The numerical simulations were performed with the hydrostatic configuration of the Massachusetts Institute of Technology (MIT) primitive equation model . The computational domain extended from 99 to 55∘ W and from 6 to 33∘ N (Fig. a); it was set up with a horizontal resolution of 1/12∘, which is well below the internal Rossby radius of deformation in this region 60–80 km,. The same topography as the Operational Mercator global ocean analysis (Mercator) of the EU Copernicus Marine Information Service was used wherein the topography is based on ETOPO1 for the deep ocean and near the coast and on GEBCO8 for the coast and slopes. In this topographic setup, the majority of the islands of the Lesser Antilles are represented (Fig. ). The islands that are too small to be captured are modeled as shallow ridges. In the vertical, the model contained 50 levels in z coordinates, increasing in depth from 1 m at the surface towards 459 m at the lower levels.

The time step was 240 s. All simulations had a total duration of 25 years, the first 5 years of which were considered spin-up and were excluded from the analysis. The model output was saved as 5 d averaged fields. The simulations were initialized with time-averaged fields for sea-surface height, temperature, salinity, and velocity. These fields were obtained from the years 2007–2017 of Mercator. The vertical diffusion of tracers was parameterized with the GGL90 mixed layer parameterization . The horizontal diffusion was parameterized with the Redi scheme diffusion coefficient of 125 m2 s-1;. The sub-grid-scale mixing was parameterized with Smagorinsky viscosities .

All simulations were forced with time-averaged surface and lateral boundary conditions. A sponge layer with a meridional width of 1.25∘ (15 grid cells) was applied at the lateral boundaries. There, the velocity fields were relaxed towards the time-averaged fields of Mercator (years 2007–2017) with a relaxation time that varied linearly from a month at the offshore side of the domain towards a day at the domain boundary. At these boundaries, temperature, salinity, and velocities in the zonal and meridional directions were prescribed. At the surface, a freshwater flux, temperature restoration, and a wind stress were applied (Fig. b–d). The temperature was restored towards sea-surface temperature (SST) averaged over the years 2007–2017 of Mercator with a relaxation timescale of 1 month. The surface freshwater flux was obtained from the diagnosed freshwater flux of a 250-year simulation described in . The Orinoco River, Magdalena River, and Mississippi River are prescribed as stationary freshwater fluxes at the open boundaries with discharges based on .

The reference simulation (Ekman100) is forced by a time-averaged zonal wind forcing computed from the wind fields of years 2007–2017 of ERA-Interim . Prescribing stationary values implied that the NBC Rings, which can trigger the formation of the anticyclones , were not represented at the boundaries. However, and showed that a realistic eddy field in the Caribbean Sea can be obtained without the presence of NBC Rings. Moreover, in Sect. 3 we will show that a realistic eddy field is obtained with these boundary conditions. Therefore, we considered these boundary conditions sufficient for the purpose of this study.

To investigate the effect of wind-driven upwelling on the westward intensification of anticyclones, only this zonal wind forcing (τx) was altered between simulations. The magnitude of the zonal wind stress was reduced or increased in each simulation by the same constant over the entire domain (Fig. ). With this approach, we ensured that we only change the upwelling strength and not the curl of the wind stress. The magnitude of the reduction or increase was determined based on the zonal wind stress in the upwelling region in Ekman100. In Ekman150, the zonal wind stress in the upwelling region was 50 % stronger than the wind stress in Ekman100, resulting in a theoretical increase in the northward Ekman transport of 50 %. The wind stress along the coast was weaker in Ekman50, leading to a theoretically weaker upwelling (50 %) in this simulation compared to Ekman100. The same principle was applied in Ekman75 and Ekman125.

