We have performed the first 17O-excess measurements on the NEEM ice core localisation on Fig. , spanning a sequence of 7 DO events during MIS 3, from GI-7 to GI-13. The method for 17O-excess measurements was first described in . Two microlitres of water are fluorinated under a helium flow and the resulting oxygen O2 is purified on a molecular sieve before being trapped in a manifold immersed in liquid helium. O2 is then measured on a DELTA V isotope ratio mass spectrometer vs. pure oxygen. Dual inlet measurements last 75 min for each converted oxygen sample. Each water sample has been converted and measured at least twice. Every day, we use two in-house water standards in our fluorination line to check the stability of the measurements and to perform calibrations. The resulting pooled standard deviation for 17O-excess measurements is 5 per meg.

For this study with δ18O varying between -43 and -38 ‰, we mainly used three in-house standards at around -32, -40 and -58 ‰. These in-house standards are calibrated vs. V-SMOW (Vienna Standard Mean Ocean Water, δ18O=0‰; 17O-excess=0permeg) and SLAP Standard Light Antarctic Precipitation, δ18O=-55.5 ‰; 17O-excess=0permeg,. The calibration of our raw 17O-excess data can then be performed in two different ways. In a first attempt, we have simply subtracted from all raw 17O-excess data the 17O-excess difference between the measured and the calibrated standard at -40 ‰. In a second attempt, we have used a two-point calibration between measured and accepted values of V-SMOW and SLAP so that the 17O-excess correction increases with the δ18O difference between O2 obtained from the sample conversion and O2 obtained from V-SMOW conversion. We show in Fig. b the comparison of 17O-excess evolution after these two corrections as well as the raw data. The general evolution of the 17O-excess profile and the particular separation in phases over GS-9 is not affected by the different corrections.

Finally, note that the measurements were done in three rounds, with a first series measured in spring 2011, a second one at the beginning of 2012 and a third one in spring 2014. Comparisons of the profiles are displayed in Fig. a after correction with the two-point calibration (V-SMOW vs. SLAP). The measurements were not necessarily performed over the same exact depth levels, which makes the inter-comparison less accurate. Both the mean levels and the variability are very coherent between the three 17O-excess profiles.

Our new δ18Oatm (δ18O of O2) data set measured on the NEEM ice core consists of 95 data points with replicates. A melt–refreeze technique has been used to extract the air from the ice samples. Our isotope ratio mass spectrometer is equipped with 10 Faraday cups to measure simultaneously the isotopic ratios δ18O and δ15N as well as the elemental ratios O2/N2 .

To reconstruct the past δ18Oatm signal, δ18O measurements have been first corrected for eventual gas loss during ice storage using the δO2 / N2 ratio measured in the same samples more details can be found in: δ18Ogas   loss   corrected=δ18Omeasured+0.01δO2/N2+10. The obtained data are then corrected for thermal and gravitational fractionation occurring in the firn. To do so, we use the isotopic composition of nitrogen δ15N measured in the very same ice samples as well as firn modelling, to reconstruct the gravitational (δ15Ngrav) and thermal (δ15Ntherm) signals contributing to the measured δ15N data and references therein. The gravitational effect being proportional to the mass difference in between the two isotopes , δ18Ograv is therefore twice as large as δ15Ngrav. Temperature gradients in the firn create thermal fractionation of nitrogen and oxygen isotopes, and this effect is 1.6 times larger for δ18O compared to δ15N due to differences in diffusivity coefficients . The final δ18Oatm is therefore obtained as follows: δ18Oatm=δ18Ogas loss corrected-1.6δ15Ntherm-2δ15Ngrav. The resulting δ18Oatm pooled standard deviation is 0.03 ‰ . The δ18Oatm data set is reported on the GICC05 timescale (Fig. f, black dots) using the NEEM gas age scale from . The obtained profile is in agreement with previously published results from measured on the Siple Dome ice core, Antarctica (Fig. f, grey dots). The advantage of using the NEEM rather than the Siple Dome δ18Oatm data resides in the smaller NEEM ice–gas synchronisation uncertainty during GS-9 (Appendix A2).

