We select 550 m to compare simulated outputs and the World Ocean Atlas (WOA) dataset, as this is a typical depth for CDW to travel in the open ocean before crossing the continental shelf break towards the Antarctic ice shelves . In the WOA dataset , the annual mean 1 °C isotherm and 34.6 isohaline encompass the Southern Ocean, separating warm and cold water masses (Fig. e, j). Similarly, the 1 °C isotherms and 34.6 isohalines of all ocean reanalyses are located at similar positions, indicating that all four reanalyses generally simulate similar hydrographic structures (Figs.  and S1 in the Supplement). Inside the 1 °C isotherms, WOA shows two local temperature minima in the Weddell Sea and Ross Sea (Fig. e), which are also present in all four ocean reanalyses. Inside the 34.6 isohaline, observations show a local maximum in the Ross Sea (Fig. j), and such structures are also simulated in ECCOv4r5, ECCO LLC270, and B-SOSE (iterations 105 and 139).

We compare the February WOA mean to the February mean state of the reanalyses (Figs.  and S2) because the WOA dataset is biased towards summer observations. All reanalyses represent the surface temperature and salinity values well. The position of isolines and isohalines agrees between the reanalyses and matches the WOA for the Southern Ocean (Fig. ). Close to the continent, the Weddell Sea matches the temperature and salinity values particularly well except for B-SOSE: the cold and relatively salty surface water mass is a feature of ECCOv4r5, ECCO LLC270, and GECCO3. However, the sector between the Antarctic Peninsula and the Ross Sea has a fresh surface bias within the coastal current, confined to the top layer. Among these model outputs, B-SOSE (iter139) shows the strongest freshening (Fig. 5h). We hypothesize that the fresh bias is related, in part, to the choice of freshwater flux from the Antarctic continent. In ECCOv4r5, ice sheet melt is computed, and iceberg calving is prescribed . In ECCO LLC270, both melt and calving are prescribed (see , their Fig. S1b). In B-SOSE, melt and calving are prescribed, also based on , and the fresh bias is partially exacerbated by too strong a sea ice seasonal cycle (i.e., too much formation in winter). GECCO3 does not prescribe any freshwater discharge around Antarctica. Local sea ice melting likely also contributes to the fresh cap creation during spring, which is not vertically mixed in summer. It is likely that the warm sea surface temperature (SST) bias in this region is also caused by the fresh bias: the stratified water column does not promote vertical mixing of the summer warming.

We also compare mixed-layer depths (MLDs) because ocean reanalyses are widely used for physical oceanography and biogeochemistry studies (Fig. 6). MLD is defined as the depth level where the density difference from the surface exceeds 0.03 kg m⁻¹ following . We select September and February, the same times as selected for sea ice concentration and volume. These months coincide with or are close to the MLD maximum and minimum. In September, observed MLD is located deep (about 500 m) in lower-latitude regions (∼ 50° S) in the Pacific sector. Similar patterns are simulated for all ocean reanalyses. We also find several spots with deep MLDs along the Antarctic coasts; these spots likely represent formation regions of AABW in the WOA. However, none of the ocean reanalyses show such deep MLDs along the Antarctic coast. In February, MLDs become shallower for all locations but maintain deep MLD (about 100 m) regions at lower latitudes (∼ 50° S). Such patterns are mostly present in all reanalyses, but simulated MLDs are biased shallow in all ocean reanalyses presented in this study. For example, February spatial mean MLDs for the region south of 60° S are 26, 24, 15, 17, and 49 m for ECCOv4r5, ECCO LLC270, B-SOSE (iter139), GECCO3, and WOA, respectively.

We show time series of MLD at the center of the Weddell Gyre and the Ross Gyre as ocean reanalyses tend to show better agreement for the open-water regions (Fig. 7). All ocean reanalyses show similar seasonal cycles with maximum and minimum in values December–January and September–October, respectively (Fig. 7c–d). Simulated MLDs show small variability, and the standard deviations of simulated MLDs are about 30 m or less. Standard deviations become large between October and December during the shallowing of MLDs. The time series of the simulated MLD at the center of the Weddell Gyre shows larger interannual variability, while the time series of simulated MLD at the center of the Ross Gyre shows smaller interannual variability (Fig. 7a–b).

