We use 15 illustrative impact metrics providing a broad spectrum of impacts of climate change on marine ecosystems, and each of them are associated with 4 limits (Table 1). Limits are ranked according to the expected severity of impacts upon exceeding them: exceeding limit 4 for a given metric is expected to result in more severe impacts than exceeding limit 1. Thus, staying below limit 1 is more ambitious because a higher emission reduction would be required to achieve this goal. Limits at a given level are not necessarily dependent, i.e., they can be exceeded at different time and global warming levels.

A literature review combining observations with simulation studies on critical limits in the ocean system is conducted to define the impact metrics and limits. While the aim of this study is to define limits based as much as possible on the literature, many metrics suffer from a lack of knowledge regarding the assessment of actual impacts that an exceedance would have on the Earth system or ecosystem functioning, especially in a multi-stressor context (Williamson and Guinder, 2021). Some physical metrics have been more thoroughly investigated, while biogeochemical metrics are less constrained. Observations and laboratory experiments suggest numerous critical limits for key ecosystem stressors. Moreover, these limits are species-dependent and can vary over a wide range. Thus, for some metrics, we favour limits based on relative changes to characterise reasonably safe levels instead of absolute changes. These choices could be refined through future research and further dialogue with stakeholders. If the literature does not permit us to define limits for a specific metric, then we use ad hoc limits that cover the simulated mitigation space from very strong mitigation to very little or no mitigation effort.

The impact metrics include six physical parameters, related to surface atmospheric warming, marine heatwaves, steric sea level rise, sea ice extent, and the Atlantic Meridional Overturning Circulation (AMOC), five chemical parameters, related to global and regional ocean acidification and deoxygenation, and four ecosystem parameters, related to productivity, biomass, organic matter export, and metabolic performance.

Our CMIP6 model ensemble is composed of 9 ESMs (Table 2). This ensemble is based on the one used in Canadell et al. (2021), but excluding model family duplicates, and using the variant r1i1p1f1 (or equivalent). All ESMs use ocean components with a nominal horizontal resolution of about 1° with grid refinements of up to about 1/3° both poleward and at the equator. We use 3 scenarios from CMIP6 covering the period 2015–2100, which are initialized from the end of the historical simulation (1850–2014) that is based on estimates of historical forcings (O'Neill et al., 2016). These scenarios cover very different possible futures: The low-emission, high-mitigation scenario SSP1-2.6 assumes that the world gradually shifts toward a more sustainable pathway, and that early and consistent climate mitigation limits the end-of-century radiative forcing to 2.6 W m⁻². In contrast, the SSP5-8.5 scenario assumes resource-intensive, strong economic growth based on the exploitation of fossil fuel reserves and no climate mitigation. The very high carbon dioxide (CO₂) emissions in this scenario lead to a radiative forcing of 8.5 W m⁻² at the end of this century. The SSP5-3.4-OS scenario follows the SSP5-8.5 pathway up to year 2040. Then, strong climate mitigation policies are implemented, including carbon dioxide removal from the atmosphere, leading to a peak and decline in surface temperature and a final radiative forcing level of 3.4 W m⁻² in 2100. To use the same model ensemble for all scenarios, we excluded models that do not provide SSP5-3.4-OS.

All ESM model outputs used in this study have been regridded to a 1°-resolution regular grid (360×180 grid cells) before analysis. Since most of the impact metrics are expressed as a change relative to the period 1850–1900, model biases would only be an issue for the analysis presented here if the response to forcing would significantly depend on the baseline state. However, we removed potential model drifts from two sensitive metrics (ΔSSL and Global ΔO₂) by calculating the difference between the projected signal and its equivalent from the corresponding preindustrial experiment (piControl). To account for carbonate chemistry biases in the present-day mean state simulated by the CMIP6 ESMs, we follow the methodology of Terhaar et al. (2020). Changes in aragonite saturation state (Ωarag) have been computed offline using mocsy 2.0 (Orr and Epitalon, 2015) from regridded annual CMIP6 model outputs of dissolved inorganic carbon (DIC), alkalinity, sea water temperature (T), and sea water salinity (S). These modelled changes have then been added to the contemporary saturation state that we derived from the observation-based GLODAPv2 data product for DIC and alkalinity (Lauvset et al., 2016), and the World Ocean Atlas 2018 for T and S (Locarnini et al., 2018; Zweng et al., 2018).

