The climate problem is highly complex, the stakes are very high, and substantial knowledge gaps remain, especially in the ocean biogeochemistry domain (Kwiatkowski et al., 2020). More broadly, the need to address complex issues related to the carbon cycle, ecosystem services and biogeochemistry through Earth System Models (ESMs), e.g., in the context of climate adaptation and resilience, has been highlighted by expert groups (Hewitt et al., 2021). Similarly, Jones et al. (2024) evaluate modelling priorities to support international climate policy and emphasise the value of “a coordinated, internally consistent set of simulations, data, and knowledge to support Intergovernmental Panel on Climate Change (IPCC) assessments” and outline multiple applications of Coupled Model Intercomparison Project (CMIP) projections. These include investigations of threats to marine ecosystems (which have consequences for the ocean's ability to buffer climate change, Tjiputra et al., 2025) and downstream services under various climate scenarios and associated risks of tipping points. Jones et al. (2024) also state that improving confidence in future projections requires models to reproduce the observed historical period. Furthermore, they identify parameter uncertainty as one of the key elements of uncertainty in climate models.

Against this background and in line with the recommendations of expert bodies, we focus on the climate priority challenge associated with marine ecosystem and biogeochemistry modelling, and particularly on marine primary production. Phytoplankton primary production (PP), the process by which marine autotrophs convert CO₂ into organic matter through photosynthesis, is a major component of the ocean and planetary carbon cycle. Currently estimated at around 50 Pg C yr⁻¹ (Kulk et al., 2020, 2021), the magnitude of marine PP is five times the estimated fossil fuel emissions of 10 Pg C yr⁻¹ in 2022 and nearly 20 times the net ocean carbon sink (Friedlingstein et al., 2025). Its magnitude is comparable to that of terrestrial primary production (Lurin et al., 1994; Longhurst et al., 1995; Field et al., 1998; Friedlingstein et al., 2025). A key question in climate research is whether the current levels of marine PP can be maintained under climate change (Tagliabue et al., 2021), when marine ecosystems are increasingly threatened by a variety of processes, including ocean acidification (Jin et al., 2020; Dai et al., 2025), rising seawater temperatures (Kwiatkowski et al., 2020), intensified storminess over the oceans (Gastineau and Soden, 2009; Young and Ribal, 2019; Gentile et al., 2023; Liu et al., 2024), ocean deoxygenation (Schmidtko et al., 2017), modified current and stratification influencing surface nutrients (Maishal, 2024), changes in aerial nutrient supply (Bergas-Masso et al., 2025), biodiversity loss (Luypaert et al., 2020), and sea-ice loss (Myksvoll et al., 2023). In this review we consider PP estimated from two types of models: “satellite-based models” that utilise remote-sensing data together with physiological models, whose parameters are informed by in situ measurements, to calculate PP; and mechanistic “ecosystem models” which use numerical methods to solve a set of differential equations representing many ecological processes, with one of them being PP.

When discussing marine PP it is important to keep track of its different components. Theoretically, PP before any of the loss terms are considered is referred to as gross PP (GPP); once the respiration by marine autotrophs is subtracted from GPP we obtain net PP (NPP). GPP can also be partitioned according to whether only the organic carbon fixed into particulate material is considered (production of particulate organic carbon), or if the exudates (dissolved organic matter) are also included in the estimate (production of total organic carbon, Regaudie-de-Gioux et al., 2014). Models can make explicit distinction between these components, though this is not always done. When it comes to in situ observations, experimental methodologies and carefully assembled protocols exist for measurement of each of these components (IOCCG, 2022); however, practical constraints may limit the extent to which the components may be differentiated from each other. Furthermore, various observational methods of the same component could yield differing values. For example, multiple methods for measuring the same component could have different intrinsic timescales that are applicable to them, making direct comparisons difficult. Improving observational tools for PP, including developing reliable PP error models, is a priority for the scientific community, in addition to the modelling issues that are the primary focus of this paper.

Here, we focus mainly on GPP as computed in many ecosystem models. For satellite-based estimates, we treat PP derived from short (1–4 h) in situ incubations as GPP, and those derived from longer (12–24 h) incubations as NPP, while fully recognising that the distinction is not that clear cut (e.g., Halsey et al., 2011). Furthermore, estimates of the magnitude of losses due to respiration vary considerably. Some estimates place it at about 30 % of GPP (e.g., Platt and Sathyendranath, 1991), while some other estimates are higher (e.g., 60 % according to Halsey et al., 2011). Platt and Sathyendranath (1988) compared daily water-column PP computed on the basis of short incubations with those measured in situ over daily time scales, and showed the two sets of independent estimates to be comparable, which points to low respiration losses. Also, satellite-based estimates of NPP (Behrenfeld and Falkowski, 1997) tend to be roughly the same or higher than GPP estimates (Longhurst et al., 1995). Since, by definition, NPP cannot be greater than GPP, these comparisons reveal a great deal of uncertainty in respiration, or in PP computed using different approaches, when compared with each other.

