Abrupt Heinrich Stadial 1 cooling missing in Greenland oxygen isotopes

Glacial wintertime sea ice mutes abrupt Heinrich Stadial 1 cooling in Greenland oxygen isotope records.

Abstract

Abrupt climate changes during the last deglaciation have been well preserved in proxy records across the globe. However, one long-standing puzzle is the apparent absence of the onset of the Heinrich Stadial 1 (HS1) cold event around 18 ka in Greenland ice core oxygen isotope δ¹⁸O records, inconsistent with other proxies. Here, combining proxy records with an isotope-enabled transient deglacial simulation, we propose that a substantial HS1 cooling onset did indeed occur over the Arctic in winter. However, this cooling signal in the depleted oxygen isotopic composition is completely compensated by the enrichment because of the loss of winter precipitation in response to sea ice expansion associated with AMOC slowdown during extreme glacial climate. In contrast, the Arctic summer warmed during HS1 and YD because of increased insolation and greenhouse gases, consistent with snowline reconstructions. Our work suggests that Greenland δ¹⁸O may substantially underestimate temperature variability during cold glacial conditions.

INTRODUCTION

Glacial periods exhibit Dansgaard-Oeschger (DO) and Heinrich modes of abrupt climate variability, both of which are linked to changes in Atlantic meridional overturning circulation (AMOC) (1, 2). Heinrich stadials are generally believed to represent more-complete AMOC shutdown and, thus, more-extreme winter cooling than DO stadials (3). Greenland ice core δ¹⁸O of water isotopes is a commonly used proxy for site temperature due to the “temperature effect” (4) and has been considered as a key index that records abrupt climate changes around the Arctic region. However, Greenland δ¹⁸O fails to differentiate the DO and Heinrich stadials from similar δ¹⁸O levels in Heinrich stadials and regular DO stadials. This contrasts proxies for Atlantic circulation and temperature (5, 6) and Asian monsoon (7) that show clearly distinct signatures of the Heinrich stadials. Consequently, the commonly used Greenland event stratigraphy (8) only identifies DO stadial-interstadial transitions but not Heinrich stadials. These contrasting observations raise a fundamental question: Was Greenland climate not affected by Heinrich variability, or was Heinrich variability not recorded in Greenland δ¹⁸O?

Northern Hemisphere climate during the last deglaciation was punctuated by abrupt events such as Heinrich Stadial 1 [HS1; ~18 to 14.7 thousand years ago (ka)], Bølling-Allerød warming (BA; 14.7 to 12.9 ka), and Younger Dryas cooling (YD; 12.9 to 11.7 ka). These events occur coherently in Northern Hemisphere proxy records (Fig. 1) (9, 10). Here, we focus on the last Heinrich event, the HS1, a period of weakened AMOC from 18 to 14.7 ka before present (BP) (11), which contains the abrupt Heinrich event 1 ice-rafted debris layers around 16.2 ka BP (12), while Greenland δ¹⁸O shows virtually no depletion signal around the HS1 onset at ~18 ka BP (Fig. 1B and fig. S1) (13, 14). This is in contrast to other proxies across the Northern Hemisphere (Fig. 1, E and F) (9, 10), from North Atlantic sea surface temperature (SST) (Fig. 1E) (15) to Asian monsoon speleothem calcite isotope δ¹⁸O_c (Fig. 1F) (7). This apparent lack of δ¹⁸O response over Greenland has led to the speculation that Greenland did not cool during HS1, implying a decoupling of the Arctic from hemispheric climate trends (16). Here, combining climate proxies with transient isotope-enabled climate model simulations, we resolve this puzzle and suggest that the HS1 cooling did occur but was not recorded in ice core δ¹⁸O.

Fig. 1. Water isotope–climate evolution during the last deglaciation.

(A) Climate forcings including June solar insolation at 60°N (red), CO₂ concentration (green), meltwater forcing at the Northern Hemisphere (blue) and the Southern Hemisphere (orange). ppm, parts per million. (B) Modeled (purple) and observed ice core δ¹⁸O at Greenland Ice Sheet Project 2 (GISP2) (black) (13) and GRIP (gray) (49). (C) As in (B) but for modeled (red) and reconstructed (see legends and text) temperature. (D) As in (B) but for modeled (blue) and reconstructed (black) (28) ice accumulation rate. (E) Modeled (red) and observed (black) SST at the Iberian margin. (F) Modeled (purple) and observed (black) (7) speleothem δ¹⁸O_c in Hulu Cave. (G) Modeled AMOC intensity (blue) and ²³¹Pa/ ²³⁰Th in sediment core GGC5 as a proxy for AMOC intensity (black) (11). (H) Model global distribution of δ¹⁸O and temperature at Last Glacial Maximum (LGM). In (B, C, E, and F), the scales are the same in the model and observation but with an offset. In (D), the scale of reconstructed ice accumulation rate is 2.5 times larger than the modeled one. In (F), the modeled speleothem δ¹⁸O_c is calculated as δ¹⁸O_c = δ¹⁸O − 0.24ΔT, with ΔT as the annual temperature anomaly from LGM (18). In (H), red (green) contours represent temperature above (below) zero (counter interval, 5°C), with the white contour as 0°C. A five-decade running mean is applied to model time series.

