We used two in situ data sets. The first is a set of full-depth high-resolution CTD profiles of temperature from the trans-basin cruise aboard the R/V Maria S. Merian conducted in January 2017 MSM60;, and from thirteen years (2009–2022) of repeated cruises conducted in support of the maintenance of the western portion of the South Atlantic MOC Basin-wide Array (SAMBA-West). The cruises were typically carried out once per year, although the sampling month varied between years, as illustrated in Fig. 2 of . Both datasets were collected along 34.5° S. We analyzed the longitudes of 48.5, 47.5, and 44.5° W. These sites correspond to the CTD cast sites referred to as Sites BB, C, and D in the literature , as illustrated in Fig. 1. For more information about the quality control and processing of the data, see .

The second data set corresponds to three historic temperature time series sampled by sensors moored within the Vema Channel at nearby locations (Fig. ). They were named: “CLIVAR mooring” ; “E2 mooring” ; and “SAMBAR mooring” . The main information about each mooring is reported in Table .

The WOA18 climatology contains a set of long-term mean fields of ocean properties . The data set consists of fields obtained through optimal interpolation of historic oceanographic data from different sources such as ship-deployed CTD, moored and drifting buoys, expendable bathythermographs (XBT), and gliders. The spatial distribution of the available hydrographic profiles within the study region is shown in Fig. in the Appendix to provide context for the observational support underlying the climatological fields. The 0.25° annual climatological fields of temperature and salinity averaged between 1994 and 2005 were used for validation of the ocean simulation outputs. These fields were also used to compute σ4 using the Thermodynamic Equation Of Seawater – 2010 software TEOS-10;. All comparisons with this data set were made utilizing its grid as reference, and hence model outputs were interpolated accordingly.

We analyzed the outputs of ECCO Version 4 release 4 (v4r4), its latest product release and the first to include the Arctic Ocean. This reanalysis is produced with the MIT General Circulation Model MITgcm;. ECCO uses a unique grid formulation, in which between 70° S and 57° N a simple latitude-longitude (LL) grid is adopted, but out of this region a cubed-sphere grid (CS) is used; this configuration is known as the Lat-Lon-Cap (LLC) grid . Its horizontal resolution varies spatially from 22 km (high latitudes) to 110 km (mid-latitudes), and it has 50 vertical levels with different resolutions, varying from 10 m near the surface to 457 m between its two deepest levels. ECCO outputs are available from January 1992 to January 2018 with a temporal resolution of hours, days, and months, of which only the monthly fields were used in this study. The atmospheric forcing of this simulation is from the European Centre for Medium-Range Weather Forecasts interim reanalysis ERA-Interim;.

ECCO assimilates data from Argo floats, CTDs, XBTs, marine mammals, gliders, Ice-Tethered Profilers, moorings, and satellites . ECCO utilizes the adjoint method which simulates what is observed in the ocean based on the governing equations of motion , therefore the model conserves heat, salt, volume and momentum .

We used SODA version 3.4.2 , downloaded from https://dsrs.atmos.umd.edu/DATA/soda3.4.2/ (last access: 2 February 2025). It is built on the Modular Ocean Model, version 5, ocean component of the Geophysical Fluid Dynamics Laboratory Coupled Model GFDL CM2.5;. SODA's native horizontal grid is a tripolar Arakawa-B grid, which varies from 0.1° × 0.25° resolution at high latitude to 1/4° × 1/4° in the tropics, with 1440 × 1070 grid points. In the present study, we used the interpolated 1/2° product. SODA spans the period 1980–2020 with monthly outputs and has 50 vertical levels, ranging from 10 m resolution in the upper 100 m to much coarser resolution in the deep ocean. Bottom topography is based on the NOAA/NCEI ETOPO5 bathymetric dataset, with modifications to improve the representation of narrow passages. The model uses the GFDL Sea Ice Simulator with the dynamics and thermodynamics of . The atmospheric forcing of this simulation is the ERA-Interim .

The main data sets that SODA assimilates are the World Ocean Database of historical hydrographic profiles and in situ and remotely sensed sea surface temperature. SODA assimilates data through the usage of an optimal interpolation method, in which the ocean state is constructed from a forecast based on the difference between the model and the observations .

GLORYS is a global ocean and sea ice reanalysis developed by the Mercator Ocean operational oceanography center. The version used in this study features a LL quasi-isotropic horizontal grid with a resolution of 1/12° and 50 vertical levels. The grid vertical spacing increases with depth, with resolution of 1 m near the surface and 450 m at 5000 m depth . Bottom topography is based on ETOPO1 and coast and continental shelf are based on GEBCO8 . The simulation is conducted using the NEMO platform and uses the ERA-Interim fields as atmospheric forcing . The vertical mixing is parameterized according to a turbulent closure model (order 1.5) and sea ice is simulated using the LIM2 model . The reanalysis spans the period from 1993 onward at a monthly resolution. For this work we used outputs from 1993 to 2020. The model assimilates satellite-derived sea level anomaly, sea surface temperature, and sea ice concentration, along with in situ temperature and salinity vertical profiles, using a reduced-order Kalman filter.

OFES version 2 is a forward ocean model based on version 3 of the Modular Ocean Model (MOM) . It features a LL grid system with a horizontal resolution of 0.1°, which covers the global ocean from 76° S to 76° N. The model has 105 vertical levels extending from the surface to 7500 m depth. The top 21 levels are uniformly spaced at 5 m, while the last 6 are spaced at 300 m. Bottom topography is based on the ETOPO1 bathymetric dataset.