Considerable differences exist in model-based estimates (here and elsewhere, we use “models” without a qualifier, to mean both satellite-based and ecosystem models) of the current and past global PP in the ocean, and in ecosystem-model based projections into the future.

Satellite-based estimates of global marine PP are converging around 45–55 Pg C yr⁻¹ (Fig. 1A). These estimates were obtained from both multi-sensor products of the Ocean Colour Climate Change Initiative (OC-CCI; version 6, Sathyendranath et al., 2019; Kulk et al., 2020, 2021), as well as from single-sensor products of the Oregon State University (http://orca.science.oregonstate.edu/, last access: 28 April 2026), which include the Carbon, Absorption, and Fluorescence Euphotic-resolving (CAFE) model (Silsbe et al., 2016, 2025), Carbon-Based Primary Productivity Model (CBPM; Westberry et al., 2008), the Vertically Generalised Production Model (VGPM; Behrenfeld and Falkowski, 1997) and the VGPM-Eppley model (which incorporates the Eppley (1972) temperature function). However, we note that much higher values (up to 67 Pg C yr⁻¹) and lower values (≤45 Pg C yr⁻¹) have also been reported from satellite-based products (Antoine et al., 1996; Behrenfeld et al., 2005; Carr et al., 2006; Uitz et al., 2010) (here we recognise that satellite products may differ in the computed PP components, as noted earlier).

Large differences also emerge in the PP trends over the last decades estimated from both the CCI and Oregon State University products (Fig. 1B), as well as associated reanalyses (e.g., those of Gregg and Rousseaux, 2019). These differences are strongly impacted by the choice of historical period and the underlying characteristics of the satellite products (e.g., single sensor or multi-sensor), but the choice of satellite-based PP model does matter: in a recent comparison (Ryan-Keogh et al., 2025) of six satellite-based primary production models applied to a common satellite product (OC-CCI) and a common period (1998–2023), four of them showed declining trends, while the other two showed an increasing trend. Interestingly, the split is along the lines of whether the models incorporated temperature-dependent production parameters, or not. Ryan-Keogh et al. (2025) also compared satellite products with those of several ecosystem models from the Climate Model Intercomparison Project (CMIP-6), and concluded that, in general, the climate models underestimated the decreasing trends seen in some of the satellite-based models.

Differences in marine PP and its trends are not limited to satellite-based products. Earth System Model intercomparisons show considerably larger uncertainty than the satellite models for the annual NPP estimate during the (recent past) “historical” period (with values reported in the 17–83 Pg C yr⁻¹ range; Bopp et al., 2013; Doney et al., 2014; Laufkötter et al., 2015; Tagliabue et al., 2021; see Fig. 2), whilst showing weak or no trends over the recent historical period (Kwiatkowski et al., 2020). Ecosystem model uncertainties are even higher in future projections where models disagree even on the sign of change up to the year 2100 under the high emission scenario, although most ecosystem models project a decline in global PP. While the uncertainty in annual NPP in the recent past has decreased in the CMIP6 (Coupled Model Intercomparison Project phase 6) ensemble compared with CMIP5, the uncertainty in projected PP trends has increased significantly in the CMIP6 ensemble compared with CMIP5 (Kwiatkowski et al., 2020). In particular, while the ensemble mean in CMIP5 suggested a significant decrease in PP at the global scale of -8.06 % ± 4.83 % (where the uncertainties are reported as the inter-model standard deviation), the CMIP6 ensemble has a much smaller mean and the larger standard deviation includes the null hypothesis of no trend (-1.76 % ± 9.01 %). Frölicher et al. (2016) have noted that ecosystem model uncertainties (missing/mis-represented processes, parameter uncertainties) dominated the total uncertainty in the 21st-century projections of PP and their relative importance with respect to scenario uncertainty does not decrease with projection lead time. Recent studies have confirmed this, highlighting the role of uncertainty in the representation of key biogeochemical processes, including diazotrophy (Tagliabue et al., 2021; Bopp et al., 2022; Doléac et al., 2025), bacterial remineralisation (Kim et al., 2023) and parameter uncertainty (Jones et al., 2024), including in zooplankton grazing rates (Rohr et al., 2023). Laufkötter et al. (2015) concluded that the projected future changes in marine PP are driven by multiple processes, including changes in circulation or mixing, leading to a stronger lateral or vertical loss of biomass; increased aggregation or mortality of phytoplankton; or higher grazing pressure. Laufkötter et al. (2015) also noted that temperature-dependent functions of PP and loss terms can affect the direction of change of PP from marine ecosystem models in climate warming scenarios. Regional variations in PP are especially sensitive to how models represent this wide range of processes (Dutkiewicz et al., 2013), and given the high uncertainty in their model representation, very few of the models agree on the direction of the trend regionally. Furthermore, global models, with their coarse horizontal resolution, struggle to capture coastal and estuarine processes that enhance PP (coastal regions account for 14 %–33 % of global PP; Gattuso et al., 1998), which makes them also prone to underestimate global PP.

