MOM6-COBALT captures the main spatial patterns of key environmental parameters (SST, SSS and sea surface nutrients) fairly well in the coastal domain (Fig. 2). The global SST field simulated by the model reproduces the strong large-scale tropical to polar SST gradients, with a global median bias of -0.2 ∘C (Fig. 2a–c), and biases at the scale of MARCATS regions ranging from 0 ∘C in the NE Atlantic (M17) to 1.3 ∘C on the US East Coast (M10, Fig. 3a and Table S1). With a global median bias value of 0.2, the model also correctly reproduces the observed SSS patterns that are mainly regulated by evaporation and freshwater inputs from precipitation, riverine runoff and ice melt, with lower SSS values in polar regions and along the coasts of Southeast Asia and higher SSS values along the coasts of evaporation basins such as in the Arabian Sea or the Mediterranean Sea (Fig. 2d–f). The SSS analysis at the MARCATS scale reveals absolute SSS biases that are generally less than or close to 1, except for five MARCATS regions where absolute biases exceed 2. These MARCATS regions are mainly located in marginal seas (the Baltic Sea, M18; the Black Sea, M21; and the Persian Gulf, M29) but also include one polar region (the Canadian Archipelago, M13) and one tropical region (Tropical West Atlantic, M7; see Fig. 3b and Table S1). Similar to SSS, largest the model–data discrepancies for nutrients are mostly found in marginal seas (Fig. 3c–e and Table S1). For instance, the largest PO4 and SiO4 biases are encountered in the Black Sea (M21, absolute biases of 3 and 75 µmol kg-1, respectively). The Peruvian Upwelling Current (M4), the Bay of Bengal (M31) and the NE Pacific (M1) also present large biases in NO3 and PO4 (e.g., NO3 bias of 8 µmol kg-1 for M4). However, the global median nutrients biases are much smaller, reaching 0.3, -0.2 and -0.4 µmol kg-1 for nitrate (NO3, Fig. 2i), phosphate (PO4, Fig. 2l) and silicate (SiO4, Fig. 2o), respectively.

The model–data seasonal evaluation reveals that MOM6-COBALT reproduces the global SST and SSS amplitudes remarkably well (median absolute bias of 0.1 ∘C and 0.0, respectively; see Table S2 in the Supplement). Some exceptions can nevertheless be diagnosed, such as in the marginal Black Sea (M21), where the bias in SST seasonal amplitude reaches -1.3 ∘C, and in three MARCATS regions (Bay of Bengal, M31; tropical West Atlantic, M7; and Siberian shelves, M43) where the SSS seasonal biases are larger than 0.4. The model–data comparison also reveals that the phasing of the SST and SSS seasonal cycles are in very good agreement (Pearson correlation close to 1) for all 45 MARCATS regions, with the exception of four for which significant deviations in SSS are found, i.e., two marginal seas (Hudson Bay, M12, and the Red Sea, M28) and along the Californian Current (M2) and Brazilian Current (M6). The nutrients analysis shows absolute global median biases in seasonal amplitude of 0.1, 0.0 and 0.7 µmol kg-1 for NO3, PO4 and SiO4, respectively. Seven MARCATS regions present absolute biases larger than 1.5 µmol kg-1 that are mainly located in marginal seas (Baltic Sea, M18; Sea of Japan M40; and Sea of Okhotsk, M41) but also in polar (Siberian, M43, and Antarctic, M45, shelves) and subpolar (NE Pacific, M1) regions and in the Bay of Bengal (M31). The model–data comparison sometimes shows significant phase shifts in their seasonal signal (Pearson coefficient < 0.5), such as for MARCATS regions located in Indian and tropical margins, marginal seas, and EBCs.

