The choice of horizontal resolution for EAC03K was based on the length scales that we are interested in resolving within the EAC system. To understand the range of dynamics that we expect to resolve, one must consider the horizontal grid spacing with respect to the internal deformation radii of the region . This relationship is approximated by 1RH=Ld,1Δ=cg2f2+2βcgΔx2+Δy22 where Ld,1 is deformation radius of the first baroclinic mode, cg is the associated gravity wave speed, f is the Coriolis parameter, and β=∂f/∂y is its meridional gradient. The approximate grid size Δ is taken as the average of the grid length along the two horizontal axes, Δx and Δy.

We show the spatial distribution of the first baroclinic deformation radius in Fig. a, whereby typical open ocean values exceed 90 km towards the equator, and contract poleward to around 20 km in the mid-latitudes – primarily controlled by the increased effect of the Coriolis parameter towards the poles. The ratio of this length scale to the respective grid spacings of the two resolutions is shown in Fig. b, where the difference between the two models is very clear. The horizontal resolution of the 10 km models (STHPAC10K and OM2-01) yields a Ld,1/Δx ratio of approximately 2.5 across most latitudes, while the EAC03K model exceeds a ratio of 6 throughout (Fig. b). According to , a ratio of at least 2 is sufficient to omit the Gent-McWilliams (GM) mesoscale eddy parameterization , as the grid is fine enough to partially resolve large-scale eddy effects. This justifies the exclusion of GM in our 10 km configuration. However, resolving the full dynamical influence of mesoscale eddies on the general circulation – particularly their energy cascades and interactions with topography – requires finer resolution. recommend a ratio of 6 for this purpose, and EAC03K meets this criterion across the entire domain (Fig. b). At this resolution, the model can begin to capture higher-order baroclinic modes, likely resolving up to the third baroclinic mode in the open ocean, and on the order of the first baroclinic mode over the shelf . Combined with the increased vertical resolution , this horizontal resolution leads to a well-resolved mesoscale regime for EAC03K throughout the domain, whereby scale interactions are free to self-regulate without diffusive parameterizations suppressing the variability. demonstrate that explicitly resolving mesoscale eddies leads to increased transport within these coherent features. We anticipate this will lead to an improved representation of the EAC extension eddy field, and subsequent improvement in temperature variability both inshore and offshore of the continental shelf along south-eastern Australia .

To ensure consistency across model configurations, the two regional models use the same atmospheric forcing as the global model (OM2-01): version 1.5 of the Japanese Reanalysis for Driving Oceans (JRA55-do v1.5; ). The historical simulations span 1990–2020, using interannual forcing (IAF) to capture the large-scale climate variability and long-term change over three decades. Surface momentum fluxes are computed online within MOM6 from these prescribed atmospheric fields using a relative-wind formulation, such that the applied wind stress depends on the difference between the 10 m wind and the model surface ocean velocity. In an ocean-only configuration, this relative-wind formulation is expected to reduce the net wind work on the ocean circulation, and can damp mesoscale eddy variability relative to simulations forced with direct winds .

Tidal forcing and river inputs are omitted in the regional configurations, as our focus is on the deep ocean circulation's influence on shelf conditions, and river runoff in the Australian region is relatively low. We are primarily interested in the ocean dynamics governing connectivity across the shelf break and not necessarily the localised processes driving variability within the inner shelf region (< 200 m). Previous studies of different sections of the shelf along eastern Australia have found very little influence from tidal and river forcing on shelf conditions away from the estuaries and inner shelf .

The first sub-grid parameterisation sensitivity experiment concerns the Mixed-Layer Eddy (MLE) scheme. In MOM6, this scheme acts to restratify the mixed layer in the presence of sharp horizontal buoyancy gradients , and has been shown to influence mixed-layer structure and sea surface temperature, particularly in coarser-resolution models where mixed-layer eddies are unresolved . Here, however, our aim is not to evaluate MLE in that established coarse-resolution setting, but to test whether it remains influential in the 3 km EAC03K configuration, where submesoscale variability is partially resolved and some mixed-layer restratification processes may begin to be represented explicitly.