Several studies have also been carried out to compare estimates from ecosystem models with satellite-based products and in situ observations, both at global scale (Carr et al., 2006; Steinacher et al., 2010; Bopp et al., 2013; Laufkötter et al., 2015; Séférian et al., 2020; Ryan-Keogh et al., 2025) and at regional scales (Friedrichs et al., 2009; Saba et al., 2010; Lee and Yoo, 2016; Doléac et al., 2025). In some cases, these comparisons (e.g., between ecosystem and satellite-based models) led to better constrained PP projections, e.g., in the tropics, using an emergent constraint approach (Kwiatkowski et al., 2017). However, it is fair to say that, overall, these comparisons have not led to convergence of model outputs that would reduce the uncertainty of marine PP estimates. All previous works have highlighted large differences between estimates (e.g., varying from <-60 % to >60 %; Séférian et al., 2020), with highly variable spatial patterns (Bopp et al., 2013). Tagliabue et al. (2021) highlighted the need for stronger constraints on NPP using new approaches that include the growing observational coverage from Biogeochemical-Argo (BGC-Argo) floats (Claustre et al., 2020; for an example of this see Arteaga et al., 2022). Field-based observations of PP, typically treated as the “truth”, are often compared with model outputs to evaluate model performance. However, this type of comparison is confounded by uncertainties in PP measurements, which can be quite high, as well as by the differences in the spatial and temporal scales of the in situ observations and the validated models. Furthermore, there are also questions around whether the PP is measured directly, or estimated indirectly. For example, BGC-Argo estimates PP indirectly, inferring it from other, more directly measured variables.

Given these challenges both in remote sensing and ecosystem modelling, the IPCC has assigned low confidence to current estimates of marine PP and its trends. The reasons cited include uncertainties in production estimates and projections, the short duration of available time series data used in the analyses, and the lack of independent validation (IPCC, 2019, 2021; Gulev et al., 2021). This assessment is of particular concern as it has major implications for ecosystem service provision, mitigation planning, enhancing adaptation and building resilience to climate change (Hewitt et al., 2021). These applications often require regional to local information, as PP determines spatial variability in ecosystem services such as fisheries (Marshak and Link, 2021; see Fig. 3), but uncertainties increase at these scales compared with global estimates (Tagliabue et al., 2021). Both remote sensing and ecosystem models can, in principle, deliver such regional insights, when used with granularity and resolution needed at the appropriate scales. Reducing uncertainty in models, ideally through a coordinated and internally consistent set of simulations, data and knowledge, would then enable us to discuss downstream services under various climate scenarios and associated risks of tipping points (Jones et al., 2024). Such improvements would support climate policy, as well as management decisions pertaining to climate mitigation and adaptation strategies, at both international and regional levels.

