Stratification, defined here as the density difference between the depth of 200 m and the surface, is a useful indicator of change in the upper ocean, as it integrates changes in both temperature and salinity. In Fig. 1a, we present the time series of global mean stratification changes for the historical period and the RCP8.5 projection. All the models project an increase in stratification (ranging from 6 to 30 % by the 2090s). However, the amplitude of stratification differs considerably across the models. The GFDL-ESM2M and MPI models are relatively close to the observed mean stratification in the WOA09 data set (red square, 1.81 kg m-3) for the present era. NorESM1-ME shows the weakest stratification (1.74 kg m-3) while the stratification in HadGEM2-ES is strongest (2.45 kg m-3). Long-term trends are in general agreement across models, but the rate of stratification increase varies, with IPSL-CM5A-MR showing the most rapid increase and NorESM1-ME the slowest increase.

Surface processes that decrease density can largely explain the stratification increase in the RCP8.5 projections. Global mean sea surface temperature (SST) warms by 2.6–3.5 ∘C, accompanied by sea surface salinity (SSS) decreases of 0.05–0.25 psu over the 21st century (Fig. 1). By 2100, the global mean SST ranges from 20.4 ∘C (HadGEM2-ES) to 21.8 ∘C (NorESM1-ME). Model spread decreases in the RCP8.5 projections in response to strong anthropogenic forcing (Fig. 1b). SSS shows a clear declining tendency from 1850 to 2100 (Fig. 1). Compared to the WOA09 observational data, most of the models are too fresh at the surface in the 1990s, especially the HadGEM2-ES, which has the lowest global mean SSS. The model spread is partly due to internal variability simulated by the climate models. Model differences in physics, but also in spin-up procedures, the way RCP scenarios are set up and model climate sensitivities all likely contribute to the model spread (Knutti and Hegerl, 2008; Szopa et al., 2013).

Vertical density profiles help to further explain the changes in stratification. Mean vertical profiles of density in the 1990s and the density change between the 1990s and the 2090s show that all the models become more buoyant at the surface as a consequence of heating and/or freshening of the upper ocean (Fig. S1 in the Supplement). The density changes at the surface vary by almost a factor of 2 among models, from -1.1 kg m-3 (HadGEM-ES2) to -0.6 kg m-3 (GFDL-ES2M), but converge to a relatively narrow range (approximately -0.2 kg m-3) at 500 m depth. Most of the density change occurs between the surface and 200 m. Below 200 m, the density change in most of the models varies linearly with depth. Thus, our definition of the stratification index, as the density difference between the surface and 200 m, is reasonable. The converging reductions in density among models at about 500 m agrees with some previous studies based on observations and CMIP3 models (Bindoff et al., 2007; Lyman et al., 2010; Capotondi et al., 2012). Compared to WOA09 data, the models generally underestimate the density of the upper ocean in the top 150 m and most models overestimate the density below 350 m (resulting in a positive stratification bias) (Fig. S1a).

Vertical profiles of temperature and salinity from each model are also shown in Fig. S1. The surface ocean exhibits strong warming of 1.7–3.5 ∘C by the 2090s and the warming magnitude declines quickly with depth, which is associated with the heat uptake capacity of individual models. For instance, the GFDL models seem to be more efficient in transporting heat downward than the IPSL models. Above 300 m, the temperature changes vary widely among the models. Temperature changes as a function of depth are complex, and model-to-model differences may be related to a number of factors including rates of vertical mixing and the seasonal thermocline dynamics. At the depth of 500 m, the mean temperature change converges at about 1.2 ∘C. The ocean heat uptake capacity is linked to ocean diapycnal mixing and other processes, such as mixing by mesoscale eddies. The weak temperature gradients in the GFDL models suggest high rate of heat uptake, and are consistent with ocean heat uptake estimates by Kuhlbrodt and Gregory (2012). The large model spread in temperature profiles suggests considerable differences and uncertainties in the parameterizations of these physical processes across the models. Vertical profiles of salinity are more scattered than for temperature (Fig. S1c). In the 1990s, most of the models underestimate salinity from the surface down to 550 m. Surface salinity is generally biased low by 0.05–0.25 psu. Most of the freshening with climate change takes place above 100 m, which also acts to increase stratification. Note that the salinity increases at 100–300 m in some models (IPSL, GFDL-ESM2M, HadGEM2-ES) partially compensates the impact of rising temperatures on density.