For all impact metrics, time series have been smoothed using a 20-year running mean before identifying the years and global warming levels when a certain limit is exceeded. The exceedance is identified by the time and global warming level at which a given limit is exceeded for the first time. This definition does not account for “overshooting” limits of an impact metric. Cases where a limit is first exceeded, but the system returns to a state below the limit later in time, is counted as an exceedance. Nevertheless, we provide analysis that allows for identifying cases where such overshooting of limit happens (Figs. 3 and 4).

Some limits might be exceeded by only a part of all available models or ensemble members, while other limits might be exceeded by all models or ensemble members within the time horizon of the scenario simulations (until 2100). If a model does not exceed a given limit, we interpret this as an exceedance in the last year of the simulation (year 2100). This approach is conservative in the sense that it assigns the earliest possible exceedance year (the unknown true exceedance year of this model is later, or the model might not exceed the limit at all), and it ensures that all information provided by the model ensemble is used. A similar approach is applied along the global warming dimension, by attributing the highest warming level reached by the model under the given scenario to models not exceeding a given limit. Consequently, our exceedance estimates are characterised in terms of uncertainty and probability. We define exceedance uncertainty as the interquartile range of an exceedance estimate in a given model ensemble for a given experiment, impact metric, and limit. We define exceedance probability as the proportion of models exceeding a limit across all models that provide data for a given impact metric and limit (i.e., here we exclude models that did not exceed the limit within the simulation period). We assign high probability to an estimate of exceedance time (or warming level) if a limit is exceeded by at least 80 % of our CMIP6 ensemble (i.e., at least 8 out of 9 ESMs) or of the ensemble members of an individual EMIC, medium probability if 50 %–79 % of the models or ensemble members exceed the limit, and low probability if less than 50 % of models or ensemble members exceed the limit. This approach avoids the use of ensemble means, and uncertainty intervals such as ensemble standard deviation to define exceedances, from which a lot of information on the distribution of exceedances within the ensemble is lost leading to an underestimation of the resulting exceedance uncertainty. We will further discuss this issue below with some examples. We note that there are small differences as to which CMIP6 ESMs can be used for which metric, since not all models provide all data necessary for all impact metrics (Table 2).

In addition to output from CMIP6 ESMs, we also analyse scenario simulations of two Earth system models of intermediate complexity (EMICs), the Bern3D-LPX model and the University of Victoria Earth System Climate Model (UVic).

Both EMICs have simulated large perturbed-parameter ensembles (PPE) to estimate the range of parametric (model) uncertainty. The EMIC PPEs were run over the historical period as well as for the SSP1-2.6, SSP5-3.4-OS, and SSP5-8.5 scenarios. Ensemble generation, sampled parameters, and calculation of ensemble member skill scores based on observations differ between the two models and are briefly outlined below.