The spatial distribution of the annual mean pCO2 simulated by MOM6-COBALT is in good agreement with the observational pCO2 values extracted from the SOCATv6 database with generally low pCO2 values (blue colors) in temperate and high latitudes and high pCO2 values (yellow and red colors) in tropical and sub-tropical regions (Fig. 4a–c). The model–data pCO2 evaluation at the regional scale shows that 33 of the 45 MARCATS present absolute biases lower than 20 µatm (Table S1). The regions where the bias exceeds this threshold include two EBCs (the Californian Current, M2, and the Peruvian Upwelling Current, M4), two marginal seas (Sea of Japan, M40, and Sea of Okhotsk, M41), and one polar region (Antarctic shelves, M45), a subpolar region (NW Pacific, M42) and the tropical East Atlantic (M23) shelf. Note that in some MARCATS regions, in particular in marginal seas and Indian seas, there are no SOCATv6 observations to perform the comparison (e.g., the Bay of Bengal, M31; see Fig. 4b and Table S1). Hence, we also evaluate the performance of MOM6-COBALT against the continuous coastal SOM-FFN pCO2 product, which uses a neural network interpolation method to fill data gaps and resolve the spatiotemporal coastal pCO2 variability globally.

Our results show that MOM6-COBALT reproduces the main spatial features of the annual mean pCO2 field captured by the coastal SOM-FFN product, as revealed by the relatively low globally averaged bias of 2.5 µatm (Fig. 4a and d). In both the model and the SOM-FFN product, low coastal pCO2 values are consistently found in temperate and high-latitude regions in both hemispheres, while high pCO2 values are largely limited to (sub-)tropical regions. The largest discrepancies (Fig. 4e) are found at high latitudes (poleward of 60∘ N and 60∘ S, negative bias), along the Peruvian and Namibian upwelling systems (high positive bias) and more locally close to the mouth of some large rivers (e.g., the plume of the Amazon or the Rio de la Plata, high negative bias). We note, however, that these regions are poorly sampled in the SOCATv6 dataset (Fig. 4b) and are thus likely weakly constrained in the coastal SOM-FFN product (Fig. 4d).

At the regional scale, differences in annual mean pCO2 between MOM6-COBALT and coastal SOM-FFN are lower than 20 µatm in 35 MARCATS (Table S1, Fig. 3f), which partly is a reflection of the low annual mean biases observed in the environmental driver variables in these regions (see Sect. 3.1.1). In EBC, WBC and subpolar coastal regions, the model tends to overestimate the regional mean pCO2 compared to coastal SOM-FFN (positive bias), except along the US East Coast (M10), in the East China Sea and Kuroshio (M39), and in the NE Atlantic (M17, Table S1). In polar regions, the model generally underestimates the mean pCO2 compared to coastal SOM-FFN, except around S Greenland (M15). In Indian, marginal, and tropical coastal regions, no general trend can be identified regarding the sign of the bias, which can be positive or negative.

Quantitatively, the 10 MARCATS regions with absolute biases > 20 µatm are mainly located in regions for which very limited or no observational data have been compiled in the SOCATv6 database (Table S1) and/or for which large discrepancies can already be identified at the level of the master environmental variables (Sect. 3.1.1). These regions mainly belong to EBCs (three out of the six EBC MARCATS regions) and marginal seas (three out of the nine marginal seas MARCATS regions), with the remaining four being either polar (the Canadian Archipelago, M13, and the N Greenland, M14), subpolar (NW Pacific, M42) or Indian margins (the Bay of Bengal, M31). The largest biases are found in the Peruvian Upwelling Current (M4), SW Africa (M24), the Californian Upwelling Current (M2) and the Canadian Archipelago (M13), with biases of 106, 79, 35 and -53 µatm, respectively.

