Permafrost thawing as a possible source of abrupt carbon release at the onset of the Bølling/Allerød

One of the most abrupt and yet unexplained past rises in atmospheric CO2 (>10 p.p.m.v. in two centuries) occurred in quasi-synchrony with abrupt northern hemispheric warming into the Bølling/Allerød, ~14,600 years ago. Here we use a U/Th-dated record of atmospheric Δ14C from Tahiti corals to provide an independent and precise age control for this CO2 rise. We also use model simulations to show that the release of old (nearly 14C-free) carbon can explain these changes in CO2 and Δ14C. The Δ14C record provides an independent constraint on the amount of carbon released (~125 Pg C). We suggest, in line with observations of atmospheric CH4 and terrigenous biomarkers, that thawing permafrost in high northern latitudes could have been the source of carbon, possibly with contribution from flooding of the Siberian continental shelf during meltwater pulse 1A. Our findings highlight the potential of the permafrost carbon reservoir to modulate abrupt climate changes via greenhouse-gas feedbacks.

C hanges in the global carbon cycle during the last deglaciation are so far not completely understood. However, based on the data and model-based interpretation, the emerging picture indicates that the rise in atmospheric CO 2 of B45 p.p.m.v. during the first half of the deglaciation (B1 p.p.m.v. per century) was probably fuelled by the release of old, 13 C-and 14 C-depleted deep ocean carbon 1,2 . The processes responsible for CO 2 rise have changed dramatically with the beginning of the Bølling/Allerød (B/A) B14,600 years before present (B14.6 kyr BP). Here the abrupt CO 2 rise recorded in the EPICA Dome C (EDC) ice core 3,4 was six times faster than before, about 10 p.p.m.v. in 180 years or B6 p.p.m.v. per century (Fig. 1). Atmospheric CH 4 rose by 150 p.p.b.v. between 18.5 and 14.6 kyr BP and then by the same amount again, but within centuries, around the onset of the B/A. The changes in both greenhouse gases (GHG) imply that a ratio of both changes DCH 4 /DCO 2 is a factor of five larger around 14.6 kyr BP than during the previous four millennia. Such a change in the ratio DCH 4 /DCO 2 might be the first indication that the wetlands identified 5 as the main contributor to the rapid rise in CH 4 at the onset of the B/A might also have contributed to the abrupt rise in CO 2 at that time.
Although this analysis of CH 4 and CO 2 changes gives some first ideas on the potential cause of the abrupt CO 2 rise around the onset of the B/A, its ultimate source was so far not identified. The d 13 C signature of terrestrial or marine carbon sources are different and might allow some source detection. However, the data uncertainty and density of the atmospheric d 13 CO 2 record did so far not allow such an identification 6 . A high-resolution U/Th-dated time series of atmospheric D 14 C derived from Tahiti corals 7 over that event offers now some new and independent insights on the exact timing and magnitude of the carbon release event and brings some suggestions on its potential origin.
Here we show that the synchronous change in atmospheric D 14 C and CO 2 derived from the Tahiti and EDC data sets at the onset of the B/A can be explained by the same process and suggest permafrost thawing being this process. We finally examine the climate impact of the GHG changes around 14.6 kyr BP using a state-of-the-art-coupled Earth system model. Special focus of these investigations is the imprint of the GHG changes on the Antarctic temperature signature and the relevance of these changes for the interpretation of bipolar climate linkages during abrupt climate changes 8 .

Results
Atmospheric D 14 C and ice core CO 2 . The new coral-based atmospheric D 14 C record from Tahiti 7 shows a prominent decline around 14.6 kyr BP, an anomaly not visible in the IntCal13 D 14 C stack 9 (Fig. 2). For comparison, we briefly discuss specific details related to IntCal13 and what other 14 C archives record at that point in time: After 13.9 kyr BP IntCal13 is based on tree rings with very little variability. For older samples, however, the various archives differ by more than the measurement errors. A D 14 C anomaly similar to the Tahiti data can be seen in speleothems from Bahamas 10 (Fig. 2). The anomaly is not seen in speleothems from the Hulu Cave 11 or in the marine sediments from Cariaco 12 (Fig. 2). The Cariaco record bears some problems-therefore, a part of it has been excluded by the IntCal13 group specifically during the Heinrich 1 event, that is, just before the B/A 9 . Necessary corrections of speleothem 14 C data for its dead carbon fraction (DCF) introduce large uncertainties to atmospheric D 14 C based on them 9 . Furthermore, the DCF is not constant but depends itself on climate 13 making the speleothems an archive difficult to interpret, especially during rapid climate changes. The signal might thus potentially be smoothed out in Hulu, as the DCF acts as a low-pass filter. The best recorder of atmospheric D 14 C available up to now might be the terrestrial plant material derived from Lake Suigetsu 14 . Here no corrections for the reservoir effect or for DCF are necessary. The Lake Suigetsu data, however, are rather scattered over the time interval of interest, show a steeper decline in D 14 C than IntCal13, but neither strongly support IntCal13 nor Tahiti (Fig. 2). Altogether, the evidences from D 14 C data are mixed and further data are necessary for a conclusive interpretation.
