Through winter 2014/15, anomalous air-sea interaction related to a strongly positive NAO drove further cooling. By the summer of 2015, a remarkably large volume of cold water persisted across the subpolar gyre, which subsequently dissipated over the following year (Duchez et al. 2016; Josey et al. 2018). Subsequent surface-intensified warming of the eastern subpolar gyre over 2017–19 has been attributed to a shift of the basin-scale ocean circulation, conveying warm and saline waters from the western subtropics (Desbruyères et al. 2021).

Surface salinities around the subpolar gyre are further subject to variable freshwater outflows through the Arctic gateways of Fram Strait and Davis Strait. Associated with enhanced fresh outflow in the late 1960s and early 1970s was the “Great Salinity Anomaly” (GSA), which advected around the subpolar gyre on a decadal timescale, the largest such event of several subsequently identified (Belkin et al. 1998). Following the 1970s GSA, additional freshening of the subpolar gyre (approximately double the GSA) was attributed to a sustained increase in precipitation over the eastern subpolar gyre during the late 1970s and 1980s (Josey and Marsh 2005). In the East Greenland Current (EGC) that conveys relatively fresh water from the Arctic, more recent ocean mooring measurements indicated relatively constant freshwater flux over 1998–2008 (de Steur et al. 2009), with episodic variability in the early 2010s related to enhanced Arctic outflow (de Steur et al. 2017). On the western side of the subpolar gyre, freshwater transports off Labrador have varied between high values during the 1950s and 1970s-80 s, and low values during the 1960s and since the mid-1990s, related to variations in the export of Arctic waters via the Canadian Archipelago and Davis Strait (Florindo-López et al. 2020).

Additional to Arctic and atmospheric drivers, salinity across the eastern subpolar gyre is subject to variable inflow from the subtropics. Episodic rises of temperatures and salinities around the subpolar gyre over 1958–2005 in an ocean assimilation system were associated with contraction/expansion of the subpolar gyre that allows more/less subtropical water to reach the eastern subpolar gyre from the southeast (Häkkinen et al. 2011). Subpolar gyre contraction, only partly attributed to NAO forcing, thus explains notable warming and salinity gains over 2001–05 and likely also explains the 2017–19 changes (Desbruyères et al. 2021). At other times, coordinated changes in both atmosphere and ocean appear to explain observed hydrographic variability. The combined effect of changes in atmospheric forcing and ocean circulation has thus been invoked to explain historically low salinities down to 1000 m in the eastern subpolar gyre over 2012–16, whereby freshwater was re-distributed eastward from the western subpolar gyre (Holliday et al. 2020).

The impacts of changing freshwater influences on the subpolar gyre, and associated processes, have been examined with in situ observations, ocean analyses and in high-resolution models. The latter may be sub-divided by resolution as ‘eddy-permitting’ (horizontal resolution 1/4º), or ‘eddy-rich’ (1/10º or 1/12º, substantially finer in nested regions). In hindcast simulations over 1958–2019 with an eddy-rich ocean model (mid-latitude Atlantic resolution 1/20º), deep convection shifts from the Labrador Sea towards the Irminger Sea during 2015–18, following the 2012–16 fresh anomaly and indicative of the growing influence of freshwater on convective mixing in the Labrador Sea (Rühs et al. 2021). Following the 2012–16 freshening of the eastern subpolar gyre, salinity anomalies rapidly circulated the boundary of the Irminger Sea while slowly reaching the interior, where the freshened surface layer arrested deep convection after winter of 2017/18 (Biló et al. 2022).

The processes by which freshwater anomalies spread from the boundary to the subpolar gyre interior have been examined with both observations and models. Limited import of surface-layer freshwater into the subpolar gyre interior via mean geostrophic and Ekman transports is evident in analysis of CTD and Argo observations along the 1000 m isobath; at most locations around the boundary, mean transports export surface waters away from the gyre interior (Jones et al. 2023). In eddy-rich ocean models, lateral eddy fluxes typically dominate the transport of freshwater from subpolar boundary currents to the gyre interior, although these fluxes are sensitive to model configuration (Pennelly et al. 2019; Gillard et al. 2022). Further insights are gained with Lagrangian diagnostics, whereby trajectories of conceptual water particles, or parcels, are computed with the time-evolving velocity and property fields of ocean models. Using the Parcels framework with data from an eddy-rich 1980–2019 hindcast, particles tracked backwards for 10 years from the east part of the OSNAP section reveal that recent cooling and freshening in the eastern subpolar gyre is primarily driven by reduced heat loss in the Labrador Sea and the eastward export of lighter upper-layer waters in the preceding decade (Fox et al. 2022). Using the TRACMASS algorithm with data from an eddy-permitting hindcast, water parcels tracked forwards from the east OSNAP section quantify the recirculation and transformation of water masses on seasonal timescales in the Irminger Sea, explaining wind-driven seasonality of the AMOC (Tooth et al. 2023).

In summary, much attention has focussed on recent changes in the subpolar North Atlantic, and the increasing role of freshwater input under global warming, of consequence for the AMOC. A particular focus has been the increasing influence of mass loss from the Greenland Ice Sheet, explored through CMIP ‘hosing’ protocols (e.g., Jackson et al. 2023). Here, we apply novel diagnostics to ocean and climate models of highest resolution currently available to the community, exploring a transition over 1950–2050 from historical surface density compensation to scenarios of future warming and freshening, the latter independent of Greenland melting. Freshwater-enhanced upper ocean stratification may amplify surface temperature variability, of consequence for the atmosphere. Noting the general utility of surface temperature and salinity in AMOC analysis (e.g., Boers 2021), we seek links between these variables and stratification. Given the likely importance of eddies in the AMOC response to a changing climate (Weijer et al. 2019) and the fidelity of AMOC representation in high-resolution models (Hirschi et al. 2020), we focus here on eddy-permitting and eddy-rich ocean simulations. With evidence for shifts in convective mixing between the Labrador and Irminger seas (e.g., Rühs et al. 2021), we focus our analysis on these two regions. Alongside an eddy-rich ocean hindcast, we analyse selected climate model simulations from the High Resolution Model Intercomparison Project (HighResMIP) of CMIP6 (Haarsma et al. 2016), for which a spread of AMOC declines to 2050 have been established (Roberts et al. 2020). In Section 2, we outline the model data along with the Eulerian and Lagrangian diagnostics used here. In Section 3, we present our analysis of the eddy-rich hindcast (1958–2021), followed by analyses of HighResMIP experiments spanning 1950–2050 with four selected models, in both control and anthropogenically forced simulations. In Section 4, we discuss our findings in the context of recent and ongoing changes in the subpolar gyre, of consequence for the AMOC and regional climate.

2 Methods

We first introduce the models and experiments selected for analysis, followed by an outline of the Eulerian and Lagrangian diagnostics, focused on methods for quantification of stratification and the backtracking of surface layer waters to source regions on decadal timescales.

2.1 Model simulations and data

We use monthly mean model output from the GO8p7-eORCA12 global ocean sea-ice hindcast simulation, which uses version 8 of the UK Global Ocean (GO) configuration, based on the Nucleus for European Modelling of the Ocean (NEMO) version 4.0.4 ocean model (Madec et al. 2019). The extended version of the ORCA12 grid (eORCA12) has an eddy-rich resolution of 1/12° and includes oceanic circulation under the ice cavities in Antarctica. The model has 75 vertical levels and a tri-polar grid with poles located at Canada, Siberia, and Antarctica. The vertical mixing of tracers in the model is parameterised using an improved version of the Turbulent Kinetic Energy (TKE) scheme (Gaspar et al. 1990; Madec et al. 2019). The NEMO model uses a nonlinear free surface in which the volume of the ocean grid cells at every vertical level are allowed to vary with time (Adcroft and Campin 2004). The sea-ice model used in this configuration is CICE (Hunke et al. 2017), which includes the effect of surface melt ponds. The hindcast simulation was initialised on January 1st, 1958, with 1995 – 2014 mean temperature and salinity profiles of the EN4 climatology (Good et al. 2013) and forced with the Japanese 55-year atmospheric reanalysis (JRA55-do) dataset (Tsujino et al. 2018). The model was integrated from 1958 to 2021 with a model time step of 300 s. The ORCA12 hindcast simulates recent variability, for comparison with that in the HighResMIP models, and provides the complete set of model output data that is needed to examine PEA tendencies (see Section 2.2).