The sensitivity of this scheme is controlled by a single dimensional parameter, the frontal length scale, which is on the order of the mixed layer deformation radius, NH/f (where N is the buoyancy frequency and H is the mixed layer depth) and acts as a lower bound upon which a mixed-layer overturning streamfunction is scaled via 2Ψ∝ΔsLf where Δs is the model horizontal grid scale, and the frontal length scale is chosen as 3Lf=max⁡NH/|f|,Lf,min⁡.

The base value for the nested simulations was Lf,min⁡=500 m , whilst we run two additional sensitivity experiments, with Lf,min⁡=1500 m and with the parameterization turned off entirely (Table ). A larger minimum length reduces the influence of the parameterization, and the mixed layer will not restratify as much in response to strong horizontal gradients. This is expected to have an impact on the SST long-term statistics. It is uncertain whether the scheme is needed at higher resolutions, hence turning it off will reveal insights into the model's ability to resolve processes associated with large mesoscale eddy strain.

In order to investigate the influence of the Fox-Kemper MLE parameterization , we additionally assessed biases in the mixed layer depth (MLD) against the Global Ocean Reanalysis and Simulations (GLORYS12V1; reanalysis product. This reanalysis was chosen due to its high resolution and duration spanning the majority of our model duration, with a start date of 1993.

The second sensitivity test assess the model response to the dynamic viscosity parameterization, referred to throughout as VISC. Numerical modelling of WBC regimes has historically revealed a large sensitivity to the horizontal viscosity parameterization used in the model. In the Euler momentum equations of geophysical fluid dynamics, viscosity acts to dampen the linear terms, and suppress growth of the nonlinear terms whilst the precise nature of the effect depends on whether a Laplacian or biharmonic velocity scaling is employed . The biharmonic approach has proven to be most effective at suppressing grid-scale noise whilst preserving energy at large scales and is used with the Smagorinsky dynamic viscosity scheme for all configurations in this paper .

note a strong sensitivity of the Gulf Stream separation in their MOM6 based regional model to the biharmonic viscosity whereby higher viscosities lead to a more poleward separation of the WBC, whilst lower viscosities lead to an earlier, equatorward separation. They note this viscosity response to be typical and consistent with earlier studies e.g.,. However , albeit using a simplified 1.5 layer reduced-gravity linear model of the EAC system, found the opposite effect – a weaker viscosity led to a more poleward separation, whilst a stronger viscosity led to an equatorward separation.

These two contradicting cases highlight the important and competing processes driving WBC separation. In the nonlinear case of , instabilities grow through baroclinic and barotropic processes, generating eddies that disrupt the coherent WBC and drive separation. Since viscosity damps these instabilities, a higher viscosity suppresses the turbulence and delays separation, whereas a lower viscosity allows faster eddy growth and earlier separation. In the linear case, however, separation is dictated by the large-scale wind stress curl and planetary vorticity balance, rather than instability-driven processes. In this system, viscosity dissipates momentum from the WBC, reducing its inertial resistance to the wind-driven gyre circulation. As a result, a higher viscosity weakens the WBC, making it more susceptible to turning eastward at a lower latitude, whereas a lower viscosity allows a stronger WBC that can persist farther poleward before separating.

We adopt the dynamic viscosity scheme chosen by whereby the imposed horizontal viscosity is the maximum of two quantities. The first is a predetermined velocity scale multiplied by the cube of the grid spacing u4Δx3, while the second is the biharmonic Smagorinsky viscosity – a dynamic flow-dependent quantity. The coefficients of the Smagorinsky viscosity were set to 0.06 in the double-nested and MLE experiments, which is the recommended upper bound for the MOM6 model . To test an alternate case, we take the MLE-1500 experiment from the MLE tests, and refer to this as the high viscosity run (VISC-H), while we run one additional experiment – VISC-L – with a Smagorinsky coefficient set at the lower bound of the MOM6 recommended range, 0.015 as also used in the NWA12 model (Table ).

Daily sea surface temperature (SST) fields from all models are compared with the National Oceanic and Atmospheric Administration SST product, developed using optimum interpolation (OISST; ). The OISST product combines satellite and in situ observations and is representative of the top 0.5 m of the ocean. Daily sea surface height fields (SSH) fields are evaluated against the Archiving, Validation, and Interpretation of Satellite Oceanographic data (AVISO) altimetry-derived SSH product.