We argue here that efforts to reduce uncertainty in estimates and projections of marine PP should include a focus on investigating model structures and parametrisations, with the goal of identifying genuine inter-model differences and reconciling apparent differences. In this review, we examine both the sources of differences between satellite-based and ecosystem models, as well as within these two types of models. We argue that there is strong scientific justification for considering how the current model parameterisations could be improved. A straightforward avenue to improvement is that parameters which are currently treated as constants (e.g., the Ai parameters from Table 1) are assigned the most appropriate values consistent with the model structure and with all the available information. A step further is to allow the currently constant parameter values to vary with spatial locations and times. Although variable parameters would increase the complexity of the functional forms used in PP models, we argue that, at least in the less complex PP models (e.g., within satellite models and ecosystem models used in ESMs for climate projections), there are good scientific reasons to expect such parameter variations to be realistic. We propose that absence of such variations is responsible for the many apparent differences between the current PP models. Parameter variability might be less important for the more complex models with large numbers of phytoplankton types and/or size-classes, but for those models it is still essential to focus on the best possible ways of optimising the existing constant parameters. Furthermore, these highly complex models could provide valuable information for estimating the spatio-temporal variability of parameters used in less complex ecosystem and satellite-based models. This would also help increase consistency among different models, making them more comparable. At the same time, caution must be applied to ensure that increased consistency and convergence is not confused with increased accuracy. For this, we would need to continue independent assessments of accuracy, for example by comparisons with in situ observations, with full recognition of the caveats that such comparisons entail, as discussed above.

In general, we highlight the importance of correct parameterisations that are valid across multiple spatial and temporal scales, and for multiple phytoplankton types. We also discuss the challenges posed by such PP model parametrisations, argue that this is the right time to rise to those challenges, and propose strategies to overcome them. Finally, we discuss uncertainties in marine PP that might persist even when improved model parametrisations are adopted.

In this section, we assess how marine PP is treated in satellite-based and ecosystem models, identifying inter-model differences.

It is useful to consider GPP as the product of a biomass-specific production, say PM, where M is a measure of phytoplankton biomass, multiplied by the biomass itself. In other words: 1P=PM×M, such that PM carries all the information on the physiological controls on PP, whereas M accounts for the role of varying phytoplankton concentrations. Since phytoplankton are complex organisms, many options exist for defining biomass, including concentrations of the phytoplankton pigment chlorophyll a (B), phytoplankton carbon (C), or nitrogen content. The choice of biomass often depends on practical considerations (such as data availability) or by the study objectives (for example, carbon is an obvious choice in models designed to investigate the biologically mediated carbon cycle in the ocean). Models can also be classified according to which measure of biomass they track as the main currency in the ecosystem.

Dimensional analysis suggests that, in its simplest form, PM can be represented in a canonical form with two parts: a scale factor PmM that carries the same dimensions as PM, and a dimensionless function fI of the scaled irradiance I* available for photosynthesis (Platt and Sathyendranath, 1993), where the scaling factor would be a model parameter with the same dimensions as light, such that I* is dimensionless. Thus, in such a canonical form, PM can be written as: 2PM=PmM×fII*. In this form, PmM is not strictly constant, but implicitly accounts for the effects of other environmental variables on primary production, such as temperature (T) and nutrients (N), or changes in species composition. Such dependencies can be made more explicit (removing T and N dependence from PmM), such that Eq. (1) becomes: 3PM=PmM×FfT(T),fN(N),fII*. The function F(fT,fN,fI,) can be specified as a simple product fT×fN×fI (e.g., Laufkötter et al., 2015; Kishi et al., 2007; Vichi et al., 2007; Yool et al., 2013; Butenschön et al., 2016), representing co-limitation by each variable, or it can follow Liebig's law of the minimum (e.g., Grégoire et al., 2008; Daewel and Schrum, 2013; Radtke et al., 2019), where the most limiting resource dictates the growth rate. Note that in Eqs. (2) and (3), the functions fi are dimensionless, and that all the dimensions are carried by the scaling factor PmM. When models resolve multiple phytoplankton groups or species, Equation 3 is specified for each group, and their contributions are added to get total PP.

Commonly-used functions in PP models that represent the modulating roles of temperature, nutrients and light are summarised in Table 1. When more than one nutrient is considered, additional terms have to be included for each nutrient. Thus, models (the combined functions F) differ depending on (i) how many environmental factors are included in the model, (ii) the explicit functional forms selected for each modulating function; and (iii) the parameter values Ai, Di assigned to those modulating functions, and whether they are allowed to vary with region and time. Finally, the functions fi would ideally have values within the [0, 1] interval; however, this is often not the case for some of the temperature fT functions (as can be seen in Table 1).

In some cases, it is necessary to track multiple measures of phytoplankton biomass within a model. For example, a unit conversion between chlorophyll a and carbon might be needed to make the exponent in the light function (fI) dimensionless, or it may be that the model tracks more than one currency. Such a conversion may also be needed to transform modelled phytoplankton carbon fields into chlorophyll a fields for comparison with satellite-based chlorophyll a products. This is typically achieved using a chlorophyll-to-carbon ratio (θ), which varies among phytoplankton and under different environmental conditions and is usually estimated using photo-acclimation models. Commonly used functions in photo-acclimation models are also shown in Table 1.

