Our original intention was to conduct a classical meta-analysis, which would have yielded the benefit of a quantitative measure of diatom responses to ocean acidification (OA), expressed as an overall effect size (i.e. combined magnitude) such as the response ratio. However, our literature analysis revealed a large variability in experimental pCO2 ranges as well as measured response variables, which cannot be directly compared among each other (e.g. microscopic cell counts, pigment concentrations, genetic tools). These limitations impede data aggregation as required for a classical meta-analysis. Furthermore, experimental setups differed widely in terms of other environmental factors such as temperature, light, and nutrient concentrations, all of which are known to modulate potential responses to pCO2 (Boyd et al., 2018), thereby further complicating data aggregation for meta-analysis. Therefore, we chose an alternative semi-quantitative approach where diatom responses to increasing CO2 are grouped in categories (see Sect. 2.2) and also allows us to account for differences in experimental setups, e.g. with respect to container volume (see Sect. 2.3). While this approach excludes the determination of effect size, it provides an unbiased insight into the direction of change of potential CO2 effects.

Before going into the details of data compilation we want to emphasize once more that the motivation for this study was not to investigate the physiological response of diatoms to OA. Such meta-analyses or reviews have already been made (Dutkiewicz et al., 2015; Gao and Campbell, 2014). Instead, our goal was to summarize how diatoms respond to OA in their natural habitat. More generally, experiments with ecological communities (as compiled in our study) do not so much aim for a mechanistic understanding of a certain process (e.g. in physiological experiments) but rather assess the general sensitivity of more natural communities to environmental drivers. Therefore, it is important to have a realistic setup because the net response of any player in the food web is composed of a direct physiological response to CO2 and by CO2-induced alterations of interactions with other species. From that point of view it is desirable to include all important ecosystem components because when trophic cascades are represented incompletely then the observed response in an experiment may not reflect the response that would occur in nature, which is what we are ultimately interested in (Carpenter, 1996). Clearly, investigating OA effects on diatoms or any other group in complex communities has the disadvantage that the actual cause for an observed response can hardly ever be determined with high certainty (Bach et al., 2017, 2019). However, experiments compiled herein investigated the development of initially similar plankton communities over time with the only difference being carbonate chemistry conditions between control and the treatments. Thus, we can at least be sure that the differences in diatom abundance or community composition between control and treatment (which is the focus of our study) is caused by simulated OA, even though the underlying mechanisms cannot be pinned down with certainty.

We explored the response of diatom assemblages to high CO2 (low pH) by searching the literature for relevant results with Google Scholar (15 December 2017) using the following search query: “diatom” OR “Bacillariophyceae” AND “ocean acidification” OR “high CO2” or “carbon dioxide” OR “elevated CO2” OR “elevated carbon dioxide” OR “low pH” OR “decreased pH”. The first 200 results were inspected and considered to be relevant when they were published in peer-reviewed journals contained a description of the relevant methodological details, a statistical analysis or at least a transparent description of variance and uncertainties, and tested CO2 effects on natural plankton assemblages (artificially composed communities were not considered). We then carefully checked the cited literature in these relevant studies to uncover other studies that were missed by the initial search. Furthermore, we checked the “Ocean acidification news stream” provided by the “Ocean Acidification International Coordination Centre” under the tag “phytoplankton” (https://news-oceanacidification-icc.org/tag/phytoplankton/, last access: 16 January 2019) for relevant updates since December 2017.

There were two response variables of interest for the literature compilation:

The response of the “bulk diatom community” to high CO2. For this we checked if the abundance of diatoms, the biomass of diatoms, or the relative portion of diatoms within the overall phytoplankton assemblage increased or decreased under high CO2 relative to the control. We distinguished between “positive”, “negative”, and “no effect” following the statistical results provided in the individual references. When the CO2 effect on the bulk community was derived from abundance data, we also checked if there were indications for a concomitant shift in the biomass distribution among species. This is relevant because, for example, an increase in bulk abundance could coincide with a decrease in bulk biomass when the species driving the abundances is smaller. We found no indications for conflicting cases but acknowledge that not every reference provided sufficient data on morphological details to fully exclude this scenario. Furthermore, we emphasize that CO2 can also shift the temporal occurrence of a diatom response (Bach et al., 2017). For example, a diatom bloom could occur earlier in a high CO2 treatment than in the control but with a similar bloom amplitude (Donahue et al., 2019). In this case we assigned a “positive” response because an earlier bloom occurrence mirrors a higher net growth rate under elevated CO2.