We also calculate the root mean square error (RMSE) of potential temperature and salinity at depth levels of 500, 2000, and 4000 m (Fig. 8). RMSE is calculated by interpolating each ocean state estimate output onto a common horizontal grid with a 1° resolution and computing the difference from WOA potential-temperature and salinity climatology fields. At 500 m, most ocean reanalyses show stable RMSE values, indicating that ocean reanalyses have nearly constant error throughout the model simulations. ECCOv4r5 and ECCO LLC270 perform better, maintaining similar RMSEs during model integration over approximately 30 years. In contrast, GECCO3 shows a rapid increase in RMSE within the first 10 years, after which RMSE stabilizes or slightly decreases (Fig. 8a, d). B-SOSE exhibits the steepest increase in RMSE among the ocean analyses presented here. At the deeper levels (2000 and 4000 m), all ocean reanalyses show increases in RMSE (except GECCO3 salinity) at similar rates (Fig. 8b, c, e, f). This consistent increase in RMSE may reflect the fact that all ocean reanalyses struggle to accurately simulate Antarctic Bottom Water (AABW) production, thereby misrepresenting key ocean processes and leading to similar responses in all ocean reanalyses' RMSE (discussed further below).

Recent observations estimated the ACC strength to be 140–170 Sv using advanced technologies . Other canonical estimates are 130–140 Sv . For ocean reanalysis, we also calculate the time-mean ACC strengths by determining the flux through the Drake Passage for each simulated period. The mean ACC strengths for the ECCOv4r5, ECCO LLC270, B-SOSE (iter139), and GECCO3 models are 154, 140, 162, and 141 Sv, respectively (Table 3). For simulated periods for each reanalysis, the ECCOv4r5, ECCO LLC270, and GECCO3 ACC strengths do not show unrealistic trends, exhibiting rather stable values that are consistent with observations . We note, however, that the B-SOSE model's ACC (iter105) increases by ∼ 30 Sv within 5 years with a maximum strength of 200 Sv (Fig. ), so the unrealistic ACC strengthening of B-SOSE iteration 105 is fixed in iteration 139. As ACC transport is a good measure for evaluating global simulations, we recommend checking ACC transport when using newer versions of ocean reanalysis products.

The estimates of WG and RG strengths are uncertain due to intense sea ice cover, lack of observations, and large seasonal and interannual variability. The estimated WG strength has a wide range of values between 30 and 100 Sv based on previous observation-based and modeling studies (e.g., ). The estimated RG strengths are between 10–30 Sv , and the most recent study using altimetry data estimated the RG strength to be 23 ± 8 Sv . Following previous studies, we calculate the strengths of the WG and RG by extracting the minimum stream function values in the Weddell Sea and Ross Sea, respectively. In the ocean reanalyses, simulated WG strengths are 56, 60, 54, and 35 Sv for the ECCOv4r5, ECCO LLC270, B-SOSE (iter139), and GECCO3 models, respectively (Table 3). The simulated RG strengths are 21, 32, 32, and 15 Sv for the ECCOv4r5, ECCO LLC270, B-SOSE (iter139), and GECCO3 models, respectively (Table 3). We note that GECCO3's WG and RG strengths are weaker by about 30 %–50 %, which may be explained by the fact that ECCOv4r5, ECCO LLC270, and B-SOSE use ERA products, while GECCO3 uses NCEP products as atmospheric forcing in initial guesses. Due to the large uncertainty in observational estimates, mean WG and RG strengths of ocean reanalyses are consistent with observations (Fig. ).

The ACC strength from the reanalyses shows seasonal and interannual variabilities with magnitudes of about 5 Sv, which is small considering the total ACC strength of about 150 Sv (Figs. 9–11). Seasonally, the ACC tends to be marginally stronger in winter than in summer, a pattern observed consistently across all reanalyses (Fig. a). For the interannual variability, simulated small fluctuations are consistent with observations, as find almost no interannual changes in ACC strengths using repeated oceanographic measurements from 2005 (Fig. a).