RESULTS

Deglacial evolution of water isotopes and climate

We simulate the coevolution of climate and water isotopes during the last deglaciation (20 to 11 ka) in the water isotope–enabled Community Earth System Model (iCESM) (17). We perform a set of four simulations with four realistic forcing factors (Fig. 1A) applied additively: first, ice sheet and bathymetry (ICE), then insolation (ICE+ORB), the greenhouse gases (ICE+ORB+GHG), and, last, meltwater fluxes (ICE+ORB+GHG+MWF) or iTRACE, for the isotope-enabled transient climate experiment (18).

iTRACE quantitatively captures many major features of water isotopes and climate variations during the last deglaciation from the North Atlantic SST to Asian monsoon δ¹⁸O_c (Fig. 1, E and F). Overall, global surface temperature rises during the deglaciation, with an interhemispheric bipolar seesaw during HS1 and YD in response to the AMOC slowdown that is forced by meltwater fluxes mainly to the North Atlantic (Fig. 1, A and G). Over Greenland, the model simulates clear BA and YD events that match reconstructed changes in precipitation δ¹⁸O (~4‰), temperature (~11°C), and the variation of snowfall accumulation rate (Fig. 1, B to D). The corresponding overall temporal ∆δ¹⁸O/∆T regression slope from 20 to 11 ka is ~0.36‰ K⁻¹ [as can also be estimated directly from Fig. 1 (B and C) as ≈4‰/11°C], broadly consistent with reconstructions from borehole temperature (19), deuterium excess (20), nitrogen and argon isotope fractionation (21), and a reanalysis technique (Fig. 1C) (22). Simulated mean annual δ¹⁸O remains stable through HS1 from central to northern Greenland (fig. S2), consistent with ice core records there (Fig. 1B and fig. S1, A and B). This muted δ¹⁸O response in Greenland summit, however, is accompanied by an abrupt cooling in model mean annual temperature of ~6°C at the HS1 onset (~18 ka), more than half the magnitude of the abrupt changes in BA and YD (Fig. 1B and fig. S2).

On the basis of the analysis of our simulation and available climate proxies, we hypothesize that HS1 cooling did occur in the Arctic but the cooling signal is virtually absent in ice core δ¹⁸O. At face value, the simulated ~6°C HS1 mean annual cooling appears to contradict existing Greenland δ¹⁸O-based temperature reconstructions that do not show such cooling (Fig. 1C). However, most existing reconstructions rely on δ¹⁸O and use either borehole temperatures (19), nitrogen isotopic fractionation (21), or reanalysis techniques (22) to calibrate the δ¹⁸O-temperature slope. These δ¹⁸O-based methods are therefore incapable of detecting any abrupt cooling that is not accompanied by a δ¹⁸O decrease. One reconstruction did find a ~3° to 4°C HS1 cooling via a method that combines δ¹⁸O and deuterium excess (Fig. 1C) (20). This cooling is contributed partly by a small δ¹⁸O decrease in the Greenland Ice Core Project (GRIP) ice core (Fig. 1B). However, this method still assumes a stable δ¹⁸O-temperature relationship over time and would consequently underestimate HS1 cooling in the case of a muted δ¹⁸O response. Our muted δ¹⁸O hypothesis, therefore, does not contradict any existing Greenland deglacial ice core temperature reconstructions.

Furthermore, an HS1 cooling in the Arctic is consistent with the global climatic fingerprint in response to AMOC slowdown. Climate models show consistently a characteristic global climate fingerprint response to AMOC slowdown that includes Arctic and circum–North Atlantic cooling, a southward shift of the Intertropical Convergence Zone and weakening of Northern Hemisphere monsoons, and a warming in the Southern Hemisphere due to the bipolar seesaw response (23–26). This fingerprint at the HS1 onset has been detected in many proxy records outside the Arctic (Fig. 1, E and F) (9, 10). The absence of HS1 cooling in the Arctic would therefore imply a highly anomalous deviation from the expected global climatic fingerprint. Last, mean annual ice core δ¹⁸O does broadly follow mean annual temperature closely after HS1. This suggests that Greenland δ¹⁸O may not be a uniformly valid proxy for temperature across the deglaciation, in particular, during periods of extreme cold.

Suppression of δ¹⁸O variability

It has been suggested that the Greenland δ¹⁸O-temperature slope ∆δ¹⁸O/∆T can be reduced by increased precipitation seasonality that is caused by winter cooling and drying associated with North Atlantic sea ice expansion (14, 19, 27–32) in response to AMOC slowdown (33, 34). Our simulations imply that the temporal slope was reduced even further to around ~0‰ K⁻¹ at the HS1 onset. The key question here is why is the δ¹⁸O muted at the HS1 onset but not during later stages of the deglaciation in BA and YD? This requires a quantitative assessment of all the effects on Greenland δ¹⁸O.

