As mentioned previously, both simulations were performed following the HighResMIP protocol, with an extension of the spin-up length to 100 years. A one-year coupled spin-up with a reduced ocean time step was performed beforehand exclusively for the MR configuration to mitigate initial shocks from the PHC3 ocean climatology initialization (PHC3.0, updated from ). From the final state of the spin-up, both a control simulation (ctl, 151 years) and a transient simulation (65 + 86 years) were initialized. During the spin-up and the control simulation, greenhouse gases and aerosol concentrations were fixed at 1950 levels. In contrast, the transient simulations followed the CMIP6 historical and SSP585 scenario forcings for greenhouse gases and aerosols .

For both simulations, a single simulation has been performed and analyzed in this study. This is primarily due to the limitations with regard to available computing resources, which make the production of multiple ensemble members in the MR configuration challenging (Table ). Further discussion of this aspect is provided in the Conclusions.

The two simulations were carried out using the same AWI-CM3 model version; however, as previously noted, the set of physical parameterizations applied varies with spatial resolution. The goal of this study is to evaluate the impact of resolution on model performance. To achieve this we have chosen to compare two simulations that are each optimized for their respective grid resolution. Parameter choices in both setups result from dedicated tuning processes designed to maximize performance at each resolution, which closely follows the approach of the HighResMip2 protocol for CMIP7 . Differences in parameter settings were anyway kept to the minimum, to ensure as clear a comparison as possible, with only a small number of technical parameters adjusted during tuning. Among the tuned parameters, albedo usually plays a particularly influential role due to its strong effect on the model’s radiative balance and surface energy fluxes. In our model, distinct four constant albedo values are prescribed for sea ice, snow, and melting ice or snow. A common set of values was identified during the tuning process and applied consistently across both configurations.

It is worth noting that the simulation presented in this study is based on a slightly updated version of AWI-CM3 compared to the setup described in , though the two remain highly comparable. The most significant difference concerns the remapping method used for ice-related heat fluxes exchanged between the atmosphere and ocean via the model coupler. In the earlier setup, the Gaussian remapping method introduced a spurious heat flux over ice-covered regions of the Northern Hemisphere leading to a reduced sea ice extent and premature disappearance of sea ice (Fig. ). In this model version that has been replaced by a bi-cubic remapping method that improves the Arctic sea ice representation. The Southern Hemisphere was not affected by the previous issue, and we do not observe a substantial change in the simulation of Antarctic sea ice.

Before evaluating the climatological performance of the simulations, we assess the degree of equilibration of the coupled system under constant forcing. While AWI-CM3-LR exhibits a clearly positive top-of-atmosphere (TOA) radiative imbalance since the beginning of the simulation, AWI-CM3-MR starts with near-zero values; both simulations, however, display a declining trend during the spin-up phase (Fig. a, b). After spin-up, the imbalance stabilizes at approximately +0.2 W m⁻² in the low-resolution configuration, reasonably on the lower of CMIP6 models average and reference estimates for the present-day climate , and at a nearly symmetric negative value (∼ −0.24 W m⁻²) in the medium-resolution case. The surface energy imbalance is smaller in the LR simulation and more pronounced in the MR configuration, but in both cases it closely follows the evolution of the TOA flux. The relatively stable TOA–surface difference indicates that the atmospheric column remains internally consistent, while suggesting that, particularly in the MR simulation, the residual energy imbalance is primarily associated with ongoing ocean adjustment.

In both configurations, the global-mean ocean temperature bias exhibits a relatively stable vertical structure (Fig. c, d). However, differences emerge in both the temporal evolution and the magnitude of these biases. In the LR simulation, a weak cold bias around 100 m is present during the initial adjustment phase and diminishes over time as warm biases at intermediate depths become more pronounced. This is accompanied by a persistent subsurface warming throughout the control period, indicating a sustained drift in the ocean interior despite the near-stationary behavior of the global energy budget, broadly consistent with . In contrast, the HR simulation shows a more stable depth-dependent pattern and less pronounced ocean temperature drift, with weaker subsurface warming and a clear upper-ocean cooling, consistent with a net upward energy flux at both the surface and the TOA.