The most realistic OA experiment would be one where all aspects of the natural habitat are represented correctly. Such setups are possible for benthic communities which can be sampled in situ along a natural CO2 gradient at volcanic CO2 seeps (Fabricius et al., 2011; Hall-Spencer et al., 2008; Johnson et al., 2011). However, pelagic communities are advected with currents so that it is very difficult to simulate OA in open waters. Thus, OA experiments where pelagic communities are exposed to increasing levels of CO2 were so far always performed in closed containers even though it is well known that confinement causes experimental artefacts (Calvo-Díaz et al., 2011; Ferguson et al., 1984; Guangao, 1990; Menzel and Case, 1977). The degree by which confinement causes experimental artefacts will differ from study to study, depending on factors such as the incubation volume, the length of incubation, or the selective removal of certain size classes from the incubation (Carpenter, 1996; Duarte et al., 1997; Nogueira et al., 2014). In our literature synthesis we had to deal with a large variety of experimental setups and there are very likely differences in how well a given setup represents the natural environment. Therefore, we aimed to develop a metric that allows us to estimate “how well the natural system (which we are ultimately interested in) is represented by the experimental setup”. This metric – termed the “relative degree of realism (RDR)” – was used to balance the influence of individual studies on the final outcomes of the literature analysis. Most certainly, we do not mean to devalue any studies but think that the highly different scales of experiments, ranging from 0.8 L lab incubations to 75 m3 in situ mesocosms, should not be ignored when evaluating the literature. In the following we will first derive the equation for the RDR and introduce the underlying assumptions. Afterwards we describe aspects that were considered while conceptualizing the RDR.

The incubation volume in the studies considered herein ranged from bottle experiments to in situ mesocosm studies with considerably larger incubation volumes. Smaller differences in incubation volumes (e.g. 0.5 vs. 2 L) were shown to have no, or a minor, influence on physiological rates (Fogg and Calvario-Martinez, 1989; Hammes et al., 2010; Nogueira et al., 2014; Robinson and Williams, 2005). However, they can influence food web composition (Calvo-Díaz et al., 2011; Spencer and Warren, 1996), e.g. by unrepresentatively including certain organism groups such as highly motile mesozooplankton. Larger differences in incubation volumes (e.g. 10 vs. 10 000 L) are considered to have a major influence on the enclosed communities, with the larger volume generally being more representative of natural processes (Carpenter, 1996; Duarte et al., 1997; Sarnelle, 1997). Therefore, our first assumption to conceptualize the RDR was that larger incubation volumes represent nature generally better than smaller ones.

Plankton communities were pre-filtered in many experiments to exclude larger and often patchily distributed organisms (e.g. copepods). This is a valid procedure to reduce noise and to increase the likelihood to detect CO2 effects but it also influences the development of plankton communities since the selective removal of certain size classes can modify trophic cascades within the food web (Ferguson et al., 1984; Nogueira et al., 2014). For example, Nogueira et al. (2014) compared plankton successions of pre-filtered (100 µm) and unfiltered communities and found that the removal of larger grazers and diatoms gave room for green algae and picophytoplankton to grow. Such manipulations make the experiment less representative for a natural food web, which brought us to the second assumption for the RDR: the smaller the mesh size during the pre-filtration treatment, the less complete and thus the less realistic is the pelagic food web.

To parameterize the two abovementioned assumptions, we first converted the volume information provided in each experiment into a volume-to-surface ratio (V/S). The underlying thought is that V increases with the third power to the surface area of the incubator and is indicative of the relation of open space to hard surfaces. Therefore, we first converted V into a radius (r) assuming spherical shape: 1r=34Vπ3. The surface (S) of the spherical volume was calculated as 2S=4πr2. The assumption of spherical shape was necessary because it allowed us to calculate V/S from only knowing V, which is usually the only parameter provided with respect to container characteristics. We are aware that this is a simplification because the majority of containers used in experiments will likely have had cylindrical shape. However, the conversion from volume to surface assuming cylindrical shape would have required knowledge of two dimensions (radius and height of the cylinder). Although shape can influence processes within the container (Pan et al., 2015), it is a less important factor to consider in our study because sensitivity calculations assuming reasonable cylinder dimensions showed that the V/S differences due to container shape will be small compared to the V/S differences due to the range of container volumes compared here.

The influence of pre-filtration treatments on the investigated plankton community is implemented by multiplying the V/S with the cube root of the applied mesh size (dmesh in microns,µm) so that the RDR is defined as 3RDR=VSdmesh3.