The coral-based D 14 C record from Tahiti is corrected for a reservoir age 15 of constantly 300 14 C years 7 to be interpreted as atmospheric D 14 C. In principle, the reservoir age might change over time, mainly due to ocean circulation changes. However, simulations with three different models [16][17][18] suggest that the   Atmospheric D 14 C based on Tahiti corals (magenta circles, mean ±1s) 7 or IntCal13 (grey area, ± 1s uncertainty band around the mean) 9 , the latter including linear trends with À 0.04% per year or À 0.10% per year; CO 2 from EDC (blue filled circles) 1,22,68 ; CH 4 from EDC (blue) 22 and Greenland (red) 69  reservoir age is relatively stable in the central Pacific around Tahiti for various ocean circulation changes (see Supplementary Note 1 for details). We therefore assume that reservoir ages did not change over the last 15 kyr in the central low-latitude Pacific and the D 14 C signal based on Tahiti corals is not based on local effects but indeed a recorder of atmospheric D 14 C changes. We date the start of this D 14 C decline seen in the Tahiti data with two different approaches (Fig. 2, methods) with a 1s uncertainty of less than a century to 14.6 kyr BP and calculate a D 14 C decline of B55% within 200 to 250 years. Having already excluded changes in reservoir age, we are left with either a modified carbon cycle or reduced 14 C production rates as potential process explaining the Tahiti D 14 C data. On the basis of available 10 Be, data changes in 14 C production rates cannot convincingly explain the D 14 C data ( Supplementary Fig. 1, Supplementary Note 1). All our tests therefore indicate that the Tahiti D 14 C drop at 14.6 kyr BP is caused by carbon cycle changes. This is our working hypothesis on which all else is based on, but note that its failure cannot entirely be ruled out.
Carbon cycle changes responsible for the D 14 C anomaly would also leave their imprints on atmospheric CO 2 . We can therefore use the absolute U/Th-dated D 14 C from the Tahiti corals as an independent time constraint on the atmospheric CO 2 rise. This is a novel new approach to synchronize atmospheric D 14 C and atmospheric CO 2 , because ice cores archive only a smoothed version of the atmospheric concentrations making an exact dating of the abrupt change in atmospheric CO 2 very difficult 6 . Furthermore, firnification and gas enclosure are still not completely understood 19 , and the age difference between ice matrix and embedded gases complicates gas chronologies 3 . On the most recent chronology 4 , AICC2012, CO 2 measured in situ in the EDC ice core rises by 10 p.p.m.v. between 14.81 and 14.68 kyr BP. This is more than a century faster when compared with previous chronologies 3 (Figs 3a and 4b), but might in detail be revised even further once the most recent understanding of firnification is applied 20 . Atmospheric changes in CO 2 need to have happened even more abruptly than what is recorded in ice cores 6 .
Here we use the Tahiti D 14 C as an independent age constraint for the start of the carbon cycle changes (14.6 kyr BP). Recently, others 21 have shown that the rise in atmospheric CH 4 and in temperature in Greenland are near synchronous (5 ± 25 years) at the onset of the B/A warming. From previous GHG records measured at the EDC ice core 22 , it is known that CO 2 and CH 4 also rise synchronously at 14.6 kyr BP. Combining this information, we have to conclude that the rise in atmospheric CO 2 and CH 4 together with the rapid warming of the northern hemisphere (NH) happened at the same time, and started at 14.6 kyr BP. This is only 35 years later than the suggested age of the onset of the B/A (14.635 ± 0.186 kyr BP, ± 1s) in the annuallayer counted NGRIP ice core 23,24 , well within the dating uncertainty of GICC05 (Fig. 5).
