Here, we evaluate the representation of dissolved oxygen content and deoxygenation trends in the IPSL-CM6A-LR Large Ensemble simulations (Figs.  and ).

To investigate the response of OMZs to ongoing ocean deoxygenation, both phenomena need to be tracked within consistent spatial domains. This approach enables a direct comparison between OMZ dynamics and its corresponding regional deoxygenation trends. We define five fixed regional boxes corresponding to the major tropical OMZs (Figs.  and ): tropical North Pacific (0° S–30° N / 120–295° E), tropical South Pacific (30° S–0° N / 120–295° E), tropical North Atlantic (0° S–30° N / 60° W–20° E), tropical South Atlantic (30° S–0° N / 60° W–20° E) and North Indian (0° S–30° N / 35–120° E). As the IPSL-CM6A-LR model fails to represent oxygen minimum in the subarctic Pacific, this OMZ region has not been included in the analysis. We also restrict our analysis to the 100–1000 m depth range, where dissolved oxygen concentrations are at their lowest.

Ocean deoxygenation is quantified by tracking temporal trends in spatially averaged, both vertical and horizontal, dissolved oxygen concentrations within each box.

OMZ metrics are commonly defined as the ocean volume with oxygen concentration below a given threshold , hereafter referred to as the fixed-threshold approach. However, as demonstrated in the model evaluation, the IPSL-CM6A-LR Large Ensemble does not accurately reproduce the observed locations and extent of OMZ core waters, hypoxic waters and low-oxygen waters, limiting the relevance of analysing OMZ changes in the geographic space (Fig. ). To overcome this limitation, we adopt the fixed volume-percentile approach, based on the “oxygen water mass framework” developed by .

In this framework, water masses are characterised by the oxygen-percentile relation O2*(p,t) . The oxygen-percentile p represents the proportion of ocean volume with oxygen concentrations below O2* at time t, computed as , 1p(O2*,t)=100⋅V(O2*,t)/VT with VT being the total volume considered and V(O2*,t) is the volume of water with O2<O2*.

The oxygen-percentile relations are computed using both the WOA18 climatology and the IPSL-CM6A-LR mean averaged over the present day period (1970–2014) (Fig. ). It is computed within each of the five OMZ regions (Fig. ). The IPSL model captures the overall shape of the oxygen–percentile relation observed in the WOA18 dataset (Fig. ), although it generally overestimates oxygen concentrations. Consequently, for a given percentile, the model corresponds to a higher oxygen threshold, except in the tropical South Pacific and tropical South Atlantic. In the former, a positive bias is observed below the 47th percentile; in the latter, a positive bias appears below the 29th percentile and a negative bias above it (Fig. ). These biases are consistent with the model oxygen content surplus and spatial mismatches previously described.

To account for these systematic biases, we define OMZ volumes using percentile thresholds derived from WOA18. In each region, we identify the percentile corresponding to three key oxygen thresholds in the WOA18 climatology: 20 µmolkg-1 for OMZ core waters; 60 µmolkg-1 for hypoxic waters; and 120 µmolkg-1 for low-oxygen waters. In the Atlantic basin, oxygen concentrations in observations do not fall below 20 µmolkg-1; therefore, the OMZ core is not defined in the tropical North and South Atlantic (Fig. ). The percentiles associated with these thresholds are then applied to the IPSL oxygen–percentile curves, allowing consistent identification of the three OMZ volume categories in both observations and simulations (Fig. , Table ). This method enables us to track the evolution of OMZ volumes and their oxygen content over time while taking into account the model’s oxygen biases.

We further analyse the spatial distribution of OMZ volumes (Fig. ). Across all OMZ volume classes, the thickest areas of OMZ volumes in the IPSL simulations align with those observed in WOA18. However, in observations, OMZ thickness decreases sharply at their boundaries with sharp oxygen concentration gradient, whereas in the IPSL Large Ensemble mean, the transition is more gradual, resulting in more extensive OMZ volumes (Fig. ). Additionally, the North Indian OMZ core and hypoxic volume in the model are predominantly located in the Bay of Bengal, whereas observations indicate a signal both in the Arabian Sea and the Bay of Bengal (Fig. ).