We decompose the ice core δ¹⁸O change approximately into the changes in annual isotopic composition Δδ¹⁸O_ann and precipitation seasonality ΔP_S as Δδ¹⁸O ≈ Δδ¹⁸O_ann + ΔP_S (Materials and Methods and Fig. 2A) (34). The change of isotopic composition Δδ¹⁸O_ann generally follows site temperature (Fig. 1C) via the temperature effect with a large temporal slope of ∆δ¹⁸O_ann/∆T ~ 1‰ K⁻¹, characterized by an abrupt isotopic depletion at the HS1 onset and clear BA/YD events. However, precipitation seasonality ΔP_S compensates the Δδ¹⁸O_ann change such that the total Δδ¹⁸O variability is suppressed relative to Δδ¹⁸O_ann over Greenland (Fig. 2A). Qualitatively, this suppression is consistent with previous modeling studies for glacial climate changes (31) and for millennium-scale events in response to the AMOC (33, 34), where it manifests as a low apparent isotope slope of ∆δ¹⁸O/∆T around ~0.4‰ K⁻¹ (21, 35). Quantitatively, however, our simulation suggests that for the Last Glacial Maximum (LGM) to HS1 transition, this compensation is almost complete, leading to negligible Δδ¹⁸O at the HS1 onset (Fig. 2A) or an apparent slope close to 0‰ K⁻¹ in central and northern Greenland (Fig. 1B and figs. S1 and S2), as well as, more generally, over much of the Arctic (fig. S3 and Supplementary Text). The difference in δ¹⁸O responses between the onsets of HS1 and YD appears to be associated with different cooling magnitudes and then different extents of δ¹⁸O_ann − ΔP_S compensation in the two climate stages: The cooling and δ¹⁸O_ann depletion in HS1 are about half of those in YD (−6‰ versus −12‰ and –5°C versus –10°C in the summit region) (Fig. 2A and fig. S4, A to D), while the ΔP_S enrichment is comparable in HS1 and YD (~+4‰) (with an additional +2‰ enrichment from a residual seasonality effect Δδ¹⁸O_Ps; see Materials and Methods) (Fig. 2A). This leads to the δ¹⁸O signal vanishing at HS1 but still being depleted at YD with half the δ¹⁸O_ann depletion. While the model correctly simulates δ¹⁸O at the HS1 onset in central and northern Greenland, the model is less successful in south and east Greenland where it simulates a greater δ¹⁸O depletion than what is seen in the data, reflecting the model deficiency, which is partly caused by the lower altitude and the inability to resolve local topography in the model relative to the observation (fig. S1 and Supplementary Text).

Fig. 2. Decomposition of ice core δ¹⁸O at Greenland (70° to 80°N, 50° to 30°W).

(A) Response of ice core δ¹⁸O (black, Δδ¹⁸O), decomposed into the terms from the precipitation seasonality effect (blue, ΔP_s), annual mean of the isotopic composition (red, Δδ¹⁸O_ann), and residual seasonality (yellow, Δδ¹⁸O_PS; see Materials and Methods for the detailed definitions). (B) Δδ¹⁸O_ann is further decomposed into summer [June, July, and August (JJA), solid] and winter (non-JJA, dotted) contributions. (C) Time series of total winter (September to May) precipitation (blue), surface air temperature (TS) (red) at Greenland, sea ice coverage (yellow), and SST over deep convection regions (green). (D) Same as (C) but for summer (JJA). In (B), the green stars correspond to the times of the moisture tagging experiments at LGM (20 ka), HS1 (15.5 ka), BA(13.2 ka), and YD(12.2 ka). In (C and D), deep convection is defined in the domain of 55° to 60°N, 40° to 20°W, where sea ice grows most in HS1 and YD, as shown in Fig. 3C.

Winter climate impact on Δδ¹⁸O_ann−ΔP_S compensation

The δ¹⁸O signal is suppressed by the enhanced precipitation seasonality ΔP_S associated with extended precipitation from winter (broadly defined here as September to May) to summer (June, July, and August) during the cooling events HS1 and YD. For the abrupt onsets of HS1 and YD, the shifts of precipitation weight are mainly caused by the abrupt loss of snowfall in winter (Fig. 2C and fig. S4, B and C), in response to the AMOC collapses that are here forced by meltwater discharge (fig. S5, A, C, and E). The winter drying increases the relative weight of summer precipitation abruptly (fig. S4C) in the mean annual signal in ice cores, incorporating more of the enriched summer oxygen isotopes (fig. S4D); the resulting ΔP_S enrichment opposes the Δδ¹⁸O_ann depletion and suppresses δ¹⁸O variability relative to δ¹⁸O_ann (31, 33, 34).

