We evaluate three ocean regions: the open-water ocean in the region of 60°S,60°N, the Arctic Ocean in the region of 60°N,90°N, and the Southern Ocean in the region of 60°S,90°S. In each region, we assess the OHC in both the upper 0–300 and 0–2000 m layers. The two polar regions are of particular interest as they are critical for climate studies but were poorly observed before the satellite era. Moreover, SST data in areas covered by sea ice are not assimilated, making it especially valuable to evaluate the system's performance under sea ice conditions.

For the OHC in the upper 0–300 m over the region of [60°S,60°N], the datasets reveal notable differences before 1940 (Fig. a). However, all datasets – except EN4.2.2 – suggest an OHC decline in the early 1900s and a warming trend over 1920–1940. EN4.2.2 uses relaxation towards the observed 1971–2000 climatology when no hydrographic profile data are available. The scarcity of profile observations during this period explains the very limited multi-decadal variability between 1900 and 1950. From the 1950s onward, all datasets align better with each other, reflecting a global warming hiatus from 1950 to 1980, followed by a strong warming trend beginning in 1990. Overall, CoRea1860+ exhibits variability that falls within the range of the other datasets. Before 1900, only SODA2.2.4 and CoRea1860+ are available and show some slight disagreement. SODA2.2.4 shows a comparable heat content to that of the 1980s, while CoRea1860+ shows a lower level (albeit one that is higher than in 1900–1920).

For the OHC in the upper 0–300 m over the region of [60°N,90°N] (Fig. c), CoRea1860+ aligns with SODA2.2.4 starting in 1880, following the spin-up period of SODA2.2.4, and with CHORE starting in 1910, after the spin-up of CHORE. It exhibits similar long-term variability to EN4.2.2 but with slightly weaker magnitudes. It should be highlighted that the hydrographic profile data availability at high latitudes is higher. Over the period of 1980–2010, EN4.2.2 depicts a higher warming trend in the Arctic than in the open-water ocean ([60°S,60°N]), which confirms the polar amplification phenomenon. Additionally, while CoRea1860+, SODA2.2.4, EN4.2.2, and CHORE display negative anomalies during 1900–1950, ORA-20C and CHOR show positive anomalies during the same period. ORA-20C captures the early warming trend during this time, whereas CHOR does not. From the 1950s onward, all datasets are in good agreement in terms of temporal variability, while EN4.2.2 shows higher variability amplitude than the other products.

For the OHC in the upper 0–300 m over the Southern Ocean ([60°S,90°S]; Fig. e), the datasets show greater discrepancies compared to the other two regions, particularly before 1960. For instance, CHOR and SODA2.2.4, which were forced by a former version of the atmospheric reanalysis 20CRv3, exhibit strong positive anomalies. In contrast, ORA-20C and CHORE, forced by the atmospheric reanalysis ERA-20C, display negative anomalies. EN4.2.2 remains close to the climatology before 1960 due to the sparsity of profile observations. After 1960, the differences among the datasets are diminished to some extent. Overall, CoRea1860+ exhibits a multi-decadal variability with a slight warming trend over the entire period. Notably, CoRea1860+ appears to align more closely with EN4.2.2 after 1960.

For the OHC in the upper 0–2000 m over the region of [60°S,60°N], significant differences are observed across datasets (Fig. b). CHOR and CHORE demonstrate similar OHC variability and exhibit a pronounced increasing trend, consistently with the findings of (their Fig. 10c). In contrast, CoRea1860+ and ORA-20C show moderately increasing trends accompanied by clear multi-decadal variability. Meanwhile, EN4.2.2 shows neither multi-decadal variability nor an increasing trend before 2000. This is again caused by the relaxation applied to climatology when no observations are available. SODA2.2.4 shows weak multi-decadal variability as it assimilates SST data to constrain only mixed-layer properties, e.g. temperature and mixed-layer depth , when no hydrographic profile data are available. Post-2000, all datasets exhibit a strong global warming trend. Notably, CoRea1860+ displays a weaker warming trend than EN4.2.2 but aligns with SODA2.2.4. The other datasets show a stronger warming trend than EN4.2.2.