Thus, as for V/S, the influence of dmesh on RDR does not increase linearly but becomes less influential with increasing dmesh. The rationale for the non-linear increase is that incubations will still have an increasing bias even if they do not have any pre-filtration treatment due to generally increasing organism motility with size. For example, when collecting a plankton community with a Niskin bottle, more motile organisms can escape from the approaching sampler so that the food web composure is still affected even without subsequent pre-filtration. For this reason, we also capped the maximum dmesh to 10 000 µm when there was no pre-filtration treatment applied since none of the studies included significantly larger organisms. The rationale for calculating the cube root of dmesh was that in this case the influence of V/S and dmesh on RDR becomes roughly similar. Figure 1 illustrates the change of RDR as a function of V and dmesh. High RDRs are calculated for large-scale in situ mesocosm studies (∼50–190), while bottle experiments yield RDRs between ∼1 and 12.

The key pre-requisite for an experimental parameter to be included in the RDR equation (Eq. 3) was that it is reported in all studies. Many parameters that we would have liked to use for the RDR are either insufficiently reported (e.g. the light environment) or not provided quantitatively at all (e.g. turbulence). We therefore had to work with very basic properties related to the experimental setup rather than to the experimental conditions.

A particularly critical aspect of the RDR we had to deal with was the duration of the experiments (Time). Time is reliably reported in all studies and therefore principally suitable for the RDR. Our first thoughts were that a realistic community experiment should be long enough to cover relevant ecological processes such as competitive exclusion and therefore also parameterized Time in the first versions of the RDR equation. However, we decided to not account for it in the final version because the factors that define the optimal duration of an experiment are poorly constrained. For example, a 1 d experiment in a 10 L container could indeed miss important CO2 effects caused by food web interactions. On the other hand, a 30 d experiment in the same container could reveal such indirect effects but at the same time be associated with profound bottle effects and make the study unrepresentative for simulated natural habitat. Thus, too long and too short times are both problematic and the optimum is hard to find. One such attempt to find the optimum Time was made by Duarte et al. (1997), who analysed the plankton ecology literature between 1990 and 1995. By correlating the experimental duration with the incubation volume of published experiments they provided an optimal length for any given volume. However, as noted by Duarte et al. (1997), their correlation is based on publication success and therefore rather reflects common practice in plankton ecology experiments and not necessarily a mechanistic understanding of bottle effects. Thus, as there is no solid ground for a parameterization of Time we ultimately decided to not consider it for the RDR.

Finally, we want to point out (and explicitly acknowledge) that the RDR approach to balance the influence of studies on the final outcome of the literature analysis is of course not the one perfect solution and most likely incomplete (see above). However, balancing a literature analysis with the RDR score may still be an improvement relative to the other case where each experiment is treated exactly equally despite huge differences in the experimental setup. Nevertheless, to account for both views (i.e. the RDR is useless vs. the RDR is useful) we will present the outcome of our literature analysis in two different ways throughout the paper: (1) by simply counting the number of outcomes (N) and adding them to yield a cumulative ∑N score (N-based approach; left columns in Figs. 3 and 4) or (2) by adding the RDR score of the experiments with a certain outcome to yield a cumulative ∑RDR score (RDR-based approach; right columns in Figs. 3 and 4).

Most studies that found effects of pCO2 on diatom communities related these changes to CO2 fertilization of photosynthesis. Concentrations of CO2 in the surface ocean are relatively low compared to other forms of inorganic carbon, especially bicarbonate ion (HCO3-) (Zeebe and Wolf-Gladrow, 2001). However, RuBisCO, the primary carboxylating enzyme used in photosynthesis, is restricted to CO2 for carbon fixation and has a relatively low affinity for CO2 compared to O2 (Falkowski and Raven, 2007). Therefore, diatoms (like many other phytoplankton species) operate a carbon concentrating mechanism (CCM) to enhance their CO2 concentration at the site of fixation relative to external concentrations (e.g. by converting HCO3- to CO2) and thereby establish higher rates of carbon fixation than what would be possible when only depending on diffusive CO2 uptake (Giordano et al., 2005). It is well known that the proportion of CO2 uptake vs. HCO3- uptake for photosynthesis varies largely among diatoms (Burkhardt et al., 2001; Rost et al., 2003; Trimborn et al., 2008) and is theoretically also a function of cell size (Flynn et al., 2012; Wolf-Gladrow and Riebesell, 1997). Accordingly, increasing seawater pCO2 may increase the proportion of diffusive carbon uptake and/or lower the energy and resource requirements for CCM operation (Raven et al., 2011). From a physiological point of view, these mechanisms could allow for increased rates of photosynthesis and cell division.