The adjustment of the zonal wind forcing resulted in changes in mesoscale variability over the full domain. Because the applied constant was optimized to alter the northward Ekman transport in the Cariaco and Guajira upwelling regions, it resulted in unrealistic magnitudes of the zonal wind stress (and thus of the northward Ekman transport) in some other regions. Therefore, we will only analyze the impact of the changes in wind forcing on the mesoscale variability close to the Cariaco and Guajira upwelling regions (Fig. c) and disregard other coastal upwelling regions.

The range of τx applied in the simulations covers the seasonal variations (Fig. ). The magnitude of the stationary zonal wind stress in Ekman50 was similar to the observed zonal wind stress in fall, while the stronger stationary zonal wind stress in Ekman125 was comparable to the wind stress during winter and summer months. The zonal wind stress in Ekman150 was stronger than the seasonal cycle but was similar to zonal wind stresses that are observed during years with anomalously high wind velocities .

To analyze the behavior of the mesoscale eddies, we analyzed both the eddy kinetic energy (EKE) and the individual eddies. Following , we calculated the EKE from 5 d averaged velocity fields that were high-pass filtered with a 125 d running mean. This 125 d period was long enough to capture the variability of the eddies, which have a characteristic period of 50–100 d , but was shorter than the interannual variability.

To gain insight into the westward intensification of Caribbean anticyclones, we used the py-eddy tracker to follow them . The py-eddy tracker used 5 d averaged sea-level anomaly fields to identify near-circular features. Anomalies were computed with respect to 20-year averaged fields. Negative anomalies were identified as cyclones and positive anomalies as anticyclones. More detailed information about the numerics of the py-eddy tracker can be found in . We set a minimum lifetime of 75 d. Taking into account the average westward propagation velocity of 0.13 m s-1 , the anticyclones propagate approximately 840 km during this minimum lifetime.

At the center of the eddies provided by the eddy tracker, we extracted the amplitude (Aeddy), swirl velocity (uswirl), radius (Reddy), and properties from the model output to assess their characteristics. The amplitude was defined as the sea-surface height difference between the eddy and the 20-year average sea-surface height at the center of the eddy. Since the py-eddy tracker identifies circular anomalies as eddies, we could define the swirl velocity as the average of the maximum northward and maximum southward velocity of the eddy. The location of both velocities was used to obtain a robust estimate of the eddy radius (Reddy), which we defined as half the distance between these locations. To characterize the local properties (T, S, σ) of the background and the eddies, these variables were averaged over the upper 50 m. This depth corresponds to the average depth of the pycnocline in the Caribbean Sea. The latter ensures that these properties not only reflect variations in surface forcing. The background properties were defined as the 125 d averaged values of the upper 50 m at the location of the eddy obtained from the py-eddy tracker. The differences between the properties of the eddies and the background are computed as Δ=eddy-background.

To gain insight into the geostrophic part of the westward intensification of the anticyclones in each simulation, we computed the strength of the horizontal density gradients (|∇σ|). These gradients were calculated over the upper 50 m of the water column for 5 d averaged density fields as follows: 1|∇σ|=∂σ∂x2+∂σ∂y2, where σ is the density. The contribution of temperature (|∇σT|) and salinity (|∇σS|) to the horizontal density gradients was computed in a similar manner, whereby the density differences were calculated assuming a linear equation of state as ΔσT=ρ0α(T1-T0) and ΔσS=ρ0β(S1-S0), respectively. Here, α, β, and ρ0 are constants: α=-3.1×10-4 ∘C-1, β=7.2×10-4 psu-1, and ρ0=999.8 kg m-3.

Since the focus of this study is to analyze the westward intensification of the anticyclones, we not only calculate the EKE and strength of the horizontal density gradients over the full domain, but we also calculate their contribution associated with the anticyclones only. We estimate the latter by considering the EKE and density gradients around the core of an eddy at each 5 d averaged field over a spatial extent of 1.5×Reddy. This procedure allows us to study only the westward intensification of the anticyclones.