For the OHC in the upper 0–2000 m over the region of [60°N,90°N] (Fig. d), all datasets show consistent multi-decadal variability and a slight warming trend from the 1950s onward. Before 1950, CoRea1860+ exhibits multi-decadal variability similar to ORA-20C, SODA2.2.4, and EN4.2.2 but with slightly stronger negative anomalies compared to SODA2.2.4 and EN4.2.2. In contrast, ORA-20C, CHOR, and CHORE demonstrate significant positive anomalies during 1900–1950, and CHOR and CHORE exhibit a declining trend over this period.

For the OHC in the upper 0–2000 m over the Southern Ocean ([60°S,90°S]; Fig. f), the datasets show great discrepancies. Before 1960, while CHOR, SODA2.2.4, ORA-20C, and CHORE exhibit strong positive anomalies, EN4.2.2 remains close to the climatology due to the sparsity of profile observations. After 1960, the differences among the datasets are diminished to some extent, but clear conclusions remain challenging. Over the entire period, CoRea1860+ exhibits a weak multi-decadal variability with a prominent warming trend.

For the annual variability of the OHC in the upper 300 m (Fig. ), the spatially averaged ACCs between CoRea186+ and the comparison datasets range from 0.27 to 0.47, with the highest mean ACC being observed in comparison with ORA-20C. CoRea1860+ demonstrates high ACCs in most ocean regions, attributable to similar SST constraints. However, ACCs are low or even negative in specific regions such as the Arctic Ocean, the Southern Ocean, and the eastern tropical Atlantic. The eastern tropical Atlantic stands out as a region where CoRea1860+ exhibits significant discrepancies compared to the other ocean reanalyses. This is primarily due to the model's difficulty in simulating eastern boundary upwelling systems . identified notable biases in CMIP5 models, attributing them to several factors, including the underestimation of stratocumulus cloud cover, weaker-than-observed wind stress, unresolved offshore transport by mesoscale ocean eddies, and an overly diffuse vertical temperature gradient separating the warm upper-ocean layer from the deeper ocean. There is a large disagreement in terms of sea-ice-covered areas, which is expected as there are nearly no observations there. The spatial patterns of ACCs with CHORE and CHOR are very similar, except in the Southern Ocean, likely because both datasets use the same model and observations but rely on different atmospheric forcings. ACCs with EN4.2.2 exhibit smoother spatial patterns and smaller magnitudes, primarily due to the absence of the model and the sparse availability of hydrographic profile data before 1950 (Fig. a). In contrast, prominent positive ACCs with ORA-20C are found in open-water regions, including the North Atlantic extending into the Norwegian Sea. However, there is a strong negative correlation in the northern Indian Ocean that is not seen with the other datasets. ACCs with SODA2.2.4 are mostly positive in the open-water regions.

Unlike the reference datasets, CoRea1860+ does not utilize hydrographic profile data. However, it demonstrates significant positive ACCs with these datasets for the OHC in the upper 2000 m across most ocean regions (Fig. ). The spatially averaged ACCs range from 0.19 to 0.35. The spatial patterns of ACCs with CHORE and CHOR are highly similar, with high ACCs observed in most open-water regions – except in the tropical Atlantic – and low ACCs in the Arctic Ocean and the Southern Ocean. ACCs with EN4.2.2 display smoother spatial patterns with smaller magnitudes. They are notably higher in the tropics and the Atlantic–Arctic region. Significant positive ACCs with ORA-20C are primarily concentrated in the tropics (except the tropical Atlantic) and mid-latitudes. Similarly, ACCs with SODA2.2.4 are high in the ENSO region, the Indian Ocean, and the Subpolar Gyre in the North Atlantic, while low ACCs are found in the South Atlantic and the Southern Ocean.