The simulation was initialized using temperature and salinity fields from OFES version 1 and covers the period 1958–2016, with monthly outputs. . A sea-ice component is included to simulate conditions in the Antarctic and subarctic seas. Atmospheric forcing is provided by the JRA55-do dataset , which is based on the JRA-55 reanalysis . Horizontal mixing is represented using a biharmonic operator, while vertical mixing follows the parameterization.

The Finite-Element Sea Ice–Ocean Model (FESOM) is the first fully developed global multi-resolution model specifically designed to simulate large-scale ocean dynamics. Its next-generation counterpart, the Finite-volumE Sea ice–Ocean Model (FESOM 2), adopts a finite-volume discretization approach in place of the finite-element discretization used in the original model. This shift significantly enhances efficiency, making FESOM 2 at least nine times faster than its predecessor .

FESOM 2 utilizes a horizontal triangular mesh with horizontal velocity components defined at the centers of the triangles, while scalar variables are positioned at its vertices. The vertical grid adopts the Lagrangian-Eulerian (ALE) vertical coordinate system , with both horizontal velocities and scalar variables defined between vertical levels, whereas inter-layer exchange velocities are located at the layer interfaces and aligned with scalar positions. The bathymetry is derived from global RTopo-2 dataset . FESOM 2 contains a sea-ice component that has been derived from the Finite-Element Sea Ice Model (FESIM) .

The simulations use a default model configuration in which each experiment is forced with the JRA55 atmospheric reanalysis and is initialized from the Polar Science Center Hydrographic Climatology (PHC; ). We use the Gent-McWilliams eddy parameterization and the isoneutral diffusion. Diffusivity is scaled linearly with resolution, with a reference diffusivity of 2000 m² s⁻¹ at 100 km horizontal resolution. As the vertical mixing scheme, we employ the K-profile parameterization (KPP; ). No additional calibration was applied to tailor the model to specific regions or processes of interest.

The reference simulation, hereafter referred to as FESOM R was designed to replicate results from existing numerical simulations, such as ECCO, SODA, OFES, and GLORYS. To achieve this, we adopted the exact vertical levels used in OFES, as it includes the highest number of vertical levels among the evaluated simulations. Hence, FESOM R employs 105 vertical levels that extend from the surface to 7350 m, with five levels distributed between 3800 and 5350 m. The horizontal resolution is approximately 110 km in the northern parts of the Pacific and Indian Oceans, 30 km over most of the global domain, and 10 km in the South Atlantic Ocean. In addition, the mesh is progressively refined toward higher latitudes (Fig. a; Fig. a).

The vertical-refinement experiment, hereafter referred to as FESOM V was designed to evaluate the impact of vertical resolution. It retains all configurations of FESOM R but increases the number of vertical levels from 105 to 125, all additional levels concentrated between 3800 and 5350 m. While shallower than FESOM R (5350 vs. 7350 m), this configuration provides ∼ 50 m resolution near the seafloor, compared to ∼ 300 m in FESOM R over a much larger depth range (3800–7350 m). Above 3800 m both simulations share the same vertical structure.

The horizontal-refinement experiment, hereafter referred to as FESOM H, was designed to evaluate the impact of horizontal resolution. This experiment retains all configurations of FESOM R, but increases the horizontal resolution in key regions relevant to the northward flow of abyssal waters in the Atlantic Ocean, with FESOM H maintaining the original 105 vertical levels while increasing the horizontal resolution in the South Atlantic sector to 7.5 km and further refining it to 4 km in the Vema Channel and its surrounding area (Figs. b; b).

The combined-refinement experiment, hereafter referred to as FESOM VH, was designed to evaluate how the two previous configurations work together. FESOM VH enhances both vertical and horizontal resolution. This experiment uses the same vertical resolution as FESOM V and the same horizontal resolution as FESOM H. The main characteristics of the numerical simulations are reported in Table .

We constructed horizontal maps of abyssal water properties (θ, salinity, and σ4), covering the Antarctic continental shelf and slope, the central Weddell Sea, and the western South Atlantic. The latter includes the Argentine and Brazil basins, defined as follows: the Argentine Basin spans 50–30° S and 60–30° W, while the Brazil Basin extends from 30° S to 0° and from 40 to 15° W (Fig. , white boxes). Data were extracted from the deepest valid level at each horizontal grid point over the same period as the WOA18 dataset (1994–2005) at monthly resolution. For each simulation, we calculated the time-averaged field and the mean bias (simulation minus reference).

The reference θ field (Fig. a) shows cold waters below -0.7 °C on the Antarctic continental shelf (Fig. , white contours), characteristic of DSW. Along the continental slope, temperatures increase, forming a band of relatively warmer water that separates the shelf from the central abyssal Weddell Sea, consistent with the presence of the mWDW . In the central Weddell Sea, temperatures drop below -0.7 °C again, indicating the presence of WSBW, which does not penetrate into the Scotia Sea, where higher temperatures prevail due to the presence of WSDW only (Fig. , red contours). In the Argentine Basin, waters with θ≤0 °C are observed, corresponding to AABW, which enters the Brazil Basin through the Vema Channel and progressively modifies its properties as it flows northward.

ECCO, SODA, GLORYS, and OFES all depict cold waters (θ<-0.7 °C) over the Antarctic continental shelf (Fig. b, c, d, and e). In the central Weddell Sea, ECCO, GLORYS, and OFES also reproduce waters colder than -0.7 °C, although the cold region in ECCO is spatially limited. SODA is the only simulation that fails to reproduce this feature, with central Weddell Sea bottom temperatures of approximately -0.35 °C. OFES, at the other extreme, depicts waters colder than -1.4 °C.

