Introduction

The North Atlantic (NA) Ocean plays a crucial role in the climate system by linking the tropics with the Subpolar NA region, connecting the atmosphere with the deep ocean, and integrating the northward and southward limbs of the Atlantic Meridional Overturning Circulation (AMOC). This region is where North Atlantic Deep Water (NADW) is formed, driving the lower branch of the AMOC1. Under a warming climate, significant changes are expected in this region2,3. Earth System Models (ESMs) and various studies predict a slowdown, or even a collapse of the AMOC due to the warming and freshening of the deep-water formation regions in the NA, driven by increased surface heat flux into the ocean, Arctic sea ice loss and shifts in the precipitation-evaporation balance3,4,5,6. Such changes could impact global climate in several ways, ranging from moderate scenarios where AMOC slowdown results in a slightly reduced pace of warming in the Northern Hemisphere7, to extreme cases where AMOC collapse could lead to temperature drops of around 10 °C over Europe8.

However, observational evidence of AMOC slowdown is conflicting. Direct measurements of AMOC strength, obtained through water transport measurements by the Rapid Climate Change-Meridional Overturning Circulation and Overturning in the Subpolar North Atlantic Program arrays, have not shown significant weakening of the upper limb9,10, while there is evidence of a slowdown of the abyssal limb11. These observations are limited to the past two decades and are confounded by internal variability of the climate system. Researchers also use indirect methods to study AMOC, such as analyzing fingerprints, estimates of the ocean’s forced response, to identify characteristic patterns associated with AMOC dynamics, commonly referred to as “AMOC fingerprints”12,13,14,15,16,17,18,19,20,21,22. For example, reconstructions of AMOC strength based on reanalysis of sea surface temperatures in the NA suggest a decline over the past 50–70 years16. While other proxy-based studies have not observed such a pronounced multidecadal decrease19,21,23,24.

Expected changes in the NA include a distinct climate change fingerprint associated with AMOC slowdown and reduced heat transport into NA — a “warming hole” pattern in subpolar temperature14,16,25 freshening trends in subpolar NA salinity26,27 and heat and salt accumulation in the NA subtropics15,16,26,27. While surface NA and subtropical trends have supported the notion of AMOC weakening, the “warming hole” pattern has not been directly observed and analyzed within the NA ocean interior.

This study presents a joint analysis of observations and models, enabling the statistical detection of climate change-related changes and their attribution in temperature and salinity fields within the upper 2000 m of the NA domain from the 1950s to the present. Our results provide insights into AMOC dynamics under climate change. We show that models historically underestimate observed trend patterns in temperature and salinity within the NA. Importantly, we show that the subpolar NA “warming hole” fingerprint can already be detected within the deeper ocean in observations while it only emerges in model projections towards the mid-21st century.

Results

Comparing changes in temperature and salinity in observations and models participating in the Detection and Attribution Model Intercomparison Project (DAMIP)

We start with analyzing (i) Argo and World Ocean Atlas (WOA) observations and (ii) twelve Coupled Model Intercomparison Project Phase 6 (CMIP6) ESMs participating in the DAMIP project28 to understand recent changes in the NA region. DAMIP project provides ESM outputs under single forcing scenarios such as: greenhouse gasses (GHG) only, anthropogenic aerosols (AER) only, ozone depletion (O3) only and natural forcing (NAT) only, covering the 1850 to 2020 period.

Observed changes in temperature (Fig. 1a, c) and salinity (Fig. 1b, d) from 1955 to 1974 (WOA climatologies) to 2002–2020 (from Argo) are shown in Fig. 1a, c and are denoted in this paper as \({Y}_{{Obs}}^{2020}\). Both depth averaged (Fig. 1a) and zonal mean patterns of change in temperatures (Fig. 1c) show warming in lower to mid-latitudes with a hot spot around ~35–45°N. Higher latitudes show a cooling signal at 50–60°N reaching into the deep ocean to at least 1000 m. This observed temperature is consistent with a decline in AMOC and weaker NADW formation. A weaker AMOC is expected to bring less heat from the southern hemisphere and tropics northward29,30, and results in a subpolar “NA cold spot” or “NA warming hole”14,25,31. Similar patterns of change are also observed in the EN4 ocean analysis32 over the same time periods (Fig. S1).