To investigate the drivers of deoxygenation within the selected OMZ spatial boxes, we decompose changes in dissolved oxygen (O2) concentrations into two components: the saturation concentration of oxygen (O2sat) and the Apparent Oxygen Utilisation (AOU). This decomposition follows: 2ΔO2=ΔO2sat-ΔAOU The saturation concentration O2sat represents the oxygen solubility, which depends non-linearly on temperature and salinity. It is computed at each model grid point using monthly temperature and salinity fields and gsw-python package . The AOU reflects the net effect of biological oxygen consumption and physical ventilation of water masses. It is computed as the residual between modelled O2 concentrations and their corresponding saturation values.

To further interpret AOU variations in the 100–1000 m depth range, we examine two key drivers: the carbon export flux at 100 m and the mean age since surface contact. The carbon export flux serves as a proxy for biological oxygen demand contributing to AOU, while the mean age reflects water masses ventilation.

All variables are extracted or computed prior to the application of the piControl drift correction and are subsequently detrended using the same method applied to the O2 fields.

We use the concept of time of emergence to detect climate-driven changes exceeding internal variability. The time of emergence is defined as the last time step at which the climate-forced signal becomes statistically distinguishable from internal variability, referred to as the noise . This condition is formally expressed as: 3ToE:SN≥2 where S represents the climate-driven signal and N represents the noise. A ratio of 2 corresponds to a confidence level of 95 %, ensuring that the detected emergence is robust .

Each member within the IPSL Large Ensemble contains both the climate-driven signal and the system's internal variability . By conducting an ensemble of 30 experiments, we are able to separate these two components. The climate signal is estimated by averaging across ensemble members, while internal variability is quantified as the Large Ensemble standard deviation . For each biogeochemical variable and for each previously defined OMZ volume, climate-driven signals are extracted from anomalies, which are computed relative to the pre-industrial period (1850–1899). Internal variability is time dependent and estimated using absolute values rather than anomalies. Anomalies are calculated relative to the pre-industrial period for all Large Ensemble members, which eliminates the advantage of the different initial conditions in sampling the model’s internal variability. As a result, anomalies underestimate internal variability during the pre-industrial period. The use of time dependant internal variability is particularly relevant given that internal variability can evolve under climate change scenarios. It is particularly true in the IPSL-CM6A-LR large ensemble where the spread across ensemble members decreases under future forcing, reflecting the influence of anthropogenic external forcings on internal modes of variability .

Each ensemble member can also be interpreted as a distinct realisation of the climate response to the SSP2-4.5 scenario. Accordingly, we also compute the time of emergence for each individual simulation. In this case, the externally forced signal is taken from a single member, while the internal variability is still estimated using the ensemble standard deviation. This procedure yields a distribution of times of emergence across the ensemble.

Here, we present the ensemble mean climate-driven changes in OMZ core volumes, hypoxic volumes and low-oxygen volumes across each OMZ spatial domain, along with their corresponding time-dependent internal variability (Fig. ). We also examine the time of emergence of the climate-driven volume signal, defined as the year when the forced signal exceeds twice the magnitude of time-dependent internal variability (Fig. , Table ).

Here, we present climate-driven changes in mean dissolved oxygen concentration and its associated drivers (O2sat and AOU) across each OMZ spatial domain, along with their corresponding time-dependent internal variability (Fig. ). We also examine the time of emergence of these climate-driven signals, defined as the year when the forced signal exceeds twice the magnitude of time-dependent internal variability, quantified here as the Large Ensemble standard deviation (Fig. , Table ).