The northward spreading of cold waters from the Weddell Sea is limited in ECCO, SODA, and GLORYS (Fig. b, c, and d). OFES (Fig. e) is the only simulation that reproduces the presence of WSBW in the South Sandwich Trench, a deep channel located east of the South Scotia Ridge. ECCO, GLORYS, and OFES capture cold waters with -0.7 °C <θ<0.0 °C in the Argentine Basin, but only OFES represents their further northward spreading through the Vema Channel into the Brazil Basin. SODA again stands out as an exception, with relatively warm waters in both the Argentine and Brazil Basins.

The results above describe the spatial extent of cold waters exported from the Weddell Sea. However, the bias fields reveal important regional contrasts throughout the study area. Despite the persistence of cold waters along the continental shelf, all simulations exhibit positive temperature biases, particularly along the eastern sector of the southern continental shelf (Fig. f, g, h, and i). In contrast, the eastern shelf of the Antarctic Peninsula, located to the west of this region, shows negative temperature biases, with the exception of ECCO, which exhibits a warm bias. In the abyssal central Weddell Sea (depths >3000 m), ECCO, SODA, and GLORYS are warmer than the reference, whereas OFES shows colder values.

In the Argentine Basin, abyssal waters are warmer than the reference in SODA and OFES. GLORYS exhibits localized colder patches in the southern and northwestern portions of the basin and warmer regions surrounding these patches. ECCO, by contrast, shows a consistent cold bias throughout this basin. In the Brazil Basin, ECCO, SODA, and GLORYS depict warmer conditions than the reference, while OFES shows a mixture of warm and cold regions, with the negative bias dominating over a larger portion of the basin. Notably, ECCO and OFES, the former more strongly, the latter more subtly, exhibit opposite bias signs between the Weddell Sea, the Argentine Basin, and the Brazil Basin, whereas SODA and GLORYS show consistently warm conditions across all three regions.

The reference salinity field (Fig. j) displays alternating regions of relatively high (>34.84) and low (<34.68) salinity on the continental shelf. In the central Weddell Sea, bottom salinities are approximately 34.68, and these values persist northward through the Argentine Basin. Within the Brazil Basin, salinity increases from approximately 34.72 in the basin interior to 34.88 near its boundaries.

ECCO, GLORYS, and OFES correctly reproduce the alternating salinity bands on the Antarctic continental shelf (Fig. k, m, and n), whereas SODA does not cover this region (Fig. l). In the central Weddell Sea, ECCO, GLORYS, and OFES depict salinities near 34.68, while SODA shows comparatively higher values. In the Argentine Basin, GLORYS and OFES also reproduce salinities near 34.68, although GLORYS exhibits saltier waters in the northeastern portion of the basin. ECCO displays the opposite pattern, with saltier waters in the central basin and fresher values toward its margins, while SODA shows the highest salinities of all simulations throughout the basin. In the Brazil Basin, SODA and GLORYS depict the saltiest conditions, OFES shows the freshest waters, and ECCO exhibits intermediate salinity values.

ECCO, GLORYS, and OFES generally exhibit positive salinity anomalies along the eastern side of the Antarctic Peninsula and fresh biases farther east on the Antarctic continental shelf, near the southern limit of each model domain (Fig. o, q, r). SODA, in contrast, shows a consistent fresh bias across these regions (Fig. p). In the central Weddell Sea, SODA and GLORYS display predominantly salty biases, with SODA presenting the largest deviations, while ECCO and OFES exhibit a mixture of positive and negative anomalies. In the Argentine Basin, all four simulations depict saltier-than-reference conditions. This signal extends into the Brazil Basin for ECCO (with some localized negative patches), SODA, and GLORYS, whereas OFES stands out by exhibiting predominantly negative anomalies across most of the basin, resulting in a marked basin-to-basin contrast, with fresher conditions in the Weddell and Brazil Basins and saltier waters in the Argentine Basin.

The reference σ4 field (Fig. a) shows high-density waters (σ4>46.3 kg m⁻³) over the Antarctic continental shelf. In the central Weddell Sea, bottom densities are slightly lower (σ4≈46.2 kg m⁻³) and extend toward the South Scotia Ridge without crossing it, as well as eastward into the South Sandwich Trench. In the Argentine Basin, lighter bottom waters are found (σ4≈46.1 kg m⁻³). Farther north, these waters leak through the Vema Channel into the Brazil Basin, where densities decrease slightly as they progress northward (σ4≈46.04 kg m⁻³). The spatial patterns described for θ are largely mirrored in the σ4 field (Fig. S4, first row). As a consequence, the latter displays positive (negative) biases where the former shows negative (positive) anomalies, indicating that density variations are primarily controlled by temperature rather than salinity in these simulations.

Dense and cold waters are present along the Antarctic continental slope in all simulations. However, none of them show clear evidence of DSW plumes cascading downslope to form new WSBW and WSDW (see video supplements, 10.5281/zenodo.19699061, ). In ECCO, the central Weddell Sea is initially occupied by relatively light waters that gradually become denser over time. SODA shows episodic pulses of denser waters in this region, and the basin as a whole becomes denser by the end of the simulation relative to its initial state, with periods of rapid water mass transformation also occurring. GLORYS, in contrast, exhibits comparatively little temporal variability. However, its temporal evolution is characterized by intermittent pulses rather than a smooth progression, as seen in the video supplements .