The percentage contribution of temperature change to the stratification change from the 1990s to the 2090s is shown in Fig. S2. Previous studies have shown that salinity contributes significantly to the stratification changes at high latitudes (> 40∘) in both hemispheres and in the North Pacific as a consequence of increases in precipitation (Bindoff et al., 2007). From our comparisons, temperature dominates the stratification changes in the tropical and subtropical regions (Fig. S2). Salinity dominates the stratification changes in the much of the Arctic Ocean and in the high-latitude North Atlantic. While stratification is a function of SSS and SST to a good approximation (Cabré et al., 2015), stratification change at high latitudes is also dependent on temperature and salinity at depth as vertical mixing and exchange are stronger.

In some regions, the spatial distributions and the driving process differs substantially across models. Generally, the models agree well in the tropics and in the subtropical gyres that surface warming drives the increase of stratification. In the high-latitude North Atlantic, the subpolar Pacific and the western Pacific Ocean, there is weaker agreement across the models. In the subtropical gyre of the South Pacific, stratification changes in the IPSL and CESM1(BGC) models have a stronger influence from temperature change, while the other models exhibit more complicated spatial patterns. In the North Atlantic, salinity contributes more in the IPSL and HadGEM2 models than in the other models. The southeastern Pacific is more dominated by salinity in the two GFDL models. In the Southern Ocean, the models show relatively large contributions from both salinity and temperature but with complicated spatial patterns that differ considerably across models. Projections for the regions where the models do not agree even on the driving factor should be viewed with more caution. Climate change and biogeochemical impacts in these regions tend to be projected with less consistency across models (Bopp et al., 2013; Cabré et al., 2015).

Stratification increases globally in all the models with climate change (Figs. 1 and 2). Nearly all the models predict large increases in stratification in the western tropical Pacific, the tropical Indian Ocean, the Arctic Ocean and in the high-latitude North Atlantic (particularly in the Labrador Sea). The Southern Ocean has weaker increases in stratification, partly because the surface layer mixing and upwelling are intensified due to the poleward shift of strengthened westerly jets (Swart and Fyfe, 2012). Our stratification index may underestimate the changes in the high-latitude North Atlantic, as the relatively deep mixing means that temperature and salinity at 200 m depth are changing much more rapidly than in other regions. Reductions in the deep winter mixing and NADW formation in this region are a common pattern seen in strong warming climate simulations (e.g., Cheng et al., 2013; Schwinger et al., 2014). Less drastic increases in stratification are seen over much of the rest of the oceans, with only a few small regions showing decreases in some models. An exception is the HadGEM2-ES model, which has large stratification reductions in the Arctic (Fig. 2).

One of the key factors determining global NPP is nutrient availability in the euphotic zone. Time series of global mean nutrient (0–100 m) concentrations for nitrate (NO3), phosphate (PO4), silicic acid (SiO4) and dissolved iron (dFe) are presented in Fig. 3. The magnitude of surface nutrient concentrations differs substantially across the models (varying by a factor of ∼ 1.5–2, and by a factor of 5 for dissolved iron). The IPSL models have relatively low surface nutrient concentrations. Compared to the WOA09, two models overestimate phosphate (CESM1(BGC) and GFDL-ESM2G) and five models overestimate nitrate. All of the models overestimate the silicic acid observations, with the exception of CESM1(BGC). The CESM1(BGC) model overestimates surface phosphate concentrations initially, due to excessive nitrogen limitation, but then shows the strongest surface phosphate declines over the 21st century (Fig. 4).

Over the entire period from 1850–2100, the models all display decreasing trends for surface nitrate, phosphate and silicic acid. Interestingly, surface iron concentrations increase modestly in all but one of the models by 4–10 %. Changes in iron concentrations impact marine productivity, nitrogen fixation rates and oceanic net CO2 uptake. In the CMIP5 simulations, iron inputs to the oceans from deposition and rivers are held constant over time, so the increasing surface iron concentrations may reflect increasing macronutrient limitation of phytoplankton growth, leading to reduced biological uptake of iron. The reductions in the sinking export flux also reduce the particle scavenging loss term for dissolved iron. In the CESM1(BGC) model, increased production in the high-nutrient, low-chlorophyll (HNLC) regions offset about 25 % of the reduction observed in the macronutrient-limited areas with climate change, and changing circulation patterns also altered the lateral transport of iron within the oceans (Moore et al., 2013; Misumi et al., 2014).