Our analysis reveals that the seasonal amplitudes simulated by MOM6-COBALT are systematically larger than the ones estimated by the coastal SOM-FFN product (Fig. 5a–b, red colors in Fig. 5c and positive biases in Table S2) for all coastal regions belonging to EBC, WBC, and Indian and tropical margins. For the majority of the polar and subpolar margins and for some marginal seas, the model simulates lower seasonal pCO2 amplitudes (blue colors in Fig. 5c and negative biases in Table S2). Quantitatively, absolute biases between the modeled and coastal SOM-FFN amplitudes do not exceed 20 µatm, except for in marginal seas where larger discrepancies are calculated (six of the nine marginal MARCATS regions, Table S2). The monthly mean pCO2 seasonal cycle simulated by MOM6-COBALT is also well in phase (Pearson correlation coefficients > 0.5) with the one extracted from coastal SOM-FFN in 34 out of the 45 MARCATS regions (Fig. 5d and Table S2). The agreement is especially good in the best-monitored MARCATS regions (MARCATS where > 50 % of the area is covered by SOCATv6 observations, Table S1). For instance, in regions with good data coverage, such as along the US East Coast (M10), the Norwegian Basin (M16), the Californian Current (M2), the Leeuwin Current (M33) or the Brazilian Current (M6), the Pearson correlation coefficient is higher than 0.9 (Table S2). In contrast, the seasonal pCO2 cycle simulated by MOM6-COBALT substantially diverges from that of the coastal SOM-FFN in four poorly monitored marginal seas and in a few of regions of EBCs, Indian margins, subpolar margins, and tropical margins (Pearson correlation coefficient < 0.5, Table S2 and Fig. 5d).

The model pCO2 seasonal evaluation against SOCATv6 is only performed in 11 MARCATS regions, namely the Californian Current (M2), tropical E Pacific (M3), the Gulf of Mexico (M9), the US East Coast (M10), S Greenland (M15), Norwegian Basin (M16), NE Atlantic (M17), Iberian upwelling (M19), Moroccan upwelling (M22), China Sea and Kuroshio (M39), and New Zealand (M36). The modeled seasonal cycle is in good agreement with the one derived from SOCATv6 (Fig. 6b–l, Table S2) with absolute biases < 20 µatm for all of the 11 selected MARCATS and Pearson correlation coefficients close to 0.5 or higher except for the Iberian Upwelling (M19, Pearson value of 0.2) and on the New Zealand shelf (M36, value of 0.3). We did not perform the SOCATv6 model seasonal evaluation for the other MARCATS regions because the vast majority of grid cells only include data for less than 4 climatological months (Fig. 6a). However, we also evaluated the simulated pCO2 seasonality against SOCATv6 in regions where this evaluation is not possible to be performed at the MARCATS scale. To do so, we selected four sites of smaller spatial extent than MARCATS for which we calculated climatological seasonal pCO2 signals from the SOCATv6 dataset and compared them with the model pCO2. These sites are located off the Antarctic Peninsula, on the Queensland Plateau in NE Australia, in coastal waters of Papua New Guinea and off Terra Nova in Antarctica (see black boxes in Fig. 6a). In those regions, the absolute biases of the seasonal amplitude between MOM6-COBALT and SOCATv6 (Fig. 6m–p) are less than 20 µatm, and the phase in the seasonal cycles presents a good agreement with a Pearson correlation coefficient value of 0.8, except for the Papua New Guinea data (value of 0.5). Note that the model SOCATv6 seasonal evaluation for Terra Nova presents a good agreement, but the MARCATS scale (Sea of Labrador, M11) evaluation to which this region belongs to reveals a low agreement, showing that a poor agreement between coastal SOM-FFN and the model does not equate to poor model skill when these regions are undersampled by SOCATv6.

Overall, the pCO2 spatiotemporal analysis model–data evaluation shows that out of 45 MARCATS regions, 29 have an absolute bias for their annual mean < 20 µatm when MOM6-COBALT-coastal SOM-FFN, MOM6-COBALT-SOCATv6 and coastal SOM-FFN-SOCATv6 are compared (Table S1). Together, these 29 MARCATS regions represent 65 % of the global coastal ocean surface area. For the 11 MARCATS regions that are best covered by observations (MARCATS regions where > 50 % of the surface area is covered by SOCATv6 observations, Table S1), absolute biases for the annual mean are always < 20 µatm for the three product intercomparison, except in the Californian Current (M2), in the Baltic Sea (M18) and along the NE Pacific (M1). The seasonal MOM6-COBALT against coastal SOM-FFN evaluation also reveals that 39 of the 45 MARCATS regions have pCO2 seasonal amplitude biases < 20 µatm and that 34 MARCATS regions have a Pearson correlation coefficient > 0.5 (Table S2).