While the low-resolution simulation can be considered reasonably equilibrated by 1950, the medium-resolution configuration still exhibits signs of residual imbalance at the time the transient simulations are branched off. This is reflected in the larger variability of the radiative balance and is consistent with the presence of an eddy-permitting ocean. Explicitly resolving mesoscale processes substantially increases the timescale required for the deep ocean to adjust . As a result, the approach to equilibrium is slower, and residual imbalances can persist over longer periods compared to lower-resolution configurations relying on parameterized processes .

Contrasts between low- and high-resolution configurations have been reported in HighResMIP and other multi-model studies, where increasing resolution leads to a reorganization, or even a reversal in sign, rather than a systematic reduction of model errors . However, most relevant for the present comparison is that both simulations exhibit an overall limited and well-constrained drift from 1950 onwards. This is evident in the weak trends of global-mean sea surface temperature and near-surface air temperature, as well as in the relatively stable sea ice extent across seasons and regions (Fig. ).

To obtain an objective overview of the overall performance of the model, we use a multivariate comparison against observational climatologies. This approach provides a broad assessment of how well the model captures key climate features with respect to the CMIP6 multimodel ensemble. Eighteen key climate variables were selected for this evaluation. For each variable, we calculate the absolute climatological error in the AWI-CM3 simulations, considering all four seasons and seven distinct regions. The absolute error is first computed at each grid cell of an interpolated grid and is then averaged over each of the seven regions and over each season. Finally, the resulting model error is compared to the average absolute error of 30 CMIP6 models and expressed as a fraction of this CMIP6 mean error. Values below 1 indicate better performances, whereas values above 1 suggest larger errors (Fig. a). Only the last 25 years of the historical simulation (from December 1989 to November 2014, in line with the seasonal analysis) have been considered here to maximize comparability with the observational datasets. It is worth noting that not all observational datasets cover the entire period; a detailed overview of the coverage period for each variable is provided in . To directly compare the two simulations, the index for AWI-CM3-MR was computed again using AWI-CM3-LR as the simulation reference for the evaluation (Fig. b). In this way the MR simulations show better performances than the LR simulations for values below 1 and larger errors for values above 1.

The overall score, computed as the average across all variables, regions, and seasons, indicates a general improvement over the CMIP6 standard (Figs. a and ) which, in the MR configuration (AWI-CM3-MR), corresponds to an average error reduction of 27.9 %. Examining the variable-specific errors, the simulation demonstrates a widespread reduction in the atmospheric absolute error relative to the CMIP6 models considered here, along with notable improvements in ocean dynamics. Large errors in sea surface temperature (tos) are primarily observed in the mid- and tropical latitudes and are associated with a persistent warm bias in the model (discussed in detail in Sect. 2.2), which is significantly reduced at higher latitudes, particularly in the Southern Hemisphere, when resolution is increased. It was previously noted that the model tends to display a substantial misrepresentation of mixed-layer depth (mlotst) in the Antarctic , however in this case it is only in the MR configuration showing this error. Despite the larger error magnitude with respect to the LR configuration, the MR setup produces more realistic sea ice concentration (siconc) values in the region in austral summer and autumn, but shows less agreement with observations in winter and spring. In this case, the comparison among simulations is complicated by the notably higher variability of the MR simulation during the analysis period. In the Northern Hemisphere, sea ice concentration performs well across all seasons, except for an overestimated summer sea ice extent at mid-latitudes (discussed in details in Sect. 2.3), an error that is most pronounced in the MR simulation. The 2 m temperature is generally well-simulated, here the increased resolution leads to more substantial improvements in the Antarctic compared to the Arctic. Higher spatial resolution also has a broadly positive impact on surface winds (uas and vas, representing the zonal and meridional components respectively) at the global scale. Similar improvements are observed for the 300 hPa zonal wind (ua), the 500 hPa geopotential height (zg), and the top-of-atmosphere outgoing longwave radiation (rlut). Precipitation (pr) is improved primarily in the tropics, while enhancements in cloud coverage (clt) are most evident at high latitudes. Ocean temperature and salinity (thetao and so) are, on average, well simulated; however, the impact of increased resolution is not uniform across regions or ocean depths. The higher resolution in MR generally helps reduce the ocean warm bias, except in southern mid-latitudes at 1000 m depth, where the MR is warmer (Fig. a–d). On the other hand the salinity bias, which is still below average in the LR simulation (Fig. ), is amplified at the surface, in the tropical upper ocean and in the Southern Ocean at 1000 m depth (Fig. e–h).