The changes of winter climate and isotope composition Δδ¹⁸O_win are further related to SST and sea ice cover over the subpolar North Atlantic (Fig. 2C and fig. S4, E and F), suggesting a role of sea ice in modulating the response of Greenland Δδ¹⁸O. The coherent variability between winter Δδ¹⁸O_win and sea ice, notably the feature of a weaker change in HS1 onset than in YD, is well represented in the leading mode of maximum covariance analysis (MCA; explaining 99% covariance) (Fig. 3, A to C). Accompanying the southward expansion of sea ice is a strong Δδ¹⁸O_win depletion and cooling, a decreased snowfall over Greenland and the Arctic, and a southward shift of the storm track that is indicated in the dipole response of precipitation (Fig. 3, B to E); the largest changes are seen surrounding southern Greenland in the subpolar North Atlantic, coinciding with the region of greatest sea ice change (Fig. 3, B to E). This coherent evolution of sea ice, Δδ¹⁸O_win, and winter climate is consistent with the mechanism of sea ice modulation of δ¹⁸O variability via its impact on Δδ¹⁸O_ann − ΔP_S compensation (32, 34), which is further confirmed in moisture tagging experiments (Materials and Methods, figs. S6 to S8, and Supplementary Text). In HS1 or YD, the collapsed AMOC reduces northward ocean heat transport, leading to SST cooling and sea ice expansion in the subpolar North Atlantic (Fig. 3A and fig. S4, E and F). This sea ice expansion cools the atmosphere by shutting off oceanic heat flux and increases the distance of moisture import from south of the sea ice margin over the North Atlantic, both effects favoring a depletion in atmospheric (vapor) δ¹⁸O_v (Fig. 3F) and then in precipitation δ¹⁸O_win over Greenland (Fig. 3B, fig. S7, and Supplementary Text). Meanwhile, the winter sea ice expansion reduces precipitation over Greenland by shielding the major winter moisture source from the North Atlantic (fig. S6 and Supplementary Text), leading to the reduction of winter snowfall (fig. S4, F to B) and ΔP_S enrichment (Fig. 2A and fig. S4C).

Fig. 3. Water isotope–climate–sea ice coherence in winter derived as first mode of MCA between δ¹⁸O_win and sea ice fraction.

(A) Normalized time expansion coefficients of the first MCA mode for δ¹⁸O_win (red), sea ice (blue), and winter SST in the deep convection region [red box in (C)]. (B) δ¹⁸O_win over Greenland regressed on the normalized time expansion coefficient of sea ice. (C to F) As in (B), but for sea ice fraction, surface air temperature, precipitation change percentage (Prcp) normalized by LGM climatology, and winter column weighted water vapor δ¹⁸O_v,win. Regression fields at 99% (P < 0.01) confidence level in two-tailed Student’s t test are plotted. In (C), the curves correspond to sea ice margin defined as the 15% sea ice fraction in the ocean at different periods (see legend). In (F), the column water vapor δ¹⁸O_v,win is weighted by the moisture in each layer. Prcp, Precipitation.

Sea ice extent at LGM mutes Δδ¹⁸O at the HS1 onset

The extent of sea ice varies with the background climate. The LGM is much colder than the BA, leaving the winter SST over much of the subpolar North Atlantic close to the freezing point (–2°C). This leads to the SST cooling and then sea ice expansion from LGM to HS1 being only about half that from BA to YD (Figs. 2C and 3C) despite the fact that both transitions have a comparable reduction in AMOC strength. This then leads to the magnitude of the LGM-HS1 transition being about half as large as that of the BA-YD transition in many aspects, including the depletion in Δδ¹⁸O_v (δ¹⁸O in vapor) that is mainly from the Atlantic source (fig. S7 and Supplementary Text), precipitation Δδ¹⁸O_win, and Δδ¹⁸O_ann (Fig. 2B). In particular, the reduction of moisture source from LGM to HS1 (0.13 to 0.07 ≈ 0.06 mm/day) is half that from BA to YD (0.21 to 0.09 ≈ 0.12 mm/day), mainly from the Atlantic source (fig. S6, E and F, and Supplementary Text). However, the total moisture source of the background climate at LGM (0.13 mm/day) is also half that at BA (0.21 mm/day) due to the larger sea ice cover at LGM, leading to the precipitation about half at LGM relative to BA in the model; this halved precipitation at the LGM relative to the BA is consistent with the observed accumulation rate (Fig. 1D). Together, the reduction of moisture source is ~50% for both the LGM-HS1 transition (0.06/0.13) and BA-YD transition (0.13/0.21) mainly from the Atlantic. This leads to a comparable (50%) reduction in winter precipitation weight (Fig. 2C and fig. S3, B and C) and then a comparable enrichment in ΔP_S (Fig. 2A) for both transitions. This enrichment effect due to the change of precipitation seasonality, mainly from ΔP_S, compensates the Δδ¹⁸O_ann depletion completely at the onset of HS1, leading to the absence of ice core Δδ¹⁸O; in contrast, this ΔP_S enrichment only compensates half the Δδ¹⁸O_ann depletion during YD, leaving another half as a depletion signal in the ice core Δδ¹⁸O.

The distinct impact of sea ice on Arctic climate for the LGM-HS1 transition is also shown in the distinctly abrupt cooling at ~18 ka. In contrast to the gradual meltwater discharge and the resulting gradual slowdown of AMOC and climate evolution in the mid-latitude and tropics that took over 1000 years (Fig. 1, A, E to G), the model cooling in the Arctic occurs abruptly in ~200 years at ~18 ka (Fig. 1C and fig. S2) because of the strongly nonlinear response of sea ice to climate forcing. In comparison, the abrupt changes associated with YD and BA are accompanied by abrupt changes of AMOC in ~100 years in response to either abrupt meltwater fluxes (at ~12.8 ka) or gradual meltwater fluxes (at ~14.6 and 11.6 ka) (Fig. 1), likely reflecting more the nonlinear response of the AMOC to meltwater forcing. Note that this abrupt cooling is not produced in the earlier study in TRACE-21 (23), which used an earlier model [Community Climate System Model version 3 (CCSM3)] of coarser resolution. The difference in temperature responses appears to be caused by a very weak sea ice sensitivity in CCSM3.