In the next two sections, we examine in more detail the variety of ways in which these concepts are implemented in satellite-based and ecosystem models.

Platt and Sathyendranath (1993) showed that we can anticipate systematic biases between satellite-based models that are structured differently, and we can theoretically, or numerically predict under what conditions the biases relative to each other will manifest themselves. For example, linear and non-linear models are expected to behave similarly under low-light levels, but to diverge as light levels increase. The authors also showed that when PP models have similar structures, it is possible to reduce all of them to a common, canonical form, revealing that apparently-different model types (available light models, absorbed light models, chlorophyll-based or carbon-based models) become equivalent when implemented with comparable model parameter values (Platt and Sathyendranath, 1993; also see Sathyendranath and Platt, 2007; Sathyendranath et al., 2020). Such comparisons also reveal systematic biases between spectral and non-spectral models of PP, arising from spectral effects in both underwater light penetration and phytoplankton light utilisation. It has been demonstrated that biases between spectral and non-spectral PP models can be minimised by tuning the diffuse attenuation coefficient of downwelling irradiance, which determines the rate of change of available light with depth (Platt and Sathyendranath, 1991; Kyewalyanga et al., 1992). Similarly, Kovač et al. (2016a) demonstrated that a locally tuned non-spectral model, with adjusted values of photosynthesis parameters, can outperform a spectral model, without locally tuned values of photosynthesis parameters. Such comparisons bring to the fore the importance of parameter assessment, assignment, and evaluation to understand model performances, uncertainties and divergences, which is at the core of this review.

To illustrate the point, let us focus, for example, on the light function (fI), which takes a wide range of forms in the literature (see Table 1). Even though the functional forms cannot be analytically transformed into each other (they are mathematically different), numerically they could still be very close to each other, in the sense that they can all fit the same observations similarly well when the parameters are chosen appropriately (Kovač et al., 2017). These different forms split into two classes: one that includes photo-inhibition and the other that does not (Amirian et al., 2025). Figure 8 shows that the fI, models without photo-inhibition (Webb et al., 1974; Jassby and Platt, 1976; Smith, 1936) are all practically identical to each other for equivalent parameter values and are therefore basically indistinguishable from each other. It should also be noted that the Webb et al. (1974) model is a special case of the Platt et al. (1980) model for the case of zero photo-inhibition. The fI, model that stands out is the one of Steele (1962), which struggles to match the other fI, models under low-light conditions. However, when photo-inhibition is important, the fI, model of Platt et al. (1980) can again nicely match the Steele (1962) model if their parameter values are chosen appropriately. What we learn from Fig. 8 is that a lot of the diversity in fI models is only apparent, as the diversity can be eliminated via model parametrisation.

In general, PP models are designed to represent limitations to phytoplankton growth (whether from light, nutrients or temperature) under different environmental conditions and for different groups of phytoplankton, as appropriate. These models have the potential to be generalised to deal with additional external conditions (which may not be explicitly included in the model) by incorporating spatially and temporally variable parameter values. This flexibility allows models to account for the diversity of phytoplankton and the processes responsible for their dynamics, which are not explicitly represented in current models. Representing the full diversity of phytoplankton species is not feasible due to lack of understanding and computational demand, which is why models typically rely on the use of phytoplankton classes to represent aggregations of multiple species based on shared characteristics or traits, such as body size, biogeochemical functions, life strategies and behaviours. This approach captures at best the average or most typical behaviour of each class (e.g., Anderson et al., 2021; Ratnarajah et al., 2023). When aggregating species according to their physiological and functional traits and behavioural patterns into a pre-defined number of modelled classes, fixed values are assigned to model parameters within each aggregated class. For ecosystem models, many of these parameters have assigned values based on laboratory or mesocosm experiments (Geider et al., 1998; Schartau et al., 2017; Ratnarajah et al., 2023), often focusing on a small number of carefully-selected species, far from capturing the full diversity of organisms or their responses and behaviours that might be expected in the natural environment across large spatio-temporal scales (Geider et al., 1998; Schartau et al., 2017; Ratnarajah et al., 2023). In contrast, in the natural environment, we can expect parameters to vary in time and space, reflecting both changes in the governing conditions and in the unresolved functional diversity in the makeup of modelled planktonic communities (Schartau et al., 2017). Such parameter variability can be observed in model calibration experiments (e.g., Leeds et al., 2013; Mattern et al., 2012, 2014), including those using data assimilation to estimate model parameters jointly with the model state (e.g., Pastres et al., 2003; Tjiputra et al., 2007; Roy et al., 2012; Doron et al., 2013; Simon et al., 2015; Gharamti et al., 2017a, b; Skákala et al., 2024).