The relative changes in nutrient concentrations (0–100 m) (normalized to 1990s means) are presented in Fig. 4. The relative changes in the historical run show a consistent pattern across the models for nitrate, phosphate and dissolved iron (except for HadGEM2-ES). In the RCP8.5 projection, the models show diverging estimates of the magnitude of the relative changes. For nitrate, the reductions range between 3 and 14 %, whereas for phosphate the reductions range between 3 and 20 %. Silicic acid and iron trends are even more variable than for nitrate and phosphate. For silicic acid, three models exhibit slight increases, while the others exhibit decreases ranging from 5–17 %. The variability in relative change in silicic acid concentration in the RCP8.5 is likely associated with changes in plankton community and variable diatom production (Bopp et al., 2005). All of the models include some representation of diatoms (Table 2) but the match to observed silicic acid concentrations for the current era is generally poor (Fig. 3).

The spatial distributions of mean nitrate concentration for 0–100 m in the 1990s are shown in Fig. S3. The CMIP5 models reproduce key observed features of the basin-scale distributions of surface nitrate. For example, all of the models exhibit elevated nitrate concentrations in the eastern equatorial Pacific, Southern Ocean, subarctic North Atlantic and subarctic Pacific. In the subtropical gyres of the Atlantic and Pacific basins, mean nitrate concentrations are low. These general patterns are consistent with the WOA09 observations. However, there are clear disagreements in some regions. For example, the details of the high-nitrate surface water distributions vary considerably in the eastern equatorial Pacific. The HNLC condition extends too far north and south of the equator in some models, and too far to the west in others (Fig. S3). The models also differ in the intensity and extent of high nitrate concentration waters in the subarctic North Pacific, where six of nine models show lower nitrate concentrations than the WOA09 data (MPI-ESM-LR, MPI-ESM-MR and HadGEM2-ES are closest to the observations). There are also differences in the Arabian Sea and Bay of Bengal, where most models underestimate nitrate concentrations except the GFDL-ESM2M and MPI-ESM-LR models.

Intermodel spread in NPP during the 1990s is pronounced, with NPP as low as 29 PgC yr-1 (IPSL-CM5A-LR and IPSL-CM5A-MR), while NPP in one model exceeds 75 PgC yr-1 (GFDL-ESM2M) (Table 3, Fig. 5). In addition, the spatial pattern of NPP is not well represented by the multimodel mean (Bopp et al., 2013). Satellite-based estimation of NPP is approximately 50 PgC yr-1 (Behrenfeld et al., 2006; Carr et al., 2006). The MPI-ES-MR and CESM1(BGC) models had NPP of 49.8 and 54.2 PgC yr-1, closer to the satellite-based estimates and the observationally constrained, model estimate of 56 PgC yr-1 by Buitenhuis et al. (2013). The magnitude of EP also varies substantially across models in the 1990s, ranging from 4.4 to 7.2 PgC yr-1 (Table 3).

All of the models exhibit decreasing trends in global NPP and EP with climate change, as shown in previous studies (Bopp et al., 2013; Dutkiewicz et al., 2013) and most models show more rapid decreases during the middle to latter part of the 21st century (Figs. 5–6, Table 3). All nine models project decreases in export production by the 2090s, exceeding 5 % relative to levels in the 1990s, whereas the response for NPP is divided into two groups after 2020. The CESM1(BGC) and GFDL models experience smaller changes in NPP (< 5 % relative to 1990s) while other models have larger decreases (8–16 %). The largest relative change for NPP is about -16 % (MPI-ESM-LR). The EP decreases range from 7 % for GFDL-ESM2G to 28 % for IPSL-CM5A-LR. Cabré et al. (2015) found reductions in NPP and EP for all biomes, except at the highest latitudes. The reductions in global NPP and EP covary with the increases in stratification (Fig. 6). By the 2090s, stratification increases by about 16 % in GFDL-ESM2M and up to 33 % in HadGEM1-ES. The rate of stratification increase is slower in the two GFDL models and CESM1(BGC), which also agrees with the slower rates of relative NPP and EP change.