The overall evaluation scores show that the LR simulation achieves a ∼ 13 % error reduction relative to the CMIP6 multi-model mean, while the MR simulation improves this further to a ∼ 28 % reduction. Therefore, approximately 55 % of the total error reduction compared to the CMIP6 standard can be attributed to the increase in spatial resolution. While this overall score provides a useful summary of model performance, it involves an inherent degree of arbitrariness, as it depends on the selection of variables and regions. It is therefore interpreted alongside the individual diagnostics presented below, rather than as a standalone measure of model skill. In the following, a more detailed analysis is conducted to assess specific aspects of the two simulations.

A visible difference between the two model configurations lies in their mean state temperature bias. Even in the control simulation, a persistent offset is evident: the LR configuration is, on global average, approximately 1.5 °C warmer than the MR configuration for both sea surface and near-surface air temperatures (Fig. a). A similar bias is also evident in the transient low-resolution simulation when compared to observations, becoming particularly pronounced from the last decade of the 20th century onward. In contrast, the MR setup more closely reproduces the observed magnitude of the global mean surface temperature. During most of the satellite era (1990 onward), the MR configuration better represents the observed near-surface air temperature trend (MR: 0.25 °C per decade; LR: 0.19 °C per decade; ERA5: 0.24 °C per decade).

For sea surface temperature, both simulations exhibit a warming trend stronger than observed, with the MR configuration showing the largest deviation (MR: 0.18 °C per decade; LR: 0.14 °C per decade; HadISST: 0.11 °C per decade). It is also worth noting that the low-resolution simulation exhibits a stronger residual model drift in its control experiment, which is substantially mitigated by increasing the spatial resolution, consistent with the suggestion of , who highlight the role of numerical methods and physical parameterizations in contributing to model drift. Over the 151-year simulations, the total surface temperature drift in the LR configuration amounts to 0.21 °C in the ocean and 0.24 °C in the atmosphere, compared to only 0.06 and 0.10 °C in the MR configuration, respectively.

The spatial distribution of the biases highlights how the surface warm bias in the lower-resolution setup is fairly homogeneous at the global scale, affecting both atmosphere and ocean (Fig. a, i). Nonetheless, specific hot spots emerge – such as the El Niño/La Niña upwelling regions, the Southern Ocean, and the so-called northwest corner of the North Atlantic Ocean, identified as common biases in GCMs and references therein) – as well as the Kuroshio region. Consistent with comparisons between CMIP6 and HighResMIP , the magnitude of these biases is significantly reduced when horizontal resolution is increased (Figs. e, m and ). The negative surface temperature bias near the equator, previously identified in the AWI-CM3 prototype , is relatively weak in this LR simulation and is mostly confined to land areas. Conversely, much of the low- to mid-latitude region exhibits a cold bias in the higher-resolution configuration. The 2 m temperature bias over the central Arctic can to a large degree be explained by a warm bias present in the ERA5 reanalysis . At higher resolution, this negative bias exceeds 4 °C and peaks in the Canadian sector of the Arctic Ocean, while it is less pronounced at lower resolution. This pattern also resembles results obtained from comparisons of CMIP6 and HighResMIP multi-model means . However, as mentioned above, this difference must be contextualized, as recent studies indicate that ERA5 itself exhibits a warm bias exceeding 3 °C in the same region when compared with satellite observations .