A simple illustration of how parameter variability emerges from aggregating species into classes is provided in Fig. 9. Although models differ from each other in the number of phytoplankton classes they resolve, for each phytoplankton class they typically use the same functional form to describe photosynthesis, with total phytoplankton PP corresponding to the sum of contributions across all classes. Figure 9 demonstrates that models with different numbers of classes become equivalent in their description of total PP, provided that the parameters in models with fewer classes are allowed to vary with space and time. In such a way, spatio-temporal parameter variations could effectively capture the influence of unresolved diversity in phytoplankton community structure, in models with only a few phytoplankton classes. The spatio-temporal model parameter variations would then be a consequence of the models' inability to sufficiently resolve phytoplankton species, which also means that such parameter variability would be expected to be especially relevant for the simpler models (e.g., ecosystem models typically used in ESMs). The more complex models currently in use (e.g., DARWIN; see Ward et al., 2012; Dutkiewicz et al., 2020) would have less reason to adopt spatio-temporally variable parameters; but these models are typically too computationally expensive to be run as part of ESMs in long-term ensemble-based climate projections. Furthermore, even as complex as they are, they still represent only a fraction of the real-world diversity. On the other hand, as the models get more complex by incorporating more ecosystem compartments, the challenge shifts to calibrating each of the large numbers of parameters to adequately capture the functions of each of the model components.

Together, these considerations suggest that investigating parameter assignment and parameter variability may be an important route to understand and potentially reduce many of the apparent differences between marine PP models, and hence in the estimated magnitudes of production. Investigation into the role of parameters should be followed by a consistent calibration against observational data. To estimate spatially and temporally varying parameters in ecosystem models, data assimilation can provide a natural tool for model calibration (e.g., Tjiputra et al., 2007; Singh et al., 2025). However, introducing spatio-temporally variable (or too many constant) parameters comes with its own challenges. For example, allowing the (often many) model parameters to vary substantially increases model flexibility, but at the risk of overfitting to observations, particularly if the number of model parameters is large or observational data are insufficient. Overfitting may reduce the model ability in predicting new phenomena, including future climate-driven changes. It is therefore essential that the introduction of variable parameters takes into account such risks and ensures that reasonable assumptions are made to simplify the parameter calibration task. These assumptions would ensure that model calibration is sufficiently constrained, so that there are sufficient observations per each calibrated model parameter value. For example, only a carefully selected subset of parameters may be calibrated, based on their relevance for PP (established, for example, through sensitivity analysis, e.g., Ciavatta et al., 2025) and lack of correlations with other model parameters.

A key consideration when exploring variable parameters is the spatial and temporal scales at which they might vary. For example, it would be important to establish whether seasonal, climatological variability in parameters would be sufficient to capture observed patterns, implying that variability at shorter (sub-seasonal) and longer (inter-annual) time scales could be negligible. If so, this would relax the requirement on the volumes of observational data needed for the calibration, and also on the need to continually update parameter values from day to day or year to year. Hypotheses about temporal variability scales for model parameters can be tested using long time-series of measurements at specific stations, such as the Bermuda Atlantic Time-series Study and the Hawaii Ocean Time-series, both of which present seasonal cycles in photosynthesis parameters (Kovač et al., 2016b, 2018). Another key question is whether parameters vary over fine spatial scales or maintain coherence over large scales such as within ocean biomes or Longhurst provinces (Longhurst, 2007). Preliminary evidence suggests that, at least for the global-scale applications, ecological provinces according to Longhurst might provide an appropriate template for mapping parameters (see Fig. 6B), and that monthly or seasonal time scales might be appropriate for modelling variability in photosynthesis-irradiance parameters (Britten et al., 2025). If province-based approaches emerge as viable candidates, it would be desirable to avoid sharp discontinuities in parameter values at province boundaries, which might require incorporation of smoothing methods to make inter-province changes seamless. Moreover, it is essential that model parameter calibration does not compensate for unrelated spatio-temporally varying model biases, such as those arising from external forcings or other ecosystem model constraints (e.g., boundary conditions). For example, given the importance of underlying physical processes, caution should be applied when calibrating parameters in ecosystem models to avoid models better reproducing the observed PP, but for the wrong reasons. Singh et al. (2025) illustrate that ecosystem parameters in global ocean biogeochemical models are likely calibrated to compensate for biases in their physics (see also Löptien and Dietze, 2019). To avoid mixing different sources of ecosystem model errors, parameters should be ideally estimated jointly with the model state, e.g., using joint parameter-state data assimilation techniques (Schartau et al., 2017). Finally, existing knowledge of acceptable ranges of parameter values needs to be incorporated into the calibration process to prevent parameters from acquiring unrealistic values.