The variability across models in NPP is substantially larger than that seen in EP (Table 3). The normalized standard deviation was ±27 % for NPP, but only ±12 % for EP in the 1990s. The large spread in simulated NPP and its response to climate change was also noted by (Laufkötter et al., 2015). Seven of the nine models have an EP between 6.0 and 7.2 PgC yr-1 in the 1990s, with the remaining two (HadGEM2-ES and GFDL-ESM2G) having considerably lower EP (< 5 PgC yr-1). EP is tightly coupled to new nutrient inputs to the euphotic zone in these models. NPP is less tightly coupled as the fraction of regenerated production varies across the models, and can vary spatially and temporally within some models. Thus, the large spread in NPP is not just a function of the different physical models and their transport of nutrients to the euphotic zone, but rather it is also strongly impacted by the phytoplankton community structure and export efficiency inherent in the models, and the resulting varying levels of regenerated production.

The sinking carbon flux out of the euphotic zone to net primary production ratio (particle export ratio or pe-ratio) is a measure of the export efficiency and also reflects the variable contribution of regenerated production to total NPP (Dugdale and Goering, 1967; Eppley and Peterson, 1979; Dunne et al., 2007). High pe-ratio values are typically associated with productive ecosystems dominated by larger phytoplankton (often diatoms; Buesseler, 1998; Boyd and Newton, 1999), while low pe-ratios are associated with oligotrophic food webs with most carbon flow through the microbial loop (Pomeroy, 1974; Azam et al., 1983). The CMIP5 models that include both large and small phytoplankton, assume a higher export efficiency for the large phytoplankton (Aumont and Bopp, 2006; Seferian et al., 2013; Tjiputra et al., 2013). The fraction of grazed material routed to sinking export is higher, often by a factor of 3 or more, than the fraction routed to sinking export for the small phytoplankton. Diatoms are also likely to dominate phytoplankton blooms in these models. This can drive additional, very efficient, export through aggregation, further enhancing the differences in export efficiency between large and small phytoplankton. Relative to the 1990s, six of the nine models show decreasing trends in the pe-ratio (up to a 10 % reduction) (Figs. 5–6, Table 3; see also Cabré et al., 2015). Diatoms account for a smaller percentage of NPP in the 2090s than in the 1990s in all the models, except for the MPI model, where nearly all of the production is by diatoms and the smallest phytoplankton are not explicitly represented (Table 3).

We quantify the relations between stratification and key biogeochemical variables with annual model output over the entire time period of 1850–2100. This approach captures physical–biological interactions over interannual to centennial timescales, and is thus more robust than the earlier work comparing two end points from the beginning and end of the 21st century (Bopp et al., 2013; Cabré et al., 2015). Changes in global NPP relative to the 1990s are plotted against the relative change in stratification in Fig. 7a. Across all the ESMs, a relatively good relationship exists with a correlation r2= 0.72. Larger relative increases in stratification correspond to larger relative declines in NPP. In addition, the globally fitted line with a slope of 0.38 separates the models into two groups. In one group (GFDL, IPSL and CESM1(BGC)), the NPP reductions are more modest as stratification increases; the other group is composed of the two MPI models, HadGEM1-ES and the NorESM model, which show more intense and linear reductions in NPP with increasing stratification. The reduction of NPP can be partly explained by nutrient changes responding to stratification increases. Across the models, surface nitrate and phosphate concentrations clearly decline as the stratification is enhanced (Fig. 7c and d, with r2 of 0.80 and 0.82, respectively). Note that all of these trends are robust across the full time series. Compared to the 1990s, the preindustrial stratification is weaker, surface nutrient concentrations are higher, and NPP and EP are elevated (Figs. 3–7).

This indicates significant climate change impacts on ocean biogeochemistry, prior to the 1990s and the modern era when most ocean observations have been made. The response of surface silicic acid to increasing stratification is more variable. The projected changes are more divided, as three models (MPI-ESM-LR, MPI-ESM-MR and HadGEM1-ES) show slight increases and the others show reductions in surface silicic acid concentrations (Fig. 7b).

EP is even more closely related to the stratification changes (r2= 0.89) than NPP (r2= 0.72; Fig. 7e). The EP change is also closely related to the NPP change (r2= 0.85). EP decreases by up to 20 % (Fig. 7e), whereas NPP decreases by 10–18 %. The models display two patterns in terms of the response of NPP and EP to climate change. The first group includes five models (the two IPSL models, CESM1(BGC) and the two GFDL models) where the relative declines in NPP are smaller than the relative declines in EP by a factor of 2 or more (Fig. 6 and Table 3). In this group, the EP drops by about 10 % and the NPP decreases by 5 %. In the remaining models, the relative declines in EP and NPP are larger and more similar in magnitude. For example, both EP and NPP decrease by about 14 % in the HadGEM2-ES model. The differential declines in NPP and EP in the first group of models documents declining export efficiency for the ocean biological pump, driven by phytoplankton community shifts and a decreased contribution to NPP by large phytoplankton (diatoms) (see below and Figs. 6–10; also Cabré et al., 2015).