The uncertainty, and probability related to the time and global warming levels at which limits are exceeded are highly variable across impact metrics, limits, and scenarios (Figs. 1 and 2). The probability in exceedance estimates generally decreases with higher limits, since generally less models exceed the higher limits. Note that the interquartile ranges in Fig. 1 are extended toward or up to 2100 in case that only some of the ESMs exceed a given limit, because of our choice to assign the year 2100 as time of exceedance for those models. All the reported median exceedances related to the most ambitious limit (limit 1) are projected to occur in the short and mid term (before 2060) for all scenarios. If limit 1 is exceeded with high probability (marked by full size black dots in Figs. 1 and 2) in all scenarios, the median time of exceedance is generally very similar across scenarios (ΔSAT, ΔSSL, SIE, Arctic Ocean Ωa < 1, ΔBiomass), because during earlier times the three scenarios share the same historical forcing or have only slightly diverged. Also, before 2040, the two scenarios of the SSP5 family are identical by construction, such that if all models exceed a limit before 2040, the median exceedance time is identical for SSP5-3.4-OS and SSP5-8.5. Metrics related to surface ocean aragonite saturation state (Arctic Ocean Ωa < 1, Southern Ocean Ωa < 1 and Global Ocean Ωa < 3) show particularly narrow uncertainties over the time dimension, but not over the global warming levels. This is consistent with the findings of Terhaar et al. (2023) showing that projections with prescribed atmospheric CO₂ yield an unrealistic small uncertainty for surface ocean acidification, because the forcing agent (atmospheric CO₂) is the same across all ensemble members. If projections target a certain temperature level, a more relevant estimate of uncertainty can be obtained because variations of climate sensitivity across ensemble members translate into different levels of atmospheric CO₂ at the targeted temperature, and hence into different levels of surface ocean acidification. Lower exceedance probability and larger difference in the timing of exceedance are found for ΔPOM, ΔNPP, global ΔO₂, and hypoxic ΔO₂. The lower probability in exceedance of ΔNPP limits across all scenarios is consistent with the high uncertainty in ΔNPP projections found by Kwiatkowski et al. (2020). In contrast, the higher exceedance probability linked to the metric ΔBiomass is consistent with the findings of Tittensor et al. (2021), emphasising that Biomass is a more robust metric for assessing impacts of climate change on marine ecosystems than NPP. Substantial uncertainties in global ΔO₂ were found in earlier multi-model studies (Cocco et al., 2013; Hameau et al., 2020), which are reflected by the lower probability of exceedance estimates. These uncertainties can be explained by the uncertain balance between O₂ supply from physical mixing and advection, and O₂ consumption from remineralization of organic matter. Uncertainties remain also on how these processes would respond to rising CO₂. Regarding subsurface O₂ projections, Frölicher et al. (2016) identified model structure and parametrization as the second source of projection uncertainty, consistent with the results presented in Figs. 1 and 2.

The 20-year moving averaging applied to the time series mostly removes interannual to decadal variability and emphasises the signal induced by the different radiative forcing. However, metrics with large internal variability, like the AMOC strength, can still show exceedance distribution not entirely consistent with the different radiative forcing levels applied in the emissions scenarios (Fig. 1, limit 2). Similarly, composite metrics such as ΔBiomass (i.e. sum of phytoplankton and zooplankton biomass) include interactive dynamic responses that lead to inconsistent responses with respect to the different radiative forcing levels for some models with exceedances only 2 to 3 years earlier in SSP1-2.6 compared to SSP5-8.5. Hence, differences between scenarios in exceedance timing of around 10 years should not be considered significantly different (i.e., they can be considered as occurring simultaneously).

For the less ambitious limits, exceedance time estimates generally move towards later times and higher warming levels, and the exceedance probability decreases (less models exceed the less ambitious limits). In the low emission scenario SSP1-2.6, limit 4 is exceeded by less than 20 % of the CMIP6 ESMs for any of the impact metrics due to the stringent climate mitigation implemented in this scenario. Exceedances occur generally at lower global warming level under scenarios with lower radiative forcing (Fig. 2). For high probability cases, this is due to the inherent inertia from some metrics (e.g., Fig. 2, limit 1, ΔSSL) that eventually leads to an exceedance even under the global warming stabilization induced by SSP1-2.6 scenario. For low probability cases, this observation is only the result of the default attribution of the highest warming level reached to models not exceeding a given limit (Fig. 2, limit 4, ΔBiomass).