Overall, when compared to the other datasets, CoRea1860+ captures key multi-decadal variability and long-term warming trends. In most cases, it shows significant positive ACCs with existing datasets in the representation of OHC variability despite negative or non-significant ACCs arising from the lack of assimilation of observational data (e.g. SST data under sea ice and subsurface data). Note that interannual variability in the OHC contributes much more to the computation of ACCs than the long-term trend (not shown).

The AMOC is defined as the zonally and vertically integrated meridional volume transport across a latitude section in the Atlantic Ocean. It is measured in Sverdrups (106m3s-1), and its magnitude varies with both latitude and depth. Due to the limited availability of long-term continuous measurements, the AMOC variability has only been consistently observed at 26° N since 2004 i.e. RAPID, and in a section of the Subpolar North Atlantic since 2014 i.e. OSNAP,. In this study, we focus on the RAPID measurement since the time series of OSNAP is too short. While CoRea1860+ exhibits a standard deviation in maximum AMOC transport at 26° N within the range of that of the reference datasets over 1950–2009 (except for RAPID over 2005–2021), its mean AMOC strength at 26° N is notably higher, with a time mean of 32.0 Sv , exceeding that of the other datasets (Table ). In the following, we focus on time series of the anomalies of the maximum AMOC transport at 26° N (Fig. ) and discuss the long-term trend and multi-decadal variability of the AMOC. Note that the contribution of the Ekman component to the ensemble mean AMOC in CoRea1860+ is expected to be minimal due to the lack of wind synchronization in the assimilation system. We do not remove the Ekman component from the other reanalyses and RAPID; however, a fairer comparison of interannual variability across datasets would require doing so.

The AMOC in CoRea1860+ shows a slight declining trend over the last 160 years (Fig. ), consistently with findings from the other CMIP5 models, which generally exhibit a weak decline throughout the 20th century , as well as with AMOC reconstructions based on proxy data . SODA2.2.4, CHOR, and CHORE reanalyses show a relatively strong increasing trend, which may relate to the choice of the initialization strategy at the start of the reanalysis and the fact that they are produced by the uncoupled models. NorESM is designed to simulate the response to external forcings, and NorCPM is initialized from its stable preindustrial control run.

The AMOC in CoRea1860+ remains relatively neutral before 1920 but begins to display pronounced multi-decadal variability thereafter, characterized by a 60-year periodicity and an amplitude of approximately 3 Sv (Fig. ). The AMOC exhibits stronger phases in the 1920s and 1990s, while weaker phases are observed in the 1970s and 2020s. These variations are consistent with the observed AMV that shows positive phases from around 1930 to the 1960s and from the mid-1990s onwards and negative phases from around 1900 to the 1920s and from 1960 to the 1980s, supporting a leading role for AMOC in driving AMV . The CoRea1860+ AMOC variations show limited agreement with common proxy reconstructions of AMOC based on oceanic and atmospheric parameters – the peaks in these reconstructions after the 1950s tend to lead those in CoRea1860+ by roughly a decade Fig. 1 of. While the other reanalysis datasets show distinct peaks in the 1910s, 1970s, and 2000s, with weaker phases occurring in the 1950s and 1990s, the multi-decadal variability of AMOC is much weaker. Disagreement among reanalysis products is consistent with previous findings . This may relate to the fact that coupled ocean–atmosphere processes are essential in representing the multi-decadal processes , while the other products are forced by atmospheric reanalysis.

All datasets including CoRea1860+ consistently capture the rapid decline in the AMOC from 2005 to 2010, which is in agreement with the RAPID array observations. Importantly, CoRea1860+, despite not assimilating ocean subsurface data, demonstrates AMOC variability comparable to RAPID since 2005. This consistency underscores the robustness of CoRea1860+ and lends confidence to its AMOC reconstruction before the availability of extensive subsurface observations.

