The float data were downloaded from the Argo CORIOLIS Global Data Assembly Center in France (ftp://ftp.ifremer.fr/argo, last access: January 2023). The conductivity–temperature–depth (CTD) and trajectory data were quality controlled using the standard Argo protocol (Wong et al., 2015). The raw BGC signals were transformed to biogeochemical variables (i.e., O2, Chl a, NO3, bbp, and pH) and quality controlled according to international BGC-Argo protocols (Johnson et al., 2018a, b; Schmechtig et al., 2015, 2018; Thierry et al., 2018; Thierry and Bittig, 2018).

In the Argo data system, the data are available in three data modes: “real time”, “adjusted”, and “delayed” (Bittig et al., 2019). In the real-time mode, the raw data are converted into state variables, and an automatic quality control is applied to “flag” gross outliers. In the adjusted mode, the real-time data receive a calibration adjustment in an automated manner. In the delayed mode, the adjusted data are adjusted and validated by a scientific expert. While the real-time and adjusted data are considered acceptable for operational application (data assimilation), the delayed-mode data are designed for scientific exploitation and represent the highest-quality data with the ultimate goal (when time series with sufficient duration have been acquired) of possibly extracting climate-related trends (Bojinski et al., 2014). However, for some variables, only a limited fraction of the data is accessible in the delayed mode. Consequently, for each variable, we selected the highest-level data mode, in which at least 80 % of the data are available (see Table 1). Note that this criterion is not applied to O2, as only delayed-mode data were selected for this variable in order to generate the pseudo-observations from the CANYON-B neural network (more detail given in the following). We removed data with missing location or time information and data flagged as “bad data” (flag = 4). Depending on the parameter and the associated data mode, we also excluded data flagged as “potentially bad data” (flag = 3) (see Table 1). Finally, it should be noted that the status of the different modes of adjustment for bbp is still very inhomogeneous in the global BGC-Argo database. A quality control procedure in real time has just been proposed to the Argo Data Management Team but is not yet operationally implemented in the database (Dall'Olmo et al., 2022). As there is no current official consensus regarding the qualification of bbp data, we decided to use all data modes for this study.

Particulate organic carbon (POC) concentrations were derived from bbp observations. First, three consecutive low-pass filters were applied on the vertical profiles of bbp to remove spikes (Briggs et al., 2011): a two-point running median followed by a five-point running minimum and a five-point running maximum. The filtered bbp profiles were then converted into POC (mgC m-3) using a simplified version of the empirical POC / bbp algorithm developed by Gali et al. (2022), i.e., for depths larger than the mixed-layer depth (MLD): 1POCbbp=c+a⋅e-0.001⋅b⋅(z-MLD)z>MLD, where c is a constant deep value and b is the slope of the exponential decrease; the aforementioned variables are set to 12 010 mgC m-3 m and -6.57, respectively, as proposed by Gali et al. (2022). The global coefficient a is set to 37 990 mgC m-3 m to be consistent with a relationship, developed for global applications (i.e., POC = 38 687.27⋅bbp0.95) (European Union-Copernicus Marine Service, 2020). In the mixed layer (ML), z is fixed at z = MLD, and Eq. (1) simplifies to 2POCbbp=c+a,z≤MLD. Finally, we complemented the existing BGC-Argo data set with pseudo-observations of NO3, PO4, Si, Alk, and DIC concentrations as well as pH and pCO2 using the CANYON-B neural network (Bittig et al., 2018). CANYON-B estimates vertical profiles of nutrients; the carbonate system variables from concomitant measurements of float pressure, temperature, salinity, and O2 qualified in delayed mode; and the associated geolocalization and date of sampling information. CANYON-B was trained and validated using version 2 of the Global Ocean Data Analysis Project (GLODAPv2) data set (Olsen et al., 2016). The CANYON-B estimates of NO3 and pH were merged with measured values based on the rationale that CANYON-B estimates have RMS errors (NO3=0.7 µmol kg-1 and pH = 0.013) (Bittig et al., 2018) that are of the same order of magnitude as those of the BGC-Argo observation errors (NO3=0.5 µmol kg-1 and pH = 0.07) (Mignot et al., 2019; Johnson et al., 2017).

Finally, we verified that the RMS errors in BGC-Argo data (both measured and from CANYON-B estimates) are lower than the RMS difference between the model and BGC-Argo data; therefore, the comparison of simulated properties with the BGC-Argo data leads to a meaningful evaluation of the model performance. We believe it is reasonable to draw conclusions on the model uncertainty from BGC-Argo data as long as the BGC-Argo errors are much lower than the model–observation RMS difference.