All region-wide metrics are evaluated over the EAC domain (Fig. , EAC03K box) and all model grids are bilinearly interpolated onto the observational products' horizontal grid. The root mean square error (RMSE) at each coarsened model grid point is calculated and the spatial average of this metric is computed in order to assess overall differences between datasets.

Marine heatwave (MHW) analysis has also been performed on the surface SST fields of all datasets, following the approach of . The SST fields are first regridded to the coarser OISST (∼ 25 km) grid using bilinear interpolation, and then we computed the daily-varying climatology and the 90th percentile threshold temperature at all grid points in the EAC domain. We then use the 90th percentile temperature as a lower bound and identify all anomalies exceeding this. If the temperature persists above this threshold for greater than 5 days, we record this as a MHW event and various metrics are stored for the event such as the anomaly above the 90th percentile temperature, and the duration.

The EAC jet flows coherently along the shelf break of southern Queensland and northern New South Wales between latitudes 14–32° S. In 2012, a mooring array was established at 27.5° S to gain insight into the subsurface structure and variability of the jet upstream of the separation point . This mooring array was in operation until 2022, allowing almost a decade of high spatial and temporal resolution in situ measurements of the jet. In this jet regime, we are interested in whether the regional models can represent the seasonal cycle of volume transport identified in observations. We compute poleward transports within the longitudinal range of the mooring array, and calculate the climatology over the simulation period. We do not take into account interannual differences here, and compare the climatologies calculated from 30 year model simulations with that from the ∼ 10 year observational period.

In order to identify where the EAC separates from the coast in each model (and observations), we established two distinct definitions that describe this process – the “separation latitude”, and the subsequent “retroflection latitude”. Adapting the approach of , we choose a SSH contour that best represents the path of the EAC jet at 28° S and follow this poleward to the “separation latitude”, defined as the variable point where this contour exceeds a zonal distance of 80 km from the 2000 m isobath (Fig. a, b). Following the contour further south, we define the retroflection point where the tangent of this contour veers due east (i.e., tangent vector transitions from the south-east quadrant to the north-east quadrant (Fig. a, b). We make the distinction between the two points since the dynamics controlling the pair of latitudes are likely semi-independent, however no distinction has been made previously between the two separate processes. Adopting this more nuanced definition of the WBC separation latitudes has revealed new insights into the controlling dynamics of both the separation and retroflection latitudes.

Additional improvements on the approach taken by were necessary to account for cyclonic frontal eddies developing inshore of the EAC jet , leading to the temporary detachment and re-attachment of the jet upstream of the main separation latitude. choose the SSH contour that aligns with the maximum poleward velocity of the upstream EAC jet, which in our 3 km model, led to approximately 20 % of the timesteps failing to identify an appropriate pair of separation/retroflection latitudes. An improvement was achieved when smoothing the nominated SSH contour using a simple coarsened linear interpolation which removed the effect of frontal eddies that pushed the contour offshore, only for it to reattach ∼ 50 km downstream. However, further improvement was achieved by deciding to choose the SSH contour based on an iterative process, similar to that of , whereby we first detect the longest positively-sloped segment of a zonal SSH transect at 28° S and assume (via geostrophy) this represents the poleward EAC jet at this latitude. Then, starting from the western extent of the segment, we follow the SSH contour originating at that point to see whether it is a good representation of the EAC jet. This decision is automated and primarily based on the length of the contour and a sensible result for the detected points. This is repeated at each timestep until a suitable SSH contour is found, which has resulted in a vastly improved failure rate less than three percent. The failed timesteps show SSH contour maps with no sign of an EAC jet present, highlighting the variable nature of the EAC system as a whole.

The most variable component of the EAC system and potentially the most important for our interests in the south-east Australian region is the eddy field aka “Eddy Avenue”; poleward of the EAC separation point. There is a direct relationship between the growth of instabilities in the EAC separation region and eddy characteristics in the EAC extension . As such, if there are biases in the separation dynamics, these are likely to cascade into the EAC extension eddy characteristics in our model. To get a synthesis of the eddy variability in this region, we perform an eddy tracking using the python package py-Eddy-Tracker which can efficiently track the size and polarity of eddies over the 30 year analysis period from the SSH fields. We then calculate the mean amplitude of eddies within coarsened grids (25 × 25 km) which provides a metric to understand the amount of eddy kinetic energy (EKE) present spatially in the model domain.