Summer warming during HS1 and YD

In contrast to the fluctuating temperature in winter and annual mean, model Arctic summer temperature increases almost linearly during the deglaciation (Fig. 2D). This summer warming appears to be decoupled from summer sea ice and SST over the North Atlantic (Fig. 2D and fig. S4) and represents an amplified continental response to increased summer insolation and GHG (fig. S5, B, D, and F). This simulated summer warming sheds light on another puzzle of the deglacial climate evolution, namely, the continued snowline retreat in Northern Europe and Greenland during the YD, despite this period being conventionally thought of as a cold period (14, 36, 37). The summer warming forces glacial retreat during the critical summer melt season, while the mean annual cooling occurred predominantly in winter and, therefore, had little impact on snowline. While this hypothesis has been proposed previously (14, 36, 37), a quantitative confirmation has remained challenging because of the potentially opposite temperature responses to different forcing factors, notably insolation, GHG, ICE, and meltwater, in addition to feedbacks within the climate system. Our simulations provide, to our knowledge, the first firm quantitative basis supporting this idea.

The summer warming is also important for the evolution of ice core δ¹⁸O during HS1 and YD. Although winter climate is responsible for the abrupt signal in ice core δ¹⁸O, the summer warming, the associated northward shift of storm track and increased summer snowfall, also contributes substantially to an increasing trend of summer precipitation weight (Fig. 2D and fig. S4, B and C). This increasing trend of summer precipitation weight and the resulting ΔP_S enrichment allow the summer precipitation weight to remain elevated such that the ice core δ¹⁸O remains absent and suppressed during the entire periods of HS1 and YD, respectively. Our calculation shows that, without this summer wetting trend, summer precipitation weight would have decreased steadily in these two periods by a total of ~20%, reducing half of the weight increase caused by the abrupt winter drying and leading to a larger δ¹⁸O signal in ice cores.

DISCUSSION

Our study suggests that a strong winter Arctic cooling occurred concurrent with hemispheric climate changes at the onset of HS1 in response to AMOC slowdown and the resulting sea ice expansion. This cooling signal, however, is virtually absent in Greenland ice core δ¹⁸O. The near absence of an ice core Δδ¹⁸O signal at the HS1 onset suggests a strong dependence of the Δδ¹⁸O/∆T sensitivity on background climate. In particular, during extreme cold winter climate of glacial period, the limited potential for further sea ice expansion mutes the ice core δ¹⁸O response substantially via its impact on the precipitation seasonality ΔP_S and its compensation on isotopic composition Δδ¹⁸O_ann. Greenland ice core δ¹⁸O variability and its paleothermometer scaling depend on the background climate as well as the magnitude of the event (34) and is therefore unlikely to have a uniform sensitivity to temperature during the deglaciation, as noted by previous studies of observations (14, 38) and models (31–33, 39). Despite the δ¹⁸O compensation, the ice core Δδ¹⁸O signal is dominated by the large fluctuating winter temperature signal and almost completely misses the summer temperature signal of an almost linear warming trend throughout the deglaciation.

Our study further implies that Arctic climate underwent even greater abrupt changes than those inferred from the Greenland ice core δ¹⁸O, not only for HS1 but also for other Heinrich stadials during the glacial periods in general. Under the glacial conditions, the large background sea ice leads to a strong δ¹⁸O compensating effect such that the depletion in the ice core δ¹⁸O tend to be limited by a lower bound despite strong cooling.

MATERIALS AND METHODS

The iTRACE simulation

The iTRACE is performed in the state-of-the-art iCESM1.3 but with water isotopes enabled (18), largely following the strategy of the previous transient simulation TRACE-21 (23). The incorporation of water isotope enables us to compare the modeled water isotopes with the observation directly to better understand past climate changes. The iCESM1.3 (17) is composed of the community atmosphere model (CAM5.3), Parallel Ocean Program (version 2), Los Alamos Sea Ice Model (version 4), and Community Land Model (version 4). The resolution of atmosphere and land is nominal 2° resolutions (1.9° in latitude and 2.5° in longitude), with 30 vertical levels in the atmosphere; the resolution of ocean and sea ice is nominal 1° horizontal resolutions (gx1v6), with 60 vertical levels in the ocean. Our iCESM1.3 reproduces the present-day physical climate reasonably well, similar to CESM1.2 (40). The spatial distribution and seasonal cycle of water isotope are reproduced reasonably well over the Arctic and Asia (17). There are, nevertheless, some model biases. For example, the mean annual precipitation δ¹⁸O at Greenland tends to be enriched compared with observations, partly, because model Greenland ice sheet (fig. S2A) is lower than that in the real world by ~300 to 400 m (fig. S1E).