The WG strength from the reanalyses shows a substantial seasonal and interannual variability (Figs. 9–11). For the interannual variability, time series of negative sea level anomalies in the center of the gyre obtained from sea surface height (SSH) observations between 2011 and 2019 suggest an acceleration of the gyre beginning in 2014 and ending in 2017 , which seems to be captured to some extent in these ocean reanalyses (Fig. b). There is, however, a substantial disparity in the interannual variability among the reanalyses (Fig. 10b). For the seasonality, the gyre tends to strengthen during winter and weaken during summer (Fig. b), which aligns with SSH observations from 2013 to 2019 .

The RG strength demonstrates substantial interannual variability according to reanalyses, though it exhibits less marked seasonal changes than the WG (Figs. 9–11). Observations of the Ross Gyre indicate a slowdown starting in 2015 and continuing into 2016 . This is replicated across all reanalyses with varying amplitudes (Fig. c). Notably, observations of the RG display distinct seasonal variability, weakening in summer and intensifying bi-annually in autumn and spring . While the winter slowdown is consistently observed in all reanalyses, there are differences in how they represent the intensity and timing of the bi-annual intensification (Fig. c).

The spatial structures of the ACC, WG, and RG are also similar across all ocean reanalyses (Fig. k–n). For example, the simulated 2 °C isotherms, which roughly represent the southern extent of the Antarctic Circumpolar Current (ACC), are consistently located in all reanalyses. Additionally, the stream functions within the ACC exhibit similar patterns across the reanalyses. To assess the agreement between the reanalyses and observations, we employ the mean dynamic topography (MDT) as a proxy. Although the MDT only represents the circulation of the surface geostrophic currents, it enables a qualitative comparison of the spatial structure of the Southern Ocean circulation. The MDT derived from the regional dataset (Fig. o) including ice-covered measurements from demonstrates good agreement with the stream functions obtained from the reanalyses within the ACC. The WG structures also exhibit similarity across all ocean reanalyses and observations, as indicated by the zero stream function contours extending from 60° W to 120° E. The simulated RG size (defined here using zero stream function contours) is larger in ECCO LLC270, extending from 160° E to 100° W, while ECCOv4r5 and GECCO3 show smaller RG sizes extending from 160° E to 140° W. While the MDT derived from satellite altimetry confirms a larger RG, it does not account for deeper currents, water column height, or ageostrophic movement.

We use the repeatedly measured oceanographic sections from Joinville Island towards Kapp Norvegia (Norvegia Point in English; section SR04 in Fig. a). Based on observations, we find four main water masses – Antarctic Surface Water (ASW), Warm Deep Water (WDW), Weddell Sea Deep Water (WSDW), and Weddell Sea Bottom Water (WSBW) – as shown in Figs. i–j and i–j e.g.,. For example, in the central part of the WG, the Antarctic Surface Water is found in the top 50–100 m, characterized by a cold (colder than 0 °C) and fresh (fresher than 34) layer. The WDW is located above the 0 °C isotherms, WSDW is located between 0 and -0.7 °C, and WSBW is located below -0.7 °C (Figs. –). At the southwestern edge of the section over the continental shelf, we find a thickened layer of ASW. Over the slope, we find a warm core of WDW (or modified CDW, mCDW) flowing towards the Filchner–Ronne Ice Shelf (see red circles in Fig. i). At the northwestern edge of this section, the thickening of the ASW over the continental shelf is also observed. Over the continental slope, we find a layer of cold (< -1 °C) and fresh (< 34.65) water over the continental slope (blue ellipses in Figs. i, i, and S3), which is the newly formed Antarctic Bottom Water originating from the Filchner–Ronne Ice Shelf.

For the ocean reanalyses along the same section, all reanalyses show similar hydrographic characteristics. They all show (a) thick ASW over the continental shelf and (b) cores of WDW at the southwestern side of the section (Figs. –). None of the simulations, however, show a layer of cold and fresh water over the continental slope at the northwestern side of this section because none of the ocean reanalyses resolve the downslope descent of Dense Shelf Water and formation of AABW from the Weddell Sea (Figs. –, S3).