So how do these theoretical considerations align with (a) the variable and species-specific physiological responses of diatoms to increasing CO2 (Dutkiewicz et al., 2015) and (b) the results from community-level experiments compiled in this study? Regarding the variability of physiological responses, progress has recently been made by Wu et al. (2014), who experimentally demonstrated a positive relationship between cell volume and the magnitude of the CO2 fertilization effect on diatom growth rates. Their findings agree well with theoretical considerations, which predict that high CO2 is particularly beneficial for carbon acquisition by larger species as they are more restricted by diffusion gradients due to lower surface-to-volume ratios than smaller cells (Flynn et al., 2012; Wolf-Gladrow and Riebesell, 1997). The outcome of our literature analysis supports this allometric concept (Fig. 3, Table 2). Twelve out of 13 experiments in which cell size was taken into account found a shift towards larger species. This is reflected in the ∑RDR score of 265 which is ∼25 times higher than the opposite result (i.e. CO2-induced shifts towards smaller diatoms, Fig. 3c). An allometric scaling of CO2 sensitivity is particularly useful for modelling since cell size is a universal trait which is relatively easy to measure and therefore frequently available (Ward et al., 2012). Accordingly, it may lead to significant improvements of ecological and/or biogeochemical model projections under CO2 forcing when more than one size class for diatoms is considered.

However, although the Wu et al. (2014) allometric approach constitutes a solid starting point to help with understanding the variable responses of different diatom species, it probably also still needs some further refinements. For example, central components of CCMs seem to be adapted to diatom cell sizes, thereby potentially alleviating a strict cell size dependency of CO2 limitation (Shen and Hopkinson, 2015). Furthermore, size dependency alone cannot account for taxon-specific differences in the mode of carbon acquisition (diffusive uptake of CO2 vs. CCM-supported uptake of HCO3-) and how this will affect the competitive ability of species under increasing CO2. OA will lead to much larger changes in dissolved CO2 than in HCO3-. Thus, species that rely to a larger extent on a resource-intensive CCM may benefit more from increasing pCO2 on a cellular level as they could increase the proportion of diffusive CO2 uptake. However, it is also possible that the same species would be disadvantaged on the community-level because their niche, i.e. being competitive at lower CO2 due to an efficient CCM, is diminished under high CO2 conditions (a scenario that is neglected in the physiological literature). Which of the scenarios occurs in nature would also depend on how flexible species are in terms of switching carbon acquisition modes, as well as resource allocation. In this regard, it is noteworthy that only few physiological studies on OA effects have taken into account the role of changing nutrient concentrations or even a transition to nutrient limitation. The available experimental evidence suggests that increasing pCO2 may reduce cellular nutrient requirements for CCM operations and therefore free resources for elevated maximum diatom population densities, particularly when running into nutrient limitation (Taucher et al., 2015). Unfortunately, however, the relevance of this mechanism has so far only been investigated in mono-clonal laboratory experiments but not on the community level.

These considerations illustrate that cell size is an important factor but is not sufficient to predict physiological or even the community level of diatoms to OA. Moreover, the allometric concept as well as the additional mechanisms described above generally presumes positive effects of CO2 fertilization, thus yielding no first-order explanations for observed negative responses of diatoms to changing carbonate chemistry. Obviously, increasing CO2 concentrations are accompanied by increasing proton (H+) concentrations under ocean acidification. High H+ concentrations may reduce key metabolic rates above certain thresholds and outweigh the positive influence of CO2 fertilization as has been observed in coccolithophores (Bach et al., 2011, 2015; Kottmeier et al., 2016).

Another pathway by which ocean acidification may alter diatom communities is the pH effect on silicification and silica dissolution. Low seawater pH should theoretically facilitate silicification as the precipitation of opal occurs in a cellular compartment with low-pH conditions (pH ∼5) (Martin-Jézéquel et al., 2000; Vrieling et al., 1999). At the same time, a lower pH should reduce chemical dissolution rates of the SiO2 frustule (Loucaides et al., 2012). While experimental evidence on this topic is still scarce and partly controversial (Hervé et al., 2012; Mejía et al., 2013; Milligan et al., 2004), it is not unlikely that OA-induced changes in the formation and dissolution of biogenic silica may alter the strength of the frustule and therefore the palatability of diatoms to zooplankton grazers (Friedrichs et al., 2013; Hamm et al., 2003; Liu et al., 2016; Wilken et al., 2011). As for the other physiological effects, e.g. carbon fixation, it is likely that OA impacts on silicification will vary among different diatom species, e.g. according to their species-specific intrinsic buffering capacity, thereby leading to further taxonomic shifts within diatom communities.