Reduced nutrient availability is a major contributor to the declines in NPP and EP. However, the relationship varies from one model to another because growth and export are complicated functions of macronutrient limitation, temperature, irradiance and iron limitation, as well as the routing of organic matter within the ecosystem that drives export efficiency. Higher metabolic rates with warming can be compensated to a large degree by changes in the supply of nutrients in terms of globally integrated productivity (Dutkiewicz et al., 2013). The NPP response is also strongly impacted by phytoplankton community structure, which modifies export efficiency, and the corresponding magnitude of regenerated primary production. For the IPSL, CESM1(BGC) and GFDL models that show larger declines in EP than in NPP, this pattern is driven by a decreasing contribution to total NPP by large phytoplankton (Table 3, Figs. 8–9). Most of the primary production in these models is by smaller phytoplankton. The GFDL models express this pattern most strongly, with minimal declines in NPP, despite declines in EP approaching 10 % (Fig. 6 and Table 3). The other models tend to have production that is dominated by diatoms, and do not capture the community shifts towards increasing small phytoplankton dominance (and reduced export efficiency) under increasing nutrient stress. The declines in NPP with increasing stratification are more linear and more similar in magnitude to the declines in EP (Fig. 7a, b and h). Thus, there are also very strong correlations between the climate-driven changes in the fractional contribution of diatoms to NPP and both the changes in stratification and the changes in EP (Fig. 7f and g, correlations of r2= 0.85 and r2= 0.95, both much higher than the correlation between changing stratification and NPP, r2= 0.71). Cabré et al. (2015) found similar patterns relating community composition, NPP and EP comparing the period of 1980–1999 with 2080–2099, across low to midlatitude biomes.

Some of these patterns are illustrated in Fig. 8, which shows the contribution of diatoms (large phytoplankton) to NPP for the 1990s. Most of the models show elevated diatom production at high latitudes and low diatom contributions in the subtropical gyres. However, there are large discrepancies in the magnitude of the diatom contribution, ranging from about 30 % to more than 90 % in the Arctic Ocean, for example. At the global scale, diatoms account for only 9.4 % of NPP in the GFDL-ESM2M model and reach a maximum of 91 % in the MPI-ESM-MR model (Table 3). The large variability across the models reflects, in part, the lack of observational data sets to constrain phytoplankton community composition at the time these models were being developed. The new globally gridded ocean atlas of plankton functional types, Marine Ecosystem Data (MAREDAT) (Buitenhuis et al., 2013) has started to fill this gap, and should lead to improved representations of plankton community structure in the future as the data set becomes increasingly populated and is entrained into model development and validation. Remote sensing estimates of phytoplankton community composition and size class structures are also providing useful constraints for global-scale modeling efforts (e.g., Alvain et al., 2005; Hirata et al., 2008; Kostadinov et al., 2009; Siegel et al., 2014).

The spatial patterns of the shifts in phytoplankton community composition with climate change are illustrated in Figure 9; which shows the change in the percentage of NPP by diatoms (2090s–1990s). There are some robust trends across the models. One of the areas with the biggest declines in diatom production is the high-latitude North Atlantic. This region typically has some of the biggest stratification increases with climate change, greatly reducing the deep winter mixing that entrains nutrients to the surface (Moore et al., 2013; Randerson et al., 2015). Nearly all the models also show large declines in diatom contributions to production in the Arctic Ocean. The CMIP5 models show consistent trends of increasing stratification, declining surface nutrient concentrations and a longer growing season with climate change in the Arctic (Vancoppenolle et al., 2013). Increasing surface temperatures and substantial declines in sea ice cover allow for a longer growing season with climate change. Thus, nutrients in surface waters are more completely used up by summer's end, leading to community shifts with decreased diatom production and an increased fraction of production by smaller phytoplankton. In the CESM1(BGC) model, this community shift allows for a small increase in central Arctic NPP, even as export production and surface nutrient concentrations decline, due to the increased fraction of NPP from small phytoplankton and the resulting increases in regenerated production (Moore et al., 2013).