All OMZ regions exhibit a decline in mean dissolved oxygen concentrations throughout the simulation (Fig. ). The northern parts of the tropical Pacific and tropical Atlantic OMZs show the most pronounced deoxygenation, with decreases in dissolved oxygen concentration of 8.1 and 12 µmolkg-1, respectively, by the end of the simulation (Fig. a and c). These losses are greater than in the southern hemisphere OMZs of the same basin, which decline by 4.4 µmolkg-1 in the tropical Pacific and 4.0 µmolkg-1 in the tropical Atlantic (Fig. b and d). With the exception of the tropical South Atlantic, all regions experience an acceleration in deoxygenation during the SSP2-4.5 period (2014–2059) compared to the present-day period (1970–2014) (Fig. ). In the tropical South Atlantic, however, deoxygenation follows a nearly linear trend, with a present-day rate of 0.035 µmolkg-1yr-1 and a slightly slower rate of 0.029 µmolkg-1yr-1 during SSP2-4.5. During the SSP2-4.5 period, oxygen loss is more severe in the Northern Hemisphere OMZ regions: the tropical North Pacific experiences deoxygenation at twice the rate of the tropical South Pacific, and the tropical North Atlantic at more than five times the rate of the tropical South Atlantic (Fig. f–i). The North Indian Ocean undergoes the greater oxygen decline of all regions, with an 8.1 % reduction relative to its pre-industrial mean by the end of the simulation (Fig. j).

In all regions, the climate-driven deoxygenation signal emerges from internal variability before the end of the simulation period (Fig. , Table ). These signals emerge after the mid-20th century in all regions and before 21st century, beginning in the tropical North Atlantic in 1957, followed by the tropical South Pacific and South Atlantic in 1972 and 1974, respectively, then the North Indian Ocean in 1982, and finally the tropical North Pacific in 1994 (Fig. , Table ).

Decomposition of the oxygen signal into contributions from oxygen solubility (O2sat) and apparent oxygen utilisation (-AOU) reveals contrasting dynamics. A decrease in O2sat emerges in all regions during the first quarter of the 21st century, while the -AOU signal shows regionally divergent trends: it declines and exceeds their internal variability in the North Pacific, North Atlantic, and North Indian Ocean, but increases in the tropical South Atlantic and the tropical South Pacific (Fig. , Table ).

In all regions, O2sat emerges after deoxygenation (Fig. , Table ). Both O2 and O2sat decline gradually until about 2000, after which the decline accelerates sharply and continues to decrease throughout the simulation period (Fig. a–e). In all regions, the decline rate increases by a factor 10 between the periods before and after around 2004 (Fig. a–e and Table ). Emergence occurs close to or shortly after this regime shift: in 2003 in the North Pacific, 2004 in the South Atlantic, 2008 in the North Atlantic, 2018 in the South Pacific, and 2022 in the North Indian Ocean (Fig. , Table ). The acceleration is accompanied by a reduction in signal variability, consistent with emergence occurring near the regime shift (Fig. ).

Although all OMZ regions display a consistent trend of deoxygenation and O2sat decline, the -AOU component shows contrasting behaviours between hemispheres. In the tropical North Pacific and tropical North Atlantic, -AOU decreases by 5.3 and 7.1 µmolkg-1, respectively, by the end of the simulation (Fig. a and c). The emergence of this signal occurs later in the tropical North Pacific (2009) than in the tropical North Atlantic (1986) (Fig. a and c, Table ). By contrast, the tropical South Atlantic initially exhibits a slow decrease in -AOU at a rate of 5.0×10-3 µmolkg-1yr-1, followed by a sharp increase after 2004 at a rate of 0.086 µmolkg-1yr-1 (Table ). This reversal results in a total increase of 2.8 µmolkg-1 by the end of the simulation. The positive -AOU anomalies in this region become distinguishable from internal variability in 2045 (Fig. b and d, Table ). In the tropical South Pacific, the -AOU signal does not show a distinct regime shift point but shows a slower decreasing trend during SSP2-4.5 period, with a rate of 6.7×10-3 µmolkg-1yr-1, compared to 0.011 µmolkg-1yr-1 before 2004. In the North Indian Ocean, -AOU follows a decreasing trajectory, with its signal emerging concurrently with that of O2sat in 2022 (Fig. e, Table ).