The global model simulation used in this study (see Appendix A1) originates from the global ocean hydrodynamic–biogeochemical coupled system, based on the NEMO–PISCES (Nucleus for European Modelling of the Ocean–Pelagic Interaction Scheme for Carbon and Ecosystem Studies) model, implemented and operated by Mercator Ocean for the Copernicus Marine Service within the framework of the European Union's Earth observation program (European Union-Copernicus Marine Service, 2019). The BGC component is constrained by the assimilation of satellite Chl a concentrations, and a climatological damping is applied to nitrate, phosphate, oxygen, and silicate with the World Ocean Atlas 2013, to dissolved inorganic carbon and alkalinity with the GLODAPv2 climatology (Lauvset et al., 2016), and to dissolved organic carbon and iron with a 4000-year PISCES climatological run. The BGC model is forced in offline mode by daily averages of ocean physics, sea ice, and atmospheric conditions. The ocean physics and sea-ice forcing come from the global ocean physics analysis and forecasting system at 1/12∘ (Lellouche et al., 2018); the aforementioned system assimilates along-track altimeter data, satellite sea surface temperature and sea-ice concentration, and in situ temperature and salinity vertical profiles. The BGC model has a 1/4∘ horizontal resolution with 50 vertical levels (22 levels in the upper 100 m, and the vertical resolution decreases from 1 m near the surface to 450 m near the ocean bottom.

We used the daily output of Chl a, NO3, PO4, Si, O2, pH, DIC, Alk, and pCO2 as well as the weekly output of two size classes of phytoplankton, the small detrital particles and microzooplankton (resampled offline from a weekly to daily frequency via constant interpolation) from 2009 to 2020. Note that the method of linear resampling, which artificially increases the number of data, could potentially bias the statistical results, especially in regions with poor data coverage. As suggested by Gali et al. (2022), the POC concentration was computed offline by adding the two size classes of phytoplankton (the small detrital particles and microzooplankton modeled by PISCES) together. This particular combination of phytoplanktonic and non-phytoplanktonic organisms has been shown to match the small POC observed by the floats. The partial pressures of CO2 values were extrapolated in the mixed layer from the surface value estimated by the model. The Black Sea was not considered in the present analysis because the model solutions are of poor quality. Finally, the daily model output was co-located in time and space closest to the BGC-Argo float positions and was interpolated to the sampling depth of the float observations. The characteristics of the model are further detailed in the Appendix.

In this section, we present the 23 metrics used for the clustering of the ocean and for the assessment of the model simulation with BGC-Argo data. The metrics are associated with the carbonate chemistry, the biological carbon pump, and oxygen levels. Most of the metrics evaluate the model state accuracy via the comparison of simulated state variables with BGC-Argo observations that are depth averaged in the mixed (hereinafter indicated with the subscript “mixed”) and mesopelagic (hereinafter indicated with the subscript “meso”) layers. This two-layer comparison between model and BGC-Argo data provides an indirect evaluation of the key processes and fluxes associated with the carbonate chemistry, biological carbon pump, and oxygen levels in the mixed and mesopelagic layers. In addition, some of the metrics assess the skill of the model with respect to capturing emergent properties, such as the nitracline depths, DCMs, and OMZs. The metrics are described below and summarized in Table 2. The definition of the metrics is the same for the model and the BGC-Argo data. The MLD is computed, following De Boyer et al. (2004), as the depth at which the change in potential density (from its value at 10 m) exceeds 0.03 kg m-3. Dall'Olmo and Mork (2014) define the mesopelagic layer as the region between the deeper of either the euphotic layer depth or the MLD and a depth of 1000 m. However, for ease of use, we adopt a simplified definition that considers the mesopelagic layer to be the region between the MLD and a depth of 1000 m. To ensure the accuracy of the metrics' calculation, we have checked the representation of the MLDs in the model. The model's MLDs closely match the observed data, as indicated by an overall mean-square difference of approximately 30 % of the total variance in the observations.