Based on this evaluation, we attribute for each MARCATS region a level of confidence in the model to coastal SOM-FFN agreement (“high”, “medium” and “low”; see Table 1 and Fig. 1a). Out of the 45 MARCATS regions, 25 are labeled with high agreement, meaning that they fulfill the following criteria regarding the annual mean and the seasonality (Table 1 and dotted MARCATS regions in Fig. 1a): a bias < 20 µatm in the annual mean pCO2 between MOM6-COBALT and coastal SOM-FFN, a bias < 20 µatm in the magnitude of the seasonal pCO2 cycle, and a seasonal phase characterized by a Pearson correlation coefficient > 0.5. Note that the some MARCATS regions, i.e., the Siberian shelf (M43), the Antarctic shelf (M45), the NE Pacific (M1), the tropical E Atlantic (M23) and the tropical W Indian Ocean(M26), also present an annual mean pCO2 bias < 20 µatm in the MOM6-COBALT-SOCATv6 and coastal SOM-FFN-SOCATv6 comparisons (Table S1). In addition, seven high-agreement MARCATS regions also show a data density > 50 % (this comes to 13 MARCATS regions if we lower the data coverage to > 30 %, Fig. 1a). These 7 MARCATS regions are located in contrasted coastal environments, i.e., three EBCs (Iberian upwelling, M19; Moroccan upwelling, M22; and the Leeuwin Current, M33), one WBC (US East Coast, M10), one Polar region (Norwegian Basin, M16), one subpolar region (NE Atlantic, M17) and one marginal sea (Gulf of Mexico, M9). These seven high-agreement MARCATS regions could also result from the very good correspondence between the data–model annual mean and seasonal patterns in environmental fields (Table S1 and Table S2, except M22, M33 and M9 for the nutrient phasing) and are therefore excellent potential candidates for an analysis of the processes controlling the coastal pCO2 dynamics. A total of six additional MARCATS regions fulfill the criteria related to the seasonal pCO2 evaluation, but they fail to fulfill the annual mean pCO2 bias threshold of 20 µatm. These medium-agreement regions (Table 1 and dashed regions in Fig. 1a) include two EBCs (Californian Current, M2, and SW Africa, M24), one marginal sea (Sea of Okhotsk, M41), two polar regions (Canadian Archipelago, M13, and N Greenland, M14) and one subpolar region (NW Pacific, M42) shelves. The majority of marginal seas are systematically associated with large biases relating to either pCO2 or the main environmental variables. These regions fulfill only one criterion (or none of them) regarding the pCO2 seasonality, and they are hence labeled as low-agreement regions (Table 1, Fig. 1a). Other low-agreement regions include one EBC (Peruvian Upwelling Current, M4), one Indian region (Bay of Bengal, M31), two tropical regions (tropical E Pacific, M3, and SE Asia, M38), two subpolar regions (Sea of Labrador, M11, and New Zealand, M36) and one WBC region (Brazilian Current, M6).

While our results show a relatively good agreement between MOM6-COBALT and coastal SOM-FFN regarding the spatial and temporal pCO2 distribution over the global coastal ocean, the comparison remains challenging for several reasons.