The analysis of the last 35 years of the historical simulations indicates that the MR experiment exhibits a substantially reduced mean temperature bias, decreasing from 0.85 to –0.24 °C for global 2 m air temperature, and from 0.88 to –0.05 °C for sea surface temperature. Moreover, the observed global warming trend during this period is well reproduced in the MR simulation. Examining regional trends in the extratropics (Fig. b–d, f–h, j–l, n–p), the benefits of increased resolution are still generally evident, though in some regions less pronounced. In the Southern Hemisphere, the LR simulation fails to capture the observed cooling trend, instead displaying widespread warming in both the atmosphere and ocean (Fig. c, k). The MR simulation, by contrast, introduces more spatial detail: coastal regions tend to show weaker positive or neutral trends, and localized cooling, especially in the Pacific sector, is better represented (Fig. d, l). In the Northern Hemisphere, the MR simulation produces an overly strong surface air warming trend-especially pronounced at mid-latitudes-when compared to reanalysis data, likely related to the more pronounced cold bias (Fig. h). Meanwhile, the LR experiment shows only modest warming across much of the Arctic Ocean, with stronger trends confined to the Greenland and Barents Seas (Fig. g). For sea surface temperatures, both simulations reveal broadly similar warming patterns in the Northern Hemisphere extratropics; the erroneous negative trend in the Labrador Sea is here more clearly visible (Fig. o, p), especially in the MR simulations. Overall, while the MR simulation significantly reduces global mean temperature biases and better captures regional details, especially in the Southern Hemisphere, its improvements in replicating observed spatial trend patterns remain limited and regionally dependent for this individual realization. It is worth noting that regional trend patterns can be strongly affected by internal variability and part of the mismatch with observations may reflect this variability rather than systematic model biases.

The medium-resolution configuration exhibits a stronger increase in air temperature per unit of CO₂ concentration already during the historical period, with this difference becoming even more pronounced starting from the late 2020s (Fig. a). A similar, though slightly less pronounced, signal is also evident in the sea surface temperature. It is noteworthy that this year also marks the onset of a pronounced regime shift in the maximum sea ice area over the Southern Ocean (Fig. b): within approximately 15 years, Antarctic sea ice loses nearly 7 million km² of its area. This rapid decline is analyzed in detail by . Here, we simply note that such a sharp decline is absent in the low-resolution simulation, which instead shows a more gradual and stable decrease typical of many CMIP6 models. Enhanced sea ice variability in the Southern Ocean is also apparent prior to this event in the medium-resolution simulation, with several fluctuations linked to more or less severe Weddell Sea polynya episodes. For this reason, despite being an average over a long period of 25 years, the error displayed when computing the performance index (Fig. a) is influenced by this strong variability.

Several studies have shown that CMIP6 climate models tend to simulate a stably stratified Southern Ocean . Consistent with this, the low-resolution AWI-CM3 simulation exhibits a shallow winter mixed layer depth (MLD) with very limited variability across most of the region south of 60° S (Fig. a, c). Notably, the Weddell Sea lacks any significant deep MLD features, showing some deepening only in a narrow coastal region. This behavior aligns with the limited polynya features in this model configuration. In contrast, the medium-resolution AWI-CM3 simulation shows pronounced MLD variations (Fig. b, d). Along the Antarctic coast, the MLD exceeds 200 m, consistent with active coastal polynya formation. In the Weddell Sea, the MLD reaches depths greater than 300 m, making it the region with the strongest variability in the Southern Ocean. This deeper and more variable MLD contributes to a more plausible representation of polynya occurrence, while at the same time leading to an overestimation of event intensity. It is important to note, however, that the observational record of polynya events is very limited, which adds uncertainty to such evaluations. The overly deep mixed layer is co-located with a localized and pronounced warm bias in 2 m air temperature (Fig. e), likely linked to enhanced oceanic heat release in this region. In contrast, the low-resolution configuration shows no such localized warm bias, instead displaying a relatively uniform positive bias across the Southern Ocean.