The uptake of ∼26 % of anthropogenic CO2 by the global ocean (Friedlingstein et al., 2022) has altered the oceanic carbonate chemistry over the past few decades (Iida et al., 2020). Therefore, assessing how models correctly represent the oceanic carbonate chemistry is critical if we aim to derive accurate climate projections of future change. The classical variables for the study of carbonate chemistry are DIC, Alk, pH, and pCO2 (Williams and Follows, 2011). These variables are assessed in the mixed (DICmixed, Alkmixed, pHmixed, and pCO2mixed) and mesopelagic (DICmeso, Alkmeso, and pHmeso) layers. The partial pressure of CO2 is only assessed in the mixed layer, as the evaluation of pCO2mixed plays a critical role in the assessment of BGC models' skill to correctly represent the air–sea CO2 flux.

The biological carbon pump is the transformation of nutrients and dissolved inorganic carbon into organic carbon in the upper part of the ocean via phytoplankton photosynthesis and the subsequent transfer of this organic material into the deep ocean. The functioning of this pump relies on key pools of nutrients and carbon as well as several processes that control mass fluxes between the pools. Changes in the biological carbon pump are now manifesting globally (Capuzzo et al., 2018; Osman et al., 2019; Roxy et al., 2016).

One way to indirectly evaluate a model's ability to accurately capture essential processes related to the biological carbon pump in the ocean's upper layer, such as primary production, respiration, and grazing, is to compare various ML pools (here the nutrients (NO3mixed, PO4mixed, and Simixed), Chlmixed, and POCmixed) with BGC-Argo observations. Similarly, the assessment of the mesopelagic nutrients and POC concentration (hereinafter denoted NO3meso, PO4meso, Simeso, and POCmeso) provides an indirect evaluation of the key mesopelagic layer processes, such as export production and respiration.

In stratified systems, a DCM is formed at the base of the euphotic layer (Barbieux et al., 2019; Cullen, 2015; Letelier et al., 2004; Mignot et al., 2014, 2011). It has been suggested that the DCM plays a key role in the synthesis of organic carbon by phytoplankton (Macías et al., 2014). Therefore, DCMs are key features to be assessed in BGC models with respect to processes involved in the biological carbon pump, such as primary production. However, the DCM layer generally escapes detection by remote sensing. Furthermore, the DCM is also an emergent feature that develops in response to complex physical and biogeochemical interactions (Cullen, 2015). Thus, its evaluation provides critical information regarding the accuracy of the model with respect to capturing complex patterns of key ecosystem processes. The depth and magnitude of the DCM (HDCM and ChlDCM, respectively) are helpful metrics for the assessment of DCM dynamics. The depth of the DCM is calculated as the depth where the Chl a maximum occurs in the profile, with the criterion that the HDCM should be deeper than the MLD. The magnitude of the DCM corresponds to the Chl a value at the HDCM.

NO3 is often depleted in the surface layers and is a limiting factor for phytoplankton growth in most oceanic regions. The vertical supply of NO3 to the surface layers depends, among other factors, on the vertical gradient of NO3 (the nitracline) and, in particular, on its depth (the nitracline depth) (Cermeno et al., 2008; Omand and Mahadevan, 2015). Therefore, the comparison of the simulated nitracline depth (Hnit) with BGC-Argo observations allows for an indirect assessment of the model performance with respect to reproducing vertical fluxes of NO3. Following previous studies (Cermeno et al., 2008; Lavigne et al., 2013; Richardson and Bendtsen, 2019), the depth of the nitracline is identified as the first depth where NO3 is detected. A detection threshold of 1 µmol kg-1 is used, which is an upper estimate of the accuracy of BGC-Argo NO3 data (Johnson et al., 2017; Mignot et al., 2019).

Oxygen levels in the global and coastal waters have declined over the whole water column over the past decades (Schmidtko et al., 2017), and OMZs are expanding (Stramma et al., 2008). Therefore, assessing how models correctly represent ocean oxygen levels as well as the OMZs is critical to monitor their change over time. Similar to the assessment of DCMs, evaluating oxygen minimum zones (OMZs) provides insight into how the model represents emergent dynamics resulting from intricate physical and biogeochemical interactions (Paulmier and Ruiz-Pino, 2009). Oxygen levels are evaluated in the mixed (O2mixed) and mesopelagic (O2meso) layers. OMZs are defined as oceanic regions where O2 levels are lower than 20 µmol kg-1 (Paulmier and Ruiz-Pino, 2009). OMZs are characterized by their depths (H2min) and their concentrations (O2min).