First, while the climatology of Laruelle et al. (2017, coastal SOM-FFN) is currently the best available product for a model–data comparison, it has its own limitations. For instance, in some regions, particularly for coastal upwellings such as the Moroccan (M22) and Peruvian (M4) upwellings, the pCO2 fields generated by the coastal SOM-FFN do not reproduce the high and variable pCO2 values measured in situ well (see, e.g., Friederich et al., 2008; McGregor et al., 2007). Such poor performance of the coastal SOM-FFN algorithm in these types of systems has already been identified by Laruelle et al. (2017). Indeed, upwelling regions are still relatively poorly monitored and expand partly beyond the coastal domain used by Laruelle et al. (2017), leading to locally skewed calibration of the SOM-FFN. Deficiencies in the observation-based product can thus partly explain the large model–data bias (106 µatm, the largest of all MARCATS regions) calculated in the Peruvian upwelling region. Moreover, although the Surface Ocean CO2 Atlas database (SOCAT) has expanded significantly over the past few years, some regions are still poorly monitored. In the coastal regions where no observational data exist (e.g., in the Black Sea, the Sea of Okhotsk, the Bay of Bengal, Fig. 4b) in the SOCAT database used here (SOCATv6, Bakker et al., 2016), it is difficult to evaluate the performance of the SOM-FFN and thus of an Oceanic General Circulation Model (OGCM) in reproducing the pCO2 field. In addition, for certain regions subjected to complex dynamic biogeochemical settings (e.g., upwelling, seasonal cover of sea ice, influenced by rivers, marginal seas), the pCO2 field reconstructed by the SOM-FFN suffers from poor performance, which can partly be explained by the lack of observational data. This lack of observations could partly explain why MOM6-COBALT-coastal SOM-FFN pCO2 biases exceed 20 µatm in these regions. The seasonal model evaluation against SOCATv6 is limited at the MARCATS scale and mainly performed against coastal SOM-FFN due to the very few coastal regions that contain a continuous climatological seasonal pCO2 cycle (Fig. 6a) in the SOCATv6 database. This study highlights the regions (Fig. 1a, e.g., Indian ocean margins, the Peruvian upwelling, marginal seas) where new observational data are most urgently needed, specifically those collected during periods of the year that are currently not covered, to improve our understanding of the CO2 exchange between coastal regions and the atmosphere at the regional and global scales. In addition, only one global continuous pCO2 climatology derived by the SOM-FFN method currently exists for the coastal ocean. It would therefore be beneficial for the community to develop other observation-based climatologies relying on other interpolation techniques, as is currently the case for the open ocean.

Second, the model–data comparison should also be analyzed in the light of the current limitations in the model itself. OGCMs have been designed for global ocean applications, and the coarse spatial resolution of these models, on the order of 0.5∘ in the present study, cannot accurately resolve mesoscale and sub-mesoscale processes and tidal mixing in shelf regions even with a model configuration including parameterizations for these processes. The coastal currents are also not always well resolved because of the coarse resolution of the shelf bathymetry. These small-scale hydrodynamic features are known to affect the spatiotemporal variability of pCO2 and the air–sea CO2 exchange (Bourgeois et al., 2016; Kelley et al., 1971; Lachkar et al., 2007; Laruelle et al., 2010). Therefore, although MOM6-COBALT runs at 0.5∘, discrepancies between coastal SOM-FFN and MOM6-COBALT in narrow EBCs such as the Peruvian Upwelling Current (M4) and along SW Africa (M33) could also be explained by the limited spatial resolution of the model. Moreover, OGCMs such as MOM6-COBALT have a relatively simple representation of biogeochemistry that does not fully capture some of the important processes of the carbon dynamics in coastal waters, such as sea ice temporal dynamics (Adcroft et al., 2019), neritic calcification (O'Mara and Dunne, 2019), or terrestrial and marine organic matter decomposition and burial (Lacroix et al., 2021a, b). Moreover, the largest biases observed in marginal seas can partly be explained by large fluvial inputs and oceanic water flows through fine-scale topography (e.g., straits) that are poorly represented in global OGCMs.

Finally, the annual mean and seasonal pCO2 biases between the coastal SOM-FFN and MOM6-COBALT can also be traced back to divergences in the environmental fields simulated by the model compared to observations (Tables S1 and S2). For instance, in most marginal seas, the model poorly resolves the annual mean and seasonal cycle of SSS and nutrients compared to the observations. These discrepancies impact the simulated pCO2 via the controls of the SSS on the CO2 solubility and of nutrients on the biological pump and CO2 uptake. In the tropical W Atlantic (M7), which is under the influence of the Amazon River, the model simulates lower annual mean SSS (and therefore lower pCO2) than the observations. In the tropical E Pacific (M3) and in Southeast Asia (M38), the poor agreement between simulated and observed seasonal pCO2 cycle could be explained by significant biases in the nutrient seasonal cycles (low Pearson correlation coefficient). Interestingly, some regions reveal significant biases in the major environmental fields but not in the pCO2 (e.g., the tropical W Atlantic, M7), while in other regions the reverse is observed (e.g., the Mediterranean Sea, M20; W Arabian Sea, M27; and in New Zealand, M36). In addition, for some regions biases in environmental fields do not affect the pCO2 as expected. For instance, along the US East Coast (M10), MOM6-COBALT simulates larger SST compared to observations, while the simulated pCO2 is lower compared to coastal SOM-FFN on an annual mean. This clearly shows that biases in environmental fields are not sufficient to fully explain the biases in pCO2 diagnosed between MOM6-COBALT and coastal SOM-FFN.