Here, we present the ensemble-mean climate-driven changes in mean carbon export at 100 m and mean age since surface contact across each OMZ spatial domain, along with their corresponding time-dependent internal variability (Fig. ). We also examine the time of emergence of these climate-driven signals, defined as the year when the forced signal exceeds twice the magnitude of internal variability (Fig. ).

By disentangling the effects of biological activity and ventilation, the abrupt shift in the -AOU signal at about 2004 in the tropical North Atlantic coincides with simultaneous changes in both carbon export and water mass age (Fig. ).

Before 2004, carbon export remains stable across all OMZ regions, except in the tropical North Atlantic, where it increases at 9.4×10-3 gCm-2yr-1 (Fig. a–e and Table ). The North Indian Ocean shows the highest mean carbon export before 2004, with an export rate of 3.3 gCm-2yr-1. Tropical Pacific and tropical Atlantic OMZ regions exhibit a hemispheric asymmetry. The carbon export rates are higher in the southern hemisphere OMZ regions, with values of 1.9 gCm-2yr-1 in the tropical South Pacific and 2.3 gCm-2yr-1 in the tropical South Atlantic, than in the northern hemisphere OMZs of the same basin, which have an export of 1.6 and 1.8 gCm-2yr-1, respectively. After 2004, a pronounced decline in carbon export is observed over approximately two decades in the North Indian Ocean, tropical South Atlantic, and tropical South Pacific, before stabilising toward the end of the simulation (Fig. b, d, and e). The cumulative decline in carbon export over this period reaches 1.7 gCm-2yr-1 in the North Indian Ocean, 0.63 gCm-2yr-1 in the tropical South Atlantic, and 0.70 gCm-2yr-1 in the tropical South Pacific. By contrast, the tropical North Atlantic exhibits stable carbon export during this period, followed by a continued gradual increase at a rate similar to that observed before 2004 (Fig. d). Meanwhile, the tropical North Pacific shows a steady post-2004 increase in carbon export, at 4.6×10-3 gCm-2yr-1 (Fig. c and Table ).

Before 2004, the age since surface contact shows no statistically significant variations in any OMZ regions, except in the tropical North Atlantic, where it begins increasing in 1925 at a steady rate of 3.8×10-2 yryr-1, continuing through the end of the simulation (Fig. a). The tropical Pacific exhibits the highest mean age since surface contact before 2004, with values of 263 years in the tropical North Pacific and 227 years in the tropical South Pacific. The tropical Atlantic shows lower values, with 122 years in the tropical North Atlantic and 166 years in the tropical South Atlantic. After 2004, the age since surface contact decreases sharply in the tropical Southern Hemisphere OMZ regions, with total reductions of 7.3 years in the tropical South Pacific and 9.0 years in the tropical South Atlantic by the end of the simulation. These represent declines of 3.2 % and 5.4 % relative to their respective pre-industrial means. The corresponding rates of decline are 0.20 yryr-1 in the tropical South Atlantic and 0.15 yryr-1 in the tropical South Pacific (Fig. b and d). A slower decline is also observed in the North Indian Ocean, with a rate of 0.028 yryr-1 (Fig. e). These changes lead to the emergence of the ventilation signal from internal variability in 2032 in the tropical South Atlantic and in 2039 in the tropical South Pacific. In the tropical North Pacific, the age since surface contact also begins to decline around 2004, although at a slower rate of 0.040 yryr-1 (Fig. c).

Here, we present the ensemble distribution of the time of emergence of climate-driven changes in mean dissolved oxygen concentration and OMZ volumes across each OMZ spatial domain (Fig. ). We show the corresponding boxplots, along with the time of emergence and associated trend sign for each ensemble member (Fig. ).

For OMZ volumes and dissolved oxygen concentrations, the sign of the trend at the time of emergence of the ensemble mean is further supported by the emergence of individual ensemble members (Fig. ). In cases where the Large Ensemble mean does not emerge by the end of the simulation, some individual members nonetheless show emergence. They correspond to members exhibiting stronger variability around the externally forced trend, and thereby corroborate the trend direction of the ensemble mean. This is particularly evident in the tropical South Pacific and tropical South Atlantic, where individual members exhibit clear emergence of OMZ core and hypoxic volume contraction (Fig. b and d).