In the Argentine and Brazil Basins, the simulations also exhibit distinct temporal evolutions. In ECCO, SODA, and OFES, the long-term trends in these basins differ from those observed in the Weddell Sea. While the Weddell Sea tends to become denser over time in ECCO and SODA, the Argentine and Brazil Basins show a general lightening trend. OFES presents a similar inter-basin contrast but with the opposite sign: the Argentine and Brazil Basins show an initial increase in density followed by a gradual decrease, yet abyssal waters remain denser at the end of the simulation than at the beginning. This contrast between the Weddell Sea and the adjacent basins in ECCO, SODA, and OFES may indicate limitations in the representation of deep-basin connectivity. GLORYS, in contrast, shows relatively small temporal changes and a similar pattern across all three basins, although the lack of temporal continuity in its outputs complicates the interpretation of its variability.

Overall, the comparison of θ, salinity, and σ4 shows that all simulations broadly capture the large-scale distribution of abyssal water masses, but exhibit consistent limitations in their representation of abyssal properties and inter-basin connectivity. ECCO and OFES reproduce the general thermohaline structure but display contrasting bias patterns between basins, while SODA shows a more spatially uniform yet persistently warm and saline deep ocean. Across all simulations, density biases closely follow temperature errors, indicating that inaccuracies in the θ field are the primary driver of the discrepancies in the σ4 field. Moreover, none of the simulations clearly represents dense shelf water cascading from the Antarctic continental shelf, and the inconsistent temporal evolution of abyssal density between the Weddell Sea and the Argentine and Brazil Basins further points to limitations in the representation of deep-basin connectivity.

To further examine the discontinuities between connected basins, we focus on the Argentine and Brazil basins. While the pathways connecting the Southern Ocean to the Argentine Basin involve complex topographic features such as the South Scotia Ridge and multiple fracture zones, the Vema Channel provides a relatively direct and well-defined passage linking the Argentine and Brazil basins. We therefore analyze the θ–S relationship in each simulation and compare it with the corresponding reference (Fig. a and b). The diagrams use all grid points from the surface to the bottom within the Argentine and Brazil basins (the two northernmost white boxes in Fig. ), binned into a 2D histogram. Model outputs were interpolated to match WOA18's horizontal resolution, and all datasets were then interpolated onto a uniform 50 m vertical grid from the surface to 6000 m. Figure  shows a zoom-in of the densest portion of the full θ–S diagram.

When comparing the two basins, WOA18 shows a well-defined θ–S curve in the Argentine Basin, with the coldest waters at approximately -0.5 °C (Fig. a). In the Brazil Basin, this curve shifts toward warmer and lighter values, with no waters colder than 0.0 °C (Fig. b). This occurs because a large fraction of the coldest waters does not reach the northern portion of the Vema Channel due to the existence of numerous sills along the channel, while part of this cold water also mixes with the overlying water masses . As a result, abyssal waters within the Argentine Basin are noticeably colder and denser than those reaching the Brazil Basin.

In the Argentine Basin, ECCO exhibits a noticeable scattered dispersion of waters lighter than 45.84 kg m⁻³ and a tilt of the curve toward higher salinity values at densities above 45.92 kg m⁻³ (Fig. c), while SODA presents a shape closer to the reference, though lacking waters colder than 0.0 °C (Fig. g). Despite these differences, both simulations share a common limitation when transitioning to the Brazil Basin: the θ–S distribution shows a marked reduction in the occurrence of the densest water classes and a systematic shift of the entire curve toward lighter densities relative to the reference (Fig. e, f, i, and j), indicating a discontinuity in abyssal water transport through the Vema Channel.

GLORYS closely follows the reference θ–S curve in the Argentine Basin, including waters colder than 0.0 °C and with only minor deviations from the reference (Fig. k and l). OFES presents a similar overall shape but with a narrower salinity range (Fig. o and p). In the Brazil Basin, GLORYS shows a striking reduction in the number of θ–S occurrences for the density range of 45.84–45.92 kg m⁻³ (Fig. m), indicating a lack of the densest bottom water masses, though the overall shape of the curve still agrees with the reference (Fig. n). OFES, in contrast, exhibits a shift of the entire curve toward lighter densities, similar to ECCO and SODA, but retains slightly denser waters at the very bottom of the basin (Fig. q and r).

The θ–S analysis reveals that, while the simulations differ in how they represent abyssal water properties in the Argentine Basin, they all exhibit a common artificial change in abyssal water properties when transitioning to the Brazil Basin. The way this artificial change manifests, however, differs between simulations: ECCO, SODA, and OFES show a systematic shift of the entire θ–S curve toward lighter densities, whereas GLORYS shows a volumetric reduction concentrated in the densest water classes while broadly preserving the overall shape of the curve. Together, these results point to a limitation in the representation of abyssal water mass connectivity across the Vema Channel.

To investigate where and how the changes in abyssal water properties develop we next focus on the Vema Channel, the main passage connecting the two basins. Figure 4 shows temperature sections extracted at the nearest available latitude to 31.3° S in each model's native grid. To avoid artificial smoothing and loss of signal, no interpolation to a common latitude was applied to the different products. Examining the temperature structure within this region provides a direct way to assess how deep-water masses are represented as they flow from the Argentine to the Brazil Basin.