The seasonal evolution of pCO2 averaged over the US East Coast (M10) is represented in Fig. 7a. Ocean pCO2 is at a minimum in winter (February–March ∼ 331 µatm), increases through spring and peaks in summer (August, ∼ 400 µatm) before decreasing again in autumn. Figure 7b reveals the complex interplay of the four ocean internal processes (thermal and biological processes, ocean circulation, and freshwater flux) on the seasonal pCO2 variability that can either act in synergy or oppose each other.

The thermal effect (thermal, red line in Fig. 7b) increases pCO2 from early spring to summer by decreasing the solubility of CO2. In contrast, the solubility of CO2 increases in autumn and winter, inducing a decline in pCO2. The largest changes in pCO2 associated with the change in SST occur during spring (29 µatm per month in June) and autumn (-26 µatm per month in November). This thermal effect was already identified by Signorini et al. (2013) in their observational study and further confirmed by Cai et al. (2020). These authors highlighted that lowest pCO2 was generally reported in winter or at the beginning of spring and highest pCO2 in summer or autumn, despite significant temporal and spatial heterogeneity between the different sub-regions of the US East Coast (Scotian shelf, the Gulf of Maine, the Georges Bank and Nantucket shoals, the Middle Atlantic Bight, and the South Atlantic Bight). The effect of biological processes above the mixed-layer depth (bio, green line) reduces pCO2 throughout the year, revealing that primary production exceeds organic matter degradation in the surface layer all year long. The largest pCO2 decrease associated with biological processes is observed in early spring (values of -68 µatm per month in April), which is well documented (e.g., Shadwick et al., 2010, 2011; Signorini et al., 2013). The transport of chemical species by ocean circulation (circ, blue line) increases pCO2 and tends to oppose biological processes year-round except at the end of autumn and beginning of winter. This pCO2 increase induced by the circulation term is at its maximum in April (26 µatm per month). Throughout the year, the contribution of freshwater fluxes (fw, pink line) remains minor compared to the other terms (maximum absolute value of 9 µatm per month in January). For each month and season, the air–sea CO2 exchange term (CO2 flux, black line) counteracts change in pCO2 associated with ocean internal processes taking place in surface seawater (sum of bio, circ, thermal and fw). The CO2 flux term increases pCO2 at the sea surface (acting as an atmospheric CO2 sink) throughout the year, except during summer (between July and September) where it decreases sea surface pCO2 and releases CO2 towards the atmosphere (acting as an atmospheric CO2 source). This year-round simulated atmospheric CO2 uptake (except for the summer season) is also in agreement with previous literature (Fennel et al., 2019; Laruelle et al., 2015; Signorini et al., 2013). The study of Laruelle et al. (2015) has nevertheless shown that in spring the southern part of the US East Coast is quasi-neutral and that in autumn some regions, such as the Gulf of Maine or the Georges Bank, act as a CO2 source. The temporal change of pCO2 (pCO2 change, cyan line) is the result of the non-perfect balance between the internal processes and the air–sea CO2 flux.