Fig. 1: Observed changes in temperature and salinity between 2002 and 2020 (mean of Argo profiles) and 1955–1974 (WOA climatology) (\({Y}_{{Obs}}^{2020}\)).
figure 1

Depth (top 2,000 m) averaged temperature (a) and salinity changes (b). The zonally averaged meridional profiles of NA temperature (c), and salinity changes (d). Gray crosses show where the observed changes fall outside the 5th to 95th percentile spread across ESM changes under ALL scenario.

Full size image

The Gulf Stream brings warm and salty subtropical waters northward, while sinking NADW brings saltier water into the deep and moves it southward. From 1955–1974 to 2002–2020, the observed surface to 2000 m changes in salinity (Fig. 1b, d) show a general pattern of salinity increase in the subtropics and mid-latitudes with maxima over the Gulf Stream pathway and a decrease in the subpolar domain. This is consistent with prior observations26, and an already observed intensification of the hydrological cycle: more evaporation-precipitation in the subtropics and more precipitation-evaporation in the subpolar region33.

The multi-model average of patterns of change in temperature from 1955–1974 to 2002–2020 is labeled as \({X}_{{ALL}}^{2020}\), where 2020 denotes the last year for trend calculation and ALL refers to the CMIP6 historical runs up to year 2014 and the Shared Socioeconomic Pathways (SSP5-8.5) scenario beyond 2014 with all forcings included. \({X}_{{ALL}}^{2020}\) shows a general warming pattern with a hotspot located at ~35–45°N and another at ~60–70°N (Fig. 2a, c).

Fig. 2: Simulated changes in temperature and salinity under ALL forcing between 2002–2020(\({X}_{{ALL}}^{2020}\)) (left-hand column), or 2031–2050(\({X}_{{ALL}}^{2050}\)) (right hand column) and the base period of 1955–1974.
figure 2

Depth (0-2000 m) averaged NA temperature (a, b) and zonally averaged meridional profiles of temperature (c, d). Depth averaged NA salinities (e, f) and zonally averaged profiles of salinity (g, h). ALL refers to the CMIP6 historical runs (up to yr 2014) and the scenario SSP5-8.5 beyond 2014 with all forcings included. Dots show where more than 60% of models agree on the sign of changes.

Full size image

In contrast to the observed trend (Fig. 1a, c), the current simulated trend up to 2020 does not show a cooling spot in the subpolar Atlantic. Additionally, the freshening observed in the subpolar Atlantic (Fig. 1b, d) is not present in the ensemble mean (Fig. 2e, g). Instead, models produce a salinity increase throughout the North Atlantic. While the current patterns of change simulated by the twelve ESMs do not show a distinct “warming hole” pattern, the future changes, from 1955–1974 to 2031–2050 (\({X}_{{ALL}}^{2050}\)), demonstrate the emergence of a cooler and fresher patch in the NA. As time progresses, we observe a gradual intensification of warming everywhere in the ocean (surface and subsurface) south of 50°N (Fig. 2b, d). Additionally, we see the gradual emergence of a “neutral region” around 50°N-60°N, which we recognize from previous literature as the “NA warming hole”25. This “neutral region” is also evident within the water column, and propagates from the surface ~55°N towards the deeper ocean ~60°N (Fig. 2c, d).

Future changes in salinity (1955 to 2050) show a much more distinct fingerprint in the NA, consistent with previous studies26,27 (Fig. 2f, h). Figure 2 shows a gradual salinification of the ocean (surface and subsurface) south of 50°N, as well as the gradual emergence of a negative salinity trend north of 50°N, both at the surface and with depth. The largest drop in salinity is observed at the surface, consistent with a gradually intensifying hydrological cycle and resulting surface freshening over the 21st century33. In contrast to temperature, patterns of change in salinity show a much stronger contrast between higher and lower latitudes both in depth-averaged fingerprints (Fig. 2e) and in zonal mean fingerprints deeper within the water column (Fig. 2h). Freshening of the subpolar NA stratifies the water column; this is consistent with fresher NADW, and also with less NADW formation29,30. This fresher NADW anomaly propagating southward at depth is clearly noticeable in the observations in Fig. 1d and future predictions of ESMs (Fig. 2h).

To summarize, while observations (\({Y}_{{Obs}}^{2020}\)) show a “warming hole” and freshening of the subpolar NA for the 1955–2020 period, these patterns are clearly underdeveloped in DAMIP multi-model simulations (\({X}_{{ALL}}^{2020}\)). The DAMIP ESMs are underestimating the current observed changes. However, as we run the models further into the future, the emerging patterns of change become increasingly more similar to the observed ones, such that \({X}_{{ALL}}^{2050}\) is visually similar to \({Y}_{{Obs}}^{2020}\).