Carbon cycle simulations. A release of 125 Pg of C into the atmosphere within a time window of 50 to 200 years was proposed before 6 to explain the rise of 10 p.p.m.v. in CO 2 measured in EDC. The true atmospheric CO 2 then shows an overshoot whose peak amplitude mainly depends on the length of the assumed release time ( Supplementary Fig. 2). Furthermore, low-resolution CO 2 time series from other ice cores with amplitudes of 17, 15 and 19 p.p.m.v. in Byrd, Taylor and Siple Dome, respectively [25][26][27][28] (Fig. 3a), indicate that the true atmospheric signal had a larger amplitude than what was measured in situ in EDC 6 .
In our previous analysis, we also investigated atmospheric d 13 CO 2 , from which the potential source of the released carbon might have been identified. The new compilation of ice core d 13 CO 2 data published in the mean time 1 offers a new look on the information contained in that record. This d 13 CO 2 compilation is now based on data from EDC, EPICA DML and Talos Dome. In our time window of interest (15-14 kyr BP), however, only data from EDC were obtained with some new data points adding to the previous record. This revised d 13 CO 2 record shows a drop by 0.1% around 14.7 kyr BP followed by a subsequent rise by B0.2% at 14.4 kyr BP (Fig. 3b). Distinguishing terrestrial from marine carbon sources for our carbon release leads in our bestguess scenarios (see below how that was chosen) to either a drop in d 13 CO 2 of 0.4% or less than 0.1% in the true atmospheric signal, respectively, but only to À 0.15% and less than À 0.05% in a time series that would be recorded in EDC (Fig. 3b). Both marine-and terrestrial-based d 13 CO 2 simulations fall within the uncertainties of the measurements before 14.6 kyr BP. The small rise in d 13 CO 2 after 14.4 kyr BP, however, indicates that directly after the onset of the B/A other processes released less d 13 Cdepleted carbon to the atmosphere, for example, CO 2 outgassing from warm oceanic surface waters. On the basis of the data uncertainty of d 13 CO 2 in EDC, it is still impossible to clearly identify if the released carbon was of marine or terrestrial origin.   NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6520 ARTICLE Using the same carbon cycle model 6 , we repeat simulations of carbon release with special focus on 14 C. The model dynamic with respect to 14 C was extensively evaluated (Supplementary Note 2). We assume a depletion in 14 C of the released carbon with respect to the atmosphere (D(D 14 C)) between À 50 and À 1,250%. This range in the 14 C anomalies covers carbon sources from the mean terrestrial biosphere potentially released by shelf flooding 6 ( À 50%), suggested signatures of old carbon of Pacific intermediate waters as measured off Baja California 29 ( À 400%) and Galapagos 30 ( À 700%) to a maximum effect of 14 C-free carbon ( À 1,250%). The Shelf Flooding Hypothesis is explained in detail in the next section, and some more details on our assumptions on 14 C are found in the methods.
The highest-simulated anomalies in atmospheric D 14 C are obtained for short release times reaching À 100% for 50 years and the largest D(D 14 C) of À 1,250% (Fig. 4c). The amplitudes drastically decline with longer release time towards less than À 35%. D 14 C anomalies are significantly smaller if D(D 14 C) was À 700% or less (Fig. 4c). Because of the distinct dynamics of the D 14 C data, release times shorter than B110 years are at odds with the Tahiti 14 C reconstruction. Simulated anomalies in atmospheric carbon records are nearly identical for Atlantic meridional overturning circulation (AMOC) in the strong or weak mode (Methods, Fig. 4c, Supplementary Fig. 3). Combining the information from both the ice core data and our analysis of the Tahiti D 14 C data leads to a range of scenarios with carbon release times between B110 and 200 years in which model results and data agree (Fig. 4c). The range of possible scenarios fulfilling the data constraints also includes some with D(D 14 C) between À 700 and À 1,250%, and so we cannot entirely exclude the possibility that the released carbon still contains some 14 C. From these possible scenarios, we selected the one with the longest release time of 200 years to be our best-guess scenario, because short release times lead to higher amplitudes in atmospheric CO 2 , which are not supported by other ice core data. This scenario pinpoints to a D(D 14 C) of À 1,250% resulting in peak amplitudes of À 42% in atmospheric D 14 C (Fig. 4c) and of þ 22 p.p.m.v. in atmospheric CO 2 (Fig. 4d). The depletion in 14 C necessary for the model output to agree with the data implies that the shelf flooding hypothesis connected with meltwater pulse 1A (MWP-1A) 6,31 seems at a first glance to be in disagreement with the Tahiti-based atmospheric D 14 C reconstructions (Fig. 4c). We discuss details on a potential contribution connected with MWP-1A further below. The release of deep ocean carbon, although so far not suggested to play a role during this rapid CO 2 rise, might only potentially be responsible here, if water masses are detected, which are even more depleted in 14 C than what is known until now 29,30 .