The entire 17O-excess record covering GI-7 to GI-13 (Fig. d) shows a general correlation with the δ18O record and an anticorrelation with d-excess, while δ18O and d-excess are themselves well anticorrelated: GI corresponds to high δ18O and 17O-excess levels and low d-excess levels, GS corresponds to low δ18O and 17O-excess levels and high d-excess levels. Following the current understanding of the 17O-excess proxy as given in introduction, such a general correlation with δ18O can be explained by (i) a synchronous change of high-latitude temperature and lower-latitude hydrological cycle (relative humidity at the oceanic source of evaporation or change in the water mass trajectory/recharge) and/or (ii) an influence of local temperature on 17O-excess through kinetic effect at snow formation. In addition to this general correlation with δ18O, the high-resolution 17O-excess record reveals periods where 17O-excess is decoupled from δ18O. The most obvious decoupling between δ18O and 17O-excess is observed during GS-9 (Fig. d, green arrows): 17O-excess decreases after the δ18O decrease at the beginning of GS-9 and increases several centuries before the δ18O increase at the beginning of GI-8. In fact, within GS-9 of constantly low δ18O, we can infer three distinct phases from 17O-excess variations (Fig. d, Table ): phase 1 (39 900–39 350 a b2k), corresponding to the beginning of GS-9, is marked by high 17O-excess values (mean = 55 per meg); phase 2 (39 350–38 610 a b2k) corresponds to low 17O-excess values (down to 34 per meg, mean = 42 per meg); finally during phase 3 (38 610–38 220 a b2k) 17O-excess reaches high interstadial levels again (mean = 57 per meg).

The decoupling observed between δ18O and 17O-excess implies that variations of 17O-excess during GS-9 cannot be due to local (Greenland) temperature changes. This conclusion is strengthened by an independent temperature reconstruction based on firn gas fractionation δ15N, Fig. b and confirming a cold and stable temperature over the entire GS-9. The 17O-excess variations during GS-9 thus necessarily reflect changes in the organisation of the hydrological cycle of the lower latitudes, either through their impact on relative humidity at the source of evaporation, or through a change in the recharge of water along transportation towards Greenland. It is counter-intuitive that such changes in the evaporation conditions or moisture source do not affect the d-excess signal. While we do not have a definite explanation for the lack of signal in the d-excess record, we already know that d-excess in Greenland is much more sensitive to local temperature changes than 17O-excess . Moreover, because of its linear definition, d-excess is not sensitive to mixing of different water masses along the trajectory. A simple mathematical calculation shows that the mixing of two water vapour masses with similar d-excess and 17O-excess but different δ18O produces a stable d-excess but a decrease in 17O-excess.

Because of the complexity of processes which may be responsible for the specific 17O-excess signal, it is thus important to confront our finding to other low-latitude proxies registered in Greenland ice cores, to independently confirm the decoupling between Greenland and lower latitudes during GS-9 as highlighted by the 17O-excess record.

Our new high-resolution δ18Oatm measurements of air entrapped in the NEEM ice core confirm a decoupling between Greenland temperature and lower-latitude water cycle over GS-9. Indeed, earlier studies have shown that millennial δ18Oatm variations are global responses to water cycle changes in the northern Tropics , where ITCZ shifts drive changes in precipitation isotopic composition . Our 95 new measurements of δ18Oatm depict a stable, low δ18Oatm during phase 1, followed by an increase over phase 2 (+0.14 ‰, Fig f, black arrow), and finally a stable high plateau during phase 3, in agreement with the Siple Dome record (Fig. f). The onsets of changes of δ18Oatm and 17O-excess occur synchronously at the beginning of phase 2 (ice–gas synchronisation uncertainty less than ±100 a for the NEEM record, Appendix A2), but the increase of δ18Oatm is much slower than the observed changes in 17O-excess; this is expected because of the long residence time of oxygen in the atmosphere 1000–2000 a,. Second, in addition to the δ18Oatm record, records of NEEM CH4 Fig. g and its hydrogen isotopic composition Fig. e, NGRIP δD-CH4, in the air trapped in ice core also exhibit significant variations during GS-9 that are synchronous with the three phases identified from the 17O-excess variations. Phase 1 is characterised by a minimum in methane mixing ratio, whereas the onset of phase 2 is marked by a 20 ppb abrupt increase (Fig. g, red arrow) accompanied by a 8±4.8 ‰ increase in δD-CH4. At the onset of phase 2, the concomitant changes in CH4, δD-CH4, δ18Oatm and 17O-excess in the absence of any Greenland temperature change can consistently be explained by a low-latitude change in water cycle as induced by a southward shift of the ITCZ.