The WOA18 reference field reveals a discrepancy between the actual location of the Vema Channel and its representation (Fig. a). Only three vertical levels cover the channel, spanning a total of five grid points, none of which capture its full depth. Despite this, the reference field indicates the presence of waters colder than 0.0 °C in the region corresponding to the channel, suggesting that WOA18 captures at least the thermal signature of the abyssal flow, even if its spatial structure remains poorly resolved.

The Vema Channel is poorly represented in both ECCO (Fig. b) and SODA (Fig. c). In ECCO, a single column of grid points is present in the region where the channel should be located, whereas the SODA grid contains no points in this region, despite being slightly deeper. In both simulations the channel is represented misaligned, wider and shallower than in reality. The temperature sections show waters warmer than the reference, with no waters colder than 0.0 °C. GLORYS provides a better representation of the channel (Fig. d), with two vertical levels inside it: the deepest containing two grid points approximately 300 m above the actual bottom, and the shallower containing four. Nevertheless, no waters colder than 0.0 °C are found within the channel. In OFES, the channel structure is slightly misaligned with its actual location (Fig. e). The number of grid points and vertical levels inside the channel matches those in GLORYS, but the abyssal waters are colder and less stratified than the reference.

In the region of the Vema Channel, ECCO, SODA, and GLORYS all exhibit positive temperature biases relative to the reference, while OFES is colder. For SODA and GLORYS, this is consistent with the warm bias observed throughout the Argentine Basin (Fig. g and h). In ECCO however, this represents an inversion of the pattern seen in the Argentine Basin, where waters are significantly colder than the reference (Fig. f). This pattern is consistent with a bathymetric control imposed by the Vema Channel sill. Figure shows that, although ECCO exhibits the coldest abyssal temperatures, these waters are confined below sill depth because of its relatively strong abyssal vertical temperature gradient, leaving warmer waters available at the sill. In contrast, OFES exhibits a weaker abyssal vertical temperature gradient. Although its deepest waters are warmer than those in ECCO, colder waters are already present at sill depth, explaining the opposite sign of the temperature anomaly within the Vema Channel and the Brazil Basin.

Here we compare simulation outputs with in situ observations from three moorings deployed within the Vema Channel (Table ). Comparing OGCM outputs against pointwise in situ observations is an inherently demanding test, given that OGCMs are not designed to resolve variability at these small scales. Accordingly, this analysis should not be interpreted as a conventional validation, but rather as an exploration of model performance under such unfavorable conditions.

During the CLIVAR period, ECCO shows the highest correlation (≈0.50) among the four simulations but its mean temperature substantially deviates from the reference (Fig. a and d). SODA shows a weaker correlation (≈0.19) and a similarly large deviation from the reference mean. GLORYS achieves a small correlation (≈0.33), and a considerably greater variability (Fig. a), while its mean temperature is closer to the reference than ECCO and SODA. OFES exhibits a slightly negative correlation (≈-0.05), but provides the closest representation of the reference mean among the four simulations.

During the E2 period, ECCO shows an improved correlation (≈0.66) relative to the CLIVAR period, but with increased variability, and its mean temperature remains similar to CLIVAR though slightly warmer than the reference (Fig. b and e). SODA exhibits a negative correlation (≈ -0.4) and a larger variability than in CLIVAR, with a mean temperature that deviates more from the reference. GLORYS shows a reduced correlation (≈ 0.22) and increased variability relative to CLIVAR (Fig. b), but its mean temperature remains closer to the reference than ECCO and SODA. OFES exhibits a more negative correlation (≈-0.33) than in CLIVAR but improved variability, and again provides the closest representation of the reference mean among the four simulations.

During the SAMBA period, only SODA and GLORYS provide outputs. SODA shows a small but positive correlation (≈0.24) and statistical metrics similar to the CLIVAR period, but its mean temperature deviates even further from the reference, suggesting a possible overestimation of the abyssal ocean warming trend . GLORYS shows a correlation (≈0.34) similar to the CLIVAR period, but its variability remains excessive (Fig. c). Unlike SODA, GLORYS does not exhibit a pronounced warming trend.

In addition to the mooring records, the simulations were compared with CTD data from the SAMBA-West array, just south of the Vema Channel at Sites BB, C, and D (Fig. , green diamonds). We used the deepest bin of each CTD cast to construct the reference time series, shown as black stars in Fig. . Mean temperatures and linear trends for each simulation and site are summarized in Table .

At Site BB, all simulations produce temperatures warmer than the reference (Fig. a, Table ). OFES is the only simulation with a mean temperature close to the reference, though still slightly warmer, whereas ECCO, SODA, and GLORYS substantially overestimate it, with GLORYS additionally exhibiting the largest uncertainty among the four simulations. The observed CTD trend at this site is not statistically significant, which excludes a meaningful comparison of warming trends with the simulations.

At Site C, ECCO and SODA overestimate the reference mean temperature, with SODA showing the largest deviation (Fig. b, Table ). GLORYS is slightly below the reference while OFES is slightly above, both with the same absolute deviation, though GLORYS exhibits considerably larger uncertainty. Regarding trends, ECCO provides the closest match to the observed warming trend, while SODA substantially overestimates it. GLORYS shows a cooling trend in contrast to the observed warming, and OFES overestimates it moderately. Overall, OFES provides the best representation of the reference mean, while ECCO captures the observed trend most accurately.