For most global ocean simulations without hundreds of years of spin-up, the simulated changes in the deep ocean are mostly caused by model drifts, a phenomenon well-recognized among ocean modelers . As the model's deep-ocean circulation, mixing, diffusion, etc., are different from those of the real ocean, the model's state gradually shifts from the observed quasi-steady state to the model's steady state. Thus, we need to pay extra attention when analyzing time evolution at depths. When users want to discuss deep-ocean changes, it is imperative to verify that (1) the magnitude of simulated changes exceeds that of model drift and (2) the identified changes stem from other external forcing mechanisms. In this section, we demonstrate the influence of model drifts for the Weddell Gyre section by comparing it with repeated observations.

studied the observed time evolution of deep-water characteristics in the Weddell Gyre (WG) and found warming trends at most sampling locations. The mean warming rate is about ∼ 0.002 °C yr⁻¹ below 3000 m, which exceeds that of the global ocean by a factor of about 5. We find a similar level of warming by comparing observations in 1989 and 2010 (Table 4). Salinity also increases at most sites below 700 m. This salinity increase is not strong enough to fully compensate for the warming effect on seawater density, and deep water masses show a general density decrease as shown in . Despite bottom-water property changes attracting scientific attention, these observed changes in temperature and salinity in the deep Weddell Gyre are small and water mass characteristics are highly stable. For example, in comparing observed sections for 1989, 1995, 2005, and 2010, it is difficult to spot differences for both temperature and salinity (Figs. i, i). When comparing temperature and salinity for 2010 and 1989, the observed temperature and salinity trends at the center of the WG below 2000 m are about ∼ 0.0019 °C yr⁻¹ and ∼ 0.0001 yr⁻¹, which is consistent with .

The simulated WG sections show all observed water masses (ASW, WDW, WSDW, and WSBW) for all simulated periods for all ocean reanalyses (Figs. , ). Year 1 shows excellent agreement with ocean reanalyses and observations since all the reanalyses are initialized with interpolated ocean observations such as the World Ocean Atlas with some adjustments by adjoint methods. However, all the reanalyses simulate larger changes compared to the observations. Such changes in deep water masses can also be seen in the entire deep Southern Ocean, as shown in Fig. 8.

For simulated potential temperature, ECCOv4r5 and GECCO3 show warming of WDW (located roughly at 1500–2000 m) and WSBW (located roughly below 3000 m) as shown in Fig.  and Table 4. The simulated warming of WDW in these two ocean reanalyses exceeds 0.0084 °C yr⁻¹ (3 times the observed value), while the observed mean temperature increase between 1500–2000 m is 0.0028 °C yr⁻¹ (Fig.  and Table 4). ECCO LLC270 shows moderate warming, but the cores of WDW inflow along the continental slope show a maximum warming rate of 0.01 °C yr⁻¹. B-SOSE (iterations 105 and 139) exhibits complex patterns with both warming and cooling, and its simulated maximum and minimum trends are much larger than those of observations (Figs. e–f and S4). Thus, in all reanalyses, simulated water mass changes for both WDW and WSBW are much larger than those based on observations.

The simulated ECCOv4r5 and ECCO LLC270 salinity increases below the halocline. These rates for both WDW and WSBW are over 10 times larger than those for observations (Fig.  and Table 4). B-SOSE demonstrates slightly different behaviors with some patches of freshening close to the coast, but deep water shows a large salinity increase that is similar to what is found for other ocean reanalyses (Figs.  and S5, Table 4). GECCO3 presents the best agreement with the smallest salinity changes. Since observations show that the temperature, salinity, and thus density of the deep WS water masses remain relatively stable, ocean reanalyses fail to reproduce realistic changes. They all show excessive temperature and/or salinity changes (Table 4).