In more extensive analysis, we select four models from the HighResMIP ensemble (Haarsma et al. 2016), each used in two coupled runs. Following short (30–50 year) periods of spin-up with each model, two 100-year experiments are undertaken:

• control-1950, driven with constant 1950s forcing for the entire 100 years;

• hist-1950, externally forced (1950–2014) followed by highres-future, subject to external forcings following the SSP585 future scenario for the period 2015–2050.

These runs are henceforth referred to as ‘control’ and ‘forced’. The models selected for analysis are HadGEM3-GC31, CESM1-3 and EC-Earth3P (Roberts et al. 2020, Table 1). We further selected two configurations of HadGEM3-GC31, with high (1/12°) and medium (1/4°) ocean resolution, both with high (50-km resolution) atmospheric resolution (HH and HM respectively); we subsequently denote these variants as HadGEM3-HH and HadGEM3-HM. The resolution of the Parallel Ocean Model (POP) in CESM1-3 is 1/10°, higher in the Arctic, while that in EC-Earth3P is 1/4°. With logistical constraints in mind, we selected this model subset to span resolution (1/12° to 1/4°), source code (NEMO vs. POP) and code implementation at the same 1/4° ocean resolution (HadGEM3-HM vs. EC-Earth). Full details of configurations and parameterizations are provided in Table 1 of Roberts et al. (2020).

Specifically relevant to our study, the HadGEM3 (-HH and -HM), CESM1 and EC-Earth3P models are characterised by structurally different AMOCs and associated poleward heat transport (Roberts et al. 2020). The AMOC in EC‐Earth3P is substantially weaker than the other models, with peak overturning notably higher in the water column. For given AMOC transport, the associated heat transport at 26°N is consistently lower than observations in all HighResMIP models apart from CESM1, in which stronger heat transport is associated with a near‐surface warm bias. As further highlighted by Roberts et al. (2020), the forced response of the AMOC likewise varies across the four models, and we also find substantial differences in forced changes of surface properties and fluxes across the subpolar North Atlantic.

2.2 Eulerian analyses – the Potential Energy Anomaly (PEA) and associated tendencies

As a model-independent index for stratification, we calculate the Potential Energy Anomaly (PEA), used extensively in shelf sea settings (see Marsh and van Sebille 2021). PEA is comparable to other integral measures of stratification in recent use, such as convective resistance, with units m²s⁻² (Gillard et al. 2022) or buoyancy content, with units J kg⁻¹ (Biló et al. 2022). It has been recently proposed as a more consistent approach to diagnosis of surface mixed layers (Reichl et al. 2022), and previously calculated as convective energy (units J m⁻³) to define the strength of stratification in a very high resolution model of the Labrador Sea (Pennelly and Myers 2020). We summarise here the basis of a PEA calculation. The potential energy per m of depth (in units of J m⁻³) is obtained for a stratified water column, ${PE}_{stratified}$, and its mixed counterpart, ${PE}_{mixed}$,

$${PE}_{stratified}=\frac{1}{h}\underset{z=-h}{\overset{z=0}{\int }}\rho \left(z\right)gzdz,$$

(1a)

$${PE}_{mixed}=\frac{1}{h}\underset{z=-h}{\overset{z=0}{\int }}\overline{\rho }gzdz$$

(1b)

where $h$ is a specified depth, $\rho \left(z\right)$ is the vertical density profile, we integrate upwards from z = -$h$ to $z$ = 0 at the surface, and $\overline{\rho }$ is the depth-averaged density (i.e., for a fully mixed water column). We then define the difference and ${PE}_{mixed}$ minus ${PE}_{stratified}$ as the potential energy anomaly, $\phi$,

$$\phi ={PE}_{mixed}-{PE}_{stratified}=\frac{1}{h}\underset{z=-h}{\overset{z=0}{\int }}\left(\overline{\rho }-\rho \left(z\right)\right)gzdz.$$

(2)

By this definition, $\phi$ > 0 for a stratified water column, and $\phi$ increases with stratification.

As emphasized, the use of PEA avoids discrepancies between models in parameterization of mixed layer depth, and accounts for stratification in a fixed depth range, here the upper 1000 m. This diagnostic is also versatile in being calculated from the potential temperature and salinity fields, which are universally available as HighResMIP data, while mixed layer depth was not available to us for all models.

Another utility of the PEA framework is the attribution of PEA time derivatives, or tendencies, to various stratifying and mixing processes. PEA tendencies are used here to establish the relative importance of these different processes on a seasonal timescale. A thermal tendency term for PEA can be derived (see Marsh and van Sebille 2021) as an incremental change (reduction) of potential energy, $\Delta PE$, per unit time, $\Delta t$, related to the surface net heat flux,

$${\left.\frac{\partial \phi }{\partial t}\right|}_{thermal}=\frac{1}{h}\frac{\Delta PE}{\Delta t}=\frac{1}{h}\frac{1}{\Delta t}\left(\frac{1}{2}gh\frac{\alpha \Delta Q}{{c}_{p}}\right)=\frac{\alpha g{Q}_{net}}{2{c}_{p}}$$

(3)

where $\Delta Q$ is heat gain per unit area (J m⁻²) in time increment $\Delta t$, giving net surface heat flux, ${Q}_{net}=\Delta Q/\Delta t$ (W m⁻²); ${c}_{p}$ and $\alpha$ are the specific heat capacity (J kg⁻¹ °C⁻¹) and thermal expansion coefficient (°C⁻¹) for seawater.

We can further attribute a PEA tendency to surface freshwater input, $f$ (positive into the ocean), which will change the salinity of a layer with thickness Δh by,

$$\Delta S=-\frac{Sf\Delta t}{\Delta h}$$

(4)

We then relate a change in density to this change in salinity, by how much seawater will contract, given $\Delta \rho =\rho \beta \Delta S$, hence,

$$\Delta \rho =-\frac{\rho \beta Sf\Delta t}{\Delta h},$$

(5)

and define the difference (reduction) in potential energy,

$$\Delta PE=-\frac{1}{2}gh\Delta \rho \Delta h.$$

(6)

Substituting (5) into (6), we relate the change of potential energy to freshwater input,

$$\Delta PE=\frac{1}{2}gh\rho \beta Sf\Delta t$$

(7)

Associated with salinity changes, a haline tendency due to the surface freshwater input thus follows,

$${\left.\frac{\partial \phi }{\partial t}\right|}_{haline}=\frac{1}{h}\frac{\Delta PE}{\Delta t}=\frac{g\rho \beta Sf }{2}$$

(8)

Surface heating and freshwater input (net precipitation, runoff, sea ice melt) act to increase stratification – positive $\phi$ tendencies. Surface cooling and freshwater loss (net evaporation, sea ice freeze) and combined mixing (winds, tides, interior diapycnal mixing), ${\left.\partial \phi /\partial t\right|}_{mixing}$, act to weaken stratification – negative $\phi$ tendencies. For a given volume, divergence of advective density transport, primarily associated with horizontal transports of heat and freshwater across volume boundaries, is a further positive or negative tendency term, ${\left.\partial \phi /\partial t\right|}_{advective}$. Four tendency terms (units W m⁻³) thus determine a total tendency, ${\left.\partial \phi /\partial t\right|}_{total}$,

$${\left.\frac{\partial \phi }{\partial t}\right|}_{total}={\left.\frac{\partial \phi }{\partial t}\right|}_{thermal}+{\left.\frac{\partial \phi }{\partial t}\right|}_{haline}-{\left.\frac{\partial \phi }{\partial t}\right|}_{mixing}+{\left.\frac{\partial \phi }{\partial t}\right|}_{advective}.$$

(9)

With limited data available for calculation of the right-hand side terms (and from only the ORCA12 hindcast – see below), we simplify (9) to a balance between surface heat and freshwater fluxes, and a residual term that combines mixing and advective terms,

$${\left.\frac{\partial \phi }{\partial t}\right|}_{total}={\left.\frac{\partial \phi }{\partial t}\right|}_{thermal}+{\left.\frac{\partial \phi }{\partial t}\right|}_{haline}+{\left.\frac{\partial \phi }{\partial t}\right|}_{residual}.$$

(10)

For the ORCA12 hindcast and the four climate models, we calculate monthly-mean PEA with monthly-averaged temperature and salinity fields in the upper 1000 m. Using monthly-averaged net downward surface heat flux and net upward water flux (including runoff and net sea ice freezing/melting) from the ORCA12 hindcast, we calculate the corresponding thermal and haline PEA tendencies. The total tendency is calculated from month-to-month changes of PEA, to infer a residual tendency following (10). As surface freshwater fluxes for the HighResMIP experiments were not available to us, this baseline PEA budget analysis was restricted to the hindcast.