The model LGM (20 ka) state was derived by a spin up for 1000 years from an earlier CCSM4 LGM simulation. This LGM simulation was spun up from 2250 years from a state in which the ocean state was initialized from the present-day observations, with a constant value of 1.05‰ added to δ¹⁸O to account for the isotopically enriched seawater caused by the increased LGM ice sheet volume. Our LGM state shows a global mean temperature about 6° to 7°C cooler than the preindustrial, an AMOC of transport of about 25 Sv close to the model preindustrial, but with a shallower depth. Over the summit (70° to 80°N, 50 to 30°W), the mean annual precipitation (snow) accumulation at LGM is ~50 mm/year, comparable to the reconstructed observation (Fig. 1D and fig. S4B). The spatial slope over Greenland δ¹⁸O is around 0.8‰/°C, consistent with the observation. The sea ice covers the Nordic Sea and the western subpolar North Atlantic (Fig. 3A), consistent with reconstructions (41, 42). Starting from 20 ka, all experiments were integrated to 11 ka. First, three parallel experiments were branched off with three forcing factors added one by one: first, the ice sheet (ICE-6G), then the orbital forcing (43), and, last, the greenhouse gases (44), forming three transient sensitivity experiments, ICE, ICE+ORB, and ICE+ORB+GHG, respectively. The baseline iTRACE simulation (also ICE+ORB+GHG+MWF) is branched off from the ICE+ORB+GHG run at 19 ka, with the MWF imposed largely similar to that in TRACE-21ka (Fig. 1A). The δ¹⁸O values in the meltwater were prescribed according to (45), and the δ𝐷 values in the meltwater were the δ¹⁸O values multiplied by a factor of 8. During the simulation, we change the ice sheet every thousand years (at 19 ka, 18 ka, 17 ka,…) and modified the ocean bathymetry twice at 14 and 12 ka on the basis of the ICE-6G reconstruction (46). These experiments allow us to estimate the impact of each forcing effect approximately as follows: ice sheet (ICE), orbital (ICE+ORB − ICE), greenhouse gases (ICE+ORB+GHG − ICE+ORB), and meltwater flux (iTRACE − ICE+ORB+GHG).

Global surface temperature in iTRACE rises with an interhemisphere bipolar seesaw during HS1 and YD, broadly consistent with reconstructions (9, 10). Our preliminary analyses further show that iTRACE is in good agreement with reconstructions in precipitation δ¹⁸O and climate over the entire pan-Asian monsoon region (18). We also note several improvements of the model climate of iTRACE simulation over TRACE-21 (23) because of the improved climate model, model resolution, and the incorporation of water isotopes. For the purpose of studying Greenland here, in comparison with reconstructions, the overall temperature evolution during the deglaciation is much improved in iTRACE (Fig. 1C) compared to TRACE-21, especially after the BA, because the latter failed in simulating the resumption of AMOC after the YD (47).

Decomposition of δ¹⁸O response

The ice core water isotope is accumulated in precipitation and therefore can be calculated as the total precipitation weighted isotope

$${\mathrm{\delta }}^{18}O={\mathrm{\Sigma }}_{m=1}^{12}{\mathrm{\delta }}^{18}{O}_m\frac{P_m}{P}$$ (1)

where δ¹⁸O_m is the oxygen isotope composition of calendar month m in precipitation, P_m is the precipitation amount of month m, with the sum as the annual precipitation amount $P={\mathrm{\Sigma }}_{m=1}^{12}{P}_m$. The response of the ice core water isotope can thus be decomposed as in (48)

$$\mathrm{\Delta }{\mathrm{\delta }}^{18}O=\mathrm{\Delta }{P}_{\mathrm{s}}+\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{s}}$$ (2)

where

$$\mathrm{\Delta }{P}_{\mathrm{s}}={\displaystyle {\sum }_{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right) {\mathrm{\delta }}^{18}{O}_m, \mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{s}}={\displaystyle \sum _{m=1}^{12}}\left(\frac{P_m}{P}\right) \mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_m$$ (3)

Here, ΔP_s is caused by the change of the season cycle of precipitation or precipitation seasonality. Since ${\sum }_{m=1}^{12}\frac{P_m}{P}=1$, the change of precipitation weight is constrained as ${\sum }_{m=1}^{12}\mathrm{\Delta }\left(\frac{P_m}{P}\right)=0$. As a result, this term can be rewritten as

$$\mathrm{\Delta }{P}_{\mathrm{s}}={\displaystyle {\sum }_{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{\mathbf{P}}\right) ({\mathrm{\delta }}^{18}{O}_m-{\mathrm{\delta }}^{18}{O}_{\text{ann}})$$ (4)

where

$${\mathrm{\delta }}^{18}{O}_{\text{ann}}={\displaystyle {\sum }_{m=1}^{12}}\frac{{\mathrm{\delta }}^{18}{O}_m}{12}$$ (5)

is the (arithmetic) mean annual of the isotopic composition. Normally, warm (cold) season has higher (lower) δ¹⁸O_m, and therefore, δ¹⁸O_m − δ¹⁸O_ann > 0 (<0). Therefore, if the seasonal cycle of rainfall shifts from winter to summer, as in the case of HS1 and YD (fig. S4C), ∆P_s > 0 is enrichment (Fig. 2A).