Concerning the measured increase in δD-CH4 at the onset of phase 2, such a shift would also be in agreement with model simulations of the clathrate release signature in ice cores , and is also supported by the synchronously lighter δ13C-CH4 anomaly recently measured in the EDML ice core . We thus cannot exclude a sudden clathrate release at the onset of phase 2 that might be caused by e.g. less dense water masses in the North Atlantic Ocean following e.g. a massive iceberg discharge, as modelled by . Higher-resolution methane isotopic data would be necessary to investigate in detail this possibility. However, the long duration of the δD-CH4 anomaly (phase 2, 740 a) as well as the stable plateau of methane mixing ratio and δ13C-CH4 throughout phases 2 and 3 (1130 a) call for an additional mechanism, such as source mix change or/and heavier isotopic composition of tropical precipitation at the onset of phase 2. The mixing ratio increase and the following plateau are consistent with the activation of new methane sources in e.g. South America as simulated by , associated with a southward ITCZ shift at the onset of phase 2. Moreover, such a southward ITCZ shift is expected to produce more depleted precipitation in the SH (Southern Hemisphere) Tropics and heavier precipitation in the NH (Northern Hemisphere) Tropics . The increase in both δD-CH4 and δ18Oatm at the onset of phase 2 are consistent with this mechanism, provided that NH Tropics remain the main source of methane and oxygen. A minor part of the δD-CH4 increase can be due to oxidation of methane in the troposphere, consuming preferentially the light methane isotopologues, if the mean atmospheric temperature decreases at the onset of phase 2 .

Finally, two abrupt ∼20 ppm increases in the atmospheric CO2 concentration recently unveiled from high-resolution Antarctic ice core records occur at the onsets of phase 2 and phase 3 Fig. h and. After Antarctic and Greenland gas record synchronisation through CH4 (Appendix A2), our study evidences that the CO2 rise at the onset of phase 2 is synchronous (±100 a) with a southward shift of the ITCZ, as suggested by the other NEEM ice core proxy variations. Different mechanisms are proposed to explain this CO2 rise, as already discussed in : most probably upwelling in the Southern Ocean around Antarctica as suggested by increased opal burial rates in marine cores , with a potential minor contribution from upwelling off the NW African coast . The causes of such modifications of the oceanic carbon storage may involve strengthening and/or southward shift of the SH westerlies, and/or AMOC slow down, and therefore remain debated e.g..

Altogether, our multi-proxy ice core data set reveals a 740 a long period starting 550 a after the onset of GS-9 (phase 2), which is marked by a southward shift of the ITCZ. The transition to phase 3 is characterised by reversed variations, with 17O-excess and δD-CH4 reaching interstadial values 480±50 a prior to the increase in δ18O marking the onset of GI-8. We explain this pattern by a gradual northward shift of the ITCZ associated with lighter precipitation in the NH Tropics and heavier precipitation in the SH Tropics. The flat methane and δ18Oatm records suggest a globally stable methane budget and biosphere productivity, possibly due to compensating effects at different latitudes.

The sequence of events identified during HS-4 contrasts with the other non-Heinrich stadials where the pattern of changes recorded in Greenland δ18O and other proxy records mostly depict a GS/GI contrast. These other GS are mostly characterised by low δ18O and 17O-excess levels, low methane mixing ratio (without any plateau) and slightly increasing δ18Oatm (compared to the strong increase observed during HS-4). We therefore underline that the three-phase sequence identified during HS-4 is really specific and should not be applied to the other stadials studied here.

Some fingerprints of the decoupling between high and low latitudes identified during HS-4 can also be evidenced during HS-5 (or GS-13) even though less proxy data are currently available. We observe synchronous δ18Oatm and methane increases about 140±40 a after the beginning of GS-13, at 48 200 a b2k on the GICC05 chronology, 1340 ±80 a before the onset of GI-12 (Fig. f and g, black and red arrows). This shift is accompanied by a small 17O-excess decrease (Fig. d, green arrow). The CO2 increase that was clear at the onset of phase 2 during GS-9 is however much less pronounced or perhaps even absent at the onset of phase 2 during GS-13 +5 ppm in TALDICE, less clear in the Byrd CO2 record,and Fig. h. Finally, based on the available data we do not observe any decoupling between Greenland and lower latitudes toward the end of GS-13 as was observed over the phase 2 to phase 3 transition during GS-9. However, δD-CH4 data are lacking during this time period and we cannot draw any firm conclusion here.

Low/mid-latitude palaeoclimate archives have already provided information on climate variations during Greenland stadials and especially over Stadial 9 e.g.. In particular, several phases have already been identified in sediment cores off the European margin and especially off Iberia, based on pollen assemblages and charcoal records : in the first phase, before the IRD layer and until the maximum IRD content is reached, the climate follows a progressively cooler and wetter trend. In the second phase after the IRD peak response, the climate is on the contrary progressively drier, with more frequent fire regime episodes.