To obtain a monthly stratification index, and associated surface temperature and salinity anomalies, we average PEA in boxes representative of the Irminger Sea (25-45ºW, 60-65ºN) and the Labrador Sea (45-60ºW, 55-63ºN). The Irminger Sea box limits are motivated by evidence from a Lagrangian study of the time-mean overturning in property space in the ORCA12 hindcast, obtained with water parcels released to follow Arctic outflow through Fram Strait and Atlantic inflow into the eastern subpolar North Atlantic and Nordic seas; this analysis indicates that the great majority of dense water formation that supports the AMOC lower limb is driven by air-sea interaction and mixing in the latitude range 60-65ºN (D. Dey, personal communication). We are further motivated by a series of studies using new observations from the OSNAP array to conclude that the mean and variability of the AMOC in the subpolar gyre is driven by buoyancy exchanges north of the OSNAP-East section (see Jones et al. 2023 and references therein), which extends across the Irminger Basin to Cape Farewell at around 60ºN. The Labrador Sea box limits are specified to encompass a broad area of deep mixing (Pennelly and Myers 2020), Although larger boxes have been previously used to define the Labrador Sea, such as 70–40°W, 45–72°N, for calculation of a ‘Deep Mixed Volume’ (Roberts et al. 2020), we prefer to limit the box dimensions to focus on locations of dense water formation.

2.3 Lagrangian analyses – TRACMASS parcel trajectories

To explore how surface waters reach the subpolar gyre at decadal timescales, we track water parcels backwards in three dimensions for 10 years, sampling southward flows across a section between east Greenland and southwest Iceland. For these calculations, we use the Lagrangian trajectory code TRACMASS v7.0 (Aldama-Campino et al. 2020) with monthly-mean ocean mass/volume transport and property fields from the HadGEM3-HH climate model. Given time-dependent gridded volume transport fields, TRACMASS simulates parcel trajectories by analytical solution of differential equations for parcel position in longitude, latitude and depth, providing exit locations and times per grid cell (Vries and Döös 2001). The calculation is ‘stepwise-stationary’, in that volume transports are assumed to be constant during intermediate time steps between updates of transport fields (Döös et al. 2017). Large ensembles of water parcel trajectories are thus calculated offline. Trajectories calculated by TRACMASS are analogous to streamtubes, describing non-divergent flow for which volume transport at one end of a streamtube is equal to that at the other end (Van Sebille et al. 2018), equivalent to the volume transport pathways of an incompressible fluid. While sub-grid scale processes are not parameterized during the TRACMASS analyses, parcel properties (temperature, salinity) evolve along trajectories due to air-sea interaction and parameterized mixing in the climate model.

To evaluate changing pathways and the associated timescales and properties, trajectory initial locations were seeded in the upper 14 m along a zonal transect across the Irminger Sea just south of Denmark Strait, in each month of 1990 and 2040. Water parcels were subsequently advected backward in time for 10 years, i.e., until 1980 and 2030 respectively. A maximum volume transport of 50 m³s⁻¹ was assigned to each trajectory. Following the orthodox TRACMASS approach, parcels are representative of incremental volume transport: when the southward volume transport in a grid box along the transect exceeds 50 m³s⁻¹, the number of seeded parcels was obtained as the volume transport divided by 50, rounded up to integral value. Parcel location data are statistically analysed to obtain maps of parcel mean salinity and depth. Transport increments are integrated to obtain total transports across selected sections and the sea surface.

3 Results

We begin with an overview of stratification in the Irminger and Labrador seas from the ORCA12 hindcast, followed with analysis of anomalies of PEA and surface properties, and their correlation, in that experiment. We then focus on PEA variability in relation to that of surface temperature, salinity and density, in control and forced experiments with the coupled models. Finally, shifting the focus onto drivers of changes in the co-variability of stratification and surface properties, and specifically the influence of freshwater on the subpolar gyre, we conclude the results with an examination of the back-trajectories of surface waters in the Irminger and Labrador seas for selected years of the forced experiment with HadGEM3-HH.

3.1 ORCA12 hindcast (1958–2021)

We first describe the spatial pattern of PEA during the approximate times of minima (weak stratification) and maxima (strong stratification), in March and September respectively. We further consider the preceding minima and maxima of tendencies attributed to net surface heat and freshwater fluxes, in January (PEA decreasing) and July (PEA increasing). In Fig. 1, we map PEA, and the thermal and haline tendencies, in 2016, a year noted for deep late winter mixing in the Labrador Sea that was preceded in late winter 2015 by deepest recorded mixing in the Irminger Sea. In March, a large expanse of PEA minima (< 50 J m⁻³) covers much of the Labrador Sea, stretching into the Irminger Sea, where PEA lies in the range 50–150 J m⁻³ across most of the Irminger Basin, west of the Reykjanes Ridge (Fig. 1a). By September, PEA is higher across the region, with minima only just below 200 J m⁻³ and 250 J m⁻³ in the southern Labrador Sea and the southern Irminger Sea respectively (Fig. 1b). Most evident in September is a signature of positive PEA (> 500 J m⁻³) associated with the stratifying influence of warm Atlantic Water to the east and west of Greenland.

Fig. 1

Example maps (mid-latitude North Atlantic) of PEA (J m⁻³) and tendencies due to surface heat and freshwater fluxes (10^–5 W m.⁻³) from the ORCA12 hindcast: PEA in (a) March and (b) September of 2016; thermal tendencies due to surface net heat flux in (c) January and (d) July of 2016; haline tendencies due to surface freshwater fluxes in (e) January and (f) July of 2016. Also indicated are the two boxes used to average PEA and surface properties across the Irminger and Labrador seas. In (a) and (b), we select March and September to show the minima and maxima of PEA, while the choice of January and July in (c)-(f) are selected to highlight the extremes of seasonal forcing. Note that the range for thermal tendency is larger than that for haline tendency by a factor of 5

Across most of the region, in both January and July, the thermal tendency dominates the haline tendency, by up to an order of magnitude (Figs. 1c-f). The thermal tendency most notably undergoes a large seasonal cycle between negative values through net cooling in January (Fig. 1c) and positive values through net warming in July (Fig. 1d). An exception to the dominance of the thermal tendency over the haline tendency is seen along the western boundary of the Labrador Sea in January, where sea ice freeze induces a negative tendency through brine rejection, adjacent to offshore sea ice melt, which induces a positive tendency (Fig. 1e). Also notable is a positive haline tendency over the Irminger Sea to the southeast of Greenland, due to locally high precipitation rates in cyclonic systems that result from interaction of large-scale westerly air flow with the topography of southern Greenland.

In Fig. 2, we present time series of monthly PEA and corresponding anomalies relative to monthly means, area-averaged across the Irminger Sea and the Labrador Sea. Superimposed on seasonal cycles in Fig. 2a (further examined below), long-term increases in PEA are evident in both seas: 12.53 J m⁻³ decade⁻¹ in the Irminger Sea; 7.13 J m⁻³ decade⁻¹ in the Labrador Sea. Superimposed on these trends is considerable interannual variability; coincident with documented episodes of renewed convective mixing in the Labrador Sea (1989–94, 2008/09, 2015/16) and the Irminger Sea (2014/15) are negative excursions of PEA in the range 50–100 J m⁻³, highlighted in Fig. 2b.