At Site D, SODA again shows the largest deviation from the reference mean, while OFES provides the closest representation with a mean temperature only slightly above the reference (Fig. c, Table ). Notably, ECCO shifts from overestimating the reference mean at Sites BB and C to underestimating it here. This pattern is consistent with the bathymetric blocking hypothesis discussed in Sect. . In ECCO, Site C is represented by the 4264 m model level, corresponding to the deepest model level connected through the Vema Channel, whereas the deeper Site D is represented by the 4640 m level. The strong temperature contrast between these two levels suggests that the coldest abyssal waters remain below the deepest model level connected through the Vema Channel. GLORYS remains close to the reference mean but with large uncertainty. Regarding trends, ECCO again provides the closest match to the observed warming trend, while SODA substantially overestimates it. GLORYS shows a near-zero trend, and OFES moderately overestimates it.

In summary, the comparison highlights that model performance varies substantially across locations and periods, with no single simulation consistently outperforming the others. SODA consistently overestimates the reference mean temperature across all sites, while GLORYS, despite reproducing mean conditions reasonably well, exhibits excessive variability throughout. OFES generally provides the closest representation of the observed mean temperature, and ECCO captures the warming trends most accurately, but its mean temperature is highly sensitive to the model level representing each site due to the strong vertical temperature gradient. Together, these results reinforce the limitations identified in the previous sections regarding the representation of abyssal water properties in the western South Atlantic.

The bottom temperature distribution in FESOM R (Fig. b) reproduces several of the recurring limitations identified in Sect. , despite being based on an independent model formulation. Cold waters (θ<-0.7 °C) occupy the Antarctic continental shelf, while only a limited area of the central Weddell Sea contains bottom waters colder than -0.7 °C, similar to ECCO. These waters do not extend beyond the South Scotia Ridge, consistent with ECCO, SODA, and GLORYS. Waters with θ≤0.0 °C are exported into the Argentine Basin but remain confined to its southern part, allowing warmer waters to dominate most of the basin. As a result, the waters entering the Brazil Basin are markedly warm, closely resembling the distributions simulated by ECCO and SODA.

The bias field of FESOM R (Fig. f) further confirms that the reference experiment reproduces the large-scale deficiencies identified across the evaluated simulations. Although positive temperature biases are widespread throughout the study region, their spatial distribution is not uniform. The Brazil Basin exhibits substantially stronger warm anomalies than the Argentine Basin, producing an abrupt inter-basin transition similar to that identified in ECCO and OFES. Unlike these simulations, however, the discontinuity in FESOM R arises from a change in the magnitude of the warm bias rather than from a reversal of its sign, suggesting that the degradation of abyssal water properties increases downstream of the principal inter-basin pathways.

We now examine how controlled changes in horizontal and vertical resolution modify these patterns. FESOM V (Fig. c) shows a substantial improvement relative to FESOM R. A broader area of the central Weddell Sea is occupied by cold waters with θ≤-0.7 °C, though these waters still do not extend into the South Sandwich Trench. Waters with θ≤0.0 °C now span the entire Argentine Basin, and a colder class of waters flows from this basin into the Brazil Basin. The bias map (Fig. g) reflects this improvement, displaying reduced magnitudes and a more spatially uniform warm bias across the study region, without the basin discontinuities present in FESOM R.

FESOM H (Fig. d) exhibits the weakest performance among the four experiments. No waters colder than -0.7 °C are present in the central Weddell Sea, and waters colder than 0.0 °C are absent in the Argentine Basin, occupying just a small portion of the Georgia Basin. A warmer class of waters, relative to both FESOM R and FESOM V, is exported northward into the Brazil Basin. The bias field (Fig. h) displays spatial patterns similar to those of FESOM R but with substantially larger magnitudes, resulting in a poorer representation of inter-basin connectivity.

FESOM VH (Fig. e) performs better than FESOM R in some aspects but worse in others. Waters colder than -0.7 °C occupy a broader area of the central Weddell Sea, and a colder class of waters flows from the Argentine Basin into the Brazil Basin, indicating an improved representation in that region. However, the connection between the Weddell Sea and the Argentine Basin remains poorly represented, as waters colder than 0.0 °C are not exported northward from the Georgia Basin, unlike in FESOM R and FESOM V. The bias field (Fig. i) shows a distinct pattern, with larger positive biases in the Argentine Basin and comparatively smaller biases in the Brazil Basin.

The spatial patterns described for θ are largely mirrored in the salinity fields (Fig. j–r), and as a consequence, the σ4 field (Fig. ) displays predominantly negative biases across most of the analyzed regions, indicating that density variations are primarily controlled by temperature rather than salinity. Over the Antarctic continental shelf, density biases are positive, in contrast to the negative values observed elsewhere. Despite the presence of relatively dense waters on the shelf, these waters are not sufficiently transferred to the abyssal Weddell Sea through downslope cascading, limiting the ventilation of the deep basin and contributing to the negative density biases observed in this region.

The video supplements support this picture, as no clear indications of denser waters reaching the abyssal Weddell Sea can be seen in any of the configurations. Unlike the previously analyzed simulations, the FESOM simulations do not exhibit inverse density trends between adjacent basins. Instead, all experiments display a general negative density trend across the Weddell Sea, Argentine, and Brazil Basins, though the rates of change differ among configurations. In the Weddell Sea, FESOM R and FESOM H lose all waters within the density range 46.16<σ4<46.20 kg m⁻³ by 2019, whereas FESOM V and FESOM VH do not completely lose this density class, though its volume decreases over time. In the Argentine Basin, FESOM R loses all waters within 46.08<σ4<46.12 kg m⁻³ by 1993, FESOM V by 2002, FESOM H by 1988, and FESOM VH by 1995. In the Brazil Basin, FESOM R and FESOM H lose all waters within 46.04<σ4<46.08 kg m⁻³ by 1968, while FESOM V and FESOM VH retain this density class throughout the simulation. These results indicate that enhanced vertical resolution promotes a more coherent density evolution across connected basins.