This discrepancy explains why the medium-resolution simulation exhibits greater variability in the inner Weddell Sea sea ice maximum compared to observations (Fig. a, e). Modeled September sea ice concentration trends in the Southern Hemisphere mirror the corresponding temperature trends (Figs.  and c, e): while the low-resolution simulation shows a largely uniform decline, the medium-resolution experiment captures some localized increases, such as in the Ross Sea and the Weddell Sea. However, both simulations display a marked retreat in sea ice concentration across East Antarctica, an inconsistency with satellite observations. In the Northern Hemisphere, the observed boreal summer sea ice loss in the central Arctic is largely absent in the low-resolution run but is reproduced, albeit with regional biases, at higher resolution, including excessive retreat in the Greenland and Barents Seas and an absent response in the Russian-Canadian sector (Fig. g, i, k). In March, sea ice extent is overestimated in the medium-resolution simulation (Fig. l) and shows indeed a more spatially distributed decline, stronger particularly at mid-latitudes (Fig. l), consistent with stronger positive temperature trends. In contrast, the low-resolution simulation exhibits spatial variability that more closely aligns with satellite data (Fig. i). During austral summer, neither simulation successfully reproduces the observed sea ice trends of the Southern Hemisphere (Figs. b, d, f and b, d, f). Overall, the main improvement associated with reduced grid spacing is a better representation of average sea ice trends (Fig. ) and an enhanced ability to capture regional variability in the Southern Hemisphere. However, both resolutions still fall short in reproducing key observed features, such as the stability of East Antarctic sea ice and the full extent of Arctic summer sea ice loss.

The AWI-CM3 simulations reproduce the expected structure of the AMOC streamfunction in its mean state (Fig. a, b). However, two main differences emerge when comparing the streamfunctions from the low- and medium-resolution simulations. In the upper cell, within the 500–1500 m depth range, the low-resolution simulation shows a peak in northward transport primarily between 30 and 40° N, whereas the medium-resolution simulation displays a more uniform maximum across the latitudes. The comparison with observations at 26° N reveals that the medium-resolution simulation aligns more closely with the RAPID profile in the upper ocean but tends to overestimate the southward branch of the AMOC, suggesting a too-strong return flow of North Atlantic Deep Water (NADW) along the Deep Western Boundary Current (DWBC) (Fig. c, d). Conversely, the low-resolution simulation, while better capturing the magnitude of the NADW transport, locate its flow too shallow compared to observations (Fig. c, d), consistent with findings by , and with a too pronounced variability. Additionally, the medium-resolution simulation shows a clearer signal of northward transport of Antarctic Bottom Water (AABW) at abyssal depths (Fig. a, b). Similar resolution-dependent differences have been reported in earlier studies (e.g., ).

Starting around 1980, both simulations show a decline in the AMOC at 26.5° N under increasing atmospheric forcings (trend absent in the control simulations), but this weakening is confined to the northward flow component of the upper cell. In contrast, the southward branch exhibits no clear trend in the low resolution simulation, indicating a deep ocean circulation that is less sensitive to surface-driven changes, at least initially during the transient response phase. On the other hand, higher resolution shows more coherent changes between northward and southward transport, showing a decline in DWBC especially pronounced in the last two decades of the 21st century. As shown in , improving atmospheric resolution can lead to significant AMOC weakening via salinity-driven suppression of deep convection in the North Atlantic deriving from altered wind and ocean circulation.

Comparing the simulated volume transport across major ocean channels with observational estimates, provides initial insight into the effects of both increased horizontal resolution and a finer vertical discretization on ocean circulation and vertical processes (Table ). In the Arctic, the medium-resolution simulation generally produces stronger volume transports across key straits compared to the low-resolution setup, with a tendency toward a positive bias relative to observations. Through the Davis and Fram Straits, the medium-resolution simulation falls within the observational uncertainty range, while the low-resolution one tends to underestimate the outflow. In contrast, the medium-resolution simulation overestimates poleward transports through the Barents and Bering Seas, leading to an excessive inflow of warm Atlantic and Pacific waters. This increased inflow likely contributes to a positive near-surface salinity bias in the Arctic Ocean, which is more pronounced in the medium-resolution case (Fig. a–d). This bias is consistent with known shortcomings of AOGCMs, often attributed to an insufficient freshwater supply from Siberian rivers . At the same time, the enhanced northward transport in the medium-resolution simulation slightly reduces the warm bias (Fig. a, e) and significantly decreases the salinity bias in the North Atlantic near-surface (Fig. a, b).

Improvements are also evident in the representation of the Gulf Stream separation, which reduces the localized fresh bias in the North Atlantic surface, and in the freshening of the Mediterranean Sea, which mitigates the excessive outflow of overly salty and warm Mediterranean waters around 40° N (Fig. c–f), present in the LR. These Mediterranean waters, in the low-resolution simulation, spread northward along the European continental shelf, contributing to a saltier and warmer NADW, which is then transported southward. The medium-resolution simulation shows improvements in ocean circulation and mixing processes, leading to a more realistic representation of deep water formation and large-scale circulation in the North Atlantic Ocean, consistent with the findings of .