Fig. 2

Time series of (a) PEA and (b) corresponding anomalies relative to monthly means, in the ORCA12 hindcast (1958–2021) area-averaged across the Irminger Sea and the Labrador Sea. The grey bars in (b) highlight episodes of renewed convective mixing in the Labrador Sea (1989–94, 2008/09, 2015/16) and the Irminger Sea (2014/15)

From the PEA time series in Fig. 2a and surface heat and freshwater fluxes (see Section 2.2), we further determine monthly PEA tendencies over 1958–2021 for the Irminger and Labrador seas (see Fig. S1). We thus obtain the long-term (1958–2021) mean and standard deviation of monthly PEA and monthly-mean PEA tendencies, shown as seasonal cycles in Fig. 3. Consistent with the spatial structures in Fig. 1, the monthly-mean PEA values for the Irminger Sea (Fig. 3a) exceed those for the Labrador Sea (Fig. 3b), with minima in March/April and maxima in September. Standard deviations are at a similar level throughout the seasonal cycle, with Irminger Sea values almost double those for the Labrador Sea. The seasonal cycle of PEA total tendency is largely attributed to the thermal tendency, which reaches positive maxima in June/July and negative minima around December/January. The haline tendency is weakly positive all year around, attributed to net surface freshwater gain. Small peaks of the haline tendency in spring are associated with increased net precipitation and/or sea ice melt.

Fig. 3

Mean (1958–2021) seasonal cycles of (a,b) PEA (thick line, mean; thin lines, ± standard deviation) and (c,d) PEA tendency (total, due to net surface heat flux, due to net surface freshwater flux, and the residual tendency), for (a,c) the Irminger Sea and (b,d) the Labrador Sea

As the thermal tendency is more negative than positive, and with PEA balanced over the seasonal cycle, the year-round residual tendencies are generally positive, except for slightly negative values for the Labrador Sea in late summer. Positive residual tendencies imply an influx of buoyancy through convergence of advective heat and freshwater transports in the surface layer (mixing would drive a negative tendency). This residual tendency peaks around January/February in the Irminger Sea, and around February/March in the Labrador Sea, consistent with in-fluxes of warm surface waters from lower latitudes at this time of year.

With a view to an anticipated connection between surface conditions and stratification, we examine monthly anomalies of sea surface temperature (SST) and salinity (SSS). Area-averaged anomalies, presented in Fig. 4, show the considerable interannual to decadal variability in both Irminger and Labrador seas. Following relative warmth in the early/mid 1960s, both regions cool until the early 1970s. Also notable from the late 1960s to the early/mid 1970s are SSS minima reminiscent of the original GSA (Belkin et al. 1998). Historically negative SST anomalies in the Labrador Sea over 1989–94 coincide with deep convective mixing. Subsequent warming of both regions led to sustained warmth in the 2000s, punctuated by cold episodes in the Labrador Sea (2008/09, 2015/16) and the Irminger Sea (2014/15) that again coincide with renewed convective mixing (indicated in Fig. 4).

Fig. 4

Time series of SST and SSS anomalies in the ORCA12 hindcast (1958–2021) area-averaged across: (a) the Irminger Sea; (b) the Labrador Sea. The grey bars highlight cold episodes in the Labrador Sea (1989–94, 2008/09, 2015/16) and the Irminger Sea (2014/15) that coincide with renewed convective mixing

To quantify links between stratification and surface properties, we calculate moving correlations between anomalies of PEA and anomalies of SST, SSS and SSσ₀. Due to the relatively short time series, we calculate correlations with data spanning short 10-year windows, although at monthly frequency, the sample sizes of 120 ensure that correlations are often significant at 99% confidence intervals (two-tailed). In Fig. 5, we present correlations for anomalies averaged across the Irminger Sea (Fig. 5a) and the Labrador Sea (Fig. 5b). Negative PEA-SSS correlations are consistent with strengthened stratification when the surface is anomalously fresh. Weak anti-correlation of PEA with SST (and SSσ) is consistent with comparison of Figs. 2 and 4, revealing that stratification often strengthens while SST decreases, and vice versa. This suggests that stratification is more sensitive to contrasting variability of sub-surface temperature (and density), such that PEA associated with cool and deep mixed layers may exceed that of warm and shallow mixed layers. We further note that PEA correlations with SST, SSS and SSσ are generally weak, often insignificant, and decline close to zero in the early 2000s and 2010s.

Fig. 5

Moving 10-year correlations of monthly PEA anomalies with SST, SSS and SSσ₀ anomalies averaged across (a) the Irminger Sea and (b) the Labrador Sea, and (c) moving correlations of SST with SSS anomalies (averaged across both seas), in the ORCA12 hindcast (1958–2021). The thick (thin) segments indicate significant (insignificant) correlations

Notwithstanding the relatively weak correlations between PEA and surface properties, a striking feature in Fig. 4 of the SST and SSS time series pre-2000 are coincident warm/saline anomalies and cold/fresh anomalies, hence density compensation. From the 2000s onwards, this coincidence is less evident. This changing character of SST and SSS co-variability is quantified with the moving correlations shown in Fig. 5c. For both regions, moving correlations are strongly positive until the early 2000s, after which they decline. Correlations become insignificant by the late 2010s and early 2020s, visually trending toward negative values. This provides a motivation for our subsequent investigation of ongoing changes in the co-variability of SST and SSS, as revealed in the coupled models, and further links to regional stratification in those models.

3.2 HighResMIP coupled simulations – Eulerian analyses

In Fig. 6, we show the spatial distributions of PEA for March (Fig. 6a-d) and September (Fig. 6e-h), in the first year of each control simulation. As in the ORCA12 hindcast, PEA falls below 50 J m⁻³ across much of the subpolar gyre in March, most extensively in EC-Earth3P. By September, PEA has increased throughout the region, with minima of around 200 J m⁻³ persisting in the Labrador Sea of HadGEM3-HH and HadGEM3-HM, with low values extending also to the Irminger Sea of EC-Earth3P; PEA is substantially higher in CESM1-3, reaching September minima of around 300 J m⁻³. The March and September PEA distributions quantify a range of stratification across the four models, being strongest in CESM1-3 and weakest in EC-Earth3P. The ‘September minus March’ differences (Fig. 6i-l) emphasize a diversity of amplitude in the seasonal cycle, smallest in EC-Earth3P and largest in CESM1-3, which are respectively the least and most seasonally stratified of the four models by this measure. We subsequently focus on surface property and PEA changes in the forced experiments.

Fig. 6

March (a-d), September (e–h) and September minus March (i-l) PEA (units J m⁻³) in 1950 of control simulations: (a,e,i) HadGEM3-HH; (b,f,j) HadGEM3-HM; (c,g,k) CESM1-3; (d,h,l) EC-Earth3P. Also indicated are the two boxes used to average PEA and surface properties across the Irminger and Labrador seas

In Fig. 7, we co-plot SST and SSS anomalies for each experiment in the Irminger Sea (left column) and Labrador Sea (right column). Warming and freshening are quantified with trends for the periods 1950–2014 and 2015–50, corresponding to the hist-1950 and highres-future periods of the forced scenarios, presented in Table 1. Trends are significance tested following the method of Santer et al. (2000). Significant trends prevail in both seas and the four models, with the exceptions of SST over 1950–2014 in the Irminger Sea of EC-Earth3P and in the Labrador Sea of CESM1-3, and SSS over 2015–50 (both seas) in CESM1-3. In the early decades (and from the 1980s in CESM1-3), the anomalies largely co-vary as in the ORCA12 hindcast pre-2000. From around 2020, the SST and SSS anomalies diverge, as anomalously warm and fresh conditions emerge. Motivated by the emergence of surface freshening in HadGEM3-HH, HadGEM3-HM and EC-Earth3P, we further examined the extent to which surface density (SSσ) anomalies – consequential for convective mixing and re-stratification – are linked to SSS anomalies, co-plotted in Figure S2 for the Irminger Sea (left column) and Labrador Sea (right column). SSσ anomalies coincide with SSS anomalies on a range of timescales, for both Irminger and Labrador seas in HadGEM3-HH, HadGEM3-HM and EC-Earth3P. In these three models, substantial negative SSS and SSσ anomalies of around 0.25–0.5 psu and 0.5–1.0 kg m⁻³, relative to early decades, are established by the 2040s. No clear relationship between SSS and SSσ anomalies is evident in CESM1-3, although long-term declines of around 0.25 psu and 0.5 kg m⁻³ are evident.

