Paris Agreement could prevent regional mass extinctions of coral species

Coral reef ecosystems are expected to undergo significant declines over the coming decades as oceans become warmer and more acidic. We investigate the environmental tolerances of over 650 Scleractinian coral species based on the conditions found within their present‐day ranges and in areas where they are currently absent but could potentially reach via larval dispersal. These “environmental envelopes” and connectivity constraints are then used to develop global forecasts for potential coral species richness under two emission scenarios, representing the Paris Agreement target (“SSP1‐2.6”) and high levels of emissions (“SSP5‐8.5”). Although we do not directly predict coral mortality or adaptation, the projected changes to environmental suitability suggest considerable declines in coral species richness for the majority of the world's tropical coral reefs, with a net loss in average local richness of 73% (Paris Agreement) to 91% (High Emissions) by 2080–2090 and particularly large declines across sites in the Great Barrier Reef, Coral Sea, Western Indian Ocean, and Caribbean. However, at the regional scale, we find that environmental suitability for the majority of coral species can be largely maintained under the Paris Agreement target, with 0%–30% potential net species lost in most regions (increasing to 50% for the Great Barrier Reef) as opposed to 80%–90% losses under High Emissions. Projections for subtropical areas suggest that range expansion will give rise to coral reefs with low species richness (typically 10–20 coral species per region) and will not meaningfully offset declines in the tropics. This work represents the first global projection of coral species richness under oceanic warming and acidification. Our results highlight the critical importance of mitigating climate change to avoid potentially massive extinctions of coral species.


| INTRODUC TI ON
Photosynthetic Scleractinian corals can form complex reefs that support diverse and productive ecosystems (Plaisance et al., 2011). However, they have experienced significant declines during recent decades as a result of local human pressures (e.g., pollution, water quality, tourism, and overfishing) and are extremely vulnerable to the ongoing climate crisis (Burke et al., 2011;Hughes et al., 2018;van Hooidonk et al., 2016). Narrow environmental requirements (Couce et al., 2012;Guan et al., 2015) coupled with the breakdown of their symbiotic relationship with intracellular algae during periods of intense accumulated heat stress (i.e., coral bleaching; Berkelmans, 2009;Heron et al., 2015) make many coral species highly sensitive to environmental change. Mass bleaching events are becoming increasingly frequent (Sully et al., 2019), while ocean acidification is reducing the ability of corals to build their aragonite skeletons and potentially making them more susceptible to erosion (Chan & Connolly, 2013). Some coral populations may be unable to withstand these changes, while others may endure or expand into new habitats with the evolving conditions (van Woesik et al., 2011;Yamano et al., 2011). Climate-induced transformations of coral assemblages have already occurred and are set to continue as global temperatures rise (Hughes et al., 2018).
Coral reef ecosystems are of great conservation importance because of their high concentration of species compared to other marine habitats (Plaisance et al., 2011). Forecasting the response of coral species richness to different climate scenarios could, therefore, provide valuable information on the feasibility of achieving conservation objectives. Species distributions are constrained by their tolerance to environmental conditions in the areas they can reach (i.e., the potential environmental niche). Novel environmental conditions under climate change (such as increased temperature and lower pH) could, therefore, alter the distributions of coral species and consequently impact the species richness of future reefs. Paleontological evidence suggests coral species distributions experienced large contractions in tropical regions during periods of intense climate change, but that coral species often avoided extinction, and persisted in suitable refugia and through range expansions into higher latitudes (Pandolfi & Kiessling, 2014). Coral species richness is, therefore, likely to decline over time at the local (i.e., reef) scale throughout much of the tropics. However, the fate of regional coral diversity in tropical areas and range expansions into subtropical zones is less clear, as the current rates of ocean warming and acidification may be unprecedented in the last ~300 million years (Hönisch et al., 2012).
Ecological niche models (ENMs) are statistical tools that analyze observations of species distribution or abundance in relation to environmental factors to define the equilibrium environmental requirements of a species and allow extrapolation in space and time (Elith & Leathwick, 2009). In this paper, we train ENMs for 710 tropical coral species based on their present-day distributions and a set of 23 relevant environmental covariates (Couce et al., 2013). We analyze predicted changes in environmental suitability over this century under two different future Shared Socio-economic Pathway climate scenarios (Meinshausen et al., 2020): SSP1-2.6 (which we refer to as the "Paris Agreement" with a radiative forcing level of 2.6 Wm −2 by 2100) and SSP5-8.5 ("High Emissions" with a radiative forcing of 8.5 Wm −2 by 2100). The potential habitat shifts of spawning species were constrained by connectivity between reef areas, using the output from an existing global coral larval dispersal model (Wood et al., 2014). We focus on temporal changes in sea surface temperature (SST), salinity, and aragonite saturation (Ω Arag ), which can be predicted reliably by climatic general circulation models (GCM) and are important factors defining the ecological niche of coral species (Couce et al., 2012). Other environmental variables were kept fixed at present-day values. Our models do not directly predict impacts on corals from extreme events (e.g., coral bleaching or storms) or local human pressures (e.g., nutrient inputs), nor do they capture physiological or ecological responses (e.g., mortality, bleaching, or adaptation) or species interactions. Instead, they identify areas where environmental conditions for coral metapopulations will and will not be favorable in the future. For brevity, we talk about coral "species richness"-inferred from the number of species with overlapping niches-and discuss "losses" and "gains" in richness. However, we acknowledge that we are referring to equilibrium richness and not the true presence or absence of the species, which often lags behind the change in environmental conditions. This work represents the first projection of future coral species richness at the global scale. Through this approach, we aim to shed light on the potential for coral diversity to persist and spread to new regions within this century of rapid anthropogenic climate change.