Consistent with , resolving mesoscale eddies in the medium-resolution simulation reduces the positive deep-ocean temperature bias overall, indicating more efficient vertical heat transport from the interior-with the exception of 1000 m depth where MR shows a slightly warmer bias than LR simulation in the mid-latitudes of the Southern Hemisphere (Fig. c, d). However, the surface ocean change from LR to MR is different from the classical eddy-heating effect in , where SST tends to warm when explicitly resolving mesoscale eddies. This pattern suggests that, while mesoscale processes improve deep-ocean realism, their effect on subsurface and surface temperatures is spatially heterogeneous, likely reflecting regional feedbacks such as stratification, sea ice dynamics, and coupled ocean-atmosphere interactions.

At higher southern latitudes, both model configurations tend to overestimate volume transports through the selected channels (Table ). However, the medium-resolution simulation generally falls closer to the observational uncertainty ranges. A clear example is the transport through the Drake Passage, where both simulations produce a stronger Antarctic Circumpolar Current (ACC) than older observational estimates (e.g., ). More recent full-depth measurements that include both baroclinic and barotropic components of the flow, however, indicate higher transport values . In this context, the AWI-CM3 medium-resolution configuration approaches the upper bound of these updated observational estimates.

A well-known bias in low-resolution models is the incorrect representation of the Agulhas Current system. AWI-CM3 displays a similar feature, with an overly strong advection of warm waters from the Indian Ocean into the Atlantic at around 40° S (Fig. a, b). This bias then extends northward at Atlantic Intermediate Water (AIW) depths (Fig. c–f). Consistent with previous studies , resolving ocean mesoscale eddies leads to a more realistic ocean vertical stratification. In line with this, the medium-resolution simulation shows a reduction in the Agulhas leakage and in the warm and salty biases near the thermocline, particularly in subtropical regions, due to improved eddy-driven mixing (Fig. c–f). The only exception to that is a slightly increased salinity bias in the water column above 2000 m, between 40 and 60° S, but is part of the Brazilian current rather than of the Agulhas one.

It has been shown that variations in the Agulhas system can influence thermohaline properties in the North Atlantic and potentially modulate the AMOC through the salt–advection feedback . In our experiments, however, the low-resolution model shows only a limited AMOC response despite the stronger Agulhas leakage, whereas in the medium-resolution configuration the more realistic representation of the Agulhas system is accompanied by a stronger AMOC. A plausible explanation is that, in the LR simulation, compensatory effects between subsurface warming and salinification in the North Atlantic dampen the AMOC response. Furthermore, suggest that under present-day climate conditions the Agulhas leakage exerts only a weak influence on the AMOC, which is broadly consistent with our findings. An additional aspect worth noting is that the more realistic Agulhas representation in MR appears to be associated with a stronger sensitivity of the deep circulation to surface forcing and coupled feedbacks.

The AWI-CM3 low-resolution simulation shows a widespread positive bias in 500 hPa geopotential height, particularly pronounced in the Northern Hemisphere extratropics (Fig. a). In contrast, a negative bias dominates high southern latitudes, with particularly strong anomalies around Antarctica. In the medium-resolution simulation, the bias pattern is largely reversed in sign and the overall magnitude of the biases is substantially reduced globally (Fig. b).

The global zonal-mean temperature cross-sections for the low-resolution simulation (Fig. c) reveal a bias structure and magnitude similar to that reported by . In the medium-resolution the cold biases in the upper troposphere (around 250 hPa), present in both hemispheres, persist but are reduced (Fig. d). An analogous, but more pronounced improvement is seen in the warm biases of the lower troposphere over polar regions as well as in the positive temperature bias in the upper stratosphere. A particularly notable change occurs in the extra-polar troposphere, where the warm bias present in the low-resolution simulation not only decreases in magnitude but also reverses sign in the medium-resolution run. The sign and vertical structure of the tropospheric temperature biases – and how they differ between the two simulations – directly explain the contrasting patterns of geopotential height bias observed in the two configurations.