Detection and attribution of observed changes in temperature and salinity

To confirm our visual intuition, we perform a detection and attribution of the observed changes34,35. Detection involves demonstrating that our observed pattern of change (changes represented by the difference of 20-year averages for the early 21st century minus the 1955–1974 base period) in the data is statistically different from natural variability. Attribution means finding which of the possible single forcings is responsible for the detected change.

Our analysis follows ref. 35 and adopts a simple regression technique to estimate scaling factors, where we regress observed patterns of zonally-averaged temperature or salinity fields (\({Y}_{{Obs}}^{2020}\) in Fig. 1c, d) onto modeled patterns for model runs with all forcings turned on (\({X}_{{ALL}}^{20{XX}}\) in Fig. 2c, d), i.e., \({Y}_{{Obs}}^{2020}\) = β * \({X}_{{ALL}}^{20{XX}}\) + Intercept, details in Methods (Eq. 1). Resulting regression ‘scaling factors’ (β in Eq. 1) are shown in Fig. 3 under blue shading. Observational changes are detected when β is close to 1 and significantly different from 0.

Fig. 3: Detection-attribution analysis.
figure 3

Temperature (a) and salinity (b) scaling factors for zonally averaged meridional profiles. See methods for details. In blue shading are results under ALL forcings for the one-signal analysis (\({X}_{{ALL}}^{2020}{,X}_{{ALL}}^{2030}{,X}_{{ALL}}^{2040}{,X}_{{ALL}}^{2050}\)), where the year listed on the caption is the last year of the period used for the trend calculation). In yellow shade are results under DAMIP scenarios: multi-signal analysis for the GHG, AER, NAT and O3 changes in temperature and salinity (\({X}_{{AER}}^{2020}{,X}_{{GHG}}^{2020}{,X}_{O3}^{2020}{,X}_{{NAT}}^{2020}\)). The 90% confidence intervals are shown by the grey bars. To attribute changes in the observed data to various forcings we look for scaling factors β that are close to unity and significantly different from 0, as well as high correlation coefficients between the observed and simulated fingerprints. Correlation coefficients shown at the bottom of the figure.

Full size image

We find a value of ~0.34/0.31 for \({\beta }_{{ALL}}^{2020}\) for historical changes in temperature/salinity fields, respectively (Fig. 3), suggesting that the observed zonal mean pattern of climate change is not well matched by the models for the 2002–2020 period. If that is the case, will the observed climate change pattern be matched better as the models are run further? We next regress \({Y}_{{Obs}}^{2020}\) onto future trend patterns \({X}_{{ALL}}^{2030}\),\({X}_{{ALL}}^{2040}\),\({X}_{{ALL}}^{2050}\) to retrieve corresponding scaling factors. As seen in Fig. 3a, b (blue shading), β gradually increases as we analyze later periods. The best match for temperature is between observed Argo patterns of change and the modeled patterns between years 2031–2050 and 1955–1974 \({(X}_{{ALL}}^{2050}\)) with \({\beta }_{{ALL}}^{2050}\) = 1.1 and a 0.8 correlation coefficient. The best match for salinity is between the observed and the modeled patterns for years 2021–2040 minus 1955–1974 \({(X}_{{ALL}}^{2040}\)), with \({\beta }_{{ALL}}^{2040}\) = 0.96, and a 0.73 correlation coefficient. This indicates a delay of ~25 years in the evolution of the fingerprint in CMIP6 models and confirms our prior visual intuition. Similar results are achieved by using a larger number of CMIP6 ESMs (Fig. S2), which confirms the robustness of our analysis. We also repeated the analysis using only ESMs with AMOC greater than 18.5 Sv under NAT scenario and found negligible changes in the main results of this paper.

We additionally consider observed surface to 2000 m depth-averaged patterns of change in temperature and salinity (\({Y}_{{Obs}}^{2020}\) in Fig. 1a, b). We regress these onto corresponding modeled patterns (\({e.g.X}_{{ALL}}^{20{XX}}\) in Fig. 2a, b, e, f). The resulting scaling factors (Fig. S3) similarly show that the observed changes are not well matched by the models for the 2002–2020 period, however, the fit improves (with scaling factors closer to 1) when considering future modeled fingerprints.

Attribution involves determining the relative contributions of various climate forcings to the total observed change and assigning statistical confidence to these contributions34. We used single-forcing simulations provided by the DAMIP project in attribution analysis (see Methods). For each DAMIP run from the multi-model ensemble we calculated the mean of temperature or salinity for 2002–2020 minus 1955–1974. These are the single forcing fingerprints due to greenhouse gases (labeled GHG), aerosols (AER), natural variability (NAT) and ozone (O3), shown in panels c-f in Figs. 4, 5.