The modeled surface current in Ekman100 enters the Caribbean Sea through the southern passages of the Lesser Antilles (Fig. a), which is similar to what is seen in observations . The flow through the Lesser Antilles is highly variable and depends on variations of the flow upstream and the wind forcing . Based on scarce data, estimated the transport at 66∘ W as 18.4±4.7 Sv. In our simulation, the transport is more stable due to the stationary forcing and is on average 13.6 Sv at 66∘ W, which is close to the lower range of the estimate of . Further westward at 80∘ W, the modeled flow accelerates over shallow topography at 17∘ N, where it continues northwestward towards the Yucatán Channel into the Gulf of Mexico.

In the southwest of the Caribbean Sea, the modeled surface currents display a cyclonic recirculation (Fig. a). This recirculation is known as the Panama–Colombia Gyre (PCG) and is wind-driven . Part of the PCG continues as an eastward subsurface countercurrent, which results from the Sverdrup circulation in the North Atlantic tropical cell . The countercurrent flows below the upwelling region, and its depth is related to the wind strength . The model is able to reproduce the subsurface countercurrent at a depth of approximately 100 m, which is slightly higher in the water column than at the 200 m observed by and the 150 m observed by ; the modeled strength (0.11 m s-1) is comparable to their observations.

Ekman100 displays a strong meridional density gradient in the mixed layer that varies between σ=25.1 kg m-3 in the south (11∘ N) and σ=22.7 kg m-3 in the north (18∘ N, Fig. b). The strongest meridional gradients are close to the surface and colocated with the Caribbean Current. The location of the dense waters coincide with the two major upwelling regions: Cariaco and Guajira. The upwelled waters are colder and more saline than the surrounding surface waters (Fig. c, d).

The Cariaco upwelling region in the southeast of the Caribbean Sea is located at 63–65∘ W, 10–12.5∘ N . The modeled minimum (24.9 ∘C) and maximum temperatures (27.4 ∘C) in the Cariaco Basin are less extreme than the observed temperatures minimum: 20.3 ∘C, maximum: 30.1 ∘C; . Although the model is not able to capture this seasonal variability, the modeled average SST (25.4 ∘C) is similar to observations 25.2 ∘C;. Note that these modeled temperatures are lower than the restoring temperature (Fig. c), which indicates that the model is able to reproduce coastal upwelling. The Guajira upwelling region is located west of the Cariaco upwelling region between 69 and 74∘ W . Here, the observed average SST is slightly higher (25.5 ∘C) than in the Cariaco upwelling region . Similar to these observations, the model displays a temperature difference between the two upwelling regions (26.1 ∘C in Guajira; Fig. c). Furthermore, the modeled temperatures are in line with , who found surface temperatures varying between 25.6 and 28 ∘C in the Guajira upwelling region.

In addition to the meridional density gradient, the model also displays a clear zonal density gradient in the Caribbean Sea (Fig. b): the Caribbean Current is relatively light as it enters the Caribbean Sea. This zonal density gradient is mainly due to a zonal salinity gradient and to a lesser extent to a zonal temperature gradient (Fig. c–d). As mentioned in the Introduction, the zonal temperature and salinity gradients are related to the offshore advection of cold and saline upwelling filaments . The magnitude of the zonal density gradient is largest in the surface mixed layer and decreases rapidly below. According to observations, the zonal salinity gradient is also related to the dispersal of the Amazon and Orinoco River plumes into the basin e.g.,. In the model, the Amazon River plume enters the domain at the southern boundary, while the fresh water of the Orinoco River enters the domain at 61∘ W, 9∘ N (Fig. d). The river plume is advected with the mean flow and becomes more saline through mixing with the saline (upwelled) surface waters in the basin, as well as through evaporation. Overall, the zonal and meridional gradients of salinity and temperature are similar to observations, indicating that the model is able to capture the upwelling, the advection of the river plume water, and the relevant mesoscale processes.