The Arctic SIE in March, as simulated by CoRea1860+, exhibits a long-term declining trend throughout the historical period, primarily attributable to anthropogenic forcings (Fig. a). Superimposed on this decline, there is notable multi-decadal variability, which reflects the internal variability of the climate system . Two periods of significant decline stand out: one in the early 20th century and another in recent decades. IAPICE1 exhibits variability closely matching that of CoRea1860+ throughout the entire period, except for a slightly more pronounced decline in the early 20th century. HadISST2.2 also displays a declining trend, which is particularly pronounced after 1980. However, HadISST2.2 is characterized by strong interannual variability but does not show comparable multi-decadal variability or the decline in the early 20th century observed in the CoRea1860+ reanalysis. On the other hand, SIBT1850 shows both a long-term declining trend and significant multi-decadal variability, with a larger amplitude than CoRea1860+ and IAPICE1. In particular, CoRea1860+ underestimates the large SIE seen in SIBT1850 and HadISST2.2 in the 1960s and 1970s. In the satellite era, CoRea1860+ is close to IAPICE1 but underestimates the rate of decline in the Arctic SIE compared to both HadISST2.2 and SIBT1850. However, CoRea1860+ reproduces interannual variability similar to that observed in the other datasets, suggesting that it reasonably captures short-term fluctuations in sea ice.

In terms of the ACCs of sea ice concentration in March (Fig. a–c), CoRea1860+ demonstrates some agreement with the comparison datasets. The spatially averaged ACCs range from 0.05 to 0.13, with the highest mean ACC being observed in comparison with IAPICE1. CoRea1860+ aligns with SIBT1850 and IAPICE1 in marginal ice regions, such as the Bering, Labrador, Greenland, and Barents seas, while it shows agreement with HadISST2.2 in the Kara, East Siberian, and Beaufort seas. Notably, the spatial coverage of the ACC with HadISST2.2 is significantly lower compared to that with SIBT1850 and IAPICE1, maybe due to its use of climatological data before 1900 and in the 1940s when missing data (Fig. ).

The September Arctic SIEs in CoRea1860+ and HadISST2.2 exhibit similar behaviour: moderate multi-decadal variability and a lack of any significant trend before 1980, followed by a strong declining trend in the satellite era (Fig. b). The offset observed in the 2010s is primarily attributed to the switch in the assimilated dataset from HadISST2.1 to OISSTV2 (Sect. ). IAPICE1 and SIBT1850 demonstrate the same variability as HadISST2.2 during the satellite era. While they show similar multi-decadal variability before 1980, SIBT1850 exhibits stronger positive anomalies and larger variability compared to IAPICE1 .

For the variability of sea ice concentration in September (Fig. d–f), CoRea1860+ exhibits good agreement with IAPICE1 in almost the whole Arctic Ocean, with a mean ACC of 0.36. It shows stronger agreement with SIBT1850 across most of the Arctic Ocean, except in the central Arctic, where ACCs are negative. Conversely, CoRea1860+ aligns with HadISST2.2 in the central Arctic but exhibits weaker agreement in the marginal ice regions.

Note that interannual variability in sea ice concentration contributes much more to the computation of ACCs than the long-term trend (not shown).

The Antarctic SIE in March (during the austral summer) in CoRea1860+ exhibits more pronounced interannual–decadal variability compared to HadISST2.2, particularly before 1980 (Fig. c). In contrast, HadISST2.2 shows constant but larger positive anomalies than CoRea1860+ before 1940. After 1940, the two datasets show a similar pattern: a pronounced decline in SIE from 1940 to 1980 followed by a slight increase in the subsequent decades. In terms of the ACCs of sea ice concentration in March (Fig. a), CoRea1860+ agrees well with HadISST2.2 in the Kong Håkon and Ross seas. In the other regions, ACCs are either not significant or significantly negative. The spatially averaged ACC is about 0.05.

The September Antarctic SIE in HadISST2.2 remains nearly constant until 1960, whereas CoRea1860+ exhibits notable interannual variability during the same period (Fig. d). Unlike the differences observed in March, the anomalies of the two datasets before 1960 are relatively close in September. While HadISST2.2 begins a pronounced decline in SIE in 1960, CoRea1860+ shows a weaker decline from 1940 to 1980. After 1980, the September SIE variability in all datasets aligns closely, indicating better agreement in the recent decades. For the variability of sea ice concentration in September (Fig. b), the spatially averaged ACC is about 0.15. CoRea1860+ shows strong agreement with HadISST2.2 in the Atlantic and Pacific sectors and strong disagreement in the Kong Håkon Sea. In the other regions, ACCs are not significant.