Across all OMZ regions, the ensemble mean time of emergence is generally earlier than the ensemble median, with large spread quantified by the interquartile range (Fig. ). The North Indian Ocean stands out as the only region where all metrics (dissolved oxygen concentrations and OMZ volumes) exhibit a narrow emergence window, with spreads limited to 17 years for oxygen signal and 25 years for OMZ core volumes (Fig. e). In other regions, broader spreads are linked either to weaker trend magnitudes at the time of emergence, as in the tropical South Pacific where deoxygenation shows an interquartile range of 54 years, or to regime shifts that change the trend sign, such as in the tropical South Atlantic where the low-oxygen volume signal displays a 71 year spread (Fig. b and d).

In all regions except the North Indian Ocean, the ensemble mean for dissolved oxygen concentration emerges earlier than, or concurrently with, the low-oxygen volume ensemble mean (Fig. ). However, due to the large spread among members in time of emergence, the distributions of deoxygenation and OMZ volumes time of emergence overlap in all regions (Fig. ).

Among the three OMZ regimes, only the low-oxygen volume exhibits a consistent expansion with a detectable emergence of the climate-driven signal from internal variability across all OMZ regions (Fig. ). This result is consistent with the findings of , who showed that OMZ volumes defined using higher oxygen thresholds expand across all OMZ regions. A comparison of their time of emergence also reveals regional offsets between ocean basins. Specifically, the tropical Atlantic OMZ shows an earlier emergence, by approximately two to three decades, compared to the tropical Pacific OMZ (Fig. ). This earlier emergence is not due to faster expansion, but rather to lower internal variability of the tropical Atlantic low-oxygen volumes, which makes the climate-driven trend detectable before the tropical Pacific low-oxygen volumes expansion (Fig. ).

Although low-oxygen volumes expand in all OMZ regions, subdividing the tropical Pacific and tropical Atlantic OMZs into northern and southern subregions reveals a consistent 30 year delay in the emergence of low-oxygen volumes in the Southern Hemisphere compared to the Northern Hemisphere (Fig. ). While internal variability remains similar across hemispheres within each basin, the magnitude of the climate-driven trend is stronger in the Northern Hemisphere in both basins (Fig. ). A similar hemispheric asymmetry is observed in the mean deoxygenation trends, where climate-driven declines in oxygen concentrations are stronger in the northern hemisphere OMZs than in the southern hemisphere OMZs of the same basin (Fig. ).

Low-oxygen volumes are critical as they influence the distribution of habitable zones for many marine species . In the IPSL-CM6A-LR model, the emergence of tropical Atlantic low-oxygen volumes occurs earliest, impacting oxygen-dependent ecosystem sooner than in other OMZ regions. Moreover, the expansion of low-oxygen waters is stronger in the Northern Hemisphere than in the Southern Hemisphere, leading to uneven impacts on marine ecosystems between hemispheres.

Moreover, the hemispheric asymmetry implies a stronger expansion of low-oxygen waters in the Northern Hemisphere than in the Southern Hemisphere, leading to uneven impacts on marine ecosystems between the two hemispheres.

The north-south subdivision of the OMZs shows asymmetric dynamics between the North and South parts of the OMZs. The core of the OMZs only expands in the northern hemisphere by the end of the simulation while they are shrinking in the southern hemisphere OMZs of the same basin (Fig. ). The tropical South Pacific and tropical South Atlantic OMZs show a contraction of their volume with lowest oxygen concentration, the OMZ core in the former and hypoxic waters in the latter, which is consistent with the findings of (Fig. ). In both tropical Pacific and tropical Atlantic, the southern OMZ core is being ventilated by younger water masses (Figs.  and ). This is consistent with the single-pipe and mixing network ventilation framework developed by . It suggests that in a warmer and more stratified ocean, the proportion of different ventilation sources shifts, favouring younger subsurface waters. In the tropical South Atlantic, ventilation is provided by younger subsurface waters originating from the South Indian Ocean, which are advected into the South Atlantic via the Agulhas Current . In the tropical South Pacific, ventilation occurs through water masses from the southern Pacific basin, which are advected northward by the Humboldt Current. This ventilation process is co-located with southern OMZ core location in the geographic-space (Figs.  and ). This oxygenation trend in the tropical South Atlantic basin is also exhibited by CMIP6 multi-model mean over the present-day period in and by the end of the 21st century over SSP1-2.6 and SSP5-8.5 in .