The four FESOM experiments indicate that vertical resolution exerts a stronger control on the large-scale representation of abyssal water masses than horizontal resolution. In both pairs of experiments (FESOM R–FESOM V and FESOM H–FESOM VH), increasing the number of vertical levels consistently produces a colder abyssal ocean and enhances the northward propagation of colder and denser abyssal waters, thereby reducing the discontinuities in abyssal properties between the Weddell Sea, Argentine Basin, and Brazil Basin. The improved performance of the vertically refined configurations is consistent with previous studies showing that enhanced vertical resolution improves the representation of deep water masses and strengthens the deep overturning circulation

Although these results highlight the dominant role of vertical resolution, they do not imply that grid resolution alone controls the simulated abyssal circulation. Other aspects of the model formulation, including atmospheric forcing, bathymetry, sea-ice representation, and subgrid-scale parameterizations, are also expected to contribute to the remaining differences among the simulations. In particular, the poorer performance of the horizontally refined configurations may partly reflect the interaction between increased horizontal resolution and the Gent–McWilliams (GM) parameterization, whose diffusivity scales with grid spacing. However, this interpretation remains speculative, as the parameterizations were intentionally kept unchanged in all experiments to isolate the effects of grid resolution. Assessing the interplay between horizontal resolution and subgrid-scale mixing schemes therefore lies beyond the scope of the present study.

In the Argentine Basin, FESOM R shows a clear absence of waters colder than 0.0 °C, with maximum densities reaching only around 45.92 kg m⁻³ (Fig. c), and a saltier–warmer bias along these density levels relative to the reference (Fig. d). In the Brazil Basin, the absence of cold and dense water classes is even more pronounced, with minimum temperatures around 1.0 °C and maximum densities near 45.84 kg m⁻³ (Fig. e and f), indicating a poor representation of the connection between the two basins, consistent with the common deterioration of abyssal thermohaline properties and the associated loss of the coldest and densest water classes from the Argentine Basin to the Brazil Basin identified in the previously evaluated simulations.

FESOM V shows subtle but notable changes relative to FESOM R in the Argentine Basin: temperatures now reach 0.0 °C and waters with densities exceeding 45.92 kg m⁻³ are present, unlike in FESOM R where densities remain at or just above this value, though the saltier–warmer bias along density levels persists (Fig. g and h). In the Brazil Basin, a colder and denser class of waters now occupies the deepest part of the water column, with properties similar to those found in the reference (Fig. i). This indicates an improved connection between the two basins, with the θ–S curves lying much closer to the reference and nearly overlapping (Fig. j).

FESOM H shows a θ–S distribution in the Argentine Basin similar to that of FESOM R (Fig. k and l), but still being warmer and lighter. In the Brazil Basin, however, the deviation from the reference is larger than in FESOM R, with a greater loss of cold waters, a shift of the distribution toward lighter densities, and a higher concentration of warm points (Fig. m and n). This indicates that increasing horizontal resolution without modifying vertical resolution does not improve the representation of abyssal waters and can even degrade it.

FESOM VH closely resembles FESOM V in both the Argentine and Brazil Basins (Fig. o–r), with only minor differences in the θ–S distribution. Compared to FESOM H, FESOM VH recovers most of the improvement associated with enhanced vertical resolution, but the additional horizontal refinement does not substantially alter the θ–S relationship relative to FESOM V. These results confirm that vertical refinement provides consistent improvements in the representation of abyssal waters, whereas horizontal refinement alone has mixed or detrimental effects.

Overall, the θ–S analysis confirms that vertical resolution is the dominant factor controlling the representation of abyssal water properties and inter-basin connectivity. Refining the vertical grid consistently improves the density structure and the connection between the Argentine and Brazil Basins, whereas horizontal refinement alone does not yield comparable benefits and can even degrade the results.

In FESOM R (Fig. f), the Vema Channel is represented but remains poorly resolved, consistent with the goal of replicating the configurations of the previously assessed simulations. Only a few grid points fall inside the channel, and the deepest ones do not reach its full depth, with minimum temperatures staying above 0.0 °C. In FESOM V (Fig. g), the channel structure becomes more clearly defined and the deepest parts contain colder water than in FESOM R, including temperatures below 0.0 °C, although the amount of very cold water remains limited. FESOM H (Fig. h), in contrast, although offering a more detailed geometric description of the channel than the reference, does not improve the abyssal thermal structure: bottom waters remain warmer than in both FESOM R and FESOM V, and the coldest isotherms bend upward toward the western side of the channel, inconsistent with the expected structure. Note that FESOM V and FESOM H do not increase the maximum depth of the channel compared to FESOM R.

FESOM VH (Fig. i), which combines the enhanced vertical resolution of FESOM V and the finer horizontal grid of FESOM H, restores and extends the improvements obtained with higher vertical resolution. The coldest isotherms are again present within the channel, and the bottom water layer bends toward the eastern wall. This eastward displacement of the cold core is consistent with in situ observations of the AABW flow through the Vema Channel , which describe a bottom-intensified current steered along the eastern flank. The mechanism behind this deflection has been attributed to the Ekman flux induced by bottom friction, which drives a lateral shift of the densest water mass toward the eastern boundary. This feature is also reproduced in regional numerical simulations of AABW flow , where the cold core similarly adheres to the eastern wall while maintaining a narrow bottom-intensified jet. The presence of this structure in FESOM VH therefore indicates that the combined refinements enable the model to reproduce both the cold core and its realistic eastward deflection, though waters colder than 0.0 °C remain absent. Note that in FESOM VH the maximum depth of the channel is increased compared to the other three FESOM experiments, which further contributes to a better representation of the channel bathymetry.