In line with observations , we find that the flow in the Caribbean Sea is highly variable (Fig. ). In the eastern part of the basin, the modeled surface EKE is relatively low (100–300 cm2 s-2; Fig. a) and similar to observations of . The EKE increases westward towards a maximum >900 cm2 s-2 at 78∘ W. The modeled magnitude of EKE is higher than found in satellite altimetry >600 cm2 s-2;, but it is more similar to estimates obtained from surface drifters >900 cm2 s-2;. This is in line with other modeling studies , and this discrepancy is mainly attributed to the coarse resolution (0.25∘) of the gridded altimetry data products . The modeled spatial variability of EKE matches analyses of satellite altimetry well . Corresponding to observations , we find that the eddy kinetic energy is surface intensified (Fig. b). In line with the modeling results of , the magnitude of EKE at depth also increases towards the west (Fig. b).

The variability in the Caribbean Sea manifests itself in the presence of mesoscale eddies (Fig. ). In line with observations , the mesoscale eddies are predominantly anticyclonic (Fig. a). As expected, the surface densities of these anticyclones are lighter than those of the surrounding surface waters (Fig. b). The density differences are due to both temperature and salinity (Fig. c, d). Figure c also displays the northward advection of a cold filament. Similar filaments have been observed, and it is known that these can be advected several hundred kilometers from the upwelling region . This advection results in the offshore cooling of surface waters .

From the py-eddy tracker , we found that on average 8.55 anticyclones and 5.70 cyclones are formed in the reference simulation Ekman100 east of 75∘ W between 12.5 and 17.5∘ N every year. These numbers are in line with the 8–12 anticyclones per year estimated from surface drifter data . To our knowledge, there are no previous estimates for the formation rate of cyclones in the Caribbean Sea.

The anticyclones have an average amplitude of Aeddy=0.17 m and swirl velocities of uswirl=0.60 m s-1 between 65–75∘ W and 12.5–17.5∘ N (Table ). These amplitude and swirl velocity values are similar to those found in hydrographic surveys and similar to velocities obtained from surface drifters uswirl=0.5 m s-1;. In general, the core of the anticyclones is warmer (ΔT=+0.15 ∘C) and fresher (ΔS=-0.18 psu) than the surrounding surface waters (Fig. c–d).

In the central Caribbean Sea (65–75∘ W and 12.5—17.5∘ N), the cyclones are less energetic than the anticyclones and have an average amplitude of Aeddy=-0.16 m and swirl velocity of uswirl=0.50 m s-1. Observations indicate that cyclones are generated near topographic features in the Caribbean Sea and can contain upwelling waters . The modeled properties of the cyclones are in agreement with these observations. It appears that the mesoscale eddies are close to geostrophic balance, with an average Rossby number of 0.15±0.01 (anticyclones) and 0.14±0.03 (cyclones), where the Rossby number is calculated as Ro=uswirl/(Reddy*f), in which uswirl is the swirl velocity of the eddy, Reddy is the radius, and f the Coriolis parameter.

There is a strong and significant correlation between the amplitude of the tracked mesoscale eddies and the surface EKE. This suggests that the westward increase in EKE is related to the strength of the eddies. We find that 57 of the anticyclones that are formed in the Caribbean Sea during the 20 years of the simulation propagate from a region with weak EKE (<65∘ W) towards a region with high EKE (>75∘ W). The paths of these anticyclones are indicated with the black lines in Fig. a. The other 96 anticyclones that are formed in the Caribbean Sea during the 20 years of the simulation are either generated west of 65∘ W or do not pass 75∘ W (grey lines in Fig. a). In contrast to the anticyclones, cyclones display relatively short tracks: only one cyclone passes both 65 and 75∘ W (black line in Fig. b). This is similar to the observations of , who showed that the cyclones follow a different path than anticyclones.