| Coral distribution data
Locations of warm-water coral reefs were obtained by combining the global dataset compiled by UNEP-World Conservation Monitoring Centre and the WorldFish Centre (UNEP-WCMC, WorldFish-Centre, WRI, & TNC, 2018) with data from " Reef-Base," version 2000(Vergara et al., 2000. The reef location data were projected onto a global study grid with a resolution of 0.25° × 0.25°. Grid cells containing coral reef locations are referred to hereafter as "reef cells." Distributions of 726 Scleractinia coral species in the Indo-Pacific were based on the ranges compiled by Hughes et al. (2013). Since species composition at individual reef locations is typically unknown, each species' range was assumed to capture the preferred environmental conditions of the species, with all reef cells within the range used as presence sites to train the models. Distributions of Atlantic coral species were based on data from the Coral Geographic interactive map (Veron et al., 2016), which divides the Atlantic (including the Caribbean) into 16 distinctive coral reef ecoregions, with the website providing the coral species known or believed to be present within each ecoregion. As above, all reef cells within the ecoregions where the species was believed to be present were used as presences for model training. Taxonomic information for all Scleractinia coral species was checked and updated using the World Register of Marine Species (WoRMS, 2019) to ensure recent taxonomic changes were captured (e.g., Huang et al., 2014).

| Species "study area"
A species-dependent "Study Area" (Figure S1.1) for spawning coral species was defined as shallow waters within 200 km from the center of all reef cells within the species' range and all other reef cells connected to them by oceanic currents. Wood et al. (2014) simulated global coral larvae transport by ocean currents, producing a matrix with the connectivity of all reef cells to each other, and found that 79% of spawning coral larvae settled within 200 km distance of the site of original release. Therefore, for the spawning coral species in our analysis the Study Area included all connected reef cells, defined as those receiving a nonzero larvae input from any reef site within the species' range, and all shallow water areas within 200 km. The aim of this was to exclude areas that were out of reach to dispersing larvae, by allowing the maximum distance that they could realistically travel from their presentday range. For 35 coral species known to be brooders according to the Coral Trait Database (https://coral traits.org; Madin et al., 2016), the Study Area was defined as the region within 100 km of reef cells within their range, to reflect their more limited dispersal range while capturing "background" conditions in nearby areas (see Data in Supporting Information for more details and examples of Study Areas). Each species' Study Area was used both to train the ENM model and to limit its potential future expansion (i.e., connectivity was assumed to remain constant throughout the study period).