The response of diatoms to increasing pCO2 in natural environments will be further modified by multiple other environmental drivers changing simultaneously. Climate change is expected to elevate ocean temperature, as well as also irradiance and nutrient availability via changes in stratification. Physiological experiments have shown that elevated pCO2 may have beneficial effects under low and moderate irradiance but this effect may reverse under high light conditions due to enhanced photoinhibition (Gao et al., 2012). Analogously, warming may have positive or negative effects on photosynthesis and metabolism in general, depending on the thermal optima of the respective species (Boyd et al., 2018). Altogether, these multiple additional drivers will also affect diatom communities, leading to shifts in their taxonomic composition and size structure, which will interact with the impacts of OA.

Diatom community responses cannot only originate from a direct CO2 effect on their physiology but also be caused indirectly through CO2 responses on other components of the food web (Bach et al., 2017; Gaylord et al., 2015). For example, if a grazer of a diatom species is negatively affected by OA then this may benefit the prey and indirectly promote its abundance. Direct OA impacts on zooplankton communities are usually assumed to play a minor role, although there is some experimental evidence that lower pH may have physiological effects at least on some sensitive species or developmental stages (Cripps et al., 2016; Thor and Dupont, 2015; Thor and Oliva, 2015). Nevertheless, much of the currently available empirical evidence indicates that zooplankton communities are affected by OA rather via bottom-up effects, e.g. via changes in primary production or taxonomic composition of the phytoplankton community (Alvarez-Fernandez et al., 2018; Meunier et al., 2017; Sswat et al., 2018). However, bottom-up effects on zooplankton biomass, size structure, or species composition may in turn trigger feedbacks on diatom communities, thereby leading to a feedback loop that may reinforce until a new steady state is reached. Such considerations illustrate that also second- or third-order effects need to be considered when assessing OA effects on the level of ecological communities. Accounting for such indirect effects requires a holistic approach considering all key players in of the food web (something that is beyond the scope of this study). Therefore, interpretations about what the observed responses could mean for entire plankton food webs or even biogeochemical element cycles (Sect. 4.3) should always be regarded with some healthy scepticism as they often neglect the potential for indirect effects.

The taxonomic composition and size structure of phytoplankton communities influences the transfer of energy from primary production to higher trophic levels. In theory, larger diatoms should support a more direct transfer because less trophic intermediates are needed and therefore less respiration occurs until prey items are in an appropriate size range for top predators (Azam et al., 1983; Pomeroy, 1974; Sommer et al., 2002). Such a size shift at the bottom of a food web might eventually lead to higher production in higher trophic levels such as fish. Indeed, recent experimental evidence indicated that fish (including commercially important species) could, under certain constellations, benefit from high CO2 due to higher food availability, although it was not tested if this response is somehow linked to the diatom size structure (Goldenberg et al., 2018; Sswat et al., 2018).

Fluxes of elements through the oceans are (like fluxes of energy through food webs) influenced by the composition of diatom communities (Tréguer et al., 2018). This is particularly well recognized in the context of organic carbon export to the deep ocean, for which diatoms are considered to play a pivotal role (Smetacek, 1985). Given that high CO2 favours large and perhaps more silicified diatoms over smaller ones (Sect. 4.1), we might expect accelerated sinking and thus a positive feedback on the vertical carbon flux. This classical hypothesis is supported by observational evidence from two consecutive years of the North Atlantic spring bloom where, despite similar primary production, particulate organic carbon sequestration into the deep ocean (3100 m) was much higher in the year when the larger diatom species dominated (Boyd and Newton, 1995). However, whether the positive relationship between size and carbon export holds under all circumstances is by no means clear (Tréguer et al., 2018). It is possible that shifts towards larger-sized species coincide with shifts in other traits that feed back negatively on carbon export. For example, when the size shift is associated with decreasing C : Si stoichiometry it may ultimately reduce carbon export (Assmy et al., 2013).

The abovementioned examples of trophic transfer and export fluxes illustrate the importance of the factor “diatom community structure” in the context of marine food production and biogeochemical fluxes. They also illustrate that our understanding of the feedbacks induced through changes in diatom communities is highly incomplete. Hence, with our limited understanding we currently cannot go further than classifying CO2-induced changes in diatom communities as “a potential risk” that may cause changes in key ecosystem services.