All of the models show some increase in the fraction of NPP by diatoms in the Southern Ocean (Fig. 9). The increase is particularly strong in the CESM1(BGC), IPSL and GFDL models. Most of the models also show some increased diatom production in the tropical Pacific. Bopp et al. (2005) reported decreasing diatom production in the Arctic and high-latitude North Atlantic, with some increases in the Southern Ocean under a strong warming climate scenario. Steinacher et al. (2010) also found declining productivity in the North Atlantic and shifts in the export ratio due to phytoplankton community shifts with decreasing diatom production. An earlier version of the CESM used in that study (CCSM3) showed only small shifts in export ratios with climate change, as the range in export ratios and the differences in export efficiencies between large and small phytoplankton were smaller than in CESM1(BGC) (Steinacher et al., 2010; Moore et al., 2013). Three models in this study (HadGEM2-ES and the two MPI models) show increases in diatom production in the low latitudes (Fig. 9). However, the diatoms dominate production nearly everywhere in these three models (Fig. 8).

There are also large intermodel differences in the spatial patterns of the pe-ratio (Fig. 10). Some of the models (GFDL, IPSL, CESM1(BGC)) show a close correlation between the pe-ratio and diatom production (compare Figs. 8 and 10), due to the enhanced export efficiency for diatoms (large phytoplankton) built into the models. Thus, there is a very high correlation between the changing contribution of diatoms to NPP and the changes in EP (Fig. 7g, Table 3). The MPI model includes one phytoplankton group and has an essentially constant pe-ratio of 0.15, explaining the linearity of the changes in NPP and EP with warming (Figs. 8 and 10). Production in the HadGEM1-ES model is dominated nearly everywhere by the diatoms (Fig. 8). Therefore, the MPI and HadGEM models cannot capture a shift towards increasing small phytoplankton dominance under declining surface nutrient concentrations. This leads to export production being closely correlated with diatom production in these models as most production is by diatoms, as well as in the other models where diatoms are assumed to export more efficiently but account for a smaller fraction of total NPP (Table 3).

There is also a strong correlation between the declines in the fraction of NPP by diatoms and declines in the pe-ratio (compare Figs. 7, 9 and 11). The largest declines in the pe-ratio are seen in the Arctic and the high-latitude North Atlantic, regions where diatom production also decreases the most. The GFDL, IPSL and CESM1(BGC) models show some reductions in pe-ratio in the subarctic North Pacific, but the spatial patterns are inconsistent (Fig. 11). The models display considerable variability in the degree of stratification increase and in the dominant factor driving these changes in the subarctic North Pacific (Figs. S2 and 2).

The correlation for the relationship between the changing percentage of NPP by diatoms vs. the changes in EP across all the models has an r2 of 0.96 and a slope close to 1 (0.94, Fig. 7g), indicating that phytoplankton community structure plays a dominant role in determining the responses of NPP, EP and the pe-ratio to climate change. The biggest declines in the fraction of production by diatoms and pe-ratios are in precisely the areas where some of the largest increases in upper ocean stratification are seen, along with declining surface nutrient concentrations, as in the Arctic Ocean and in the high-latitude North Atlantic (Figs. 6–8; see also Steinacher et al., 2010; Moore et al., 2013).

At the global scale, the CMIP5 models show considerable stratification biases for the 1990s when compared to the WOA09 data (Fig. 1, Table 3). Only the GFDL-ESM2M model is within 10 % of the observed value. From the density profiles as well (Fig. S1), it is apparent that most of the models have stronger stratification in the 1990s than seen in the observations. Liu et al. (2014) argue that climate bias is important when projecting the impact of climate change on land surface processes and Hoffman et al. (2014) document this for atmospheric CO2 mole fractions. Here, we examine how stratification biases in the 1990s may affect model projections of NPP and EP in the 2090s.

Models with stronger bias in the 1990s for surface stratification tend to predict larger climate-induced declines in both NPP and EP (Fig. 12, r2= 0.47 and r2= 0.54, respectively). The slopes are plotted when the correlation is significant at > 95 % level. Five of the models have positive biases in stratification for the current era that exceed 20 %. These models also show the largest relative increases in stratification with climate change of 26–30 % (Fig. 12, Table 3). The remaining four models (GFDL models, CESM1(BGC) and NorESM1-ME) do a better job of simulating observed stratification for the current era and predict relative increases in stratification over the 21st century that are roughly half as large, ranging ∼ 15–18 %. This suggests that the more biased models (for the 1990s) may be overestimating the projected reductions in NPP and EP for the end of the century.