The term Δδ¹⁸O_s can be further decomposed as

$$\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{s}}=\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{ann}}+\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{PS}}$$ (6)

where

$$\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{ann}}={\displaystyle {\sum }_{m=1}^{12}}\frac{\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{m}}}{12}, \mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{PS}}={\displaystyle {\sum }_{m=1}^{12}}(\frac{P_{\mathrm{m}}}{P}-\frac{1}{12}) \mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{m}}$$ (7)

Δδ¹⁸O_ann is the (arithmetic) mean annual of the change of isotopic composition and is completely independent of seasonal cycle. The residual seasonality term Δδ¹⁸O_PS reflects the change of isotopic composition weighted by the seasonal cycle of precipitation climatology. Since ${\sum }_{m=1}^{12}(\frac{P_m}{P}-\frac{1}{12})=0$, this term can be rewritten as

$$\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{PS}}={\displaystyle {\sum }_{m=1}^{12}}(\frac{P_m}{P}-\frac{1}{12}) (\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_m-\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{ann}})$$ (8)

The sign of this term is determined by the dominant season of Δδ¹⁸O_m response. In the case of HS1 or YD, the response of isotopic composition is dominated by the depletion in winter Δδ¹⁸O_m − Δδ¹⁸O_ann < 0, when the snowfall is also below the annual mean, i.e., $\frac{P_m}{P}-\frac{1}{12}<0$. This leads to an enrichment Δδ¹⁸O_PS > 0 similar to the effect induced by the response of the precipitation seasonality ΔP_s > 0. However, as the product of two relatively smaller deviations, Δδ¹⁸O_PS is usually small, relative to either Δδ¹⁸O_ann or ΔP_s over Greenland (Fig. 2A), and, more generally, the Arctic. Thus, Δδ¹⁸O_s is dominated by the change of the annual composition

$$\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{s}}\approx \mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\text{ann}}$$ (9)

Last, our decomposition is calculated as follows. Assuming two states, 1 and 2, the difference between the two states is decomposed as

$$\begin{array}{c}\mathrm{\Delta }{\mathrm{\delta }}^{18}O={({\mathrm{\delta }}^{18}O)}_1-{({\mathrm{\delta }}^{18}O)}_2={\displaystyle \sum _{i=m}^{12}}{\left(\frac{P_m}{P}\right)}_1 {({\mathrm{\delta }}^{18}{O}_m)}_1-{\displaystyle \sum _{m=1}^{12}}{\left(\frac{P_m}{P}\right)}_2 {({\mathrm{\delta }}^{18}{O}_m)}_2\\ \equiv {\displaystyle \sum _{i=m}^{12}}{\left(\frac{P_m}{P}\right)}_2\mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_m)+{\displaystyle \sum _{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right) {({\mathrm{\delta }}^{18}{O}_m)}_1\\ \equiv {\displaystyle \sum _{m=1}^{12}}{\left(\frac{P_m}{P}\right)}_1\mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_m)+{\displaystyle \sum _{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right) {({\mathrm{\delta }}^{18}{O}_m)}_2\\ \equiv {\displaystyle {\sum }_{m=1}^{12}}\left[\frac{{\left(\frac{P_m}{P}\right)}_1+{\left(\frac{P_m}{P}\right)}_2}{2}\right]\mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_m)+{\displaystyle {\sum }_{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right)\left[\frac{{({\mathrm{\delta }}^{18}{O}_m)}_1+{({\mathrm{\delta }}^{18}{O}_m)}_2}{2}\right]\end{array}$$ (10)

Here, we adopt the form

$$\mathrm{\Delta }{P}_{\mathrm{s}}={\displaystyle {\sum }_{m=1}^{12}}\left[\frac{{\left(\frac{P_m}{P}\right)}_1+{\left(\frac{P_m}{P}\right)}_2}{2}\right]\mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_m)$$ (11)

$$\mathrm{\Delta }{\mathrm{\delta }}^{18}{O}_{\mathrm{s}}={\displaystyle {\sum }_{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right)\left[\frac{{({\mathrm{\delta }}^{18}{O}_m)}_1+{({\mathrm{\delta }}^{18}{O}_m)}_2}{2}\right]$$ (12)

with state 1 as that during the deglaciation and state 2 as the LGM.

Given that

$$\begin{array}{c}\mathrm{\Delta }{\mathrm{\delta }}^{18}O={\displaystyle \sum _{m=1}^{12}}{\left(\frac{P_m}{P}\right)}_2\mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_{\mathrm{m}})+{\displaystyle \sum _{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right) {({\mathrm{\delta }}^{18}{O}_m)}_2+{\displaystyle \sum _{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right) \mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_m)\\ \approx {\displaystyle {\sum }_{m=1}^{12}}{\left(\frac{P_m}{P}\right)}_2\mathrm{\Delta }({\mathrm{\delta }}^{18}{O}_m)+{\displaystyle {\sum }_{m=1}^{12}}\mathrm{\Delta }\left(\frac{P_m}{P}\right) {({\mathrm{\delta }}^{18}{O}_m)}_2\end{array}$$ (13)

Our decomposition (Eq. 10) is an approximate decomposition of perturbations on the LGM state.