The Tahiti D 14 C data show an excursion from the long-term declining trend of IntCal13 (ref. 9) (Figs 1 and 4a). Depending on the time window of interest, IntCal13 might be approximated by a linear fit with a slope of À 0.04% per year (the whole Mystery Interval, 19-14 kyr BP) or À 0.10% per year (15.0-14.3 kyr BP), respectively. These long-term changes are probably caused by a mixture of changes in 14 C production rate and the carbon cycle 32 . While we are able to force our model with changing 14 C production rates ( Supplementary Fig. 4), all relevant processes in the carbon cycle are not yet identified. We therefore compare the D 14 C data with our original simulation results based on constant 14 C production rate, but also with some results that are corrected for the trend seen in IntCal13. If corrected accordingly, our bestguess scenario finally meets the amplitude in the Tahiti D 14 C data (Fig. 4a,c).
Evidence for permafrost thawing. The synchronicity of the NH warming and the carbon cycle change together with our suggested hypothesis for the injection of nearly 14 C-free carbon into the atmosphere make permafrost thawing and a subsequent release of   (Fig. 6b). We assume that most of the NH follow this temporal changes in temperature observed for Greenland, although with warmer temperatures closer to the freezing point further south. It might then be that large areas of the NH were permanently frozen after D/O event 3, thus about 13 kyr before thawing induced by the onset of the B/A. The D 14 C of that permafrost carbon would be depleted by À 900% with respect to atmospheric D 14 C during release around 14.6 kyr BP (Fig. 6a). However, soil carbon might age significantly in high latitudes before freezing, for example, present day North American peatlands are up to 17-kyr old 34 . Such soil ageing reduces 14 C even further. If the precursor material of the permafrost soil carbon was photosynthetically produced during D/O event 12 around 47 kyr BP (the next preceding period comparable in temperature to the B/A, Fig. 6b), it would be essentially free of 14 C and depleted with respect to atmospheric D 14 C by nearly À 1,250% (Fig. 6a). Permafrost thawing would Agrees with 14   then contain a depletion in D 14 C, which is more negative than for all other suggested processes 6,29,30 . An alternative scenario based on the destabilization of gas hydrates, which also contain 14 C-free carbon, can be rejected based on CH 4 isotopes [35][36][37] . For the present day, a rise in global mean temperature by 5 K, which because of polar amplification might represent a northern high latitude warming of 10 K, was proposed to lead to the release of more than 130 Pg of soil carbon from permafrost thawing within 200 years 38 . Greenland ice core data 33 and simulations 39 suggest that temperatures in the B/A rose by 10-15 K to near preindustrial levels in central Greenland and throughout most of the NH land areas. A large inert terrestrial carbon pool consisting of permafrost soils containing 700 Pg more C at the Last Glacial Maximum (LGM) than at present day has been proposed 40 , which needs to release its excess carbon during deglaciation. The areal extent of continuous permafrost at LGM (Fig. 7) was calculated from models 41 in PMIP3 to 26 Â 10 12 m À 2 , agrees with reconstructions 42 , and is twice as large as for preindustrial times 41 .