In CMIP6 models, tropical tropospheric warming is typically overestimated, especially between 200 and 500 hPa. This bias has been linked to surface warm biases , as well as to excessive vertical development of convection and underestimated entrainment rates in the troposphere . In the medium-resolution simulation presented here, surface warming is substantially reduced, and a similar reduction is seen in the tropical troposphere. This suggests that the improved representation of convective and dynamical processes – including tropical waves, mesoscale eddies, and smaller-scale convection – helps mitigate these persistent biases.

Because of the warm bias, the low-resolution simulation underestimates the equator-to-pole temperature gradient in the troposphere (Fig. c), as a consequence the subtropical jet streams are displaced too far toward the poles and are simulated with excessive strength and vertical extent compared to reanalysis data (Fig. e). This bias is particularly pronounced in the Southern Hemisphere and helps explain the apparent decoupling between the geopotential height and tropospheric temperature biases. The overly strong, poleward-displaced, and vertically extended westerlies intensify the circumpolar trough, leading to lower geopotential heights around Antarctica despite the presence of a warm tropospheric bias.

In the medium-resolution simulation, the reduction of tropospheric temperature biases is accompanied by a substantial decrease in zonal wind biases across most latitudes and pressure levels (Fig. f), indicating a more realistic thermal wind balance. In contrast, biases in equatorial stratospheric winds worsen with increased resolution. The medium-resolution simulation continues to exhibit stronger-than-observed westerlies in the upper stratosphere, indicating that AWI-CM3 still struggles to adequately reproduce the easterly phase of the Quasi-Biennial Oscillation (QBO), still a challenging aspect for climate models . This likely affects the representation of the Semiannual Oscillation (SAO), whose variability is strongly modulated by the QBO . Accordingly, the negative wind bias in the uppermost stratosphere points to an underestimated westerly phase of the SAO.

As shown by , most of the positive surface temperature biases can be co-located with negative biases in cloud area fraction, and consequently with an overestimation of surface shortwave radiation (Figs. a, i and b,c). This is particularly evident in the upwelling regions off the west coasts of southern Africa and South America. Interestingly, while off the coast of South America the consistent reduction in temperature bias with increasing resolution corresponds to an improved representation of clouds and, consequently, surface radiation fluxes, the situation is different off the coast of southern Africa. There, the warm bias is reduced despite a slight deterioration in the model’s representation of both clouds and radiation fluxes. A similar correspondence of patterns can be identified in the Pacific Ocean, where the temperature bias associated with the double Intertropical Convergence Zone (ITCZ) is mirrored in the bias structures of clouds, precipitation, and shortwave radiation-also as pinpointed by . In this low-resolution configuration of AWI-CM3, the precipitation bias is larger than the CMIP6 ensemble mean . However, increasing the model resolution leads to a considerable improvement, particularly in the tropical belt where we see an error reduction close to 20 % across all seasons (Fig. ). A weakening of the double ITCZ bias is also visible in the medium-resolution configuration (Fig. a, e). Mid-to-high latitudes show a limited difference with respect to observations, in both the simulations. On a global scale a slight reduction is also observed in the shortwave radiation bias, when moving from low to medium resolution. Conversely, negative biases in clouds and longwave radiation increase in magnitude. This is primarily because the improvement in these two fields is more substantial for positive biases. Nevertheless, all the variables considered show a reduction in root mean square error (rmsd), indicating an overall improvement in model performance.