The decline in Antarctic SIE from 1940 to 1980 is observed in both datasets during both summer and winter seasons, consistently with previous studies e.g.. demonstrated statistically significant decreases in seasonal SIE during the early and middle 20th century. Their reconstructions have relied on a principal component regression model that incorporated observations of land surface pressure and temperature across the Southern Hemisphere extratropics and mid-latitudes (1905–2020), as well as indices of major climate modes, such as the Interdecadal Pacific Oscillation, the Southern Oscillation Index, and the Pacific Decadal Oscillation. used paleoclimate records from ice cores, tree rings, and DA to reconstruct long-term sea ice variability from 1700 onwards. They identified a declining trend in the Weddell Sea throughout the 20th century, with the largest decrease occurring before the 1960s. combined atmospheric pressure and temperature observations from 27 ground-based weather stations with outputs from nine large ensembles of coupled climate model simulations using offline DA. Their analysis revealed a significant drop in Antarctic SIE at the end of the 1970s. presented a dataset of marine climate, sea ice, and icebergs derived from logbooks of Norwegian whaling factory ships operating in the Southern Ocean between 1929 and 1940. This dataset includes approximately 4000 sea ice and/or open-sea records from the austral summer, suggesting a potentially higher seasonal SIE during the early 1930s. While these findings have provided important context, further investigation is needed to fully understand the strong decline in the middle 20th century, which lies beyond the scope of this study.

In terms of global mean SATs, CoRea1860+ aligns well with ERA-20C, CERA-20C, and 20CRv3 during overlapping periods, particularly in boreal winter (Fig. a). In boreal summer, while there is a larger global warming trend and some differences between the datasets before 1900 and after 2000, they still exhibit very similar year-to-year variability (Fig. b).

The good agreement in terms of global mean SAT variability among the datasets is largely due to their reliance on similar SST datasets. CoRea1860+ assimilates historical SST data from HadISST2.1, while CERA-20C relaxes towards HadISST2.1. ERA-20C is forced by HadISST2.1, and 20CRv3 is mostly forced by SODAsi.3 . This connection is even more evident in the spatial distribution of ACCs (Fig. ).

CoRea1860+ exhibits consistency with ERA-20C, CERA-20C, and 20CRv3 in ocean regions and some land regions (Fig. ). The spatial patterns of ACCs are highly consistent across the reference datasets, with spatially averaged values ranging from 0.50 to 0.66. The highest mean ACC is obtained with CERA-20C, which is also a coupled reanalysis. In both boreal winter (DJF) and summer (JJA), ACCs are generally higher over oceans than over land where the oceanic influence is reduced. Due to stronger air–sea interactions in the tropics, ACCs are higher in lower latitudes compared to at high latitudes, particularly over oceans. For instance, SST directly influences SLP, which drives changes in surface winds and precipitation . In DJF, regions with lower or negative ACCs include central Eurasia, central Europe, western North America, southeastern South America, and areas near the Ross Sea in Antarctica. In JJA, lower ACCs are observed in North America, likely relating to weaker large-scale atmospheric teleconnections from the tropics than in DJF , and northern Asia. Notably, ACCs are generally higher with CERA-20C and ERA-20C compared to 20CRv3. This is likely because CoRea1860+, CERA-20C, and ERA-20C all use the same SST dataset (i.e. HadISST2.1), whereas 20CRv3 uses mostly SODAsi.3 .