We anticipate a higher resolution model to have improved connectivity across the shelf break, due to important scales of variability being better represented at 3 km resolution. Here we are interested in assessing how well the different models produce the SST warming rates along the shelf of eastern Australia. We take polygons of all the bioregions that are designed to capture quasi-homogeneous oceanographic and ecological conditions in the nearshore domain . Sea surface temperatures within each bioregion are then spatially averaged to obtain a daily SST time-series for each bioregion and then the linear trend is calculated. We further group the time-series by season and re-calculate the SST trends to understand whether or not the trends are ubiquitous across seasons. In order to further validate the model results, we use an additional observational dataset produced specifically for the Australian region, namely Integrated Marine Observing System SST (IMOS-SST) satellite product . The IMOS-SST product has a resolution of 0.02° and has proven to be more reliable than OISST in representing temperature trends and variability closer to the coast .

The mean SST biases in the global model are relatively weak and spatially patchy, whereas STHPAC10K exhibits minimal bias and closely aligns with observations (Fig. b, c). In contrast, EAC03K shows a pronounced cold bias in the EAC extension, accompanied by a warm bias in the mid-latitude eastern sector of the domain (Fig. a). This pattern may indicate a weaker poleward transport of the EAC, with an associated redistribution of heat toward the eastern boundary. In terms of the SST trends, all models capture the broad pattern of faster warming in the EAC extension (not shown), however EAC03K shows a much more concentrated, coastal-bound warm pool off south-east Australia (Fig. d).

When comparing the SST variance from a model, which represents day-to-day variability, to OISST, which represents a smoothed 5 d window of variability, we expect to find a systematic positive bias for all modelled SST variance due to the reduced spatial and temporal interpolation in the models . This expectation holds true for our 3 km model, EAC03K, whereby there is a tendency for larger variance (positive bias) across the entire domain (Fig. g). Importantly, in the EAC extension region, both 10 km models produce less variability than the already dampened OISST, suggesting that these models may under-represent the intensity or frequency of eddy-driven temperature fluctuations off south-east Australia (Fig. h, i). Both regional models exhibit an expansive north-west to south-east oriented band of high variance across the northern half of the domain, having the appearance of an imprinted south Pacific convergence zone (SPCZ) that is not present in observations or the global model (Fig. g, h). Given all models use the same version of JRA55 atmospheric interannual forcing, this is likely associated with the updated surface flux numerics in MOM6. However it appears to be independent of other biases and hence is not further explored in this paper.

At 10 km resolution, MOM6 appears to slightly improve on MOM5 in representing the mean SST pattern compared with observations (Fig. b, c). However the strong variance bias in MOM6 (Fig. g, h), the nature of which remains uncertain, prevents us from designating the updated model as unequivocally superior. The question around resolution, in comparing STHPAC10K to EAC03K is also not straightforward. For EAC03K, we have an EAC extension region with juxtaposing cold mean bias, and strong warming bias (Fig. a, d), both of which are weak or not present in STHPAC10K (Fig. b, e). Furthermore, EAC03K has a much higher SST variance throughout the extension region compared to STHPAC10K (Fig. g, h). Thus, there is likely weaker mean transport in the EAC jet, whilst the eddies in the extension region are strengthening at a faster rate, leading to the faster warming trends in the 3 km model compared to the 10 km model.

In a similar region to the cold mean SST bias for EAC03k in the poleward extension region (Fig. a), we see a very strong negative SSH mean bias (Fig. j). This SSH bias is most intense in the typical separation region of the EAC and extends to the east and south into the EAC extension (Fig. j). As a result, EAC03K shows a substantially higher root mean square error (RMSE) relative to the coarser models, which both exhibit only minimal and spatially patchy mean SSH biases. The strong coherent SSH bias for EAC03K suggests a fundamental shift in the latitude and structure of the EAC separation at 3 km resolution. How this shift translates into specific dynamical changes will be explored in the region-specific analysis.