The Hovmöller diagrams (Fig. ) show simulated changes in water mass characteristics for the middle of the WG (see blue dot in Fig.  for the location). All ocean reanalyses considered in this study show gradual change, which is the model adjustment to the model steady state from the initial condition. Among all the ocean reanalyses considered in this study, ECCO LLC270 shows the smallest temperature and salinity changes, which are still larger than those of observations (Table 4). For ECCO LLC270 temperature, we find simulated cooling above 1500 m and warming deeper than 1500 m. ECCO LLC270 salinity increases for all depths below the near-surface halocline by approximately 0.02 after 26 years. Other reanalyses show larger trends than that of ECCO LLC270 (Table 4). In ECCOv4r5, two warming peaks exist at depths of approximately 1000 and 3000 m, representing the warming of WDW and AABW. B-SOSE simulates warming for all depths, with maximum warming below the mixed layer. For GECCO3, excessive warming can be found at most depths, and the maximum warming of 0.63 °C is simulated at a depth of ∼ 1700 m. The ECCOv4r5, B-SOSE, and GECCO3 salinity show freshening for depths shallower than the halocline, and salinity increases below the halocline.

We utilize the S4P oceanographic section (Fig. ) that traverses the Ross Gyre and the Pacific sector of the Southern Ocean. We identify three main water masses (Fig. ): Antarctic Surface Water (ASW), CDW, and AABW. In the western part of the section (corresponding to the Ross Gyre), we observe CDW or modified CDW at shallower depths and AABW at deeper depths. On the other hand, in the eastern part of the section, a warmer core of CDW is observed for the top 2000 m, and we can find AABW below CDW (Fig. i). We compared the hydrographic characteristics of the ocean reanalyses along the same section and found that all simulations exhibited similar features (Figs. , ). Similarly to the Weddell Gyre section (Figs. , ), none of the simulations show a layer of cold and fresh water over the continental slope at the western side of this section because all MITgcm-based ocean reanalyses fail to simulate the descent of Dense Shelf Water and the formation of AABW.

In this section, we present an additional example where temporal evolutions of the deep ocean are not in good agreement with observations. Specifically, observations from the Ross Sea exhibit more significant warming and freshening than those of the Weddell Sea. Thus, our example in this section highlights how both model drifts and external forcing influence deep-ocean circulation and hydrography. It appears that model drifts predominantly drive the simulated changes, as most simulations consistently show changes that are substantially larger than observed changes.

According to and , observations of the S4P section suggest the Ross Sea has warmed and freshened below 2000 m at a depth mean rate of 0.002–0.004 °C yr⁻¹ and ∼ 0.001 yr⁻¹, respectively. Along S4P we find smaller changes between 1989 and 2018 (Figs.  and , Table 4) compared to observed changes because we also include the eastern part of the S4P section for our analyses where water masses present smaller changes.

Here, we plot simulated RG sections for different years (1, 5, 10, and the final year) for each ocean reanalysis (Figs. 15–16). Hovmöller diagrams (Fig. 17) also show simulated changes in water mass characteristics for the middle of the RG (see blue dot in Fig.  for the location). The simulated sections display three water masses (ASW, CDW, and AABW), which is consistent with observations. Again, year 1 demonstrated excellent agreement with observations similar to those for the Weddell Sea. For the modified CDW temperature changes (between 1500–2000 m), most ocean reanalyses show larger changes than observations (Fig. 15 and Table 4). Potential-temperature changes for B-SOSE (iter139) seem to show the best agreement with observations (Table 4), but this is an artifact of spatial averaging. B-SOSE (iter139) shows strong warming and cooling, and spatial patterns do not agree with observations. For the AABW (below 3000 m), most ocean reanalyses show warming that is too strong. The simulated changes exceed observed changes except for B-SOSE (iter105). Again, potential-temperature changes for B-SOSE (iter105) seem to show the best agreement with observations (Table 4), but this is an artifact of spatial averaging, with strong cooling and warming on the western and eastern sides, respectively (not shown). The reason for the simulated excessive warming is likely the model drift and the lack of AABW formation in these ocean reanalyses. For the modified CDW salinity changes, all simulations show larger changes in magnitude compared to observations. For depths between 1500 and 2000 m, ECCOv4r5, ECCO LLC270, and B-SOSE (iter139) simulations show a salinity increase, while B-SOSE (iter105) and GECCO3 show freshening (Fig. 16 and Table 4). Deeper than 3000 m, most ocean reanalyses show increases in salinity, while observations show freshening.