| Environmental variables
The environmental covariates used for training the models were based on Couce et al. (2013), and include SST, aragonite saturation, salinity, nutrient levels, photosynthetically active radiation levels, attenuation coefficient at 490 nm wavelength, light penetration depth, and dust. For many of these, annual average values were considered alongside monthly extremes and other measures of interannual variability (see Table S1.1 for a complete list of all variables and their sources). In the case of SST, in addition to average annual values the models also included minimum and maximum monthly SST, standard deviation of monthly averages, and the annual range (difference between monthly maximum and minimum), in order to capture seasonality and temperature extremes, which often play a key role on habitat suitability for coral species (e.g., bleaching sensitivity at the upper threshold or reproductive success at the lower threshold). See the Data S1 for more details on the variables used and their data sources.

| Environmental niche models
We trained 710 MaxEnt (version 3.4.1; Phillips et al., 2006) models for individual Scleractinia coral species, after excluding species that were present in less than 20 reef cells. MaxEnt is an environmental niche model (ENM) that compares environmental conditions in areas where a species is known to be present (the "realized niche") with background conditions of areas the species could potentially reach (the "potential niche"). It assumes a species occupies all suitable habitat in a way that is as uniform as possible (maximum entropy), testing different constraints that environmental variables could exert and choosing those that maximize the system's entropy. It is a machine-learning technique popular in recent ecological studies that often outperforms traditional methods (Elith et al., 2006). The MaxEnt models were trained within each species' Study Area, with the reef cells within the species range considered as "presences" and all other cells (including connected reef cells) used as background.
While the reef cells within the range were expected to statistically capture the environmental preferences of the species, it would be unrealistic to assume 100% occupancy throughout the range, and instead once each model was trained a presence/absence threshold was chosen so that the suitability was considered insufficient in a certain percentage of reef cells in the range (see below for details).
Model accuracy was assessed via the area-under-the-curve (AUC) scores of the receiving operating characteristic (ROC) curves (Metz, 1986), evaluated on a random 25% subset of the data not used for model training. Similarly to model training, for the ROC calculations, presences were all the reef cells within the species range, with all other cells considered "absences" or "background." Models with AUC values below 0.65 were excluded, resulting in a total of 684 of the 710 species that were initially modeled ( Figure S2 provides the distribution of model AUC values). Model training was carried out in R, version 3.5.3, using package "dismo" version 1.1-4 (Hijmans et al., 2017). We used MaxEnt's logistic output and converted it to binary presence/absence predictions by selecting a presence threshold that assumed the species would be absent at a certain percentage of reef cells within its range. Here we present results corresponding to 80% occupancy of each species of the reef cells within its range, but equivalent results for 50%, 70%, and 90% are included in the Supporting Information, to illustrate that our main conclusions are largely maintained irrespective of what percentage occupancy is assumed (see Figure S2.5 and discussion in Data S2). with additional components that model key biogeochemical, chemistry, aerosol, and vegetation processes. We considered two different emission scenarios for the future projections, the shared socio-economic pathways (Meinshausen et al., 2020) SSP 1-2.6 (which we refer to as the "Paris Agreement") and SSP 5-8.5 (referred to as "High Emissions"). Developed for the Coupled Model Intercomparison Project Phase 6 (CMIP6), SSPs represent different future socio-economic projections and political environments leading to different concentrations of atmospheric greenhouse gases. SSP 1-2.6 is one of the scenarios in the "sustainability" SSP1 socio-economic family, with a radiative forcing level of 2.6 Wm −2 by 2100, approximately corresponding (although not identical) to the previous scenario generation Representative Concentration Pathway (RCP) 2.6. It is one of the high-priority ('Tier 1') scenarios for the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report (AR6), previously referred to as the "2°C scenario," representing a strong societal focus on sustainability and high-mitigation policies often considered in the context of the Paris Agreement goals. Meanwhile, SSP 5-8.5 represents the upper edge of the SSP scenario spectrum, with an emphasis on high fossil-fuel development throughout the 21st century leading to a radiative forcing of 8.5 Wm −2 by 2100 (roughly corresponding to the previous generation RCP 8.5 scenario). Therefore, the two SSPs chosen can be considered from a climatic perspective as approximating bestand worst-case scenarios.