If we focus on exceedance estimates that have a high probability (i.e., where at least 80 % of models exceed the limit), we can provide an estimate of the time when limits are likely to be exceeded (Figs. 3 to 6). The first limit of Arctic Ocean Ωa < 1 (20 % of undersatured surface waters) is likely already passed in the CMIP6 model ensemble in all scenarios, consistent with the findings of Terhaar et al. (2021). The first limit of SIE (4×106 km²) is expected to be passed during 2013–2034 (median year 2023). So far, according to satellite-based estimates, Arctic summer sea-ice extent fell below 4×106 km² in 2012 and 2020 (https://nsidc.org/, last access: 21 May 2025). The fourth level of limits is exceeded with high probability only for SIE in SSP5-3.4-OS and SSP5-8.5, as well as for Southern Ocean Ωa < 1 in SSP5-8.5.

Avoiding emissions as high as in the SSP5-8.5 scenario and following an emission pathway similar to SSP5-3.4-OS will likely avoid an exceedance of any of the limits for ΔMHW_shift, Southern Ocean Ωa < 1, Global ΔO₂ , and ΔPOM during this century. In addition, avoiding the SSP5-3.4-OS scenario by early mitigation (as in SSP1-2.6) will likely avoid exceedances of any limit related to ΔMHW_fix, ΔAMOC, ΔBiomass, and ΔΦ until year 2100. The limits of ΔSAT, ΔSSL, SIE, Arctic Ocean Ωa < 1, AΩ<3, and ΔBiomass are likely to be exceeded across all scenarios. The effect of ambitious mitigation in SSP5-3.4-OS after 2040 can be seen for ΔSAT, ΔSSL, SIE, and ΔAMOC. For these metrics, some of the limits are first exceeded around mid-century, but later this exceedance is reversed. The uncertainty of hypoxic ΔO₂, global ΔO₂, ΔNPP, and ΔPOM projections does not allow us to conclude with confidence that none of the corresponding limits will be exceeded due to their attributed low probability (Fig. 4).

In Figs. 5 and 6, the summary of high probability exceedance estimates for all impact metrics is quite conservative, since (1) high probability is defined as at least 80 % of models exceeding a limit (i.e., at least 8 out 9 models, which is practically 88 %) and (2) medium probability exceedances (where up to 79 % of the CMIP6 models would show an exceedance of a given limit) are not included. For the low-emission scenario, already the most ambitious level of limits of 5 impact metrics (ΔSAT, ΔSSL, SIE, Arctic Ocean Ωa < 1, and ΔBiomass) is exceeded with high probability, two of them as early as in the near-term period (2021–2040). In contrast, even the third and fourth set of limits are exceeded with high probability in the high-emission SSP5-8.5 scenario, particularly toward 2100. The absence of high probability in the exceedance for hypoxic ΔO₂, global ΔO₂, ΔNPP, and ΔPOM is explained partly by a model disagreement within our CMIP6 ensemble. Another reason for low-to-medium probability in the exceedance for global and hypoxic ΔO₂ and ΔSSL before year 2100 is that changes in subsurface and whole ocean parameters have been shown to accrue beyond year 2100 and aggravate over many centuries due to the long overturning time scales of the ocean (Battaglia and Joos, 2018). There is a very clear effect of the ambitious mitigation assumed in SSP5-3.4-OS after 2040 in all timeseries of impact metrics, such that the significantly lower exceedance rate of limits, particularly by the end of the century, clearly illustrated the benefits of stringent and ambitious mitigation. Some metrics show strong hysteresis in response to cumulative carbon emissions (Boucher et al., 2012; Jeltsch-Thömmes et al., 2020; Samanta et al., 2010; Santana-Falcón et al., 2023). Due to that hysteresis behavior, sustained negative emissions are required to return to and stay under a specific limit, particularly for high climate sensitivities and peak-and-decline scenarios with carbon dioxide removal (Jeltsch-Thömmes et al., 2020). This aspect needs to be emphasized in the case of our study due to the use of simulations ending in 2100, implying that some metrics could return below more ambitious limits beyond 2100 under strong mitigation scenarios.