Fig. 7

Time series of SST and SSS anomalies averaged across the Irminger Sea (a,c,e,g) and Labrador Sea (b,d,f,h), in forced simulations with: (a,b) HadGEM3-HH; (c,d) HadGEM3-HM; (e,f) CESM1-3; (g,h) EC-Earth3P

Table 1 Decadal SST and SSS linear trends (^oC decade⁻¹ and psu decade⁻¹) plus t statistics (t, italic font) used to test for significant trends (indicated with bold font), in the Irminger and Labrador seas, over 1950–2014 and 2015–50 in the forced experiments with the four climate models. Trends are significant at a 95% Confidence Level, assuming a two-tailed distribution, where |t|> 1.962

We next consider the consequences of warming, freshening and declining surface density, comparing here the control and forced experiments. In Fig. 8, we plot monthly anomalies, relative to mean seasonal cycles per model, of area-averaged PEA for the Irminger Sea (left column) and Labrador Sea (right column). In Table 2, we present corresponding PEA trends for 1950–2014 and 2015–50. In all four models and for both regions, PEA increases substantially and significantly over the forced experiments, by 100–200 J m⁻³ over 1950–2050. Most of these increases emerge over 2020–50. Also notable is a high degree of interannual to sub-decadal variability in both the control and forced experiments, and an absence of seasonality in the amplitude of the anomalies, which implicitly persist on timescales longer than seasonal. Variability reaches higher amplitudes in the Irminger Sea for all models and is relatively subdued in EC-Earth3P (Fig. 8g-h).

Fig. 8

Time series of monthly PEA anomalies averaged across the Irminger Sea (a,c,e,g) and Labrador Sea (b,d,f,h), in control and forced simulations with: (a,b) HadGEM3-HH; (c,d) HadGEM3-HM; (e,f) CESM1-3 (HH); (g,h) EC-Earth3P (HR)

Table 2 Decadal PEA linear trends (J m⁻³ decade⁻¹), plus t statistics (italic font) used to test for significant trends (indicated with bold font), in the Irminger and Labrador seas, over 1950–2014 and 2015–50 in the control (Ctl) and forced experiments with the four climate models

Co-variability of PEA with SST, SSS and SSσ, for the entirety of control and forced experiments in both regions, is quantified with correlation coefficients (Table 3) and linear regressions (Table 4). The comparative magnitude of SST-PEA correlations and (where negative) SSS-PEA (anti-)correlations may be related to the relative thermal or haline influences on stratification. In the Irminger Sea, SST-PEA correlations exceed in magnitude the SSS-PEA correlations, in both experiments with all four models. In the Labrador Sea, SST-PEA correlations are likewise greater in the control experiment with HadGEM3-HH, and in both experiments with HadGEM3-HM, but SSS-PEA (anti-)correlations are of larger magnitude in both experiments with CESM1-3 and EC-Earth3P. It is notable that SSS-PEA correlations become more strongly negative in all forced experiments, compared to control experiments, in both regions. PEA-SSσ correlation coefficients are all substantially more negative in the forced experiments, as stratification becomes more strongly associated with surface warming and/or freshening. Regression coefficients indicate that the sensitivity of PEA to SSS correspondingly increases in the forced experiments, most strikingly in HadGEM3-HH and HadGEM3-HM. In contrast, PEA dependence on SST increases more modestly in HadGEM3-HH, HadGEM3-HM and CESM1-3, although more substantially in EC-Earth3P.

Table 3 Correlation of PEA with surface temperature (SST), salinity (SSS) and density (SSσ), for control and forced experiments with the four climate models. For each experiment, the dominant correlation, of SST and SSS, is highlighted in bold font

Table 4 Regression of PEA with surface temperature (SST), salinity (SSS) and density (SSσ), for control and forced experiments with the four climate models. For each experiment pair, the strongest regression, of PEA with SST, SSS and SSσ, is highlighted in bold font

To further quantify the changing relationships of PEA with SST, SSS and SSσ through the forced experiments, we calculate moving correlation coefficients, presented in Fig. 9 (corresponding correlations for the control experiments are presented in Figure S3). With 100-year time series and given the strong interannual to sub-decadal variability in PEA, we now calculate correlation coefficients with monthly anomalies in 30-year windows.

Fig. 9

Moving 30-year correlations of monthly PEA anomalies with SST, SSS and SSσ₀ anomalies, averaged across the Irminger Sea (a,c,e,g) and Labrador Sea (b,d,f,h), in forced simulations with: (a,b) HadGEM3-HH; (c,d) HadGEM3-HM; (e,f) CESM1-3; (g,h) EC-Earth3P. The thick (thin) segments indicate significant (insignificant) correlations

In the Irminger Sea (Fig. 9, left panels), SSσ-PEA correlation coefficients are strongly negative throughout, although notably strengthening over time in HadGEM3-HH (Fig. 9a), HadGEM3-HM (Fig. 9c) and EC-Earth3P (Fig. 9g). Most striking is the transition of SSS-PEA correlations in HadGEM3-HH and HadGEM3-HM, from significantly positive to significantly negative, reaching approximate parity with SST-PEA correlations by around 2025 in the HadGEM3-HM experiment. SSS-PEA correlations in EC-Earth3P also transition from weak to strong negative values by the early 2000s. In contrast, SSS-PEA correlations in CESM1-3 (Fig. 9e) substantially weaken from strongly negative values in early decades, to approach zero by 2035. Compared to the widely ranging changes in SSS-PEA correlations, the SST-PEA correlations are relatively stable, and strongly positive throughout, most notably in HadGEM3-HH. In broad terms, similar correlations and transitions are obtained in the Labrador Sea (Fig. 9, right panels), although with some exceptions worthy of comment. Compared to the Irminger Sea, negative SSS-PEA correlations almost throughout all control and forced experiments indicate that SSS more strongly influences stratification in the Labrador Sea. However, in later decades of the forced experiments with HadGEM3-HH, HadGEM3-HM and EC-Earth3P, SSS-PEA anti-correlations in the Irminger and Labrador seas of each model approach parity.

In the Labrador Sea (Fig. 9, right panels), the most striking transition is again the increasingly negative SSS-PEA correlations in HadGEM3-HH (Fig. 9b) and HadGEM3-HM (Fig. 9d), suggesting that SSS becomes progressively more influential in controlling stratification of the Labrador Sea in these models. In HadGEM3-HH, SSS-PEA correlations almost double from the late 1990s to 2010, from around -0.35 to around -0.7; in HadGEM3-HM, SSS-PEA correlations more dramatically transition from insignificant values until around 2000, to reach around -0.7 by 2035, when SSS slightly dominates SST in PEA variability. In considerable contrast, SSS-PEA correlations are relatively invariant throughout the forced experiments with CESM1-3 (Fig. 9f) and EC-Earth3P (Fig. 9h), while SST-PEA correlations decline from positive values to become more insignificant in CESM1-3 and rise from slightly negative to slightly positive in EC-Earth3P. Corresponding with the increasing influence of SSS and stable influence of SST on stratification in HadGEM3-HH, negative SSσ-PEA correlations slightly strengthen to 2035. Strengthening of negative SSσ-PEA correlations in HadGEM3-HM is more emphatic, changing from around -0.4 to -0.7 as the growing influence on PEA of SSS reinforces the background, more variable, influence of SST.

The corresponding moving correlations through control experiments (Fig. S3) provide context for the changes evident in forced experiments and are instructive in relation to differences between models. Positive SST-PEA correlations in all four models are generally stable throughout. SSS-PEA correlations are generally weaker and less in absolute terms than the SST-PEA correlations. Slightly positive correlations in HadGEM3-HH and HadGEM3-HM are consistent with elevated PEA during warm and saline intrusions to the Irminger Sea. In contrast, negative correlations in EC-Earth3P and CESM1-3 indicate more of a freshwater influence on Irminger Sea stratification in these two models. As one would expect, the SSσ-PEA correlations are strongly negative throughout the control experiments, although weakest in HadGEM3-HM, for which an anomalously warm surface layer is likely also anomalously saline. Strongest SSS-PEA correlations in EC-Earth3P are accompanied by negative SST-PEA correlations throughout most of the control experiment, and SSσ-PEA correlations are correspondingly weaker than in control experiments with HadGEM3-HH and CESM1-3, in which positive SST-PEA correlations and negative SSS-PEA correlations suggest that SST and SSS act together to control stratification in the Labrador Sea. A long-term decline is also apparent in the SST-PEA correlations in control experiments of HadGEM3-HM and CESM1-3, suggesting some model drift or slow natural variability.