| Future climate projections
UKESM1 data for 2021-2090 were averaged over decadal intervals and compared to averages for a historical run covering the decade 1986-1995. Differences between present-day observations and the model projected annual means of SST and salinity were computed and added to projected future annual averages. Future monthly maximum and minimum SST and salinity values were computed by adding observed present variability to the projected annual averages (i.e., assuming future variability will remain the same, similarly to what was done by e.g., Donner et al., 2005, van Hooidonk et al., 2016. Aragonite saturation was calculated from UKESM1's SST, salinity, total alkalinity, and pCO 2 data using the R library "seacarb" (Gattuso et al., 2021). All other environmental fields were kept fixed at training values and did not change for future predictions.

| Out-of-range extrapolation for novel SST and aragonite
Since Scleractinian corals already occupy some of the warmest areas of the tropical oceans, it is expected that as temperatures further rise throughout the 21st century and seas become increasingly acidic, environmental conditions in some coral reef areas will soon move beyond present-day analogs. While corals appear to be very sensitive to the upper thermal limits (e.g., Glynn, 1993), some corals also show adaptation and acclimatization to increased thermal stress (Guest et al., 2012;Matz et al., 2020) and, therefore, may be able to cope with conditions warmer than the present day. Therefore, establishing how the models assign environmental suitability when working outside the limits of the data used for model training is essential. Some techniques frequently used by habitat suitability models for out-of-range extrapolation include simple linear extrapolation (maintaining the rate and direction of the response that was observed at the limit values in training data), constant response (suitability remains fixed at the values at the limit values in training data and does not change), or fade-to-zero (as conditions move outside training the model will gradually reduce suitability until reaching zero). We used fade-to-zero extrapolation for corals believed more sensitive to bleaching (Loya et al., 2001;McClanahan et al., 2007), as these are less likely to be able to persist and adapt beyond their upper thermal limit (see further discussion and a complete list of species in Table S1.2), and constant response for the others, which may be better able to adapt and acclimatize. We also provide alternative versions of our main results wherein the models for all coral species are extrapolated either by fade-to-zero or constant response (see Figure S2.4 and discussion in Data S2), to simulate different adaptation potentials. Changing the extrapolation method did not meaningfully affect our main conclusions.

| Geographical range expansions
We further considered potential range expansion in five subtropical regions subject to warm ocean currents containing high numbers of coral larvae: (1) Northern Florida, which receives larvae input from the Caribbean via the Gulf Stream, (2) East coast of South Africa, where the Agulhas current may bring larvae, (3) Southern Japan, with input from the Kuroshio current, and both (4) Eastern and (5) Western Australian coasts, with the Eastern Australian and Leeuwin currents, respectively (Wood et al., 2014). For all these regions, there are present-day reports of new coral species establishing (Baird & Sommer, 2012;Greenstein & Pandolfi, 2008;Precht & Aronson, 2004;Schleyer et al., 2008;Vargas-Ángel et al., 2003;Yamano et al., 2011). As the larval connectivity dataset we used lacks connectivity data for non-reef areas (i.e., it is limited to reef cells only), our approach did not effectively capture range expansion into these new coral regions, since they often fell outside the species' Study Areas. Therefore, we defined a pool of species that could potentially get to one of these five range expansion regions as those present in the nearest reef cells (i.e., the most plausible direct sources of larvae) and those in more distant reef cells connected to them (from Wood et al., 2014 dataset). Habitat suitability for these species was then assessed in the respective range expansion region, notwithstanding the species' respective Study Areas.

| RE SULTS
Changes in coral species richness were investigated at two scales: local reef richness (i.e., average coral species richness at reef locations; Figures 1 and 2, top plots) and regional richness (i.e., total number of species in any reef location within a region; Figure 2  cause for optimism, with regional richness levels in some cases experiencing much smaller declines (e.g., in the Coral Triangle (CT) there was a 20% loss of regional richness as opposed to a 72% reduction in average local richness) or even maintaining present-day levels (less than 5% change for the Red Sea and Caribbean). For the remaining areas, the regional loss in coral richness was moderate (e.g., 20%-30% for WIO and CIO) or high in the case of the GBR/CS (50% loss).
Under the High Emissions scenario, regional richness is set to drastically decrease, with losses of 80%-90% coral species by 2080-2090 in nearly all areas (WIO, CIO, CT, GBR/CS), but a somewhat lower loss of 45%-50% for the Caribbean and Red Sea.
In the five subtropical regions investigated for potential range expansions our models predict low levels of coral richness, with less