Comparison of two methods of “filtering” the effect of precipitation seasonality

We have attempted to filter the effect of precipitation seasonality by removing the seasonality terms of ΔP_s and Δδ¹⁸O_PS and then considering the sensitivity of Δδ¹⁸O_ann to ∆T. In comparison, some previous works have attempted to filter the effect of precipitation seasonality by considering the sensitivity of ice core Δδ¹⁸O to the temperature weighted by precipitation T_p as $\mathrm{\Delta }{\mathrm{\delta }}^{18}O/\mathrm{\Delta }{T}_{\mathrm{p}}$ (31, 33). Our analysis of the deglaciation simulation shows that these two methods are qualitatively similar. However, quantitatively, our method is more favorable. Our calculation during the deglaciation shows that Δδ¹⁸O/ΔT_p tends to be larger than Δδ¹⁸O/ΔT (not shown), as in previous work (31, 33), implying a suppression from Δδ¹⁸O/ΔT_p to Δδ¹⁸O/ΔT. This suppression is qualitatively consistent with the suppression from Δδ¹⁸O_ann/ΔT to Δδ¹⁸O/ΔT in our decomposition method (fig. S3, A, C, and E). However, the magnitude of suppression from Δδ¹⁸O/ΔT_p is substantially smaller than from Δδ¹⁸O_ann/ΔT (not shown). Furthermore, the suppression relation of Δδ¹⁸O/ΔT_p is much less robust than Δδ¹⁸O_ann/ΔT: Over Greenland, the correlation is ~0.5 between Δδ¹⁸O and ΔT_p compared with ~0.9 between Δδ¹⁸O_ann and ΔT. The reason is that T_p overweights the summer precipitation effect into temperature, and therefore, Δδ¹⁸O/ΔT_p tends to be dominated by the summer slope Δδ¹⁸O_sum/ΔT_sum. In contrast, deglacial annual temperature variability ΔT is dominated by winter variability, and therefore, our Δδ¹⁸O_ann/ΔT is similar to the winter slope Δδ¹⁸O_win/ΔT_win and, therefore, is more robust.

Tagging experiments at LGM, H1, BA, and YD

Four water tagging experiments are performed at LGM (20 ka; ExpLGM), H1 (15.5 ka; ExpHS1), BA (13.2 ka; ExpBA), and YD (12.2 ka; ExpYD) in the atmospheric component model of iCESM1.3 (iCAM5.3) to examine the response of moisture source and δ¹⁸O. Each experiment is integrated for 40 years with the past 20 years’ data analyzed. For each experiment, the model is forced by the climatology of monthly sea ice distribution, SST, and sea surface δ¹⁸O and δ𝐷 extracted from the iTRACE experiment, as well as the continental ice sheet, orbital parameters, and GHG concentration, the same as in iTRACE. We tagged 25 regions over the globe, including 13 regions in the ocean and 12 over the lands. For the purpose here, we regrouped the 25 regions into four domains, representing the Atlantic Ocean, Pacific Ocean, land in the mid-latitude [including the entire North America, Greenland, Europe (35° to 90°N, 10°S to 60°E), and North Asia (45° to 90°N, 60° to 180°E)], and the rest of globe.

The tagging experiments track the moisture and oxygen isotopes from each source region during its life cycle from evaporation (source) to condensation (sink) such that the precipitation at a grid point of a calendar month m is the sum of precipitation contributed by all the four regions (j = 1,2,3,4)

$$P_m\mathbf{=}{\mathrm{\Sigma }}_{j=1}^4{p}_{m,j}$$ (14)

and the δ¹⁸O at month m is

$${\mathrm{\delta }}^{18}{O}_{\mathrm{m}}={\mathrm{\Sigma }}_{j=1}^4{\mathrm{\delta }}^{18}{O}_{m,j}\frac{p_{m,j}}{P_m}$$ (15)

where the p_{m, j} and δ¹⁸O_{m, j} denote the precipitation and the oxygen isotope composition in the precipitation from source j at month m, respectively.

Sea ice sensitivity experiment

To separate the impacts of sea ice coverage and SST in forcing the climate-isotope response, we performed an extra sensitivity tagging experiment that is the same as ExpBA, except that the sea ice is replaced by that of YD, denoting ExpBA_YD-seaice. Hence, the response between YD and BA can be decomposed into two parts

$$\begin{array}{c}\text{ExpYD}–\text{ExpBA}=({\text{ExpBA}}_{\text{YD}-\text{seaice}}-\text{ExpBA})+\\ (\text{ExpYD}-{\text{ExpBA}}_{\text{YD}-\text{seaice}})\end{array}$$ (16)

On the right-hand side, the first term represents the impact of sea ice forcing, as the setting of ExpBA_YD-seaice is identical to that in BA except for the sea ice cover (fig. S8, A to E). The second term represents the role of other forcings, including SST, ice sheet, orbital forcing, ocean surface δ¹⁸O and 𝛿𝐷, etc. (fig. S8, F to J). Since ExpBA_YD-seaice and ExpYD are performed at 13.2 and 12.2 ka, respectively, the difference of other forcings are minimal except for SST because of the change of AMOC, and the difference in ocean surface δ¹⁸O and δ𝐷 is also negligible. The second term, therefore, mainly reflects the role of SST. Consistent with the sea ice impact discussed, the sea ice expansion alone leads to a cooling, precipitation reduction, and depletion in oxygen isotopic composition δ¹⁸O_ann over Greenland, which is suppressed in the precipitation δ¹⁸O (fig. S8, A to E). The SST cooling alone also forces a similar response but with a smaller magnitude.

SUPPLEMENTARY MATERIALS

Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/7/25/eabh1007/DC1