Fig. 4: North Atlantic zonal mean changes in temperature.
figure 4

North Atlantic zonal mean changes in temperature for observed (a) and simulated fingerprints (b-f). Observed difference (a) refers to the difference between Argo (2002–2020) and WOA (1955–1974) measurements. Simulated fingerprints are multi-model differences between the (2002–2020) and (1955–1974) periods from separate runs: the ALL forcing runs in CMIP6 models \({X}_{{ALL}}^{2020}\) (b) and the single-forcing runs under DAMIP scenarios: GHG trends, \({X}_{{GHG}}^{2020}\)(c); AER trends, \({X}_{{AER}}^{2020}\)(d); NAT trends\({,X}_{{NAT}}^{2020}\)(e); and O3 trends,\({X}_{O3}^{2020}\)(f).

Full size image
Fig. 5: North Atlantic zonal mean changes in salinity.
figure 5

Same as Fig. 4 but for changes in salinity.

Full size image

The modeled GHG fingerprints (\({X}_{{GHG}}^{2020}\)) are very similar to the observed patterns \({Y}_{{Obs}}^{2020}\) for both temperature (Fig. 4a, c) and salinity (Fig. 5a, c). The \({X}_{{AER}}^{2020}\) fingerprints oppose the \({X}_{{GHG}}^{2020}\) to a large degree (Figs. 4d, 5d). The \({X}_{{NAT}}^{2020}\) and \({X}_{O3}^{2020}\) fingerprints are very small (Figs. 4e, f, 5e, f). Next, we regressed \({Y}_{{Obs}}^{2020}\) onto our four single forcing fingerprints (\({X}_{{GHG}}^{2020}{,X}_{{AER}}^{2020}{,X}_{O3}^{2020}{,X}_{{NAT}}^{2020}\)) via a multiple linear regression (Eq. 2). Figure 3 (yellow shade) displays the calculated attributions (scaling factors \(\beta\)) for different single forcing fingerprints. The GHG contribution explains most of the observed trend in temperature with with \({\beta }_{{GHG}}^{2020}\) slightly above 1. For salinity fingerprints \({\beta }_{{GHG}}^{2020}\) is close to 0.6, however the uncertainty ranges also include 0, which does not allow statistically significant attribution of the salinity fingerprints. The spatial depth averaged fingerprints also indicate a successful attribution of \({X}_{{GHG}}^{2020}\) for temperature (with \({\beta }_{{GHG}}^{2020}\) ~ 0.95), while salinity has \({\beta }_{{GHG}}^{2020}\) ~ 0.5 with an uncertainty range including 0 (Fig. S3). The scaling factors for the other single-forcings — \({X}_{{AER}}^{2020}\), \({X}_{{NAT}}^{2020}\), and \({X}_{O3}^{2020}\), for both temperature and salinity — are close to 0 and have a very large spread of uncertainty indicating insignificant attribution results.

The correlation coefficients between each of the simulated fingerprints and the Argo fingerprint are shown in Fig. 3 (bottom row). Like the scaling factors, correlation coefficients gradually increase as we extend the ESM runs further into the future —rising from 0.32 to 0.80 for temperature and from 0.69 to 0.78 for salinity. The GHG single forcing fingerprints show high correlation with \({Y}_{{Obs}}^{2020}\) for both temperature and salinity fields. While the rest of single forcing fingerprints do not show significant correlations for changes in temperature, the fingerprint under aerosol (AER) single forcing exhibits a strong negative correlation for changes in salinity (r = −0.67). This suggests that the observed changes in salinity may have been offset by aerosol forcings in recent decades.

Discussion

In response to GHG forcing, the pattern of a subpolar “warming hole” emerges gradually in the DAMIP modeled fingerprints (Figs. 4c, 5c; Fig. S4c; Fig. S5c). The historical changes under single-forcing GHG (\({X}_{{GHG}}^{2020}\)) are very similar to the Argo patterns of change (\({Y}_{{Obs}}^{2020}\)) in both temperature and salinity. Both \({Y}_{{Obs}}^{2020}\) and \({X}_{{GHG}}^{2020}\) results exhibit a dipole structure over the historical period, with warming and salinification in lower latitudes and freshening and cooling in higher latitudes. Additionally, depth-averaged trend patterns match well between \({Y}_{{Obs}}^{2020}\) and \({X}_{{GHG}}^{2020}\), especially in the salinity field (Fig. S5). These patterns align with current dynamics in the NA and anticipated consequences of AMOC slowdown. A strong decrease in salinity could lead to reduced NADW formation, which in turn weakens the upper branch of the AMOC36, further resulting in less heat and salt transport to higher latitudes29.