The vertical refinement in FESOM V and FESOM VH brings the abyssal waters closer to the observational reference, resulting in a colder abyssal ocean and explaining the smaller temperature bias in the Argentine Basin (Fig. g, i) relative to FESOM R and FESOM H (Fig. f, h). This cooling is not restricted to the deepest model level. Instead, it extends upward through the lower water column, allowing colder waters to reach the depth of the Vema Channel sill (compare Figs.  and A4). This likely explains the reduced temperature bias in the Brazil Basin (Fig. g, i), as colder waters are able to propagate northward through the channel. Note that, although FESOM V is slightly colder than FESOM VH in the deepest layers, FESOM VH provides a more realistic representation of the Vema Channel. This likely facilitates the northward propagation of cold abyssal waters, explaining why FESOM VH exhibits a similarly reduced bias in the Brazil Basin despite its slightly warmer abyssal temperatures.

These results show that vertical resolution is the key factor enabling the representation of cold abyssal waters within the Vema Channel, whereas horizontal refinement alone distorts the thermal structure without recovering the coldest water masses. The combined refinement of both grids, however, goes beyond simply improving the temperature field: it also allows the model to reproduce the eastward deflection of the AABW core, a dynamical feature that cannot be captured without sufficient resolution in both dimensions. As in the previous subsections, no adjustments were made to the mixing parameterizations, and the degradation observed in FESOM H may partly reflect this limitation.

During the CLIVAR period (Fig. a and d), FESOM R shows the highest correlation (≈0.75) among the four experiments, but its mean temperature substantially overestimates the reference. FESOM V, despite a near-zero correlation (≈0.02), reproduces the observed variability more faithfully (small RMSE) and exhibits the smallest warm bias, with a mean temperature close to the reference. FESOM H displays a weak correlation (≈0.08) and the largest warm bias among the four configurations. FESOM VH, like FESOM V, has a small RMSE, it also improves upon FESOM H in the mean temperature, but still exhibits small correlations (≈0.20). Overall, vertically refined configurations (FESOM V and VH) better capture the variability and mean state of the temperature within the channel, while the horizontal refinement seems to deviate the temperature mean to warmer values.

During the E2 period (Fig. b and e), FESOM V achieves the highest correlation (≈0.64) among the four experiments, with small variability, a near-zero RMSE, and the smallest warm bias, making it the configuration most similar to the reference. FESOM VH also shows small variability and RMSE, but with a worse correlation (≈-0.23) than FESOM V. FESOM R shows no correlation (≈0.00) and a larger variability, while FESOM H remains the warmest and most variable configuration, its correlation is (≈0.22). As in the CLIVAR period, vertically refined configurations outperform those with horizontal refinement alone when representing the mean state and variability of the temperature in the channel.

During the SAMBA period (Fig. c and f), FESOM V and VH remain consistently closer to the reference than FESOM R and H, both in mean temperature and variability. FESOM H shows the largest warm bias and variability, while FESOM V and VH exhibit smaller biases and RMSE values. None of the four simulations achieve high correlation values: FESOM R (≈0.24), FESOM V (≈0.15), FESOM H (≈-0.010), and FESOM VH (≈0.12).

Across all three mooring periods, vertical resolution consistently emerges as the key factor controlling model performance. FESOM V and VH systematically reduce warm biases and better reproduce the observed variability, while FESOM H generally performs worse than FESOM R in terms of mean temperature and variability. However, the correlation analysis across the three mooring sites does not allow a single FESOM configuration to be identified as systematically superior. FESOM R shows the highest correlation at the CLIVAR mooring (r=0.75), while FESOM V performs best at the E2 mooring (r=0.64), and all configurations show weak correlations at the SAMBA mooring. This suggests that vertical resolution primarily controls the mean state and variance of the simulated temperature, but does not guarantee phase agreement with the observed temporal variability at individual mooring locations.

The simulations were also compared with CTD data from the SAMBA-West array at Sites BB, C, and D (Table , Fig. ). At Site BB, all FESOM configurations produce temperatures warmer than the reference, with FESOM V providing the closest match and FESOM R the largest deviation. The observed CTD trend at this site is not statistically significant, which excludes a meaningful comparison of warming trends with the simulations.

At Sites C and D, FESOM V again provides the closest representation of the reference mean temperature, while FESOM H shows the largest deviations. FESOM VH performs comparably to FESOM V in terms of mean temperature, while FESOM R shows intermediate values. Regarding trends, all configurations overestimate the observed warming, with FESOM H showing the largest overestimation and FESOM R the smallest. These results confirm that enhanced vertical resolution is the primary factor driving improvement in the representation of abyssal temperatures along the SAMBA-West line, while horizontal refinement alone has little systematic benefit.

Taken together, the comparison with in situ data reinforces the conclusions drawn from the previous subsections: vertical resolution is the dominant factor controlling the representation of abyssal water properties in the Vema Channel and along the SAMBA-West array, while horizontal refinement alone does not yield systematic improvements and can even degrade model performance.