Unlike previous findings by , our results do not show a contraction of the OMZ core in the North Indian Ocean. This discrepancy likely stems from misrepresentation of oxygen concentrations and deoxygenation patterns in the IPSL-CM6A-LR model. The Large Ensemble simulations of the IPSL model misrepresent oxygen levels in the North Indian, failing to capture the low-oxygen conditions observed in the Arabian Sea (Fig. ). This bias is partly attributable to an overestimation of oxygen concentrations in marginal seas (the Persian Gulf and the Red Sea), combined with poorly constrained outflow parametrisations and a ventilation of the Arabian Sea OMZ by Southern Ocean water masses, where oxygen levels are overestimated . Furthermore, the model fails to reproduce the specific spatial patterns of oxygenation and deoxygenation identified in (Fig. ). Specifically, describe a distinct dipole structure with an oxygenation pool between 10° N and 10° S along the western boundary, and a deoxygenation pool between 10 and 30° S along the eastern boundary. This dipole structure supports a double-pipe ventilation mechanism within the North Indian basin. The IPSL-CM6A-LR model simulates a broad deoxygenation trend across the basin and fails to capture this ventilation pattern (Fig. ).

In the IPSL Large Ensemble, the OMZ regions mean deoxygenation signals emerges from natural variability before the end of the 20th century (Fig. ). However, tropical Pacific and tropical Atlantic OMZs exhibit an asymmetry in the relative timing of emergence between deoxygenation and OMZ volume signals (Fig. ). In the tropical North Pacific and tropical North Atlantic, the emergence of deoxygenation coincides with the expansion of low-oxygen volumes (Fig. ). In these regions, regional mean deoxygenation provides an indicator of OMZ boundaries expansion, with implications for the distribution of marine ecosystems. In the southern parts of the tropical Pacific and tropical Atlantic OMZs, the regional mean climate-driven deoxygenation signals emerge earlier than the expansion of low-oxygen and other OMZ volumes (Fig. ). Hypoxic waters, which are harmful or lethal to many marine organisms, emerge later than mean deoxygenation, with delays of 27 years in the tropical North Pacific and 19 years in the tropical North Atlantic (Fig. ).

Although the emergence of regional mean dissolved oxygen captures the regional-scale evolution of low-oxygen waters in IPSL simulations, the emergence of local oxygen trends exhibits temporal heterogeneity. Consistent with the multi-model analysis of , who computed time of emergence at each grid point using CMIP5 models under the RCP8.5 scenario , the IPSL-CM6A-LR spatial median of local emergence occurs later than the regional mean deoxygenation signal in each OMZs regions (Fig. , Table ). Except for the tropical North Atlantic, where the median emergence occurs in 1983, spatial median emergence times are in the first half of the 21st century (Table ). This is consistent with , who exhibited a global mean oxygen emergence in 2019 for the CMIP5 IPSL-CM5A-LR model. Within OMZ regions, local emergence times shows a standard deviations ranging from 7 to 24 years (Fig. , Table ). This is consistent with results of , who used the Community Earth System Model large ensemble (CESM Large Ensemble, ) under RCP8.5 to assess local deoxygenation time of emergence. The standard deviation of time of emergence within the OMZ regions reflects the delay of emergence between the three OMZ volumes of interest.