In this study, we use the k-means clustering algorithm (Hartigan and Wong, 1979) to regionalize the ocean based on the modeled climatological monthly time series of the 23 metrics described previously. The k-means clustering algorithm is an unsupervised machine learning technique that groups similar objects together in a way that maximizes similarity between objects within a group and minimizes similarity between objects in different groups. This clustering tool has been successfully used to classify marine BGC regions in different oceanic basins based on the seasonal cycle of satellite chlorophyll (Kheireddine et al., 2021; Mayot et al., 2016; Lacour et al., 2015; D'Ortenzio and d'Alcala, 2009). The step-by-step methodology used in this study is described in the following.

The first step in the analysis involved computing monthly climatological time series for the 23 metrics at each model grid cell. These time series were derived from the monthly climatological time series of state variables predicted by the model from 2009 to 2020. To account for the lognormal distribution and the wide range of values for metrics in units of Chl a or POC, a log-10 transformation was applied to these metrics. Second, to take the 6-month shift in seasons between the Northern and Southern hemispheres into consideration, the dates for grid cells located in the Southern Hemisphere were shifted by 6 months (Bock et al., 2022). Third, to classify model grid cells based on the seasonality and amplitude of the 23 metrics, each metric was standardized by subtracting the global mean and dividing by the global standard deviation. This ensured that each metric had a mean of 0 and a standard deviation of 1, enabling comparison across metrics with different units. Fourth, to reduce the dimensionality of the data set, a principal component analysis was applied to the scaled data. Only the components that explain 99 % of the variance in the data set were kept, thereby reducing the dimensions of the data set by 85 %. A k-means clustering analysis was then performed on the resulting data set. Following Kheireddine et al. (2021), the number of clusters was determined based on a silhouette analysis (Rousseeuw, 1987), which yielded a value of eight for the number of clusters.

To quantify the model predictive skill, a model efficiency statistical score (me) was computed for each metric and in each BGC region: 3me=1-∑i=1Nmi-oi2∑i=1Noi-o‾2, where mi and oi are the model and BGC-Argo matched values, respectively; o‾ is the BGC-Argo climatology; and N is the number of matchups. Assuming that the spatial variations are small in a given BGC region, o‾ represents the temporal average and ∑i=1Noi-o‾2 represents the variance due to temporal fluctuations. The model efficiency tests whether the model outperforms the BGC-Argo climatology (me>0; Fennel et al., 2022) or, stated differently, if the model–data mean-square difference is lower than the observation variance, i.e., 1N∑i=1Nmi-oi2<1N∑i=1Noi-o‾2. To ensure the robustness of me, we verified that the number of matchups for each metric and in each BGC region was greater than 100; outliers were then removed using Tukey's fences (Tukey, 1977).

The k-means clustering algorithm identified eight distinct BGC regions (Fig. 2). Six of the eight BGC regions correspond to well-defined spatial regions and are, thus, named accordingly: “Arctic”, “equatorial” (Equ.), “Mediterranean Sea” (Med. Sea), “OMZs”, “subtropical gyres” (Sub. Gyres), and “Southern Oceans”. The other two BGC regions, i.e., “low nutrient bloom” (Low Nut. Bloom) and “high nutrient bloom” (High Nut. Bloom), are located in the North Atlantic and North Pacific oceans as well as in the lower latitudes of the Southern Oceans region. These two BGC regions correspond to ocean basins that experience a phytoplankton bloom in the springtime (Westberry et al., 2016). The main difference between these regions is that macronutrients such as nitrate and phosphate are abundant in one of them throughout the year due to phytoplankton growth being mainly limited by iron (Williams and Follows, 2011). Finally, it should be noted that outlier grid cells were not removed from the analysis; these outliers are mainly present in grid cells close to the coast. Additionally, grid cells with bathymetry shallower than 1000 m were not included in the clustering analysis, as metrics associated with mesopelagic processes cannot be calculated in these shallow grid cells.

The BGC regions found in our study are generally coherent with the biomes estimated in Fay and McKinley (2014) (hereinafter denoted FM2014). The Arctic and Southern Oceans regions correspond to the FM2014 ice biome. The Sub. Gyres region corresponds to the FM2014 subtropical permanently stratified biome. The Equ. BGC region represents a larger area than the equatorial biome in FM2014. The Low Nut. Bloom and High Nut. Bloom regions correspond to the FM2014 subtropical seasonally stratified and subpolar seasonally stratified biomes, respectively. These two BGC regions are coherent in the North Pacific and the Southern Oceans in both studies. They differ, however, in the North Atlantic: in FM2014, the subpolar North Atlantic is divided between the subtropical seasonally stratified and subpolar seasonally stratified biomes, whereas in our study this area is only represented by one BGC region – Low Nut. Bloom. Finally, the Med. Sea and OMZs BGC regions are not represented in FM2014. The main differences between our study and FM2014 are due to variations in the methodology used to identify BGC regions. In our study, we used 23 input variables to identify BGC regions, whereas clustering was based on only one BGC input variable (Chl a) and three physical variables (sea surface temperature, MLD, and sea-ice fraction) in FM2014. Our method allows the identification of specific BGC regions whose main characteristics are determined by variables other than Chl a, such as the OMZs BGC region. Furthermore, our method includes coastal areas and identifies the Med. Sea, which is not included in FM2014 because it is considered a coastal region, as a BGC region.