There is broad agreement for a range of variables on the exceedances of limits between the CMIP6 and the two skill-weighted EMIC ensembles under SSP5-8.5, but also major disagreements are identified (Figs. 7 and 8). Note that the median and interquartile ranges are only shown for ensembles that pass the limits with a probability of 50 % or more. For a good agreement between ensembles both probability and median needs to match. For the limit 2, the median value of exceedance agrees within 25 years and 1 °C for ΔSAT, ΔAMOC, Southern Ocean Ωa < 1, Arctic Ocean Ωa < 1, Global Ocean Ωa > 3, and ΔPOM between the CMIP6 and the two EMIC ensembles. However, probability is variable among ensembles. The two EMIC ensembles generally show systematically high probability in limit exceedances in the first two limit sets (except limits 2 of UVic's ΔΦ and Bern3D Hypoxic ΔO₂) while the CMIP6 ensemble shows only 4 exceedances with medium probability over the same limits. The CMIP6 ensemble shows somewhat larger warming than the EMIC ensemble with earlier exceedance. This is consistent with the fact that the CMIP6 ensemble includes models with climate sensitivity larger than observation-constrained estimates (Nijsse et al., 2020; Tokarska et al., 2020). For the third and fourth limit sets, the CMIP6 models show medium probability in the exceedance of the hypoxic ΔO₂, whereas the EMIC ensembles show no exceedance or with little probability. The finer spatial resolution used in CMIP6 models compared to the EMIC ensemble could explain this difference.

Regarding the uncertainties related to the exceedance of limits, the three ensemble agrees generally well over many metrics (Fig. 7). Regarding ΔPOM and hypoxic ΔO₂, the uncertainty range from the CMIP6 ensemble is larger than the EMIC ensembles. We hypothesize that the parametric uncertainty sampled in the perturbed-parameter EMIC ensembles is a significant underestimation of the full uncertainty signal of some metrics as it lacks the structural model uncertainty inherent in the CMIP6 ensemble. The ESMs' uncertainties of exceedances related to Southern Ocean Ωa < 1, Arctic Ocean Ωa < 1, Global Ocean Ωa > 3 are narrower than the one from EMICs. We interpret this difference by an efficient bias correction of the present-day mean state applied to these metrics for CMIP6 ESMs. Another explanation could be how EMIC ensembles are constrained. For UVic, global mean profiles of ocean tracers have been used as observational constraints. The latter averages large regional variations that compensate each other resulting in similar global mean. Such globally averaged constraints might inefficiently reduce uncertainty for ice-dominated polar regions, especially since we did not constrain sea-ice area. Large variations in sea-ice cover could influence air-sea gas exchange and, as a result, Arctic Ocean Ωa < 1.

Despite such differences, robust conclusions emerge. First, both the CMIP6 and EMIC ensembles show medians passing most of the stringent limits of set 1 and set 2 with high probability within this century for global warming of 2.5 and 3.5 °C, respectively (Fig. 8). Exceptions are ΔΦ, ΔSSL (known to lag surface warming and continues to increase over centuries), and ΔO₂ metrics, for which the CMIP6 and EMIC ensembles disagree. Second, both the CMIP6 and EMIC ensembles demonstrate that many less stringent limits of set 3 and set 4 are not passed with high probability within this century for global warming of 1.5 °C and 2 °C, respectively (Fig. 8). Taken together, the results of the model ensembles collectively show that limiting global warming below 2 °C avoids passing the considered Earth system limits during this century with potentially dangerous impacts on eco- and socio-economic systems.