To assess the contribution of the anticyclones with the long tracks to the total EKE variability, we calculated their EKE from the zonal and meridional velocity fields by taking into account the EKE within a region of 1.5×Reddy around each eddy as described in Sect. . The spatial distribution of EKE due to these long-lived eddies (Fig. c) is similar to the spatial distribution of the total EKE (Fig. a). Notably, the magnitude of the EKE of these selected eddies is approximately 55 % of the total EKE (Fig. ), while it is computed from only 37 % of the eddies (57 out of 153 anticyclones in 20 years). This shows that the modeled westward increase in EKE is dominated by the westward increase in EKE of a small number of long-lived eddies, which is also confirmed by the weaker eddy kinetic energy of the eddies with shorter tracks (Fig. d). Because our results show that the eddies with long tracks dominate the mesoscale variability in the Caribbean Sea, we will focus on these 57 energetic anticyclones in the remainder of this study.

The zonal wind stress in Ekman50 and Ekman75 was reduced compared to Ekman100 such that the wind stress in the upwelling region was 50 % and 25 % less than Ekman100, respectively. The decrease in zonal wind stress in Ekman50 and Ekman75 results in weaker upwelling and consequently warmer SST in the Cariaco Basin (Fig. b, c). The highest coastal temperatures are present in Ekman50. Sea-surface salinity decreased in both Ekman75 and Ekman50 compared to Ekman100 (Fig. b, c). This freshening is related to the presence of a subsurface salinity maximum in the Caribbean Sea due to the presence of a water mass referred to as Subtropical Underwater. This water mass is located at approximately 100 m of depth and leads to more saline upwelled waters compared to the fresher surface waters . Weaker upwelling thus results in warmer and fresher surface waters (Fig. b, c), and stronger upwelling results in colder and more saline surface waters (Fig. d, e).

Besides differences in the mean properties (σ, T, S), the width of the upwelling region also changes in the simulations: in Ekman125 and Ekman150, the upwelling is confined to a narrower region than in Ekman75 and Ekman50. Moreover, we find that the core of the subsurface countercurrent is displaced upwards in Ekman50 and Ekman75 compared to Ekman100 (red contour in Fig. ). This upward change in line with observations in which the core was also displaced upwards in weaker wind conditions . The increased northward Ekman transport in Ekman125 and Ekman150 leads to colder and more saline surface waters. In these simulations, the core of the subsurface countercurrent is positioned lower in the water column (red contour in Fig. ). The changing position of the subsurface countercurrent and the changes in sea-surface properties in the upwelling regions indicate that the model is able to reproduce realistic upwelling properties, as also shown in Sect. .

The upwelling strength impacts the offshore EKE (Fig. ). In the interior of the Caribbean Sea, the surface EKE and subsurface EKE decrease (increase) for weaker (stronger) zonal wind strength. For example, the magnitude of the mean EKE in the interior of the Caribbean Sea decreases by 27 % in Ekman50 compared to Ekman100 (Fig. a). In Ekman75, a slightly weaker decrease of 15 % is present (Fig. b). In Ekman125 and Ekman150, the surface EKE increases by 13 % and 28 %, respectively (Fig. c, d). Overall, the EKE variations indicate a positive correlation between the coastal wind stress and the strength of the mesoscale eddies. A similar correlation is found in the subsurface EKE. This is in line with the finding that the upwelling was shallower in the simulations with weaker upwelling and deeper in simulations with stronger upwelling.

To gain insight into the sensitivity of the westward increase in EKE to variations in upwelling, the maximum of the total EKE between 12.5 and 17.5∘ N was computed as a function of longitude (Fig. a). The largest changes in EKE are seen in the western part of the basin. In Ekman100, the EKE increased by 123 % from 65 to 75∘ W, while the weaker upwelling in Ekman50 and Ekman75 resulted in a smaller westward increase of the EKE by 62 % and 97 %, respectively. Ekman125 and Ekman150 have a stronger westward increase in EKE compared to Ekman100. Overall, there is a clear positive correlation between the upwelling strength and the westward increase in EKE.