For PRECP, the spatial patterns of ACCs are broadly similar in boreal winter (DJF) and summer (JJA), with some notable differences over land, such as in North America (Fig. ). ACCs are particularly high in the tropics because of strong air–sea interactions there, especially over the tropical Pacific due to the strong influence of the ENSO. The agreement is higher between CERA-20C and CoRea1860+ (0.25 in DJF and 0.22 in JJA), which may relate to coupled processes. In the mid-latitudes, ACCs are generally higher in the Southern Hemisphere compared to in the Northern Hemisphere, largely due to the greater ocean coverage in the south. Additionally, higher ACCs are observed near western boundary currents , such as the Gulf Stream extension, the Kuroshio–Oyashio Current, the Agulhas Current, and the Brazil–Malvinas Current, where air–sea interactions modulate atmospheric variability . ACCs are generally lower over land faraway from the ocean, except for regions influenced by tropical Pacific teleconnections, such as North America in DJF and western Africa in JJA . In the Arctic, ACCs in DJF are higher than in JJA, particularly in the Norwegian–Barents seas, primarily due to the relatively stronger influence of the Norwegian Current in winter e.g.. Despite differences in reference datasets, ACCs in the Southern Ocean remain high during both DJF and JJA, primarily due to the Antarctic Circumpolar Current.

For SLP in both boreal winter (DJF) and summer (JJA) (Fig. ), ACCs are particularly high in the tropics, especially over the tropical Pacific, due to the strong influence of the ENSO and its teleconnections e.g.. In contrast, ACCs are lower in the mid-latitudes compared to in the tropics, especially in regions such as North America, the North Atlantic, and northern Asia. ACCs are low in the Arctic during both DJF and JJA. However, in the Antarctic, ACCs are relatively high during DJF but notably low during JJA. The spatially averaged ACCs are between 0.30 and 0.35 in DJF and between 0.20 and 0.27 in JJA.

It is worth noting that ERA-20C, CERA-20C, and 20CRv3 all assimilate surface pressure observations, whereas the CoRea1860+ reanalysis only assimilates SST observations. The influence of SST assimilation on SLP relies primarily on thermodynamic processes and air–sea interactions, which are dynamically consistent.

The North Atlantic Oscillation (NAO) is a predominantly atmospheric mode of climate variability and represents one of the most prominent patterns of climate fluctuations over the North Atlantic and surrounding regions e.g.. Following the definition in , we calculate the winter (December–March) NAO index as the difference between normalized SLPs in Lisbon and Stykkisholmur (Fig. ). The NAO indices from ERA-20C, CERA-20C, and 20CRv3 show strong agreement, reflecting the constraint provided by the assimilation of surface pressure observations (Fig. a). In contrast, CoRea1860+, which assimilates only SST observations, exhibits weak interannual variability in the ensemble mean and a relatively large ensemble spread. The correlations between the NAO indices from CoRea1860+ and those from the reference reanalyses are all below 0.16. This indicates that SST assimilation alone is not sufficient to synchronize the NAO's year-to-year fluctuations with those in the other reanalyses. Nevertheless, CoRea1860+ shows the NAO's decadal variability (Fig. b), suggesting a potential oceanic influence on longer timescales. This highlights an area of interest for future investigation, particularly in exploring how ocean–atmosphere coupling may contribute to low-frequency NAO variability.

For Z500, the spatially averaged ACCs range from 0.36 to 0.46, with no substantial differences between DJF and JJA. In the tropics, CoRea1860+ shows strong agreement with ERA-20C, CERA-20C, and 20CRv3 in both boreal winter (DJF) and summer (JJA) (Fig. ). Notably, ACCs with 20CRv3 are higher in JJA than in DJF, whereas ACCs with CERA-20C and ERA-20C are higher in DJF than in JJA. In the mid-latitudes, ACCs in the Southern Hemisphere and North America are higher in DJF compared to in JJA, while ACCs in Eurasia are lower in DJF than in JJA. As for PRECP, the tropical Pacific teleconnection signal over North and South America in DJF is evident , while land regions far from the ocean and upstream of ENSO teleconnections, such as Eurasia, have lower ACCs.

In the Arctic, positive ACCs are observed across different comparison datasets, except for ACCs with 20CRv3 in JJA. In the Antarctic, prominent ACCs are evident primarily in the Amundsen and Bellingshausen seas, linked to ENSO teleconnections.