Our study assesses a wide range of mitigation metrics over three very distinct scenario pathways and uses a large set of model results. Nevertheless, the global viewpoint that our study takes imposes certain limitations. Most important, the definitions of our metrics use averages over the global ocean or large oceanic regions neglecting spatial heterogeneity of changes in the climate system. For instance, while our metric Arctic Ocean Ωa < 1 averages across the whole ocean north of 70° N, there is significant regional variability in the extent and rate of aragonite undersaturation across the Arctic Ocean. Areas with high freshwater input from sea ice melt, such as the Canadian Arctic Archipelago, experience more rapid acidification compared to regions influenced by Atlantic inflow, which have greater vertical mixing and primary production (Popova et al., 2014). Secondly, due to data limitations, we use only one ensemble member per ESM and scenario (many ESMs provide only one ensemble member). This is a limitation for metrics where changes in internal variability are important, as the marine heatwave metrics. For these metrics, the relative contributions of long-term warming trends versus anthropogenic changes in internal variability vary geographically, with the warming trend often accounting for more than 90 % of the total changes (Deser et al., 2024; Frölicher et al., 2018; Frölicher and Laufkötter, 2018). By using a single simulation per model, we can provide a general assessment, but we fold together both model structural uncertainty and sampling uncertainty. To better assess MHW metrics under a moving-mean baseline, future studies should incorporate initial-condition large ensemble simulations (Burger et al., 2022; Deser et al., 2024), which would allow for a more robust evaluation of evolving forced responses in individual models at specific locations and times. Finally, for the metabolic index, we limited ourselves to illustrating the scenario- and model dependent changes in viable habitat based on metabolic traits for one median ecophysiotype (Fröb et al., 2024), while including a broader range of species' thermal and hypoxia sensitivities would allow for a broader multi-species assessment of habitability.

With these limitations in mind, we can draw several conclusions from our results, firstly for physical metrics. For SSL rise, models agree relatively well on the timing of when certain limits are crossed. Also, the value of stringent mitigation is clearly visible in our results, since the less ambitious limits are only breached in the high emission scenario within this century. This also illustrates the value of late mitigation (compared to no mitigation) in the overshoot scenario, because also here the crossing of the two least ambitious limits is avoided. This is consistent with previous work showing that the rate of SSL rise is quite reversible in overshoot scenarios (Schwinger et al., 2022). It should be stressed, however, that sea level rise has a high inertia and will continue beyond the end-of-century time-horizon considered in this study, even in the strong mitigation scenarios.

Our results highlight committed severe risks for marine ecosystems related to summer Arctic sea ice retreat whatever the emissions scenario. Even the least ambitious limit (ice-free Artic during summer; September sea-ice extent < 10⁶ km²) is passed with medium probability (50 %–79 % of models) in the low emission scenario. The summer Arctic sea ice reduction affects marine biota, particularly species like the Arctic cod, which are crucial for the diet of upper trophic level predators such as the black guillemot. This leads to changes in diet composition and increased nestling starvation rates among seabirds (Divoky et al., 2015). The decline in Arctic sea ice directly threatens the food security and cultural continuity of Indigenous Peoples. Many Arctic communities rely on sea ice for hunting and fishing, which are essential for their livelihoods. The sea ice loss disrupts these traditional practices, leading to food insecurity and cultural disruption (Huntington et al., 2022). Further, the decline in sea ice extent is opening new shipping routes, such as the Northern Sea Route, which connects North-East Asia with North-Western Europe. This route reduces shipping distances by approximately one-third compared to the Southern Sea Route. The opening of the Northern Sea Route could lead to significant economic benefits for global trade but raises geopolitical concerns and environmental pressures related to Arctic shipping and global supply chain reorganization (Bekkers et al., 2018).

AMOC decline in CMIP6 models has been shown to be relatively insensitive to the emission scenario, at least up to 2060 (Weijer et al., 2020). This is also reflected in our results, since the median exceedances of limits 1 (20 % decline) and 2 (25 % decline) are not very different for the strong mitigation scenario SSP1-2.6 (albeit with a somewhat lower probability) compared to SSP5-8.5 and SSP5-3.4. Nevertheless, the inter-model spread of projected AMOC decline remains large.