On the one hand, the response of the bulk NA ocean to AMOC weakening could be different than that for the upper ocean alone. For example, a weaker AMOC could result in reduced formation of NADW, which may, in turn, lead to deep ocean warming37. On the other hand, reduced heat transport could lead to negative temperature anomalies that extend into the deeper ocean17. Our study shows a column-wise decrease in temperature north of approximately 50°N, and warming to the south (Figs. 1, 2; Fig. S1 and S6), consistent with recent studies reporting column-wise negative temperature and heat content trends in this region17,38,39,40.

Furthermore, salinity changes in the bulk Atlantic Ocean, which generally follow temperature patterns, also serve as indicators of AMOC weakening26,27,40,41. A weaker AMOC contributes to the salinification of the subtropical Atlantic, which eventually leads to accelerated ocean heat uptake and the sequestration of this heat into the deeper ocean down to 2000 m42. In contrast the subpolar North Atlantic becomes fresher and cooler due to reduced heat and salt transport.

A recent study demonstrated that ESMs that simulate a weakening of the AMOC over the historical period also exhibit a ‘warming hole’ pattern, unlike ESMs that do not simulate AMOC weakening40. This provides evidence that a weaker AMOC is the primary driver of the NA warming hole in ESMs, and shows that the associated cooling and freshening signals also propagate into the deeper ocean. Our results align with these findings, indicating general deep ocean warming outside the subpolar NA, while a weakened AMOC leads to cooling and freshening within the subpolar NA.

In response to AER forcing in the models in 2002−2020 compared to 1955−1974, we observe cooling (a direct effect of aerosols43) and freshening (due to aerosols affecting the precipitation-evaporation balance33) in the subtropical domain up to 1000 meters. There is a noticeable warming and increased salinity around 50°N throughout the water column, as well as in the subtropics below 1000 meters south of 50°N (see Figs. 4d, 5d). This NA response to aerosols is also evident in the spatial top 2000 m depth-averaged patterns, spanning longitudes 60°W to 20°W along the Gulf Stream path (see Figs. S4 and S5).

The effects of aerosols on the AMOC exhibit a two-stage pattern in DAMIP models44: an initial strengthening of the AMOC due to increased aerosols and associated cooling before 1970, followed by a decline in AMOC due to decreasing aerosol emissions from North America and Europe, partially offset by increasing emissions from Asia43. This two-stage aerosol effect aligns with a stronger northward advection of Gulf Stream water from the south during 1955–1974 and a reversal during 2002–2020. Stronger AMOC under higher aerosol conditions leads to increased deepwater formation, which more efficiently transports heat and salt to the deeper ocean, resulting in reduced salinity and temperature in the upper subtropical ocean45. The temperature and salinity fingerprints in Figs. 4d, 5d agree with this aerosol-driven mechanism.

In response to natural variability and ozone effects, the scaling factors for simulated patterns of change are approximately 0 for both temperature and salinity, except for a negative scaling factor for NAT forcing in the case of salinity, which however has a very large uncertainty interval (Fig. 3a, b). For salinity, changes driven by AER, NAT and O3 are all negatively correlated with \({Y}_{{Obs}}^{2020}\), offsetting the strong impact of GHG changes on the observed patterns (Fig. 3b). In contrast, the observed temperature trend can be attributed solely to GHG forcing, since the scaling factors and correlations for the other single-forcing fingerprints are insignificant and close to zero.

As we move further into the 21st century, the impact of GHG is expected to gradually surpass the effects of AER, with the other components (NAT, O3) potentially becoming smaller. This is evident as the changes projected for 2050 (\({X}_{{ALL}}^{2050}\)) under full forcing scenario closely resemble those associated with single forcing under greenhouse gases (\({X}_{{GHG}}^{2020}\)) for the historical period, both for temperature and salinity (compare Fig. 2d, h with Figs. 4, 5a, c). Additionally, the spatial depth averaged changes between historical observational patterns of change (\({Y}_{{Obs}}^{2020}\)), historical changes under greenhouse gases forcing (\({X}_{{GHG}}^{2020}\)) and future changes (\({X}_{{ALL}}^{2050}\)) match quite well, especially for salinity fields (Fig. S5a,c; Fig. 2f; correlation coefficients on Fig. S3b).