Figures 3, 4, and 5 display the model efficiency (me) calculated for each assessment metric and BGC region. To enhance clarity, the me values are grouped by process, namely carbonate chemistry, biological carbon pump, and oxygen levels. The results are presented as bubble plots (Figs. 3b, 4b, 5b), where the size of the bubble is proportional to the me value. A bar plot (Figs. 3c, 4c, 5c) shows the median me value for a given assessment metric across all BGC regions, while another bar plot (Figs. 3a, 4a, 5a) shows the median me value for a given BGC region across all assessment metrics. Due to the limited number of assessment metrics associated with oxygen levels in most regions (i.e., 2), the mean is used instead of the median. The x and y axes in Figs. 3b, 4b, and 5b are arranged in descending order based on the median me value across all assessment metrics (as shown Figs. 3a, 4a, and 5a) and the median me value across all BGC regions (as shown in Figs. 3b, 4b, and 5b), respectively.

The model demonstrates improved performance with respect to predicting certain carbonate chemistry metrics (i.e., Alkmeso, DICmixed, Alkmixed, DICmeso, and pHmeso) compared with the BGC-Argo climatology, as indicated by median me values significantly greater than 0 (Fig. 3b, c). However, the model's ability to reproduce instantaneous variability in pHmixed is more limited, with a me value close to 0, indicating no improvement over a simple average of observed values. Furthermore, the model underperforms the BGC-Argo climatology for pCO2mixed across all regions. Despite these limitations, the model provides an overall better estimate of carbonate chemistry dynamics in all BGC regions compared with the BGC-Argo climatology, as evidenced by Fig. 3a.

The efficiency of the model with respect to estimating the biological carbon pump metrics varies across both metrics and regions (Fig. 4a, b, c). The model outperforms the BGC-Argo climatology with respect to estimating PO4 and NO3 in the mesopelagic and mixed layer as well as with respect to estimating Simeso and Hnit. However, the model's ability to predict Si decreases significantly as one moves from the mesopelagic to the mixed layer. Additionally, the metrics associated with the first trophic level, such as Chlmixed, HDCM, ChlDCM, POCmixed, and POCmeso, are systematically outperformed by the BGC-Argo climatology, with median me values less than 0 in nearly all BGC regions (Fig. 4b). Regional analysis of the median me values (Fig. 4a) shows that the model performs better than the observational mean (median me values greater than 0) in only a few regions (i.e., High Nut. Bloom, Low Nut. Bloom, Med. Sea, and OMZs), indicating that the model performs relatively well in these regions but may not be as accurate in the other regions.

The model provides better estimates of mixed and mesopelagic O2 concentrations in most BGC regions compared with the BGC-Argo climatology, as evidenced by consistently positive me values in Fig. 5b. However, the BGC-Argo climatology provides a better representation of the magnitude of O2min compared with the model, while the model performs better than the climatology with respect to predicting HO2min, although only in the OMZs BGC region. These results suggest that, while it performs well with respect to estimating mixed and mesopelagic O2 concentrations in most BGC regions, the model does not accurately capture the depth and magnitude of OMZs.

The model outperforms the BGC-Argo climatology for DIC, Alk, NO3, and PO4 in the mesopelagic and mixed layers as well as for Si in the mesopelagic layer. We attribute this good performance to the effective application of climatological damping. As described in the Appendix, climatological damping mitigates the effects of physical data assimilation in the offline coupled hydrodynamic–biogeochemical system, which can lead to unrealistic drift of various biogeochemical variables. Specifically, we used the World Ocean Atlas 2013 (Garcia et al., 2013, 2014) for NO3, PO4, O2, and Si, whereas we utilized GLODAPv2 climatology (Lauvset et al., 2016) for DIC and Alk. However, our analysis revealed that the model's performance with respect to estimating Si in the mixed layer is significantly degraded compared with the mesopelagic layer, indicating the presence of additional factors affecting the model's ability to accurately estimate this variable. Further investigation is required to identify these factors and improve the model's performance with respect to estimating Si in the mixed layer.