Hereafter, we combine North Atlantic marine cores from different latitudes along the European margin, located between 5 to 10∘ W and 40 to 57∘ N (Figs.  and ). Note that in Fig. , each marine sediment core is on its own depth scale. The available chronological control points are 14C measurements (in orange) and the NAAZ-II (North Atlantic Ash Zone II) tephra layer also identified in the NGRIP ice core (in violet, see figure for references). We have only reported the 14C measurement values and associated analytical uncertainties when available, without reservoir age correction or calibration to calendar ages. Indeed, for this time period the 14C measurement uncertainties (minimum ±380 years here) and their scarcity already prevent any centennial-scale synchronisation of these marine sediment cores.

We hereafter compare in each core the relative timing of changes in Neogloboquadrina pachyderma sinistral percentage (a proxy for sea surface temperature, black dots in Fig. ) with the potential Laurentide IRD deposition blue bars in Fig. , seefor details. As for the multi-proxy study of GS-9 in ice cores, the multi-proxy study of these marine cores in the European margin shows some latitudinal decoupling during stadial 9: (i) prior to the Laurentide IRD delivery, cold conditions are already depicted in the eastern part of the so-called Ruddiman belt , at 45–55∘ N e.g. Fig. a, core MD95–2006,, while interstadial conditions persist around the Iberian Peninsula at 35–42∘ N e.g. Fig. d, core MD95-2040,; (ii) the onset of the Laurentide IRD delivery coincides with a possible further cooling at high latitudes, while Iberia enters into cold stadial conditions; (iii) once low Laurentide IRD levels are registered again, mild to interstadial conditions are recorded around the Iberian Peninsula, while cold temperatures persist at 45–55∘ N. Between 40 and 57∘ N, cores MD04-2845 and MD95-2002 suggest that the situation might be intermediate, with a slight lag of the Laurentide IRD layer behind the polar species N. pachyderma sinistral increase Fig. b, c,. Higher-resolution data as well as climate-independent synchronisation tools in between marine cores (such as measurements to detect specific palaeomagnetic events and/or volcanic ash layers) are necessary for a better description.

Over stadial 9, chronological uncertainties prevent us to link ice cores to other archives at a centennial year resolution, due to uncertainties in the GICC05 timescale ±1574 a at 40 000 a b2k, and in 14C measurement, reservoir age and calibration curve for marine sediment cores Fig.  and e.g.. However, we still note that such a southward ITCZ shift can explain both the fingerprints in CH4, δD-CH4, 17O-excess and δ18Oatm, as well as the colder conditions registered in the Iberian Peninsula that correspond to the purge of the LIS, i.e. the beginning of H4. Indeed, the H4 iceberg discharge could lead to European cooling as well as southward ITCZ shift, as simulated by climate models . With this assumption, the inferred start of H4 would be 39 350 ± 1520 a b2k on the GICC05 chronology.

During phase 3, we note a progressive shift towards interstadial conditions first indicated by lower-latitude ice core proxies (Fig. , δD-CH4, 17O-excess). This would be consistent with a progressive northwards progression of the climate recovery towards interstadial climate conditions. This is in agreement with marine cores from low/mid-latitudes suggesting the start of a progressive AMOC recovery before full interstadial climate conditions are reached e.g.. The combination of ice and marine core data therefore supports the following sequence of events in phase 3: end of the iceberg delivery in the North Atlantic, northwards shift of the ITCZ and progressive AMOC restart, and finally the abrupt Greenland warming.

Finally, while some common characteristics can be drawn for HS-4 and 5, further studies are needed to assess whether this can be applied to the other Heinrich stadials or not. Considering the different background climate states and origin of the ice sheet collapse for the other HEs (Fig. ), a different sequence identification in the ice core records can be expected.

Our study shows that a change in the mid/low latitudes of the North Atlantic ocean happens after the beginning of GS-9, with a time lag of 550±60 a, i.e. when the North Atlantic region has already entered a cooling trend. Some features of our study suggest that the same may be true for GS-13 with a decoupling between Greenland temperature and lower-latitude climate 140±40 a after the beginning of GS-13. We have hypothesised that this change may be linked to iceberg deliveries, i.e. H4 and H5. This assumption is actually supported by recent studies investigating the triggering mechanisms of HEs. Indeed, it has been suggested that the Northern Atlantic temperature decrease (beginning of GS in our case) is concomitant with the onset of a progressive AMOC slow-down and shoaling accompanied by sea ice extent that isolate the subsurface ocean from the cold atmosphere and produce a subsurface warming . In addition, this subsurface warming could lead to a sea-level rise of ∼0.5 to 1 m by reorganisation of the North Atlantic dynamic topography and by oceanic thermal expansion as modelled in. Both subsurface warming and sea-level rise then contribute to destabilising ice shelves from the massive Laurentide ice sheet, causing the massive iceberg delivery of Heinrich events occurring several centuries after the onset of the Greenland stadials .