Since parameter spatio-temporal variability results from poorly resolved species types or ecosystem processes, interesting insights into its scale and patterns can be also obtained by comparing models of different complexity. For example, high-complexity ecosystem models (such as the DARWIN ecosystem model) could be used in some cases to deduce the degree of parameter variability of simpler ecosystem models or help inform spatio-temporally varying parameter calibration of those models (always keeping in mind that inter-model consistency does not automatically imply model quality). Comparison studies across models of different complexity would be desirable in this case (for some examples see Friedrichs et al., 2007; Xiao and Friedrichs, 2014). Similarly, emergent properties of ecosystem models can be leveraged to provide specific information for satellite-based models, such as vertical and class-distribution of phytoplankton (e.g., Stock, 2019), or information about nutrient distributions. Such inter-calibrations of models against each other could potentially improve satellite PP products and conversely make the satellite PP data more useful for ecosystem model development. However, one has to be cautious here: model-model intercomparisons and tuning would help models look more like each other, but independent information would be needed to ensure that the simulations are also getting closer to key features in the real world that the models are designed to reproduce.

Even after successfully overcoming the challenges associated with spatio-temporal parameter calibration, significant PP uncertainty is likely to remain in both historical estimates and future projections. For satellite-based models, residual uncertainties could be associated with inherent observational biases, e.g., gaps in data due to cloud cover or adverse viewing geometry, or inaccuracies in satellite products associated with bio-optical conditions in water, or biases inherited from calibration against in situ PP observations with their own inherent uncertainties. For ecosystem models, additional sources of uncertainty include the forcing data and the physical model driving biogeochemical processes, e.g., its vertical and horizontal resolution, and its ability to represent currents and mixing responsible for nutrient supply and export of organic material. For example, differences in how models treat external forcing, such as micro- and macronutrient depositions from the atmosphere, could still contribute to the growth of uncertainties as models become more complex. Furthermore, the spread in the underlying environmental changes such as warming, stratification, changes in irradiation, and ocean circulation among others, contributes significantly to uncertainties in the PP projection.

Further constraints are inherent to ecosystem models themselves. Traditionally, plankton are divided into phototrophic phytoplankton and phagotrophic zooplankton. However, recent research emphasises ubiquitous presence of mixotrophy in the global ocean (Mitra et al., 2023), which not only differs in its physiology and ecological role, but also in its complex interactions with other types of plankton (Flynn and Mitra, 2023). Despite certain commonality in their approach to modelling PP, as discussed above, models differ significantly in their approaches to representing various biogeochemical processes such as grazing and associated fluxes, deposition of organic matter and its remineralisation. For many of those processes (e.g., zooplankton grazing), lack of data, and variability and high uncertainty of available data, are major issues. Focusing on biological ocean carbon storage, Henson et al. (2024) identified key areas where improved understanding of processes is required to support future modelling efforts. For PP, the processes that were ranked highest were: resource limitation for growth, nitrogen fixation, zooplankton processes and phytoplankton loss processes. Current ecosystem models differ considerably in their formulation and parameterisation of these processes, contributing to uncertainties in model outcomes. Moreover, nitrogen-fixation is often not included in these models. Even when these key processes are included, spatial parameter estimation through assimilating observed state variables (such as water column nutrients and oxygen) in ocean biogeochemical models does not necessarily lead to an improved estimate of PP, suggesting that current ecosystem model parameterisations may still be oversimplified compared with the real world (Singh et al., 2025).