Regarding the combined influence of SST and SSS variability, we also calculate moving correlations of monthly anomalies in SST and SSS. For control and forced experiments of each model and for both regions, these are shown in Fig. 10. In the Irminger Sea, these SST-SSS correlations are mostly positive throughout the four control experiments (Fig. 10, left panels), being weakest in CESM1-3 (Fig. 10e) and most stable in EC-Earth3P (Fig. 10g). In the forced experiments, relative to the control experiments, correlations decrease over 1965–2035 in HadGEM3-HH (Fig. 10a), HadGEM3-HM (Fig. 10c) and EC-Earth3P. Gradual long-term declines in HadGEM3-HH and EC-Earth3P contrast with a rapid positive–negative transition from the 2010s to the 2020s in HadGEM3-HM, in which the strongest negative SST-SSS correlations are established by the mid 2020s. In contrast, SST-SSS correlations in the forced experiment with CESM1-3 trend from insignificant to weak but significant positive values.

Fig. 10

Moving 30-year correlations of SST and SSS anomalies, averaged across the Irminger Sea (a,c,e,g) and Labrador Sea (b,d,f,h), in control and forced simulations with: (a,b) HadGEM3-HH; (c,d) HadGEM3-HM; (e,f) CESM1-3; (g,h) EC-Earth3P. The thick (thin) segments indicate significant (insignificant) correlations

Similar moving correlations are obtained for the Labrador Sea in control and forced experiments (Fig. 10, right panels). Generally positive SST-SSS correlations are obtained in control experiments with HadGEM3-HM (Fig. 10d), CESM1-3 (Fig. 10f) and EC-Earth3P (Fig. 10h), while weak negative correlation trend toward insignificant values in HadGEM3-HH (Fig. 10b). As for the Irminger Sea, SST-SSS correlations trend towards negative values in forced experiments with HadGEM3-HH and HadGEM3-HM, while trending from high to low positive values in EC-Earth3P. In contrast, SST-SSS correlations trend towards more positive values in CESM1-3.

Associated with positive–negative trends of SST-SSS correlations are implied transitions from density compensation (warm/salty or cold/fresh anomalies) to warm/fresh anomalies that reinforce each other in lowering surface density. Warm/fresh events do indeed become more commonplace in each region over later decades of the forced experiments with HadGEM3-HH (Figs. 7a-b), HadGEM3-HM (Figs. 7c-d) and EC-Earth3P (Figs. 7g-h). To further emphasize the emergence of these warm/fresh events in later decades, in Fig. 11 we plot monthly SST and SSS anomaly pairs through the forced experiments, colour-coding year. Notable in HadGEM3-HH (Fig. 11a-b), HadGEM3-HM (Fig. 11c-d) and EC-Earth3P (Fig. 11g-h) are shifts towards the upper left quadrant after around 2020, indicative of coincident warming and freshening. Consistent with previous evidence, this is less apparent in CESM1-3 (Fig. 11e-f), for which warming is the more dominant signature of change in later decades.

Fig. 11

Paired monthly SST and SSS anomalies in forced experiments with (a,b) HadGEM3-HH; (c,d) HadGEM3-HM; (e,f) CESM1-3; (g,h) EC-Earth3P, averaged across the Irminger Sea (a,c,e,g) and the Labrador Sea (b,d,f,h). Anomaly pairs are colour-coded by year (1950–2050). Three of the four quadrants are labelled to denote anomalous conditions that are ‘spicy’ (warm and salty), ‘minty’ (cool and fresh), and ‘springy’ (warm and fresh). The dashed arrows highlight minty-spicy variability that is density compensating and springy variability that reduces surface density

Three of the four quadrants are labelled to denote anomalous conditions that are: warm and salty (‘spicy’); cool and fresh (‘minty’); warm and fresh. In the latter case, associating warmth with a spring season, freshness with a natural spring, and noting the increasing amplitude of variability (larger SST and SSS anomalies), we describe this new character of T-S variability as ‘springy’. Dashed arrows highlight the springy variability that reduces surface density, orthogonal to the minty-spicy variability that is density compensating.

To further examine spatial patterns of change across mid-latitudes, in Fig. 12 we present ‘2040s minus 1990s’ differences in annual-mean SST and SSS in the four forced simulations. In all four models, the surface warms by around 1ºC across the basin, although slight cooling (by up to 0.5ºC) is evident in the eastern subpolar gyre of HadGEM3-HH (Fig. 12a), with minimal change in an isolated southeast region of the subpolar gyre in HadGEM3-HM (Fig. 12c) and CESM1-3 (Fig. 12e). Surface freshening of around 0.5 psu across much of the region is clear in three of the models. The exception is CESM1-3 (Fig. 12f), for which interior gyre SSS is little changed over 50 years, with an increase of around 0.5 psu in the East and West Greenland Currents that contribute to the offshore Labrador Current; the only freshening is found in the inshore Labrador Current, traced upstream to Davis Strait. Considering the areas defining the Irminger and Labrador seas, it is further evident how both surface warming and freshening arise in HadGEM3-HH (less so in the Irminger Sea), in HadGEM3-HM and in EC-Earth3P, while these seas in CESM1-3 are dominated by warming only. The patterns of SSS change further appear to be strongly shaped by the surface circulation, associated with enhanced Arctic freshwater transport and/or reduced Atlantic salt transport. This motivates the Lagrangian analysis in Section 3.3.

Fig. 12

Maps (mid-latitude North Atlantic) of 2040s minus 1990s differences in annual-mean SST (a,c,e,g) and SSS (b,d,f,h) in forced simulations: (a,b) HadGEM3-HH; (c,d) HadGEM3-HM; (e,f) CESM1-3; (g,h) EC-Earth3P. Also indicated are the two boxes used to average surface properties across the Irminger and Labrador seas

Before considering the changing lateral freshwater influences on SSS, we examine the basin-scale forcing of SST and SSS by air-sea heat and freshwater fluxes, and sea surface height above geoid (SSH), over selected forced simulations. In Fig. 13, we map the surface net downward heat flux (Wm⁻²) averaged over 1990–99 and 2040–49, along with the ‘2040s minus 1990s’ differences, in forced simulations with HadGEM3-HH and HadGEM3-HM (the simulations for which these fluxes are available to us). In both models, extensive heat loss is declining through the experiment. Annual-mean heat loss of around 100 Wm⁻² across most of the subpolar region in the 1990s has weakened by around 50 Wm⁻² in the 2040s. To first order, the surface warming (Fig. 12a,c) is driven by reduced surface heat loss. An exception to this mechanism in the eastern subpolar gyre of HadGEM3-HH, where slight cooling develops (Fig. 12a), suggests a secondary influence of reduced ocean heat transport convergence in this region that is likely associated with more substantial AMOC weakening in that model.

Fig. 13

Surface net downward heat flux (Wm⁻²) averaged over 1990–99 (a,b), over 2040–49 (c,d) and 2040s minus 1990s differences (e,f), in forced simulations with HadGEM3-HH (a,c,e) and HadGEM3-HM (b,d,f). Positive differences correspond to reduced heat loss or increased heat gain, in the 2040s relative to the 1990s. Also indicated are the two boxes used to average surface properties across the Irminger and Labrador seas

In Fig. 14, we map the corresponding changes in rates of precipitation, evaporation, and precipitation minus evaporation (P-E). Mean fields for HadGEM3-HH are provided in Supplementary Information (Figure S4); similar fields are obtained with HadGEM3-HM. Common to both HadGEM3-HH and HadGEM3-HM is a strengthened freshwater input (Fig. 14e,f), largely due to reduced evaporation (Fig. 14c,d), most evident in the eastern basin. Differences in the extent and spatial pattern of this strengthening are substantial. In HadGEM3-HH, P-E anomalies exceed 50 cm year⁻¹ across the eastern basin and much of the Irminger Sea, which corresponds to an approximate doubling of P-E over 2040–49 compared to 1990–99 (see Fig. S3). P-E anomalies are more subdued in HadGEM3-HM, with eastern basin extremes roughly half of those in HadGEM3-HH, and consistent with local cooling in the eastern subpolar gyre of HadGEM3-HH. These different responses to forcing are notably obtained with the same atmospheric model, the only difference being in ocean model resolution. Of further note, while large increases in P-E are evident in the eastern basin, relatively modest reductions are evident over the Labrador Sea. We therefore examine SSH as a proxy for the circulation.