Previously, methane isotopes 36 suggested that a rise in boreal wetland CH 4 emissions by þ 32 Tg CH 4 per year would explain the CH 4 rise into the B/A. These findings 36 have been challenged by new methane isotope data 37 , but so far no revised CH 4 emissions from boreal wetlands have been calculated for the B/A. An alternative interpretation 5 of the CH 4 cycle based on its interhemispheric gradient suggests that the rise in CH 4 by 150 p.p.b.v. at the onset of the B/A was largely driven by the increase in CH 4 emissions from both tropical ( þ 35 Tg CH 4 per year) and boreal ( þ 15 Tg CH 4 per year) wetlands. The CH 4 change at the onset of the B/A is thus clearly dominated by tropical wetlands and its conclusive interpretation is beyond the scope of this study. However, the rise in CH 4 emissions from Greenland CH 4 (p.p.b.v.) Figure 5 | Implication on the timing of abrupt climate change as obtained in various ice core records from Greenland and Antarctica. Our results suggest that anomalies in Tahiti D 14 C and true atmospheric CO 2 are caused by the same process. This information is used here as an independent age constraint. The onset of the abrupt rise in atmospheric CO 2 (black bold, this study) is thus tied to 14.6 kyr BP. From previous ice core analysis 22 , it is known that the rise in CO 2 and CH 4 (red circles, Greenland composite 69 ) occur synchronously here. A new study 21 on the NEEM ice core tied the CH 4 rise to be near synchronous to Greenland temperature rise. This synchronicity of the start of the abrupt changes in atmospheric CO 2 , CH 4 and Greenland temperature tied to 14.6 kyr BP led to the age alignments in CH 4 and NGRIP d 18 O (high 24 (black thin line) and low 70 (red line) resolution). We furthermore show some Antarctic climate records on their own independent chronologies to illustrate the temporal north-south offsets. WD 54 d 18 O, original (blue) and 100 years running mean (yellow) and stack 3 of temperature change from five ice cores in East Antarctica, DT ANT , original (black) and 100 years running mean (orange). boreal wetlands is nearly identical to the rise in emissions of up to þ 14 Tg CH 4 per year projected from deep permafrost thawing of the next century 43 . If this rise in the boreal CH 4 flux is integrated over the 200-year time window of our carbon release scenario, a total of 3.0 Pg of CH 4 (or 2.25 Pg C in the form of CH 4 ) might have been released. This is B2% of our total estimated carbon emissions of 125 Pg C, and in line with an expert assessment on the future vulnerability of permafrost 44 estimating that 2-3% of carbon released by thawing might enter to the atmosphere in the form of CH 4 . Although the contribution from boreal wetlands to the CH 4 rises at the onset of the B/A is small, the nearly 14 C-free signature connected with our proposed permafrost thawing might be tested by 14 C measurements on CH 4 derived from ice cores 45 . So far, we suggested that NH permafrost is the responsible source of the released carbon. In the following, we hypothesize which region might have been affected in detail by permafrost thawing and how this can be tested in future studies. The PMIP3based map on the LGM permafrost extent clearly indicates that the largest areas with continuous permafrost are found in northern Siberia (Fig. 7). Thus, evidences of permafrost thawing connected to the NH warming should be expected in outflow originated from the southern edge of the LGM permafrost area (around 40-50°N), which thawed first. A lot of these areas are drained via the Amur river into the Sea of Okhotsk and into coastal seas towards the south (Caspian and Black Sea). Indeed, a combination of terrestrial biomarkers that clearly indicate the thawing of permafrost was found at the onset of the B/A in a sediment core drilled in the Black Sea that records the drainage from the Fennoscandian Ice Sheet 46 . Here variations in the normalized concentrations of different long-chain molecules provide information on changes in the abundance of peatforming plants. Such data are useful indicators for the variation of permafrost thawing and of wetland extension as well as for fluvial periglacial soil erosion in its drainage basin. In details, this study 46 provides evidence that the permafrost melting was very intense only during the initial part of the Bølling corresponding indeed to the sharp NH warming.