In this study, we determine the emergence of OMZ volumes relative to pre-industrial using the IPSL-CM6A-LR model. Model simulations provide access to a pre-industrial baseline that is not available in current observational records, making it possible to characterise the full magnitude of climate-driven changes. However, observational oxygen datasets, such as the IAP dataset , only cover the period from 1960 to the present and contain limited data within OMZs. As a result, using these datasets would only allow us to determine the emergence of deoxygenation and reconstructed OMZ volumes relative to the 1960–1970 period. The IPSL Large Ensemble simulations allow us to adopt the 1960–1970 period as a reference to compute time of emergence, providing an estimate of when climate-driven signals in OMZ volumes may become detectable using observation datasets. Using the same method as our previous analysis, times of emergence of dissolved oxygen and OMZ volumes anomalies are computed relative to the 1960–1970 period for both the ensemble mean and individual members. Internal variability is estimated by the Large Ensemble standard deviation.

Except for the hypoxic volume in the South Atlantic, all time of emergence distributions shift towards later emergence times, with the ensemble-mean emergence delayed by up to 59 years (Fig. ). However, all ensemble-mean signals that emerged with a pre-industrial reference also emerge before the end of the simulation (2059) when the 1960–1970 reference is used (Fig. ). Thus the climate-driven emergence of OMZ volumes is ongoing, although it remains challenging to detect in observations. The detection challenge arises from the difficulty of measuring OMZ volumes and accurately estimating their natural variability, which is entangled with their anthropogenic response.

The IPSL-CM6A-LR model exhibits a positive bias in dissolved oxygen concentrations, particularly in the OMZ regions. This oxygen overestimation in tropical OMZs is linked to the IPSL-CM6A-LR coarse spatial resolution, which leads to weaker equatorial currents . To assess large-scale physical ventilation, we compare age since surface contact at the end of the piControl simulation (years 3840–3850) with two data-driven inverse models: the Total Matrix Intercomparison (TMI) and the Ocean Circulation Inverse Model (OCIM) . The IPSL model reproduces shadow zones in the tropical regions of interest, indicating the representation of large-scale physical ventilation of tropical regions (Fig. ). The oxygen overestimation is then also linked to too strong deep convection in the Southern Ocean . Thus we used fixed-percentile approach, rather than fixed-threshold approach, to define OMZ volumes derived from the WOA observational product.

However, this approach does not correct for biases inherent to WOA18, which is known to underestimate hypoxic volumes . To account for these uncertainties, we evaluate the sensitivity of our results to alternative OMZ definitions (fixed-threshold versus fixed-percentile), such that the resulting range of responses encompasses uncertainties from IPSL oxygen overestimation and WOA18 underestimation of OMZ volumes (Fig. ). The sign of the climate-driven emergence trends are consistent across both definitions (Fig. e–j). However, the percentile-based approach allows to detect the emergence of OMZ core and hypoxic volume signals, where the fixed-threshold method fails (Fig. e–j). This reflects the fact that the classical method underestimates or misses the climate-driven signal due to the biased baseline in oxygen levels. The time of emergence of low-oxygen volumes is relatively insensitive to the OMZ definition used, reflecting the smaller oxygen bias at higher concentrations in IPSL-CM6A-LR (Fig. k–o).

The time of emergence is highly sensitive to the representation and the quantification of the internal variability of the signal of interest . In this study, internal variability was measured by the standard deviation of the IPSL Large Ensemble. The 2000 years piControl simulation provides an alternative estimate of internal variability, capturing decadal to millennial fluctuations. However, as it is conducted under pre-industrial forcing conditions, it does not account for changes in internal variability induced by historical and SSP2-4.5 forcing.

To evaluate the impact of the variability measure, we computed time of emergence using, for the noise term, the Large Ensemble standard deviation and the constant standard deviation derived from the full piControl simulation. The emergence the ensemble mean signal is not substantially affected by the choice of variability estimate (Fig. ). However, when internal variability is estimated from the piControl simulation, the emergence distribution across ensemble members shifts towards later times (Fig. ). Consistent with previous findings of using the CESM1 model under the RCP8.5 scenario, piControl-based variability tends to overestimate internal variability, resulting in a later time of emergence.