Next, we assessed whether the changes in the westward increase in EKE can be attributed to changes in the eddy field and to changes in the westward intensification of anticyclones. To this end, we studied the behavior of individual eddies in each simulation with the py-eddy tracker. The number of anticyclones that are formed each year in the interior of the Caribbean Sea is highly variable and varies between 8.55 (Ekman100) and 10.85 per year (Ekman50, Table ). These numbers are in line with a formation rate of 8–12 anticyclones per year as observed by . In all simulations, between 34 % and 43 % of the anticyclones that are formed in the Caribbean Sea have long tracks (Table ). The paths of these anticyclones with long tracks are, similar to Ekman100 (Fig. a), located in the southern part of the basin (not shown).

Similar to Ekman100, we computed the EKE associated only with the long-lived anticyclones that were formed in the eastern part of the basin (<65∘ W) and propagate beyond 75∘ W (Fig. b). Similar behavior is visible as for the total EKE in Fig. a: the EKE increases towards the west in all simulations, and stronger (weaker) upwelling corresponds to a larger (smaller) westward increase. In the simulations with strong upwelling (Ekman100, Ekman125, and Ekman150), the anticyclones with long tracks (Fig. b) are responsible for more than half of the total EKE (Fig. a), even though they constitute only 34 %–43 % of the total number of anticyclones in this region (Table ). In Ekman50 and Ekman75, the anticyclones with long tracks become less dominant. In these simulations, the westward increase in EKE is also less pronounced than in the stronger wind conditions. Overall, this shows that in all simulations, a substantial part of the EKE and the longitudinal variations therein are governed by the evolution of a small number of anticyclones and that this effect is stronger in simulations with stronger upwelling.

In general, the anticyclones are fresher and warmer than the surrounding waters in all simulations. The freshest and warmest anticyclones are present in Ekman50, while the most saline and coldest anticyclones are found in Ekman150 (Table ). It is interesting to note that the variation of the properties of the anticyclones is very small in each simulation (see the standard deviation of the anticyclone properties in Table ). This suggests that, like shown for Ekman100 in Fig. c, for the simulations in which the upwelling was varied the properties of the anticyclones also remain approximately constant during their propagation westward.

In line with the observations of , we found fewer cyclones than anticyclones in each simulation (Table ). The lowest average formation rate of cyclones is found in Ekman150 with 4.15 cyclones per year, and the highest average formation rate is present in Ekman50 with 6.15 cyclones per year. These formation rates were highly variable, and differences in formation rates between simulations were not significant. In none of the simulations was the py-eddy tracker able to track multiple cyclones from east to west (65–75∘ W). This implies that the cyclones are either deformed or dissipated too much such that the py-eddy tracker could not track their sea-level anomaly. Overall, the behavior of the mesoscale eddies is similar in all simulations, and the spatial pattern and magnitude of the surface EKE is governed by the anticyclones with long tracks. The westward intensification of these anticyclones is discussed in the next section.

We showed in Sect. 4 that the westward increase in EKE in Ekman100 coincided with a westward strengthening of the vertical shear of the long-lived anticyclones (Figs. c and a). This increase in vertical shear was related via the thermal wind relation to the strengthening of the horizontal density gradients at the western edge of the anticyclone caused by the advection of upwelling filaments (Fig. a). Here, we analyze the thermal wind relation in the sensitivity simulations and its relation to the westward intensification of the anticyclones (Fig. ).

In all simulations, the vertical shear of the anticyclones increases from east to west (Fig. a). This westward increase in vertical shear is similar to the westward strengthening of the horizontal density gradients (Fig. b). Similar to Ekman100, the magnitude of the horizontal density gradients is slightly smaller than the vertical shear in all simulations, but their variation is similar. In combination with the low Rossby number of the anticyclones (Table ), this implies that the anticyclones are in near-geostrophic balance during their evolution.