We evaluate the rate of change tied to each process during the marked peak-to-peak pCO2 increase observed between winter and summer (from 331 µatm in February to 400 µatm in August, Fig. 7a). A positive rate of change (in µatm per month) indicates that the process contributes to an increase in pCO2 between winter and summer (February–August). This process-based analysis reveals that the winter-to-summer pCO2 increase on the US East Coast (M10) mainly results from thermal (rate of change =+5 µatm per month) and ocean circulation (rate of change =+4 µatm per month) influences combined with a large reduction of the biological CO2 uptake (rate change of +7 µatm per month, Fig. 7b). The importance of the thermal and circulation effects and the presence of a strong biological drawdown are in line with results from past studies (e.g., Laruelle et al., 2015; Shadwick et al., 2010, 2011; Signorini et al., 2013; and Cai et al., 2020). Our results identify the reduction of biological carbon uptake as a key control of pCO2 seasonality and thus agree with the studies of Shadwick et al. (2010, 2011) but slightly diverge compared to those of Signorini et al. (2013) or Laruelle et al. (2015), which found that the thermal effect was the dominant driver. This difference is largely explained by the different levels of details in the decomposition method. While most model studies, including ours, use seasonal change in SST, SSS, DIC and ALK, observational approaches cannot isolate the compounding changes tied to biological activity from those of ocean transport.

The pCO2 seasonal cycle in the Norwegian Basin (M16) and the Californian Current (M2) simulated by MOM6-COBALT are represented in Fig. 7c and e, respectively. The Norwegian Basin shows a near-constant pCO2 value (∼330 µatm) throughout the year, except in spring when it drops by 30 µatm (minimum pCO2 value of 300 µatm in June). The phasing of the seasonal pCO2 cycle in the Californian Current is similar to that along the US East Coast, with a minimum pCO2 value of 366 µatm in March followed by an increase that reaches a maximum pCO2 value of 433 µatm in August and then decreases again at the beginning of autumn.

The decomposition of the seasonal cycle into different processes for both the Norwegian Basin and the Californian Current (Fig. 7d and f) reveal patterns that are qualitatively similar to those already diagnosed for the US East Coast (Fig. 7b). For both shelf regions, the biological and circulation effects remain negative and positive, respectively, throughout the year, while the thermal effect increases pCO2 in spring and summer but decreases pCO2 in autumn and winter. The freshwater term is also minor compared to the other terms. Quantitatively, however, the amplitude of the different terms points to different first-order control in the pCO2 seasonality for each region. The amplitudes are calculated here using the marked peak-to-peak change in pCO2, which occurs between February and June in the Norwegian basin and between March and August in the Californian Current.

In the Norwegian basin, the strong winter to summer pCO2 decreases (43 µatm, Fig. 7c) are mainly associated with the large and rapid CO2 uptake associated with the spring phytoplankton bloom (biological rate of change =-45 µatm per month on average between February and June, with a maximum pCO2 uptake of -175 µatm per month in June, Fig. 7d). This biological drawdown is only partly compensated for by the supply of high pCO2 water masses from the ocean circulation (rate of change =+24 µatm per month). These dynamics are consistent with the fact that the Norwegian Basin is one of the most productive regions of the world and is characterized by a well-documented, intense spring bloom (e.g., Findlay et al., 2008). In addition, the effect of thermal changes only plays a comparatively minor role here (rate of change =+7 µatm per month).

In contrast to the US East Coast and the Norwegian Basin, the analysis performed in the Californian Current reveals that circulation is the main driver of the winter-to-summer pCO2 increases (68 µatm, Fig. 7e). The upwelling of high-pCO2 waters increases year-round surface pCO2. However, its influence is weaker in winter than in summer, thereby explaining the pCO2 increase observed between February and August (rate of change =+12 µatm per month, Fig. 7f). This large contribution from circulation is consistent with the simulations of Turi et al. (2014), who identified the ocean transport associated with upwelling in the Californian Current as the dominant process, and the higher intensity of the summer upwelling and its impact on pCO2 were also reported in prior work (e.g., Evans et al., 2015; Fiechter et al., 2014; Turi et al., 2014). In this region, biological processes also oppose the effect of ocean circulation, with upwelled deep water bringing nutrients to the surface and stimulating phytoplankton productivity (e.g., Evans et al., 2015; Fiechter et al., 2014; Turi et al., 2014). However, it plays a minor role in the pCO2 increase (rate of change ∼ 0 µatm per month) and the thermal effect (rate of change =+4 µatm per month).