The discrepancy and delayed evolution of the warming hole in \({X}_{{ALL}}^{20{XX}}\) may be attributed to the fact that the CMIP6 experiments analyzed in this study do not include the effects of Greenland meltwater. The inclusion of Greenland ice melt has been shown to have a more pronounced influence on AMOC strength compared to simulations without additional freshwater forcing, leading to stronger response in NA salinity and temperature patterns46,47,48,49. Incorporating Greenland meltwater in model simulations could therefore result in an earlier emergence of the warming hole in NA and potentially allow successful detection for \({X}_{{ALL}}^{2020}\) patterns of change.

The strength of the AMOC under different single-forcing scenarios also reveals clear opposing effects of AER and GHG (see Fig. 6a). Furthermore, the AMOC strength under NAT matches closely the historical output under ALL forcings until year 2020 (Fig. 6a), highlighting the difficulty of trend detection in models for 1955–2020 period. Notably, the decline in AMOC strength in the historical scenario begins around 2010. Therefore, in models we will not see major AMOC-related fingerprints emerging in the 2002–2020 decade; instead, we have to wait until later decades, e.g., 2011–2031; 2021–2040, 2031–2050. This is consistent with our attribution analysis for NA temperature and salinity fingerprints (Fig. 3). Using the mean AMOC under NAT forcing, we computed the anomalies of AMOC under AER and GHG single forcings, as well as AMOC under the ALL forcing scenario (Fig. 6b). The sum of AMOC anomalies under AER and GHG single forcings closely resembles the historical ALL output. This also confirms that AER and GHG have opposing effects on AMOC for some of the decades in the analysis (Fig. 6b), a point also emerging in recent studies27,43.

Fig. 6: AMOC strength in 3 separate DAMIP experiments and CMIP6 simulations.
figure 6

a Time series of multi-model ensemble-mean AMOC strength at 30°N for the 1950-2020 time period for single-forcing DAMIP runs: AER (orange), NAT (green), GHG (red). Lines represent the average of the 12 DAMIP models in Table S1. Also shown in grey is the “ALL” simulation for 1950–2050. “ALL” includes all forcings and represents the historical simulation up to 2014 followed by the SSP5-8.5 scenario for CMIP6 models (average over the 12 models in Table S1). b AMOC historical anomalies relative to the NAT forcing (full black line) is well approximated by the sum of AMOC anomalies due to aerosols and greenhouse gas (dotted black line). The latter is calculated as (AER – NAT) + (GHG – NAT). The 1955–1974 and 2002–2020 periods used in the paper to calculate climate fingerprints are shown with gray shading. Note the clear trend in the GHG from one period to the next.

Full size image

The AMOC variability complicates the assessment of its state. The “warming hole” fingerprint, discussed in the literature, is often associated with the AMOC slowdown14,16. However, the North Atlantic “warming hole” is also believed to be significantly influenced by climate variability, such as the Atlantic Multidecadal Oscillation (AMO27). Our analysis spans the periods 1955–1974 and 2002–2020, which encompass multiple positive and negative phases of the AMO. 1955–1974 is a principally positive AMO period, while 2002–2020 is a positive AMO period. To clarify the role of AMO, we additionally analyzed the changes from the negative AMO 1975–1994 period to the positive 2002–2020 period. Despite choosing a different AMO state as a starting point, our analysis still reveals a distinct “warming hole” pattern in the bulk (surface to 2000 m averaged) response of the NA Ocean, evident in both temperature and salinity (Fig. S6). We argue that considering bulk temp and salinity indices for the warming hole allows us to circumvent the “noise” of interdecadal variability which is typically associated with surface NA fingerprints27.

The evolution of the NA salinity pattern south of 50°N from 1955 to 2020 has opposing responses in the GHG and AER fingerprints (Fig. 5c, d). As discussed earlier, a stronger AMOC under AER forcing leads to negative salinity changes in the subtropics, as it enhances the salinity transport northward. In contrast, under GHG forcing, a weaker AMOC results in salinification of the subtropics due to reduced northward salt transport. These results suggest that salinity changes in the NA may be predominantly influenced by AER effects, which can counteract and obscure the salinity signals driven by GHG, thereby increasing attribution uncertainty. By comparison, although temperature changes are also affected by AER forcing, it is still possible to confidently attribute the observed NA temperature fingerprint to GHG forcing.