The time is right to address the problem of parameter estimation in PP models, both for ecosystem models and satellite-based models. Novel and rapidly expanding observations such as BGC Argo profiles, other types of autonomous data collected by in-water vehicles and also large marine mammals (Chai et al., 2020; Claustre et al., 2020) have been providing large volumes of biological and bio-optical data that complements in situ data from long time series stations, and could be harnessed for this purpose (Fig. 10). Complementary observations from satellite remote sensing, now available over multiple decades and merged into climate-quality, consistent data streams (e.g., Sathyendranath et al., 2019), is another rich data source, along with novel satellite products from emerging capabilities such as geostationary, lidar, cubesat and hyperspectral data. When these are combined with more traditional in situ platforms, including long-term gridded climatology from sources such as World Ocean Atlas (WOA, e.g., Garcia et al., 2019), and potentially complemented by the intercomparison of models with different complexity, there is in several cases already enough data to support a suitably-constrained spatio-temporally varying parameter calibration. This opportunity is further enhanced by new advances in artificial intelligence (AI) and machine learning (ML), giving us an historically unprecedented capability to exploit large and growing datasets to address long-standing questions about marine PP. AI can be used in a variety of different ways, either as a direct prediction approach to optimise model parameterisation and also to emulate models, allowing it to explore a range of model behaviours at reduced computational cost for parameter sensitivity analyses and model calibration (e.g., Mattern et al., 2012; Schartau et al., 2017). Furthermore, recent statistical approaches unique to the ML field enable insights into what the ML model has learned, for example, using Explainable AI, or physically constrained machine learning.

However, crucial to this endeavour would be a clear focus on data quality, and on data validation, following community-wide accepted protocols and reliable uncertainty characterisation. Moreover, some regions, such as sea-ice margins, coastal margins, and high latitudes in winter, which are often regions experiencing long-term rapid changes and include some of the most productive areas of the global ocean, also tend to be regions that are difficult to observe, and hence suffer from sparse data coverage. More observations are needed in such locations to understand the behaviour of model parameters in such regions, including their future changes. Even if constrained spatio-temporally varying calibration is possible in these regions with the available datasets, the importance of further investing in data quantity and quality cannot be overemphasised.

We have argued that, given the growing abundance of observations from diverse platforms, such as satellites and BGC-Argo, combined with rapidly advancing capabilities in ensemble data assimilation techniques and artificial intelligence, the time has now come to address explicitly the importance of parameter assignment in primary production models, and in exploring the spatial and temporal variability in the parameters. We have theoretically justified why such parameter variability is to be expected both in the satellite-based models (where some models already employ variable parameters albeit in a simple fashion) and ecosystem models (where assignment of variable parameters is still quite rare), at least in models are of high complexity. In the case of primary production, the number of phytoplankton classes that are included in the model is a key differentiator of the model complexity. Relatively simpler models, such as the ecosystem models used as part of ESMs in climate projections, have limited capability to resolve phytoplankton communities. For such models, spatio-temporally varying parameters could provide a means to account for the unresolved phytoplankton variability and processes.

Spatio-temporally variable parameter calibration can shed light on the sources of differences between low or medium complexity ecosystem models typically used in ESMs and satellite-based primary-production models. Since variable parameters can capture, in a simple manner, processes or conditions that are not explicitly included in a model, analysing the drivers of parameter variability could help identify how best to overcome current model drawbacks. Furthermore, providing those models with spatio-temporally varying parameters could remove many apparent differences between models, both potentially reducing the spatial and temporal biases in model parameter calibration and enabling the simpler ecosystem models to better represent the effects of unresolved processes or phytoplankton classes. It would also create opportunities for improved intercomparison across models of different complexity, including the ability to understand more about unresolved variability in simpler models by comparing them with the higher-complexity models. One could argue that the spatio-temporally varying parametrisation could help reduce the existing high uncertainty both in historical estimates and future projections of marine PP, provided that independent information is used to avoid all models converging towards a systematically biased outcome. Due to the importance of primary production for climate research, improving its prediction can have a major impact on both climate mitigation and adaptation planning.

In the context of our climate, we need to understand how marine ecosystems in general, and phytoplankton in particular, respond to change. Three types of changes need investigation: changes in (i) phytoplankton biomass (whether they be measured as chlorophyll a, carbon or nitrogen concentration, or all of them); (ii) the rates of biological processes, with marine primary production being a key process in the global carbon cycle; and (iii) community structure. All these objectives are intimately linked to parameter variability, with the third one in particular calling for resolution of parameter variability at the level of major components of the phytoplankton community.

For many decades, we have relied on comparisons and analyses of (both satellite and ecosystem) model outputs with each other, and with in situ data, for insights into model performance, and for identifying the way forward. It is now time to shift the emphasis toward understanding the behaviour of model parameters, across models, across multiple phytoplankton types, and across multiple spatial and temporal scales. This focus has the potential to reduce uncertainties, unify divergent model results, and provide a stronger foundation for predicting marine primary production under changing climatic conditions.