Results for the metrics related to ocean acidification show a high probability of crossing ambitious limits for the Arctic and the global ocean even in the low emission scenario SSP1-2.6 while the limits specific to the Southern Ocean are only crossed for the high emission scenario. Seen from a perspective of scenarios with prescribed atmospheric CO₂ concentration, the uncertainty in these results is small. However, consistent with the study of Terhaar et al. (2023), uncertainties increase considerably when relating the ocean acidification related metrics to a specific global warming level. Ocean acidification poses a severe threat to early life stages of calcifying organisms which can dominate surface water communities in polar regions, e.g. the polar pteropod (Limacina helicina antarctica), leading to shell dissolution and fragility, high mortality, and reduced recruitment (Bednaršek et al., 2012; Gardner et al., 2018). This is of major importance as the pteropods contribute significantly to the pelagic food web and carbon export fluxes in this region (Hauri et al., 2016). Aragonite undersaturation in the Arctic region threatens calcifying organisms such as plankton and invertebrates, which depend on calcium carbonate for their structural integrity. This can lead to changes in the composition of the Arctic ecosystem, affecting both planktonic and benthic communities (Bates et al., 2013; Yamamoto-Kawai et al., 2009). At Ω < 3, the global ocean experiences widespread negative biological and ecological effects, including reduced survival, growth, and calcification in many marine species, especially those that build shells or skeletons from calcium carbonate (e.g., corals, molluscs, some plankton species; Kroeker et al., 2010). Together with long-term warming these can lead to declines in primary production and carbon export (Moore et al., 2018), resulting in lower fishery yields and reduced ecosystem productivity.

Maybe not surprisingly, we find the largest uncertainties for the biogeochemical metrics such as the one related to O₂, NPP, plankton biomass and POM. In all the scenarios used in our study, models do not agree on the projected sign of change of NPP while models agree on the projected decline of plankton biomass, albeit with significant uncertainties on the amplitude of this decline. These results are consistent with previous literature (Kwiatkowski et al., 2020; Tittensor et al., 2021). Change in NPP and biomass are influenced by complex interactions among nutrients, temperature, and ecosystem dynamics which is often beyond model capacity (Tagliabue et al., 2021). Regarding the metabolic index, viable habitat for marine organisms is lost if species-specific thresholds of metabolic demand and oxygen availability are crossed. Future projections show a decline in ocean habitability due to the combined threat of ocean warming and deoxygenation, leading to high extinction risk for polar species and loss of biological richness in the tropics (e.g., Penn and Deutsch, 2022)

Assessing the exceedance of limits for multiple impact metrics requires large model ensembles to obtain high-probability signals and corresponding uncertainties for the exceedance estimates (in years and global warming levels) linked to the projection pathways. Our analysis clearly indicates the need for better constraining and/or weighting the CMIP6 ensemble to reduce the large uncertainties found for exceedance estimates of many of the impact metrics. Simulations beyond year 2100 are needed to assess the long-term impacts of anthropogenic emissions, especially for the volume of hypoxic waters, global oxygen inventory, AMOC response, and steric sea level rise.

Our results show that ambitious limits will be exceeded with high or medium probability even if a low-emission pathway is followed, but that exceeding less ambitious limits associated with a higher risk for severe impacts is unlikely in a low-emission scenario. In contrast, under the high-emission scenario, many of the less ambitious and more risk-prone limits are exceeded with high to medium probability. The benefit of strong mitigation efforts in the overshoot pathway is clearly measurable as a decrease in the exceedance probability of the least ambitious and most risk-prone limits. Nevertheless, our analysis clearly indicates a risk of more severe impacts in the overshoot scenario compared to the strong mitigation scenario, particularly in the mid-term, highlighting the benefit of early mitigation strategies to avoid an overshoot scenario.