| DISCUSS ION
The High Emissions scenario paints a bleak picture for future coral reefs, with SST and Ω Arag levels becoming increasingly unsuitable for corals throughout the century (Figure 4), leading to projected catastrophic species loss nearly everywhere (Figures 1 and 2). However, under the Paris Agreement scenario, changes in SST and Ω Arag in tropical reef regions are projected to stabilize and corals may start to recover in the second half of the century (Figure 4), corresponding to a levelling off and decrease in atmospheric greenhouse concentrations (Meinshausen et al., 2020). This illustrates that limiting anthropogenic climate change to 2°C warming would play a vital role in maintaining coral species diversity this century, with our projections suggesting that 70%-80% of species would be retained in most key regions. Even under the Paris Agreement scenario, coral reefs will likely experience considerable declines in diversity at the local scale, but species persisting in refugia within the region will allow for potential recovery of community composition once oceanic conditions start to improve in the second half of the century. Coral species have persisted during geological periods of intense warming and, despite F I G U R E 2 Changes in local (b, c) and regional (d, e) coral species richness over time in key coral regions (a) under the Paris Agreement (SSP1-2.6; b, d) and High Emissions (SSP5-8.5; c, e) scenarios. Richness is defined as the number of species with suitable environmental conditions out of 684 Scleractinian coral species modeled.
global declines in the accretion rates and diversity of reef ecosystems, the extinction rate of coral species remained similar to background levels, with species surviving in offshore refugia (Pandolfi & Kiessling, 2014). It is unclear what will be the impact of a loss of coral richness on ecosystem functions and services. Some presentday coral reefs have experienced dramatic declines in coral richness F I G U R E 3 Changes in regional coral species richness over time in subtropical areas with potential for range expansion (a) under the Paris Agreement (SSP1-2.6, b) and High Emissions (SSP5-8.5, c) scenarios. As before richness is defined as the number of species with suitable environmental conditions out of 684 Scleractinian coral species modeled. The number of coral species that could potentially seed each area is indicated in parentheses on the figure legend.