The map (Fig. 7) also shows that a large fraction of the Siberian shelf in the Arctic Ocean was during the LGM covered by permafrost. It might thus be possible that MWP-1A, which was recorded as a rise in sea level at Tahiti 31 from about À 105 to À 85 m between 14.65 and 14.31 kyr BP might be partially responsible for the carbon release as initially suggested 6 . The flooding of continental shelves was also proposed to contribute to the CH 4 rise during deglaciation and during D/O events 47 . In our earlier study 6 , we proposed that mainly the flooding of the Sunda Shelf followed by tropical rain forest decay might have been responsible for the carbon release. This shelf flooding scenario was here addressed with a D(D 14 C) of À 50% for mean terrestrial carbon, which failed to meet the D 14 C data. In our earlier study 6 , we also discussed that the existing time series of sea level change suggest that before MWP-1A the shelves were last flooded around 30 kyr BP, 15 kyr earlier, leaving ample time for 14 C in permafrost carbon on the shelf to decay and to produce a D(D 14 C) in the released carbon of down to À 900% (Fig. 6). Most recent sediment data 48 on iceberg discharge in Antarctica during Termination I found a significant Antarctic contribution to MWP-1A. Fingerprint analysis 49 of different water sources for MWP-1A indicate that sea level would rise locally by up to 50% above global average on the Siberian Shelf for freshwater released in Antarctica. When considering the source-depending overprint 49 , we calculate, based on the present day bathymetry 50 , a maximum areal extent of 0.4 Â 10 12 m À 2 of the Siberian Shelf, which might have been flooded by MWP-1A. This is the same order of magnitude as the present day Siberian Yedoma deposit extent 51 from which an organic carbon content of 30-140 Pg C has been proposed 51 . Coastal erosion and sub-sea permafrost release in Arctic Siberia are also observed for modern times 52 with a D 14 C signature of the released organic carbon as low as À 800%. Modern organic carbon content in Eurasian Arctic 53 river runoff have D 14 C ages of up to 10 kyr. All these modern data indicate that old carbon in permafrost exists nowadays, and potentially was more abundant and older during glacial times.
In which region the thawing of permafrost finally happened might be verified by future 14 C measurements on terrigenous organic material that are retrieved from marine sediments in the suggested coastal seas. It will then be possible to finally attribute the size of the released carbon to either a pure thermodynamically thawing at the southern edge of the permafrost area or to a contribution from flooding the Siberian Shelf during MWP-1A.

Discussion
The rapid CO 2 rise at the onset of the B/A is contained with different amplitude in various ice cores (Fig. 3a). However, the uncertainty in the proposed age distribution of the CO 2 in EDC is still large 6 ( Supplementary Fig. 2c) and the assumed carbon release history and the applied carbon cycle model influence the amplitude of the proposed true atmospheric CO 2 rise. Future CO 2 measurement from the WD ice core 54 might refine some of these aspects. The WD ice core has an order of magnitude higher present day accumulation rate than EDC (20 versusB3 g cm À 2 per year), thus offers temporally higher resolved gas records. A potential WD CO 2 record still needs to be corrected for the smoothing during firn enclosure, although this effect will be a lot smaller than for EDC. Only by considering the 20% uncertainty in the mean exchange time of CO 2 before enclosure in EDC ( Supplementary Fig. 2c) would result in a CO 2 release and an amplitude in the true atmospheric CO 2 rise, which are also 20% smaller than in our best-guess scenario, for example, releasing 100 instead of 125 Pg C leading to a true atmospheric CO 2 amplitude of 18 instead of 22 p.p.m.v. (Supplementary Fig. 2d). Results are then still within the range given by the EDC CO 2 data ( Supplementary Fig. 2d). Furthermore, we calculate that a 33% reduction in the proposed carbon released to the atmosphere (85 instead of 125 Pg C) still fulfils the data constrains given by the Tahiti D 14 C record.
Other rapid CO 2 rises are detected in EDC at the end of the Younger Dryas and in Marine Isotope Stage (MIS) 3 in other ice cores 55 . Whether they are also caused by permafrost thawing is not investigated here. However, these other rapid CO 2 jumps are not always connected to a warming of the NH. Furthermore, the missing of 14 C-depleted CH 4 during the CO 2 rise at the end of the Younger Dryas around 12 kyr BP 45 suggests that other processes are responsible. Moreover, our suggested old soil carbon release during permafrost thawing at 14.6 kyr BP, requires that the same carbon sources were not tapped during other events within Termination I. Again, 14 C measurements on terrigenous material might clarify how old the carbon released from permafrost was or if earlier CO 2 rises might already have consumed the old, 14 Cdepleted carbon.