Fig. 14

Changes in rates of precipitation (a,b), evaporation (c,d) and precipitation minus evaporation (e,f) (cm year.⁻¹), subtracting means over 1990–99 from means over 2040–49, in forced simulations with HadGEM3-HH (a,c,e) and HadGEM3-HM (b,d,f)

In Fig. 15, we map SSH averaged over 1990–99, over 2040–49, and 2040s minus 1990s differences, in forced simulations with HadGEM3-HH, HadGEM3-HM and EC-Earth3P (SSH data for CESM1-3 were not available to us). Differences (Fig. 15g,h,i) reveal complex and varying patterns of change between 1990s and the 2040s. In HadGEM3-HH and HadGEM3-HM, SSH falls by more than 10 cm across much of the eastern basin while rising by 5–10 cm in the central Labrador Sea, indicative of an eastward shift in the subpolar gyre. In contrast, SSH rises by 10–20 cm across the entire subpolar region of EC-Earth3P, although less so in the eastern basin. Notable in HadGEM3-HM (Fig. 15h) and EC-Earth3P (Fig. 15i) are increases of SSH along the Labrador coast of 5–10 cm relative to offshore change that can be traced northward as far as Davis Strait. Anomalously raised sea level along the western boundary can be associated with strengthened southward barotropic flow and an increased flux of fresh Arctic waters into the subpolar gyre. That this is not the case in HadGEM3-HH suggests that by the 2040s, fresher waters are exported from the Arctic of this model, in a flow that has not substantially strengthened.

Fig. 15

Sea surface height above geoid (m) averaged over 1990–99 (a-c), over 2040–49 (d-f) and 2040s minus 1990s differences (g-i), in forced simulations with HadGEM3-HH (a,d,g), HadGEM3-HM (b,e,h) and EC-Earth3P (c,f,i). Positive differences correspond to increased sea surface height, in the 2040s relative to the 1990s. Note the shifted and increased SSH range and the doubled SSH difference ranges, for EC-Earth3P

3.3 HadGEM3-HH forced simulation – Lagrangian analysis

With evidence for an increasing freshwater influence on surface density in the subpolar gyre in three of the four forced experiments, we further explore the extent to which this is remotely forced in HadGEM3-HH, in which the shift in surface properties is most pronounced. To specifically examine the changing provenance of surface waters in the region, we use TRACMASS (see Section 2.3) to obtain decadal back-trajectories for surface waters in the forced experiment, for contrasting periods: starting in 1990, prior to a decade when T and S anomalies compensate, and stratification is more dependent on SST; starting in 2040, when T and S anomalies are progressively combining to lower density, and stratification is more dependent on SSS.

The results of this backward tracking are summarized in Figs. 16 and 17, with additional background information provided in Supplementary Information (Figures S5 and S6). In Fig. 16, we show salinity differences tracking backward southward flows across the Irminger Sea (indicated in blue), between parcels sampled in 2040 and 1990. Negative differences correspond to fresher upper ocean conditions in 2040, evident across subpolar latitudes; in contrast, higher salinities prevail across subtropical latitudes in the decade preceding 2040. From our spatial analysis, it is evident that that, by 2040 (compared to 1990), fresher source waters are reaching the subpolar gyre (Fig. 16, Fig. S5) from shallower depths (Fig. S6), while higher salinity waters (transported poleward across ~ 32ºN) are more strongly freshened en route to subpolar latitudes. In Fig. 17, we summarize the corresponding transports (of all parcels) in 1990 and 2040, southward across the Irminger Sea section in Fig. 16 and partitioned within the 10-year timescale of back-tracking between four possibilities:

• surface input (precipitation)

• Arctic provenance (import through Davis Strait)

• subtropical provenance (import across ~ 32ºN)

• residual recirculation within subpolar and subtropical latitudes

Fig. 16

Salinity differences along particle trajectories obtained for the forced experiment with HadGEM3-HH, back-tracking southward flows across the Irminger Sea (indicated in blue) in 1990 and 2040; trajectories are terminated if reaching Davis Strait (black section) or the southern subtropics (red section), within 10 years; differences are taken as particle salinity obtained from back-tracking from 2040 minus that obtained back-tracking from 1990

Fig. 17

Total transports southward across the Irminger Sea section in Fig. 16 and partitioned within the 10-year timescale of back-tracking between: precipitation: southward through Davis Strait; northward across the subtropical Atlantic; residual recirculation within subpolar and subtropical latitudes. Percentage changes for 2040, relative to 1990 are indicated above the 2040 bars

The percentage changes of transport for 2040 relative to 1990, indicated above the 2040 bars in Fig. 17, emphasize the drivers of change in relative terms. Associated with an overall 17% increase in southward transport of surface waters across the Irminger Sea in 2040 (compared to 1990), is a 10% increase in the relative contribution of precipitation and a 50% increase in the contribution of inflow via Davis Strait, while the relative contribution of inflow from the southern subtropics decreases by 33%. The shift from higher salinity inflows to freshwater influx and lower salinity inflows is augmented by a 39% increase in the residual circulation. Associated with this is a likely increased residence time of surface waters in the subpolar region of net precipitation, partly accounting for the surface freshening implicit in Fig. 16.

No parcels drifting southward in the surface layer of the Irminger Sea are tracked back to Fram Strait. In a previous Lagrangian study of the ORCA12 hindcast (D. Dey, personal communication), focused on source waters for the AMOC lower limb, it was apparent that southward flow through Fram Strait mixes in a sub-surface layer with Atlantic Water. Those parcels back-tracked to Davis Strait have been transported via the Labrador Current, then entrained into the southern flank of the subpolar gyre, reaching the Irminger Sea via the cyclonic gyre circulation.

4 Discussion

We have focused here on stratification in two sensitive regions of the subpolar North Atlantic, in relation to surface temperature and salinity, the latter associated with low-salinity Arctic outflow and high-salinity Atlantic inflow. In an eddy-rich ocean model hindcast, we examined the spatial and time-varying character of PEA, as a metric for stratification of the Irminger and Labrador seas. In the hindcast, no clear link between PEA and surface temperature and salinity was obtained, although we established a shift away from surface density compensation since 2000. Through further analysis of surface properties and stratification in the Irminger and Labrador seas of four representative climate model simulations from the HighResMIP ensemble, we have identified more substantial transient changes in the first half of the twenty-first century.

Increasing freshwater input under global warming, of consequence for dense water formation and the AMOC, is now a well-established theme of climate model studies (Jackson et al. 2023). Previously unaddressed is the likely transition from historical surface density compensation to surface warming and freshening, evidenced here in three of four different high-resolution climate models. The changing freshwater influence is due to changes in surface freshwater fluxes, including sea ice freeze/melt, and changes in ocean freshwater transport. Net mass loss from the Greenland Ice Sheet (GrIS), typically invoked in model hosing experiments, is not included in the HighResMIP simulations diagnosed here. Forcing an eddy-rich ocean model with GrIS freshwater fluxes over 1990–2019, Böning et al. 2016) find that LSW formation rates only begin to reduce (relative to a control experiment) in the late 2010s; they conclude that cumulative freshening of the Labrador Sea sufficient for substantial (> 5 Sv) reduction of the AMOC may be reached around 2040.

In three of the four models, hydrographic variability in the surface layer of the Irminger and Labrador seas undergoes a fundamental transition. During the simulated second half of the twentieth century, coincident SST and SSS anomalies were ‘warm and salty’ or ‘cold and fresh’, combined as ‘spicy-minty’ T-S variability that ensured a high degree of density compensation. Corresponding anti-correlation of freshwater and heat content is evaluated in the ECCOv4 reanalysis product over 1992 to 2015 (Tesdal and Haine 2020). Into the twenty-first century, this variability is increasingly replaced by a coincidence of anomalous warm and fresh surface conditions – defined here as ‘springy’ variability. Although restricting our analysis to selected HighResMIP models of the highest available resolution, it is evident that our findings are model-dependent, with evolving T-S variability summarized as follows:

• HadGEM3-HH and HadGEM3-HM are both susceptible to a major shift away from surface density compensation, to warming and freshening, across both of the Irminger and Labrador seas;

• EC_Earth3P is less susceptible to warming and freshening, although density compensation more gradually disappears from the Irminger Sea;

• CESM1-3 is largely resilient to such change, with density compensation in fact strengthening in the Labrador Sea by the 2040s

The reasons for different changes in T-S variability are beyond the scope of this study, although it would be informative to further examine the initial state of the upper ocean heat and freshwater budgets in each model, prior to the 1950–2050 forced experiments, and the subsequent evolution of these budgets under forcing. This undertaking requires more model output than is currently available to us. Internal variability on decadal or longer timescales may also influence model responses. To resolve this influence would demand small ensembles with each model, currently too computationally expensive to undertake.