The suggestion that HEs 4 and 5 occur after the beginning of Greenland stadials calls for another mechanism to trigger stadial conditions. Recent studies listed hereafter have proposed a local (Arctic) control on DO events, without the need of a freshwater flux into the North Atlantic Ocean. (i) The studies of Norwegian Sea marine sediment cores from and proposed that stadials were induced by progressively increasing winter sea ice cover in the Nordic seas and the formation of a halocline isolating warm subsurface waters from the North Atlantic drift from the atmosphere. (ii) These studies also showed IRD from the Fennoscandian ice sheet during the warm phases of DO events. The authors proposed that the corresponding freshwater flux helped to establish a halocline and hereby facilitated sea ice formation, inducing a cooling towards a stadial climate. A model study from has moreover shown that the overturning cell in the Nordic seas is very sensitive to freshwater perturbations, even more than the Ruddiman belt in the North Atlantic. (iii) Finally, a third mechanism (that could be combined with the two previous ones) has been proposed by : the onset of GI could be induced by the collapse of an ice shelf east of Greenland extending towards Denmark Strait, due to subsurface melt. The slow regrowth together with the expansion of regional sea ice cover would create the progressive cooling towards GS. Moreover, these three mechanisms could explain why the Nordic seas (and Greenland) remain in a stable stadial state, while low/mid-latitudes are gradually shifting towards interstadial conditions (as suggested by the increased 17Oexcess during phase 3): during stadials, the heat advected along the North Atlantic Drift accumulates in subsurface, below the sea ice in the Nordic seas or below the ice shelf east of Greenland, until the low-density warm subsurface water abruptly reaches the surface, thereby melting the sea ice and/or ice shelf, which results in a rapid warming of the Nordic seas (and Greenland).

These different mechanisms calling for a local Arctic control on DO events are supported by model studies. Indeed, atmospheric simulations suggest that Greenland temperature is sensitive to changes in sea ice cover anomalies in the Nordic seas because of “a strengthening of the easterly flow over the Nordic seas” impacting Central Greenland. In contrast, in these simulations, NW Atlantic sea ice cover anomalies (supposed to increase during HEs originating from the Laurentide) only affect the western flank of the Icelandic low and the associated atmospheric circulation anomaly does not noticeably impact Central Greenland. Supporting these results, the modelled regional pattern of Greenland warming in response to Nordic seas sea ice retreat is in agreement with regional quantitative reconstruction of Greenland temperature and accumulation rate during MIS3 .

From our data, we found that central Greenland temperature is not sensitive to lower-latitude climate changes during stadials. This is actually not a unique case. Other palaeoclimate archives show the same insensitivity, e.g. no specific sea ice change is depicted during the entire stadial 9 as inferred from deep sea core records obtained from the Irminger Sea and south of the Faroe Islands .

However, bio-indicators in marine cores from the Norwegian Sea robustly record warmer temperature e.g. based on foraminifers, and reduced sea ice cover based on dinocysts, during stadials than during interstadials. To reconcile these conflicting model–data results, and proposed a perennial sea ice cover in the Norwegian Sea during all stadials of MIS3, separated by a halocline from warm subsurface waters in the North Atlantic Drift, to where foraminifers might have moved their depth habitat. Dinocysts on the contrary may have stayed at the surface, and their observed assemblages could be due to productivity anomalies in zones of polynya. According to this scenario, the lack of specific sea ice anomaly in the Nordic seas during GS-9 and GS-13 may explain why no cold anomaly is recorded in Greenland during the same periods.

The proposed decoupling between cold, stable Greenland temperature and lower-latitude climatic changes during stadial 9 has implications for synchronisation of climate archives from different latitudes. For the last glacial period, chronologies of sediment cores from the North Atlantic and Europe are usually constructed by wiggle matching of DO-like variations to the NGRIP δ18O record on its GICC05 chronology. While it has been underlined by several studies that there were no independent chronological support for this hypothesis , the decoupling between 17Oexcess and δ18O variations during GS-9 suggests that this would actually not be correct for this specific time period. We therefore would like to emphasise the need of high-resolution climate-independent synchronisation tools, such as volcanic ash layers, as already widely encouraged by the INTIMATE working group e.g..