Our study confirms that the “warming hole” pattern recorded in the observational dataset can largely be attributed to GHG emissions. We show that the CMIP6 models underestimate the current observed changes in (top 2000 m) NA ocean temperature and salinity. We find that models only ‘catch up’ to the current ocean state by year 2040 for salinity or 2050 for temperature. This result also agrees with our findings for AMOC dynamics. ESMs under historical and SSP scenarios project a weak decline in AMOC from 1950 to 2020, with a more abrupt decline happening in further years, while AMOC forced only by GHG forcing shows an accentuated decline over this period (compare gray and red lines in Fig. 6a).

Correctly modeling the spatial temperature and salinity fingerprints of the NA is essential for understanding AMOC dynamics, the interaction between the ocean and atmosphere, and the resulting climate modes27,50,51,52. The temperature pattern has been shown to influence climate sensitivity in CMIP5 and CMIP6 models53,54. Reference 55 demonstrated a direct response of climate feedbacks to the NA warming hole under abrupt CO2 forcing. Reference 42, in turn, showed that salinity changes in the NA affect ocean heat uptake and also contribute to temperature trends in the region. These changes affect the buoyancy budget in the NA, which controls deep water formation, AMOC, and ocean-atmosphere variability in the entire Atlantic Ocean26,27,50,56. Since climate models struggle to reproduce recent NA change patterns, this introduces significant uncertainty in climate projections that needs to be addressed.

Observational data on AMOC strength is limited to the last ~20 years and does not show a clear trend9,10. In contrast, the observational fingerprint of North Atlantic temperature and salinity is based on a longer time span, capturing changes over the last ~70 years. This longer timeframe may indicate a slowdown of the AMOC, as has been suggested by recent studies14,16,27.

Methods

Observations

The observational period for this study is limited by the availability of Argo profiles. We utilized a gridded Argo product from the In Situ Analysis System (ISAS), which employs an optimal interpolation method to produce gridded temperature and salinity data57,58. The product spans from 2002 to 2020 and incorporates all available Argo profiles. To calculate changes in temperature and salinity, we used the World Ocean Atlas 2018 (WOA) climatology as a reference base period59. We averaged the WOA climatologies for the periods 1955-1964 and 1965–1974 to estimate NA temperatures at the beginning of the observational record, which we identified as our base period. We then subtracted this base period estimate from the average of the 2002–2020 Argo data to calculate observed changes in the NA region. The result is the observed fingerprint for temperature (Fig. 1a, c) and salinity (Fig. 1b, d), which we refer to as \({Y}_{{Obs}}^{2020}\). To confirm the \({Y}_{{Obs}}^{2020}\) patterns of change we also used EN4 temperature and salinity data32 (Fig. S1).

Historical and SSP5-8.5 scenario

DAMIP is the Detection and Attribution Model Intercomparison Project28, a component within the larger Coupled Model Intercomparison Project Phase 6 (CMIP6). We used up to 36 ensemble members (Table S1) from 12 Earth System Models (Table S1) participating in DAMIP to detect and attribute climate change signals in temperature and salinity changes within the NA Ocean.

For each of these models we combined the historical (up to year 2014) and SSP5-8.5 scenarios (starting in 2014) to calculate changes in temperature and salinity. We call these runs “ALL” to point out that these simulations include both natural and anthropogenic forcings.

For each ensemble we calculate the temperature (or salinity) difference between years 2002–2020, 2011–2030, 2021–2040 and 2031–2050 and the base period of 1955–1974. We then average across all ensembles available to produce a single “modeled fingerprint” for each period, which we label \({X}_{{ALL}}^{2020}{,X}_{{ALL}}^{2030}{,X}_{{ALL}}^{2040}{,X}_{{ALL}}^{2050}\), respectively. The fingerprints \({X}_{{ALL}}^{2020}\) and \({X}_{{ALL}}^{2050}\) are shown in Fig. 2.

This analysis was expanded to include an additional 12 non-DAMIP models in the Supplementary Material to verify the consistency of our results to the broader CMIP6 model set (Table S2, Fig. S2).

DAMIP scenarios

The DAMIP project provides single-forcing simulations for the following scenarios: greenhouse gasses (GHG) only, anthropogenic aerosols (AER) only, ozone depletion (O3) only and natural forcing (NAT) only, covering the 1850 to 2020 period. Since very few ESMs provide output for ozone-only forcings, we use a larger amount of ensemble members for the few models that do provide it (Table S1). These simulations were used to attribute the detected NA fingerprints.