The anticyclones display the strongest westward intensification in the simulations with stronger upwelling (orange and red lines in Fig. a, b). These simulations have relatively low vertical shear at 65∘ W. In the simulations with weaker upwelling, Ekman50 and Ekman75, the anticyclones intensify less than in Ekman100. Moreover, in Ekman50, the vertical shear even weakens between 70 and 73∘ W. At the same location, the EKE also does not increase in Ekman50 (Fig. b).

It is remarkable that, at 65∘ W, the strongest vertical shear and density gradients of the anticyclones are present in Ekman50, while the weakest gradients were located in Ekman150 (compare the blue and red lines in Fig. a, b). This suggests that the early development (east of 65∘ W) of the anticyclones differs between the simulations. The strong gradients in Ekman50 at 65∘ W are surprising because we expected stronger gradients in simulations with stronger upwelling. An explanation for the strong horizontal density gradients at 65∘ W in Ekman50 can be found after separating the total horizontal density gradients into density gradients induced by temperature and salinity (Fig. c, d). While the density gradients driven by temperature gradients have a similar magnitude in each simulation in the eastern part of the basin (Fig. c), the density gradients driven by salinity gradients differ substantially between the simulations at 65∘ W (Fig. d) and are negatively correlated with the upwelling strength.

To understand why the salinity gradients in the eastern part of the basin are stronger for weaker upwelling, we analyzed the average spatial variation of the total density gradients. Figure  shows the magnitude of the time-averaged horizontal density gradients |∇σ| of Ekman100 (middle column) and the strengthening or weakening of these gradients in Ekman50 (right column) and Ekman150 (left column) compared to Ekman100. The strongest horizontal density gradients are located at the Guajira and Cariaco upwelling regions (Fig. b). These density gradients are weaker in Ekman50 than in Ekman100 (Fig. a) and stronger in Ekman150 (Fig. c).

The horizontal density gradients at the upwelling region are due to both horizontal temperature and salinity gradients (Fig. e–h). Because upwelling brings relatively cold and saline waters towards the surface (Fig. ), the strength of these density gradients increases for stronger upwelling (Fig. c) and decreases for weaker upwelling (Fig. a). In Ekman150, the strengthening of the horizontal density gradients is mainly due to temperature-induced density gradients in the upwelling region (Fig. f, i). In contrast, in Ekman50, the weakening of the temperature gradients (Fig. d) and salinity gradients (Fig. g) contributes roughly equally to the weakening of the total horizontal density gradients (Fig. a).

In the interior of the basin, the horizontal density gradients strengthen in Ekman50 compared to Ekman100 (Fig. a), while they are weaker in Ekman150 compared to Ekman100 (Fig. c). It appears that this response results from the increasing importance of the Amazon and Orinoco River plumes for weaker wind conditions. The wind affects both the propagation and spreading of the river plume. First of all, the northward Ekman transport advects the river plumes towards the north as previously observed by . This northward Ekman transport is weaker in Ekman50, and consequently the river plumes follow the mean flow towards the west (Fig. g). The stronger Ekman transport in Ekman150 results in a more northward path of the river plumes (Fig. i). The second consequence of changing wind conditions is the impact of the wind-driven vertical mixing. The strong zonal winds in Ekman150 induce more wind-driven mixing of the surface waters than the winds in Ekman50. As a result, salinity anomalies propagate less far westward in Ekman150 than in Ekman50. The farther advection of the river plume in Ekman50 explains why the anticyclones have stronger salinity gradients in Ekman50 than in the other simulations (Fig. d).

From this analysis we conclude that a weakening of the zonal wind in Ekman50 leads to an increased influence of the Amazon and Orinoco River plume compared to the other simulations. More specifically, in Ekman50, the upper ocean density gradients in the interior of the Caribbean Sea are dominated by salinity gradients (Fig. j), while the density gradients are predominantly temperature-driven in Ekman150 (Fig. l). So, depending on the zonal wind strength, the total horizontal density gradients in the Caribbean Sea are either driven by temperature or salinity gradients, which suggests that both the river plumes and the upwelling affect the variability in the basin.