All models show a positive SSH trend in the EAC extension that is consistent with the observed trend towards more intense eddies off south-east Australia. However there are mesoscale variations to the intensity of this trend, as shown by the patchy positive and negative trend biases in Fig. m–o. EAC03K shows a stronger negative trend bias in the EAC separation region as well as a smaller region of positive SSH trend bias off the north-east coast of Tasmania. The SSH trend biases for the 10 km models are weak and spatially patchy without clear dynamical explanation. Underlying these mesoscale variations is a domain-wide negative bias shared by all models relative to altimetry. This arises from a combination of reference offsets and the omission of global mean sea level rise contributions from land-ice melt.

The SSH variance bias highlights regions where modeled eddy activity deviates from the satellite-derived observations. As with OISST, AVISO represents a smoothed realization of the true SSH field due to interpolation and observational limitations, meaning we would generally expect models to exhibit a systematic positive bias in SSH variance. However, as noted in Sect. , these simulations use a relative-wind stress formulation, which includes current feedback and is known to damp mesoscale eddy variability relative to a direct-wind forcing. This may therefore contribute to a systematic negative bias in SSH variance. Across all models, there is indeed a negative bias in SSH variance in the upper extension region between latitudes of 32 and 35° S (Fig. p–r). While this negative bias extends further north in the 10 km models, there is a shift to positive SSH variance bias to the north in EAC03K located in the separation region. This provides further evidence for a potential earlier separation of the EAC in the 3 km model, which in turn generates higher SSH variance (eddy activity) further north than typically observed.

Marine heatwaves (MHWs) are a key focal point for future use of the EAC03K model, hence it's important that this model not only represents the temperature variability, but also the presence of persistent extremes that arise in the satellite observations. Figure  summarizes the MHW results focusing on two key MHW metrics – mean intensity and mean duration. Figure a–c present the mean intensity bias averaged over all events at each grid point across the EAC domain. Here we see a similar spatial pattern to the SST variance bias (Fig. e) for the two regional models, which emerges as the broad north-west to south-east South Pacific convergence zone imprint of more intense MHW events than OISST produces. This is expected, given the intensity is essentially a derived metric from the SST variance, albeit relative to the daily-varying 90th percentile rather than the long-term mean . Furthermore, there are noticeable differences in mean intensity for the separation and extension region. The regional models both have a coherent region of weaker intensities near the separation point (∼ 32° S) while the global model has higher intensities at the same location (Fig. a–c).

MHW durations are slightly shorter in EAC03K than STHPAC10K, particularly towards the south-east of the domain, while both models show longer-than-observed durations within the EAC system. The global model shows much longer durations throughout the domain as previously shown by . The spatially averaged probability density distributions (Fig. d, h) highlight the strong agreement between the two regional models, which consistently suggest higher mean intensities and shorter mean durations than the observational distribution. The global model shows broader distributions for both intensity and duration, while the peak of the distributions describe marine heatwaves that are less intense, and of longer duration than observations over the EAC model domain. This suggests that while the regional and global models may each align with observations at different times or under specific conditions, their inherent numerical differences prevent them from converging on a consistent representation of MHW characteristics.

We now assess the ability of each configuration to capture the temporal occurrence of MHWs. While MHWs in the region have a strong stochastic element due to the dominant control of the flow by mesoscale eddies we note that all configurations have identical surface forcing. As such, while individual events may be largely stochastic, we expect large-scale forcing to impact the general interannual variability in extreme events similarly across both OISST and the each simulation.

We see a strong overall agreement between models and OISST in capturing interannual variability of MHW duration, with distinct peaks during strong El Niño years (e.g., 1998, 2010, and 2016) (Fig. a). As the probability distributions suggested (Fig. h), the two regional models agree strongly for the interannual MHW durations while MHWs in the global model tend to be consistently longer than observations across all years. Similarly, mean intensity patterns are broadly consistent across datasets, with a notable shift to higher intensities after 1994. The regional models generally produce more intense MHWs than the global model and observations, likely due to higher SST variance in the northern half of the domain (Fig. b). Interestingly, the 3 km model occasionally diverges from the 10 km model, such as in 2016, where it better matches observed mean durations. This suggests that resolution may play a more critical role during dynamically intense events, such as the advection-driven MHW of 2015/16 . Overall, while model version differences dominate the results, resolution-specific effects emerge during periods of heightened dynamical complexity, highlighting the nuanced interplay between model numerics and resolution in simulating MHWs.