Representing cloud cover over high latitudes remains a challenging aspect for AOGCMs. In agreement with most CMIP6 models , AWI-CM3 tends to overestimate cloud cover over the Southern Ocean (Fig. b, f). However, this bias coexists with an overestimation of surface shortwave radiation (Fig. c, g), suggesting that the issue involves not only the cloud amount but also cloud properties – particularly an underrepresentation of low-level clouds or an underestimation of cloud albedo, as also noted in other CMIP6 studies . Given the low surface reflectivity of the Southern Ocean, the surface albedo is particularly sensitive to cloud characteristics. Consequently, due to the strong shortwave cloud radiative cooling effect, the net cloud radiative forcing in this region is dominated by the shortwave component. This bias is likely linked to sea ice extent errors, especially during summer. For example, in the region off West Antarctica, where the positive surface shortwave radiation bias is reduced, the medium-resolution simulation shows a slight improvement in sea ice extent compared to the low-resolution configuration (Fig. b, d, f). Over the Arctic Basin, cloud cover overestimation has a different impact on the surface energy budget. In this region, clouds primarily act to warm the surface during winter by enhancing downward longwave radiation . Accordingly, AWI-CM3 shows a positive bias in surface net longwave radiation due to the excess cloud cover, while the net surface shortwave flux is underestimated (Fig. b–d, f–h). At lower Arctic latitudes, going from low to medium resolution, the longwave bias becomes weaker while the shortwave bias becomes more pronounced (Fig. c, d, g, h). This enhanced cooling effect likely contributes to the co-located overestimation of sea ice extent mentioned earlier.

Overall, it is worth noting that for all variables, the bias in the medium-resolution model configuration relative to observations is smaller in magnitude than the CMIP6 multi-model mean bias (Fig. ). Moreover, the improvement compared to the low-resolution configuration, though modest, appears consistent across seasons and geographic regions.

As mentioned earlier, our medium-resolution simulation shows a significant drop in September sea ice area in the Southern Ocean around year 2030, a feature not present in the low-resolution simulation (Fig. b). This pronounced decline, which is indicative of a strong non-linearity in the climate system, is centered in the Weddell Sea. Here approximately half of the total sea ice area is lost during the decade analyzed (Fig. a). In contrast, the low-resolution simulation maintains a relatively stable sea ice cover (Fig. i), exhibiting only slightly increased variability after mid-century and no clear decline before the last two decades of the century, consistent with the CMIP6 multi-model ensemble mean.

Prior to the onset of sea ice loss, upper ocean temperature increases steadily, though the trend differs between the two configurations: the low-resolution run shows a monotonic rise, while the medium-resolution simulation exhibits a step-like pattern (Fig. b, j). As detailed by , this marks the onset of a cascade of ocean–atmosphere feedbacks. Increased heat release from both sea ice and the ocean – via latent and sensible heat fluxes (Fig. e, f) – warms the lower atmosphere. Through the lapse rate and the albedo feedbacks both tropospheric temperature and humidity shift, enhancing the surface net shortwave radiation flux (Fig. g) and modifying the vertical structure of the atmosphere. This affects cloud cover (Fig. c, d), and cloud optical depth, ultimately influencing the surface downward longwave radiation flux (Fig. h). Figure  illustrates that this chain of radiative feedbacks is well captured in our medium-resolution simulation. The AWI-CM3 model version used here closely aligns with that employed by , and both medium-resolution simulations use the same grid resolution. This allows our simulation to be considered essentially a second realization of the one presented by , supporting the robustness of their findings. As shown in Fig. a, the sea ice regime shift appears in both simulations and occurs at a similar point in time. Importantly, similar feedback signatures are also identifiable in the observational and reanalysis data analyzed (Fig. q–x), particularly during the well-documented polynya event of 1975/1976. These processes are evident both during smaller polynya-like events that occur before the major sea ice decline around 2030, and during the large-scale decline itself. In the latter case, the magnitude of the variability is markedly stronger, and unlike the earlier events, no recovery of sea ice extent is observed in the Weddell Sea region thereafter. In contrast, the low-resolution simulation fails to reproduce this sequence of feedbacks. Across all examined fields, it shows a persistently monotonic and stable behavior, lacking the coupled dynamical and radiative variability seen in both the medium-resolution run and observational data.

It can be noted that, when comparing the timeseries, the medium-resolution simulation tends to overestimate the amplitude of variability relative to reanalysis data. However, studies have shown that ocean reanalyses often disagree in their representation of seasonal and interannual variability , particularly on regional scales, and that Southern Ocean variability can exceed what is captured by existing reanalyses or sparse observations . In conclusion, these results support the suggestion of a potential sudden, accelerated Antarctic sea-ice decline in the coming decades, but also show the evidence for a nonlinear sea ice sensitivity to greenhouse warming, which can contribute to explain the apparent mismatch between the current generation of climate models and observations.