For the three Chl a-related metrics, the model performs worse than the BGC-Argo climatology. This is unexpected, as the model incorporates a reduced-order Kalman filter (Lellouche et al., 2013) that assimilates daily L4 remotely sensed surface Chl a, providing a mixed-layer correction to the modeled Chl a (see Appendix). We verified that the assimilation of satellite Chl a improves the model's ability to predict Chl a, as the model–BGC-Argo-data misfit is lower compared with a simulation without assimilation (not shown). Furthermore, the model–satellite misfit was also found to be lower than the variability in the satellite data (European Union-Copernicus Marine Service, 2019). These results suggest that discrepancies between the assimilated satellite Chl a product and the BGC-Argo data may be responsible for the observed model–BGC-Argo-data misfit. Therefore, we suggest that future studies investigate the consistency between ocean color products and BGC-Argo Chl a products at a global scale, as these two products are expected to be assimilated together in future operational BGC systems (Ford, 2021).

Overall, the model also performs worse or no better than the BGC-Argo climatology with respect to predicting POC concentrations, the magnitude and depth of OMZs, pHmixed, and pCO2mixed. The poor performance of PISCES-based simulations relative to BGC-Argo POC observations has been extensively studied in Gali et al. (2022). They pointed out that the large model–data misfit could be the result of an imperfect BGC-Argo POC–bbp conversion factor, unsuitable model parameters associated with POC dynamics, and missing processes in the model structure. Similarly, the poor model skill in capturing the OMZs' dynamics has also already been documented in several studies (Busecke et al., 2022; Schmidt et al., 2021; Cabré et al., 2015). All of these studies suggested that improving the ocean circulation in physical models may be the most important factor to improve the accuracy of OMZs' model predictions. Finally, the negative model efficiencies for pHmixed and pCO2mixed can be attributed to the fact that these variables are driven by DIC, Alk, temperature, and salinity. Therefore, even small uncertainties in the model estimates of DIC, Alk (as shown in Fig. 3b), temperature, and salinity (Lellouche et al., 2018) can result in poor model performance in capturing the variability in pH and pCO2. This highlights the importance of accurately modeling these four variables to improve model estimates of pH and pCO2.

Observing system simulation experiments (OSSEs) have been the primary tool used to provide information about the design of the BGC-Argo observing system (Ford, 2021; Biogeochemical-Argo Planning Group, 2016). OSSEs typically comprises a realistic “nature run”, which represents “the truth” from which synthetic observations are sampled. The synthetic observations represent the observing system to be designed. To test its impact on improving model's predictive skill, the synthetic observations are then assimilated in an “assimilative run”. The accuracy of the assimilative run is then evaluated against the nature run. Here, we use the real BGC-Argo observations to provide information about the design of the BGC-Argo network. More specifically, our aim is to gain information about the regions where the model errors are greater than the variability in the BGC-Argo data and, consequently, where BGC-Argo observations should be enhanced to improve the model accuracy via BGC-Argo data assimilation or process-oriented assessment studies.

For a given BGC region, we compute a single multivariate score corresponding to the median of the 23 me values associated with each assessment metric (Fig. 6). This is consistent with the fact that the BGC-Argo floats which are now deployed observe the five variables used to derive the assessment metrics, i.e., O2, Chla, NO3, bbp, and pH. In the Arctic and Southern Oceans BGC regions (typically north of 60∘ N and south of 60∘ S, respectively), the median me is barely greater than 0, suggesting that, in these regions, the model performs slightly better than a simple mean of the observed values. In these two regions, the model is not well constrained by the assimilation of remotely sensed Chl a because satellite observations of ocean color are not possible for most of the year due to ubiquitous cloud cover. In addition, the scarcity of in situ observations probably makes the climatological forcing less efficient in these regions. Together, these factors are likely to lead to poorer model performance compared with other regions. Consequently, we strongly recommend enhancing the Arctic region where BGC-Argo observations are scarce (Fig. 1) and where the winter–spring months are particularly undersampled (not shown). We also recommend maintaining the existing high density of BGC-Argo observations in the Southern Oceans. These observations are critical to better constrain the model in these two regions where the constraint of models via the assimilation of satellite observations is not possible for most of the year.