All models show interannual variability in the poleward transports at the EAC jet zonal transect (Fig. a, c), and while the timing of enhanced (e.g., year 1998) and weakened (year 2000) transports track well between models, the magnitude of this yearly averaged transports differs substantially between the 3 and 10 km models (Fig. c). In the year 1998 for example, the yearly averaged transport for the 10 km models is between 15–18 Sv (1 Sv ≡ 10⁶ m³ s⁻¹) while for EAC03K, this peak reached 24 Sv. Although there are large interannual fluctuations in the modelled poleward transports, the long-term trend over the 30 year model runs appears relatively stable as shown by the dashed linear trend lines in Fig. c). The observational transports for the overlapping years with models shows similar magnitude to the 10 km models, while the EAC03K model produces transports weaker than observations in the first overlapping year (2012), and then transports almost twice as high for the year after (2013).

The climatology of poleward transports in the EAC jet for all datasets is shown in Fig. d. All models follow observations with a poleward strengthening of transports in summer and a minimum in winter. The two 10 km models have very little variability around the daily-varying mean throughout the climatological year which allows for a clear seasonal cycle to be identified. The EAC03K climatology has a much higher sub-seasonal variability, however there is still an overall weakening in the winter months and strengthening in the summer. Interestingly, both the observations and EAC03K appear to have a two-stage strengthening from their winter minima whereby an initial rapid increase in transport is followed by a weakening before a second strengthening into the maximum summer peak. The timing of this sub-annual cycle is earlier for EAC03K than for the mooring observations by about one month. Nevertheless, the transition period is distinctly similar between the two configurations.

Next we move south to the separation point of the EAC where our primary interest is understanding the latitude where the EAC jet separates from the shelf break, and additionally, where the main jet retroflects, i.e., where it changes direction from a poleward flow to a predominantly eastward or recirculating flow. The spatial maps in Fig. a, b demonstrate the typical case of the EAC before and after an eddy shedding event, and clearly demonstrate the importance of defining two distinct latitudes of interest for this region. Prior to eddy shedding (Fig. a), we have a separation point near 30° S and a retroflection much further poleward of 34.5° S. Two days later, after the eddy detaches from the main jet, we see little change in the separation latitude, while the retroflection latitude retracts equatorward by around 200 km. The separation latitude distribution of AVISO (Fig. c) shows two distinct peaks – a result previously reported by . However, the exact locations of the peaks between that observed by our analysis (30.0 and 33.0° S) and that reported in their study (28.7 and 30.8° S) do indeed differ. Meanwhile, the retroflection point of AVISO in Fig. d has only a single peak two degrees of latitude below the poleward peak of the separation latitude at 35.1° S. The distinct difference between the separation and retroflection distributions of AVISO support the possibility that these critical points are governed by independent dynamics and forcing mechanisms – a finding not captured in earlier studies of the separation region, but one that may have important implications for the model behaviour explored below.

There is substantial difference between models in representing both the separation and retroflection latitudes. The global model, OM2-01, shows a tendency to separate toward the poleward (southern) extent of the latitudinal range, while STHPAC10K best represents the double peak distribution shape of AVISO (Fig. c). EAC03K separates much further north than the 10 km models and observations, and while there are occasions when it separates near the poleward peak of the observations, this rarely occurs. STHPAC10K captures the retroflection latitude distribution remarkably well, with a single peak and positively skewed tail (Fig. d). Similar to the separation latitude, EAC03K also shows an equatorward shift in the retroflection point relative to observationals and the 10 km models. These findings for the 3 km model support our earlier hypothesis that the positive SSH variance bias north of the typical separation region (Fig. o) could result in an earlier (equatorward) separation.