With a focus on the major shift in HadGEM3-HH, we undertook additional Lagrangian analysis of near-surface freshwater pathways. In Fig. 18, we summarize the drivers of the changes of upper ocean heat and freshwater budgets, T-S variability and hence stratification in this model, contrasting conditions in the 1990s and 2040s. The Lagrangian analysis (Section 3.3) revealed how, by 2040, high-salinity inflow to the subpolar gyre from subtropical latitudes has declined relative to fresh inflow from higher latitudes, lowering SSS. Meanwhile, surface net heat fluxes trended positive from 1990 to 2040s, raising SST. The consequence is reduced surface density and stronger stratification.

Fig. 18

Schematic representation of the drivers of heat and freshwater budgets, and stratification, in the subpolar North Atlantic of HadGEM3-HH: (a) 1990s; (b) 2040s

As stratification strengthens, there are consequences for the energetics of turbulent mixing, encapsulated in the PEA framework. More turbulent kinetic energy is lost in working against stronger buoyancy forces and the associated potential energy demand. This will reduce the downward mixing of heat and freshwater. On this basis, we might expect that the impact on stratification of small changes in surface heat fluxes or horizontal freshwater import may be amplified through further suppression of vertical mixing. We may regard this process as ‘Arctification’, reminiscent of the Beaufort Gyre, where freshwater stratification is currently strong enough to suppress vertical mixing. SST and SSS anomalies are thus the combined result of changing surface fluxes, advection, and vertical mixing. In coupled context, SST and SSS anomalies play distinct roles. While SST anomalies directly impact the atmosphere through air-sea interaction (longwave, latent and sensible heat fluxes), SSS anomalies may further impact SST and such feedback through changes in stratification.

The changing surface conditions and stratification may relate more generally to dense water formation and the large-scale circulation. The clear trend towards higher PEA values in all four forced experiments (see Fig. 8) indicates shoaling winter mixed layers and reduced dense water formation rates in the Irminger and Labrador seas. Expectation of a corresponding reduction in the AMOC is confirmed by the downward trends reported in the HighResMIP experiments examined here (Roberts et al. 2020), with % declines by 2050 as follows: by around 30% in HadGEM3-HH and HadGEM3-HM; by around 10% in CESM1-3; by around 20% in EC-Earth3P. In broad terms, the varying extent of these declines are consistent with varying impacts of warming and freshening (in particular) on stratification in the Irminger and Labrador seas; in those experiments subject to more freshening, AMOC reductions are larger.

A limited number of model studies have addressed how ongoing and future changes in the Arctic may drive changes in the ocean and atmosphere at lower latitudes. In a pioneering study, the ECHAM5/MPI-OM climate model, with minimum ocean grid dimension of about 15 km around Greenland and 40 vertical layers, was used in a 3-member ensemble through the twentieth and twenty-first centuries (subject to IPCC scenario A1B) to evaluate changing Arctic freshwater export (Koenigk et al. 2007). While export through Fram Strait changed little, that through the Canadian Archipelago steadily increased during the twenty-first century, the Labrador Sea progressively freshened, convection was strongly reduced, and the AMOC steadily weakened from an initial 22 Sv to 16 Sv by 2100.

Using an early version of the EC-Earth model with the ORCA1 ocean component of resolution 1º, in a 12-member ensemble of 1850–2100 simulations, Brodeau and Koenigk (2016) obtained a stepwise shutdown of convection in the Labrador Sea and a progressive weakening of the AMOC; this was primarily associated with reduced surface cooling, although this ‘combines with an increasing influx of freshwater from the Nordic Seas to rapidly strengthen the surface stratification and prevent any possible resurgence of deep convection in the Labrador Sea after the 2020s.’ Analysing Arctic freshwater storage and fluxes in seven climate models from the Coupled Model Intercomparison Project Phase 6 (CMIP6), over the historical period (1980–2000) and under future emissions scenarios SSP1-2.6 and SSP5-8.5, Zanowski et al. (2021) find considerable disagreement between models, and with observations over 1980–2000, but general agreement on increased liquid freshwater export through Fram and Davis straits by the end of the twenty-first century; increases in multi-model mean freshwater export start around 2000 at Fram Strait, but later, around 2050, at Davis Strait.

Following evaluation of the ECCOv4 reanalysis, Tesdal and Haine (2020) proposed that Arctic Ocean freshwater will become ‘an increasingly important source of future freshwater increases in the subpolar North Atlantic’. A recent review of past and future Arctic-Subarctic ocean linkages highlights that Arctic outflow salinity fell to record lows in the Fram and Davis straits during the 2010s, while CMIP6 models project further increases in freshwater export through Fram Strait associated with both increasing volume transport and decreasing salinity (Wang et al. 2023). Given the clear and growing influence of the Arctic on the subpolar Atlantic, substantial differences in the mean state and the response of Arctic sea ice to forcing are likely to be important factors in explaining differences between the four models analysed here. In an analysis of HighResMIP models (Selivanova et al. 2023), present-day sea ice is substantially thicker in EC-Earth3P than HadGEM3-HM, in which thicknesses are closer to observations. In the 1950–2050 forced experiments, rates of decline in sea ice area and volume are substantially faster in EC-Earth3P over 1950–2014 and substantially slower over 2015–50, than in HadGEM3-HM (see Table 2 in Selivanova et al. 2023).

Past and future simulations are clearly also sensitive to model resolution. Following previous findings in a study of resolution dependence (ORCA1 and ORCA025 ocean configurations; N96 to N512 atmospheric configurations) in historically forced simulations, Fuentes-Franco and Koenigk (2019) identified that the impact of Arctic freshwater export on North Atlantic convection increased with atmospheric resolution. However, higher resolution comes with some caveats. In recent evaluation of Labrador Sea stratification and convection in seven HighResMIP models at high, medium, and low resolution, Koenigk et al. (2021) concluded that stratification was more realistic, but convection over-estimated, in the high-resolution configurations. Mixed layers are excessively deep in HadGEM3-HH and HadGEM3-HM, while closer to observations (ARGO, EN4) in EC-Earth3P. Although CESM1 is not included in the Koenigk et al. (2021) study, lower resolution counterparts are included in the recent study of Liu et al. (2024), in which CESM2 (closest to the high-resolution counterpart that we analysed) is characterised by upper ocean warm and salty biases, and excess January-May surface heat loss over the Labrador Sea (Fig. 9 in Liu et al. 2024) that leads to too-extensive sea ice and no convection, which is instead excessive in the Irminger Sea. Of the four models that we analysed, smallest Sep-Mar PEA differences – in 1950 of the control simulation – in the Labrador Sea of CESM1 (see Fig. 6k) suggest that some of these biases persist at higher resolution in this model family.

Given a limited number of available simulations in the HighResMIP ensemble, with only two featuring eddy-rich oceans and the challenging logistics of data analysis with such high-resolution output fields for selected variables, our analysis is necessarily limited in statistical terms. Strong natural variability on interannual to sub-decadal timescales is evident in the single HighResMIP ensemble members selected here. It would be prudent to scale up the present analysis as more high-resolution climate simulations are available, emphasizing again – with a focus on sensitive buoyancy budgets – the imperative to resolve narrow boundary currents and eddy property fluxes.

In summary, three of four representative HighResMIP simulations indicate that the Arctic and its freshwater may play a more substantial role in subpolar North Atlantic variability in coming decades, of consequence for winter convective mixing, formation of dense water, and by implication the meridional overturning circulation. Associated impacts on atmospheric state and dynamics, via air-sea interaction, are also likely to emerge. Where additional freshwater strengthens stratification, associated surface warming of a shallower mixed layer may increase heat and moisture fluxes to the atmosphere. This may in turn enhance cyclogenesis and precipitation rates. ‘Springiness’ in variability of sea surface temperature and salinity may thus be integral to a positive feedback loop in a warming climate.