Our study suggests that for Termination I, abrupt warming in the NH might lead to massive permafrost thawing, activating a long-term immobile carbon reservoir. The abruptly released carbon then amplified the initial warming as a positive feedback. Our best-guess scenario generates together with the rise in the other two important GHG CH 4 and N 2 O a radiative forcing 6 of B0.7 W m À 2 . It is important to quantify the feedback of this GHG forcing on climate to better understand the impact of the abrupt GHG changes during the last deglaciation. Since the abrupt GHG changes are contemporaneous with the onset of the B/A (in the North) and the beginning of the ACR (in the South), the sequence of the associated bipolar climate linkages are of particular interest. Therefore, we have conducted transient simulations with our best-guess reconstruction of atmospheric GHG changes during the beginning of the B/A and the ACR, using the Earth System Model COSMOS 16,56 in a coupled atmosphere-ocean configuration as outlined in detail in the Supplementary Note 3. To evaluate the global impacts of the GHG changes it is instructive to analyse Antarctic temperature changes, since temperature changes obtained from Antarctic ice cores have been shown to reflect global scale climate changes associated with CO 2 variations particularly well 57 . On the basis of our climate model investigations, we find that the abrupt rise in GHG concentrations provides an important impact on the Antarctic temperature signature associated with an abrupt AMOC strengthening at the end of Heinrich Stadial 1 ( Supplementary  Fig. 5). This highlights a potential contribution of abrupt GHG changes on the bipolar climate signature during deglaciation. In this sense, the abrupt GHG changes would be a factor that would offset the timing of the temperature maximum leading into the ACR, compared to the onset of the B/A. As layed out in detail in the climate feedback section of the Supplementary Note 3, also smaller GHG spikes bear the potential to have a substantial effect on the Antarctic temperature response when compared with impacts caused by AMOC changes.
So far a synthetic record of Greenland temperature changes 8 seems to indicate that rapid climate changes in the north might indeed have been a universal feature of deglaciations during the last 800 kyr. Hence, similar to the last deglaciation, abrupt permafrost thawing might have also occurred regularly during earlier terminations, although further studies are necessary here.
Termination II also contains 58 an abrupt rise in CO 2 , synchronous to a rise in CH 4 . A massive drop in atmospheric d 13 CO 2 accompanying this event 58 is consistent with the release of d 13 C-depleted CO 2 that might indicate a terrestrial source. However, new d 13 CO 2 data 59 did not confirm this negative d 13 CO 2 anomaly and the revised data give no indication on the source of this CO 2 rise. A synchronous change in deuterium excess 60 , a proxy for moisture source shifts, has been used to suggest that abrupt shifts in southern westerlies might be connected with the CO 2 rise 61 , but a compelling explanation remains elusive and further testing of permafrost thawing as a possible alternative interpretation is needed.
In conclusion, we here suggest that the processes responsible for the abrupt CO 2 rise at the onset of the B/A is also the underlying cause for the drop seen in atmospheric D 14 C based on Tahiti corals. This connection offers a U/Th-dated tie point for the start of the massive release of carbon at 14.6 kyr BP. Using a carbon cycle box model, and assuming the release of 125 Pg of nearly 14 C-free carbon, we are able to explain observed anomalies in atmospheric CO 2 and D 14 C. On the basis of the 14 C signature of the released carbon and the synchronicity to the warming of the NH, we suggest that the thawing of permafrost was this responsible process. A potential contribution from MWP-1A flooding the Siberian Shelf, which might have contained a large amount of permafrost, is also possible. Future 14 C measurements on terrigenous material might further constrain the source region. Our interpretation not only provides conceptual insights into the source of the excursions in the atmospheric carbon records around 14.6 kyr BP, but also offers an alternative to explanations 62,63 for the interhemispheric timing of the B/A and the ACR as found in ice cores from both hemispheres. Taken together, our findings highlight a potential climate feedback that might be obtained from abrupt CO 2 release during deglaciation. This analysis furthermore indicates that the proposed carbon cycle feedback from an anthropogenic driven permafrost thawing in the near future 38,43,44,64 may already have happened in a similar way in the past.