The eddy tracking in the EAC extension has given us insight into the typical pathway of eddies and the amplitudes of these transient features that are critical to the heat transport along south-east Australia's nearshore region (Fig. ). Here we focus specifically on anticyclonic eddies for the eddy tracking results. The 3 km model best reproduces the pathway and amplitude of anticyclonic eddies found in observations over the study period with a broad band of high amplitude eddies extending well below and to the south-west of Tasmania. The 10 km models capture the general pathway, however the band of high amplitude eddies is much narrower and eddies below Tasmania appear much weaker than in EAC03K and observations. Finally, the number of eddy tracks found in EAC03K (10 550 tracks) is much closer to that detected in the observational record (9812 tracks) while the 10 km models detected at least 20 % fewer anticylonic eddy tracks over the same period. The eddy tracking results suggest that EAC03K produces more high amplitude eddies than the 10 km models, highlighting the improvement gained from higher resolution when modelling geographical regions dominated by nonlinear turbulent ocean features. As above, these eddy-amplitude diagnostics are obtained from simulations using relative-wind stress, which is expected to suppress mesoscale kinetic energy to some degree through current feedback. Although this effect may itself vary with model resolution, it is not expected to substantially alter the qualitative differences shown here.

Australia's eastern continental shelf is exceptionally narrow, with some locations spanning less than half the width of a single 25 km resolution grid cell from the widely used observational datasets of AVISO and OISST. As a result, these global products do not resolve the sharp oceanographic gradients that define the shelf break, leading to a significant smoothing of critical nearshore features in addition to the well-known land contamination near coastlines. For this section we have introduced a new observational dataset, the IMOS-SST product at 2 km resolution which will provide further insight into the uncertainty between observational datasets and allow for more confidence when interpreting the differences between models and satellite-derived observations.

In this section we are primarily interested in assessing the agreement between models and observations in capturing the long-term SST trends along the shelf. We use the metric of relative warming in order to highlight regions that may be warming much faster than the global average, compared to those that might be warming on par with or slower than the global rate. This is a useful baseline, as it highlights regions where the ocean dynamics are changing, superimposing a localised ocean heat contribution onto the global SST warming trend.

The broad SST warming patterns within the continental shelf bioregions of eastern Australia are presented in Fig. a. The shading, which shows the long-term average warming relative to the global average rate across all datasets for each bioregion, reveals a distinct shift between those upstream of the separation range, and those downstream. North of the Manning Shelf, the warming weaker than or equal to the global average rate. From the Manning Shelf and southward, the warming rates increase to well above twice the global SST trend. This supports previous work of which also find the coastal seas downstream of the separation point to be warming at twice the rate of those upstream.

Analysis of the bioregion SST trends by season and dataset (Fig. b) shows that the north/south contrast in warming rates are nuanced. Seasonally, regions upstream of the separation point show very little change across the year, while downstream, we see much lower warming rates in winter than all other seasons. This seasonal difference is particularly pronounced in the bioregions within Bass Strait, whereby winter warming rates are on par with global averages, while summer warming rates can be greater than three times the global rate. To further emphasize the local-scale differences in warming, bioregions directly adjacent to Bass Strait, such as the Freycinet bioregion show warming rates year-round that are around twice the rate as those in Bass Strait, e.g., Flinders bioregion. The winter cooling within Bass Strait is likely dominant over the continuous heat advected down the EAC extension during this season, while for locations more exposed to the EAC, weaker winter trends are less prominent (e.g., Freycinet).

The two observational products are largely in good agreement, apart from in spring and summer for bioregions within the latitude range of the typical separation point. During these two seasons, the IMOS-SST shows warming rates at least 1.5 times greater than the OISST product. The largest disagreement between models and observations occurs in summer for the upstream bioregions, whereby all models produce warming rates much slower than the global rate (shown by the blue shading), while the observations are close to or faster than the global average. In the downstream region, there is much less inter-model agreement, except in theFreycinet bioregion. EAC03K produces warming rates in autumn that are 5.1 times the global average rate, while the observations suggest relative warming rates of 2.5 and 2.2 for IMOS-SST and OISST respectively. This surprisingly large warming for the 3 km model raised some further questions as to the validity of this signal, which was also present in the surface biases presented in Sect. 3.1.

To further investigate the warming rate in the Freycinet bioregion, we analysed the average SST by decade across all datasets (Fig. ), The 3 km model produces mean SSTs closer to the observed values for the last two decades than the global model, OM2-01. Hence, it is clear that although EAC03K is initially cold-biased in this region, as also identified in Fig. a, it adjusts and then better estimates the mean SST climate on the shelf of eastern Tasmania than the global model for the last two decades of the simulation.