For a given DAMIP scenario we take several ensembles available to calculate the temperature (or salinity) difference between years 2002–2020 and the 1955–1974 base period. We then average across all ensembles to produce a single “forcing fingerprint” per scenario, and we label these \({X}_{{AER}}^{2020}{,X}_{{GHG}}^{2020}{,X}_{O3}^{2020}{,X}_{{NAT}}^{2020}\).

Output from all models, Argo and WOA products was regridded to a regular 1 × 1 degree grid and onto the vertical resolution of the FGOALS-g3 ESM, which has the coarsest resolution of all DAMIP models. The study area was limited to the upper 2000 m, as deep Argo profiles below 2000 m are very sparse.

Detection and attribution (D&A) analysis

The task of detecting and attributing climate change signals depends on having observational data with a sufficiently long time record to detect trend patterns, as well as the accuracy with which these patterns are represented by ESMs34. Detection involves demonstrating that an observed fingerprint (for us, the difference between the mean of a ~ 20-year period from the 21st century and the mean over 1955–1974) in the data is statistically different from natural variability. Attribution seeks to determine the relative contributions of different factors to the detected change.

We use the D&A approach of ref. 35, and modify it by using the more robust Huber regression60. Detection is performed via a so-called “one signal analysis”, a linear spatial regression of the simulated temperature or salinity field under “ALL” forcings (\({X}_{{ALL}}^{20{XX}}\)) onto the observed fingerprint (\({Y}_{{Obs}}^{2020}\)):

$${Y}_{Obs}^{2020}({\rm{x}})={\beta }_{ALL}^{20XX}\ast \,{X}_{ALL}^{20XX}({\rm{x}})+{\rm{I}}{\rm{n}}{\rm{t}}{\rm{e}}{\rm{r}}{\rm{c}}{\rm{e}}{\rm{p}}{\rm{t}}$$
(1)

where x combines the 2-D latitude-longitude or latitude-depth spatial dimensions. We detect climate change when the scaling factor of the regression β is significantly different from 0 and close to 1. Scaling factors are calculated separately for several time periods (\({X}_{{ALL}}^{20{XX}}\)=\({X}_{{ALL}}^{2020}\), \({X}_{{ALL}}^{2030}\), \({X}_{{ALL}}^{2040}\), \({X}_{{ALL}}^{2050}\)).

As an example, we regress the observed temperature fields \({Y}_{{Obs}}^{2020}\) (Fig. 1c) onto the simulated \({X}_{{ALL}}^{2020}\) (Fig. 2c). Resulting β values are shown in Fig. 3a under blue shading.

Next, to achieve attribution, we perform a “multi-signal” analysis, i.e. a multilinear Huber regression of the observed temperature or salinity field (\({Y}_{{Obs}}^{2020}\)) onto the single forcing fingerprints \({X}_{{GHG}}^{2020}{,X}_{{AER}}^{2020}{,X}_{O3}^{2020}{,X}_{{NAT}}^{2020}\). We retrieve scaling factors (β) corresponding to individual forcings:

$$\begin{array}{l}{Y}_{Obs}^{2020}(x)={\beta }_{GHG}^{2020}\,\ast \,{X}_{GHG}^{2020}(x)+{\beta }_{AER}^{2020}\,\ast \,{X}_{AER}^{2020}(x)+{\beta }_{O3}^{2020}\,\ast \,{X}_{O3}^{2020}(x)\\ +\,{\beta }_{NAT}^{2020}\,\ast \,{X}_{NAT}^{2020}(x)+\mathrm{Intercept}\end{array}$$
(2)

To attribute changes in the observed data to various forcings we look for scaling factors β that are close to unity and significantly different from 0. Scaling factors are shown in Fig. 3a (for temperature) and Fig. 3b (for salinity) under yellow shading.

The uncertainty of the coefficients was computed as follows. First, the residuals between each realization and the ensemble mean for each experiment (e.g., \({X}_{{AER}}^{2020}\)) were calculated, resulting in a total of 267 realizations of internal variability (see Table S1). To account for the subtraction of the ensemble mean, the residuals were rescaled by a factor of \(\sqrt{36/35}\) 61. These residuals were then used in the regression models instead of the \({Y}_{{Obs}}^{2020}\) variable. Next, we repeated the regressions (Eqs. 1 and 2) 267 times. The resulting 267 parameters were used to calculate the uncertainty range (5th to 95th percentile) of scaling factors. This uncertainty range was centered around the original scaling factors and is shown as gray bars in Fig. 3.