Methods
Analysis of the D 14 C data. For analysis of the drop in the atmospheric D 14 C data based on Tahiti corals, we used two different approaches (Fig. 2). First, we used a linear statistical model Breakfit 65 , which calculates the break points in time series. Breakfit searches for two linear functions that are joined at the break point. To determine the break points, the model is fitted to the data applying an ordinary least-squares method with a brute-force search for the break points. A measure of the uncertainty of the break points is based on 2,000-block bootstrap simulations, applying a moving block bootstrap algorithm with a block length of 1. We were searching for two break points in the time intervals between 16 and 13 kyr BP. The two subintervals (one for each break point) were ranging from the outer boundary next to the break point of interest to the other break point. Subintervals were finally identified after at least two iterative applications to (in kyr BP): break point 1 [15.74, 14.45] and break point 2 [14.67, 13.16]. Breakfit identified the start in the D 14 C drop at 14.66 ± 0.07 kyr BP followed by its decline by 54 ± 8% within 207±95 years. Because of the very distinct dynamics of atmospheric D 14 C including a rebound after its minimum (that is, after the carbon release to the atmosphere stopped), we also analysed the data more subjectively with a non-linear approach. Here we only calculated the mean time and mean D 14 C right at the start of the carbon cycle changes around 14.6 kyr BP (two data points) and at its minimum (eight data points) assuming that D 14 C followed a non-linear pathway between both and included a rebound thereafter. The D 14 C data then starts to decline at 14.59 ± 0.04 kyr BP and stop after 258 ± 53 years with a maximum drop of 58±14% followed by a rebound of atmospheric D 14 C. This non-linear dynamic is seen in the Tahiti data but also in our carbon cycle simulations (Figs 2 and 4a). Combining the linear and non-linear approach brings high confidence that the D 14 C drop started at around 14.6 kyr BP. All uncertainties are given as 1s.
Possible D 14 C signature of permafrost carbon. The maximum possible D(D 14 C) of carbon released from permafrost thawing is a function of age and of atmospheric D 14 C during time of production. From the D 14 C signature (IntCal13) (ref. 9) of the precursor material (atmospheric CO 2 ), which varies before the B/A roughly between 250 and 550% (Fig. 6a), we first subtract the mean D 14 C value of ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6520 terrestrial carbon at the LGM in the model ( À 50%) before a further reduction in D 14 C signature is realised by the radioactive decay of 14 C (half-life time of 5,730 years).
Carbon cycle model. We use the carbon cycle box model BICYCLE in transient mode to simulate changes in atmospheric CO 2 and D 14 C. The model setup is identical to an earlier study, which already proposed the magnitude of the CO 2 overshoot during the B/A 6 .
We simulate the release of 125 Pg of carbon into the atmosphere with a constant rate that varies inversely with the time length of the event between 0.42 Pg C per year (300 years) and 2.5 Pg C per year (50 years) and configured the AMOC in either its strong or its weak mode. Both AMOC configurations differ in the strength of the overturning cell in the Atlantic with 16 Sv deep water production in the North Atlantic in the strong mode and 2 Sv in the weak mode. We repeated our previous comparison 6 of simulated atmospheric d 13 CO 2 to ice core data from the EDC because new d 13 CO 2 data were published in the mean time 1 . For this modeldata comparison of d 13 CO 2 , we distinguished terrestrial and marine sources of the released carbon by assuming a d 13 C signature of À 22.5 and À 8.5%, respectively (Fig. 3b). More details on these assumptions are found in our previous article 6 . D(D 14 C) of the simulated carbon release is depleted with respect to the atmosphere between À 50 and À 1,250%. Initial conditions of 14 C production rates influence simulated D 14 C over several ten thousand years 32 . All simulations therefore start at 60 kyr BP. In our standard case, 14 C production rates are assumed to be constant and 15% higher than present day, leading to atmospheric D 14 C of þ 250% at 14.6 kyr BP in agreement with IntCal13 (ref. 9). Long-term trends in 14 C production rate as suggested by the geomagnetic field data 32 only slightly impact our simulations ( Supplementary Fig. 4).
For model evaluation, BICYCLE is (a) compared in its oceanic carbon uptake dynamic resulting in a model-specific airborne fraction with other models, (b) used to simulate the Suess effect (years 1820-1950 AD), (c) the bomb 14  Filtering true atmospheric CO 2 into signals recorded in EDC. The smoothing effect of the gas enclosure process in ice cores that transforms a potential true atmospheric CO 2 into a time series comparable to EDC ice core data is performed with a log-normal probability density function with an assumed mean value or width E of 400 ± 80 years (mean ± 1s) 6 ( Supplementary Fig. 2c with x (in years) as the time elapsed since the last exchange with the atmosphere. From the two free parameters, m and s of the equation, we chose for simplicity s ¼ 1, which leads to E ¼ e m À 0.5 . The application of such a filter function for the transformation of true atmospheric signals into those that might be recorded in ice cores during rapid climate change was compared with results from firn densification models and extensively validated with CH 4 data from both hemisphere 6 .