F I G U R E 4
Changes in sea surface temperature (SST; a, b) and aragonite saturation (Ω Arag ; c, d) in key coral regions (continuous lines) and in possible subtropical range expansion areas (dashed lines) under the Paris Agreement (SSP1-2.6; a, c) and High Emissions (SSP5-8.5; b, d) scenarios as predicted by the UKESM1 general circulation model. The values of training conditions (circa. the 1990s) at reef cells are shown on the right-hand side. The exact geographical extent of each of the areas is shown in Figures 2 and 3. and cover to the extent that they no longer function ecologically as coral reef ecosystems (Wilkinson, 2008), including many reefs in the Caribbean (Jackson et al., 2014), and increasingly in the GBR (Hughes et al., 2018). However, low coral richness does not necessarily result in poor ecosystem functioning, with examples of reefs in areas of naturally low richness, such as Clipperton Atoll (seven species) maintaining high levels of coral cover and a healthy reef ecosystem (Glynn et al., 1996).
This paper, has presented the first global ENM analysis of the future environmental niche of individual coral species under warming and acidification, which represents a significant advancement over global ecosystem level modeling (Couce et al., 2013;Descombes et al., 2015;Jones et al., 2019). Moreover, by accounting for larval dispersal limitations our models represent an improvement over regional modeling that does not account for reef connectivity (e.g., at the species level, Rodriguez et al., 2019or ecosystem level, Guan et al., 2015, where the choice of arbitrary regions may lead to environmental limits not related to real physiological tolerances. Using a similar approach to ours, Rodriguez et al.'s (2019) trained MaxEnt models for individual coral species of the Atlantic Ocean, projecting strong declines in coral species richness in the central Caribbean by the end of the century, with increases at the outer edges of the region. This is largely consistent with our projections, although our predicted potential increases in the possible range expansion area of Northern Florida appear more moderate, likely due to the inclusion of connectivity constraints in our study. Förderer et al. (2023) use MaxEnt to model global diversity of large benthic foraminifera, which have present-day diversity patterns not dissimilar to those of coral species. For a climate scenario intermediate to the ones we looked at, they also project substantial losses of diversity throughout the tropics, particularly significant at lower latitudes.
Ecological niche models are typically trained with equilibrium conditions and do not predict coral response (Guisan & Zimmermann, 2000), such as bleaching, mortality, adaptation, drowning by raising sea level, etc. Although we speak in terms of species "gained" and "lost," we refer specifically to environmental suitability. A species "lost" from an area may still be present but in terminal decline, or its continued presence may be due to climatic adaptation and/or acclimation (van Woesik et al., 2011); "gains" in suitable areas may not be physically present due to low or infrequent connectivity, lack of suitable substrate for larvae to settle, or competition with other species (Guisan & Zimmermann, 2000;Wood et al., 2014). Since our models project a general decline in the number of species with suitable habitat, a delay in coral response whereby species temporarily survive in adverse environmental conditions could buy enough time for these conditions to recover (as in the improved forecast for the late part of the century under the Paris Agreement scenario; Figure 4).
The climate projections in coral regions (Figure 4) reflect a well-known positive relationship between Ω Arag and SST (Jiang et al., 2015), with regions experiencing the greatest warming (e.g., Red Sea, Caribbean, CIO) also expected to maintain the highest Ω Arag levels. Interestingly, our results show that regions with the greatest warming are projected to retain a high proportion of their coral species, particularly under the Paris Agreement scenario, whereas regions with low average SST and low rate of warming, but also low levels of Ω Arag (e.g., GBR/CS), are projected to experience the greatest species loss. This implies that Ω Arag (i.e., ocean acidification) may have a critical impact on coral species richness. Previous ENM work investigating future environmental suitability for coral reefs found that ocean warming was the most important stressor, driving a poleward shift in suitable habitat (Couce et al., 2013;Descombes et al., 2015;Jones et al., 2019). However, these studies are not directly comparable to ours, as they modeled the presence of suitable coral reef habitat (irrespective of species richness). Additionally, by training ENMs at a global scale it is difficult to disentangle the effects of SST and Ω Arag , which are strongly correlated (Jiang et al., 2015), whereas the correlation often breaks down at the local and regional scales (Guan et al., 2015) at which the majority of our individual-species ENMs were trained. Putting methodological differences aside, the results from previous ENM studies taken with ours suggest that future changes to the distribution of coral reefs will be driven primarily by warming, whereas acidification may be critical in determining their potential species richness.
Potential range shifts will play an important role in coral community responses to projected warming (Hoffmann & Sgró, 2011).
Despite hopeful present-day observations of poleward range expansions for some corals with increasing SSTs (Yamano et al., 2011), our projections suggest that other factors such as acidification and light attenuation may limit coral diversity at high latitudes (Jiang et al., 2015;Muir et al., 2015). The well-documented high potential for self-recruitment also limits dispersal capabilities in corals, even those with an obligate planktonic phase (Ayre & Hughes, 2000;Figueiredo et al., 2013;Gilmour et al., 2009). Regional-scale shifts have already been observed in the stock-recruitment relationships in the GBR following recurrent mass bleaching and mortality events (Hughes et al., 2019), thus impacting recovery/range expansion potential further still (although corals surviving in thermal refugia may enable future recovery, see Cheung et al., 2021). Our species-specific Study Areas were based on global estimates of maximum potential for coral larvae dispersal and do not consider self-recruitment levels.
Therefore they provide for optimistic range expansion estimates, yet even with this optimism, our projections of future species richness at the edges of the present-day distribution do not show significant potential for growth (Figures 1 and 3; Figure S2.1).
We have made inferences about future coral reefs using present-day empirical associations between species and their environment, but there are complexities that may influence how coral species ultimately respond to climate change. Currently, bleaching is the most significant climate impact on coral richness and health (Hoegh-Guldberg et al., 2018), with bleaching thresholds determined by the local maximum of monthly mean SST (Heron et al., 2015). Our model constrains coral species' habitat suitability by modeling changes in maximum monthly summer temperatures and assumes that present-day seasonal variability is maintained into the future. This is similar to the approach used by van Hooidonk et al. (2016) to predict future coral bleaching, although they focused on the upper thresholds of thermal tolerance (i.e., degree heating weeks). While the two studies are not directly comparable, van Hooidonk et al.'s (2016) prediction of significant increases in severe bleaching in the 2040s under the High Emissions scenario matches our projections of when species loss will accelerate. However, the bleaching mechanism is a complex response, with diverging behavior at the species and even at the colony level, driven by different experiences of thermal stress and variations in symbiotic clades (Obura, 2009;Safaie et al., 2018;van Oppen et al., 2017). Matz et al. (2018 (Guan et al., 2015), but small-scale regional variations in nutrient concentrations are likely to be important (Rabalais et al., 2009). Although these factors were not incorporated into our models, it is clear that they are causing stress to coral reef ecosystems in most regions of the world (Wilkinson, 2008). If this stress were to intensify this century, then the future for coral reef diversity is likely to be bleaker than suggested by our projections.
One of the main obstacles of applying ENM approaches to coral species is the lack of reliable coral distribution data for many species, particularly at the global scale. Instead, we have based our models on inferred ranges and distribution limits, rather than field observations of the species. While we believe the statistical analysis of environmental conditions at all reef cells in their range is likely to provide an accurate picture of each species' requirements, there is little information of the prevalence of each species throughout its range, and this can impact the projected richness levels. We attempted to constrain the effect of this by testing different presence/ absence thresholds ( Figure S2.5) and found that while the magnitude of the trends changed, our main conclusions still applied. However, this sensitivity analysis used a fixed percentage occupancy for all species, whereas in reality the value would be species specific.
Nevertheless, the lack of data on species presences makes this assumption necessary, and this is unlikely to change in the near future, considering the challenges of carrying out field surveys at the global scale.
A final source of uncertainty arises from the ongoing major revision of coral taxonomy using genetic techniques (e.g., Huang et al., 2014). Although the most up-to-date taxonomy was used at the time of the analysis (July of 2019), continuing taxonomic work on corals may redefine species identities and ranges in the future, requiring reanalysis for these taxa's environmental preferences.
However, our models' projected declines of species habitats under climate change have similar impacts on wide-ranging species and those with narrower distributions, and therefore we do not expect taxonomic revisions to have significant impact on our main results and conclusions. This paper set out to predict changes in the diversity of coral reef ecosystems in the near future, as conditions become warmer and more acidic. Our results suggest substantial losses in richness are likely to take place over the coming decades, with considerable geographical disparities in the predictions, and a near-total loss of coral species by the end of the century under the High Emissions scenario. Other regional and human stressors not considered here (e.g., overfishing, land-based pollution, macroalgae competition, increases in predator abundance, disease, marine litter) are compounding the pressure experienced by reefs globally, with presentday observations already showing evidence of dramatic declines in coral cover and diversity (Wilkinson, 2008). Additional work is necessary to understand and predict coral responses to acute thermal stress and cumulative pressures, as well as to assess the impact of a loss of richness on ecosystem functioning and services. Our results demonstrate the crucial importance of achieving Paris Agreement emissions targets to avoid catastrophic reductions of coral species richness within this century and to give coral communities a chance of recovery.

ACK N OWLED G M ENTS
The authors wish to thank Oliver Hoggs (Cefas) for collating the environmental datasets, Chris D. Jones (MET Office) for providing the UKESM1 climate data and Sally Wood (University of Bristol) for providing the reef connectivity data. The research idea first originated during EC's doctoral thesis and she acknowledges the contribution of her former PhD supervisors Erica Hendy (University of Bristol) and Andy Ridgwell (University of California, Riverside). The manuscript benefitted from feedback from John Pinnegar (Cefas) and from two anonymous reviewers. This work was funded by Cefas Seedcorn Projects DP418 'Quantitative Ecology' and DP441D 'Coral Reefs: Ecosystems in Transition'.

CO N FLI C T O F I NTE R E S T S TATE M E NT
The authors have no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data with the projected species present and absent at reef locations globally is available to download from the Cefas Data Hub at: