The Antarctic Ice Core Chronology 2023 (AICC2023) chronological framework and associated timescale for the European Project for Ice Coring in Antarctica (EPICA) Dome C ice core

. The EPICA (European Project for Ice Coring in Antarctica) Dome C (EDC) ice core drilling in East Antarc-tica reaches a depth of 3260 m. The reference EDC chronology, the AICC2012 (Antarctic Ice Core Chronology 2012), provides an age vs. depth relationship covering the last 800 kyr (thousands of years), with an absolute uncertainty rising up to 8000 years at the bottom of the ice core. The origins of this relatively large uncertainty are twofold: (1) the δ 18 O atm , δ O 2 / N 2 and total air content (TAC) records are poorly resolved and show large gaps over the last 800 kyr, and (2) large uncertainties are associated with their orbital targets. Here, we present new highly resolved δ 18 O atm , δ O 2 / N 2 and δ 15 N measurements for the EDC ice core covering the last ﬁve glacial–interglacial transitions; a new low-resolution TAC record over the period 440–800 ka BP (ka: 1000 years before 1950); and novel absolute 81 Kr ages. We have compiled chronological and glaciological information including novel orbital age markers from new data on the EDC ice core as well as accurate ﬁrn modeling estimates in a Bayesian dating tool to construct the new AICC2023 chronology. For the ﬁrst time, three orbital tools are used simultaneously. Hence, it is possible to observe that they are consistent with each other and with the other age markers over most of the last 800 kyr (70 %). This, in turn, gives us conﬁdence in the new AICC2023 chronology. The average uncertainty in the ice chronology is reduced from 1700 to 900 years in AICC2023 over the last 800 kyr (1 σ ). The new timescale diverges from AICC2012 and suggests age shifts

This study A correction has been applied on the datasets obtained in 2006 and 2007 because they were obtained on a new mass spectrometer.We found that the calibration of the δO2/N2 data at the time has not be done correctly when switching from the old to the new mass spectrometer and that a shift of +1.5 ‰ should be applied to the δO2/N2 data.

Figure S1
. δO2/N2 series from the EDC ice core after different storage conditions.Note that the ice quality was very bad for the ice samples cut at the bottom of the ice core (corresponding to the age range of 800 -700 ka BP).The orange rectangle frames the zone with only bubbly ice.The years of measurement are indicated and correspond to the different colors of the series.
Figure S1 shows the evolution of the mean level of δO2/N2 after different storage conditions.We do not notice differences in the δO2/N2 mean level for the samples stored at -50°C even after 14 years of storage (2022 vs 2008).
This result is similar to the one obtained at Dome Fuji by Oyabu et al. (2021) even if we are working with smaller sample (20-30 g before removing the outer part).On the contrary, ice storage at -20°C has a strong effect, especially on clathrate ice.The samples analyzed in 2022 after storage at -20°C during more than 18 years exhibit δO2/N2 values as low as -80 ‰.The bubbly ice analyzed here has been stored at -20°C.The associated mean level of δO2/N2 is not significantly different from the one measured for samples stored at -50°C but the scattering is much larger as already observed on other series from bubbly ice (e.g.Oyabu et al., 2021).

Study of the impact of filtering on δO2/N2 -insolation tie point identification
Figure S2 shows the smoothed δO2/N2 dataset using a low-pass (rejecting periods below 15 kyr) or a band-pass filter (keeping periods between 100 and 15 kyr periods, used by Bazin et al., 2013).The choice of filter does not alter the peak positions in the δO2/N2 curve We compare tie point identification performed without (method a) and with (method b) filtering of the highly resolved δO2/N2 record between 260 and 180 ka BP (Fig. S3).The signal is first interpolated every 100 years.For the method a, we identified the mean maximum (or minimum) position age max (),a (or age min (),a ) as the middle of the age interval [ 1 ;  2 ] in which δO2/N2 is superior (or inferior) to a certain threshold.The threshold is defined as 95 % (or 5 %) of the amplitude difference   between the considered maximum and the minimum immediately preceding it (or between the considered minimum and the maximum immediately following it).The process is reiterated every ~10 kyr (precession half period) when an extremum is reached in the δO2/N2 signal.For the method b (described in the main text), we detected the peak positions (age max (),b and age min (),b ) in the δO2/N2 via an automated method using the zero values of the time derivatives of the low-pass filtered δO2/N2 compiled signal.After comparison of the peak positions identified by methods a and b (Table S2), we found an average disagreement of 700 years, with the largest value, 2150 years, observed between age min (+1),b and age min (+1),a at about 230 ka BP (Fig. S3).This period coincides with abrupt variations in the EDC δD record (Fig. S3), reflecting changes in surface climatic conditions which may have impacted high resolution variability of the δO2/N2 signal in addition of the insolation effect.Over periods of lower resolution of the δO2/N2 signal, the extrema positions are not affected by the filtering by more than 600 years (Table S2).For the construction of the new AICC2023 chronology between 800 and 590 ka BP, the EDC δ 18 Oatm record is aligned with the climatic precession delayed or not by 5,000 years depending on the occurrence of Heinrich like events, reflected by peaks in the IRD record from the North Atlantic Ocean (Sect 3.2.3 in the main text).Potential errors may arise from aligning δ 18 Oatm to precession (Oyabu et al., 2022).To support the use of our approach, we test three methodologies to align δ 18 Oatm and precession.Four test chronologies are built: 1) The test chronology 1 is obtained by aligning δ 18 Oatm to 5-kyr-delayed precession as in Bazin et al. (2013).
2) The test chronology 2 is obtained by aligning δ 18 Oatm to precession as it would be expected if only precession is driving the δ 18 Oatm signal.
3) The test chronology 3 is obtained by aligning δ 18 Oatm to precession delayed if IRD counts are superior to 10 counts g -1 and to precession without delay if IRD counts are inferior to 10 counts g -1 .

Between 810 and 590 ka BP
We first evaluate the impact on the chronology whether δ 18 Oatm is aligned with the precession with or without delay between 810 and 590 ka BP.The age mismatch between test chronologies 1 and 2 is of 3,000 years on average, reaching its maximum value of 3,700 years at 712 ± 2.6 ka BP (red arrow in Fig. S4).

Between 300 and 100 ka BP
Then, we test the three methodologies to align δ 18 Oatm and precession over the 100-300 ka period, where we have high confidence in our chronology.
Over this time interval, the test chronology 3 appears to be the best compromise as it agrees well with both the AICC2023 age model and the chronology derived from δ 18 Oatm-δ 18 Ocalcite matching (Fig. S5).This is why we believe that it can faithfully be applied to the bottom part of the EDC ice core while keeping large uncertainties in the tie points (1σ uncertainty of 6 kyr).
This agreement is particularly satisfying over the 120-160 ka BP time interval.Over this period, Oyabu et al.

Background lock-in-depth (LID) scenario at Dome C
The background LID scenario can be derived either from the δ 15 N data (i.e.experimental LID), or from firn modeling (i.e.modeled LID).We favor the use of δ 15 N data when there are available.Over depth intervals where no measurements of δ 15 N were made, the LID can be deduced from firn modeling or from a synthetic δ 15 N record using the δD-δ 15 N relationship (Bazin et al., 2013).In this work, we assess the credibility of three composite LID scenario (Table S3) constructed using the firn model (Bréant et al., 2017) or the synthetic δ 15 N record when no data are available.The credibility is defined by the criterion Δ as per: Δ represents the average absolute value of the mismatch between the background LID (i.e.prior LID provided in input in Paleochrono) and analyzed LID (i.e. the posterior LID given by Paleochrono) scenarios of LID.The weaker is Δ, the closer the background scenario is to the analyzed scenario, meaning that the background scenario is in relatively good agreement with chronological information compelling the inverse model in Paleochrono.On the contrary, the larger is Δ, the more Paleochrono is forced to significantly modify the background scenario which is incompatible with the chronological constraints.Therefore, the larger Δ is, the less credible is the prior LID scenario.It should be noted that the relative error in the prior LID scenario and the age constraints input in Paleochrono are equal in each test, so that the mismatch Δ is only impacted by the value of the prior LID from one test to another.Three background scenarios of LID are tested (Table S3).
Table S3.For the construction of the AICC2012 timescale, the background LID scenario at EDC was derived from a synthetic δ 15 N record using the δD-δ 15 N relationship (Bazin et al., 2013).Yet, this scenario (A, Table S3) is associated with the largest mismatch criterion over the last 800 kyr, reaching Δ = 5 m over the 578-1086 m depth interval where no δ 15 N data are available (Fig. S6).Hence it is believed to be the least pertinent among the three tested scenarios and we decided not to use the δD-δ 15 N relationship to construct the prior LID scenario in this work.
Modeled LID scenarios (B and C, Table S3) are characterized by smaller mismatch criteria Δ than LID A regardless of the depth interval considered (Fig. S6), hence we believe that firn modeling estimates reproduce well the evolution of past LID at EDC site.In the firn model, the creep factor can be either dependent on impurity inclusion inducing firn softening (giving LID B) or not (giving LID C).The LID sensitivity to the impurity parameter is evaluated by comparing LID B and LID C performances.Even though LID B is associated with a smaller criterion Δ between 578 and 1086 m, LID B and LID C show comparable values for Δ over the last 800 kyr (Fig. S6).Bréant et al. (2017) argued that implementing the impurity dependence in the model reduces the δ 15 N data-model mismatch at Dome C.This is particularly verified over deglaciations where significant LID augmentations inferred from δ 15 N are well reproduced by the modeled LID when the impurity parameter is included (panel a in Fig. S6).
We thus follow the recommendation of Bréant et al. (2017) and use the composite LID B scenario to constrain the new AICC2023 chronology.
Discontinuities are visible when switching from experimental to modeled values when no data are available (grey rectangles on Fig. S7).To avoid these discontinuities, we test a LID scenario where the modeled LID is fitted to experimental LID values (orange curve in Fig. S7).In other words, the firn modeling estimates are adjusted, by standard normalization, to the scale of LID values derived from δ 15 N data.Adjusting the modeled LID to experimental LID values induces a modification of 4.7 m at most (see red arrow) which remains within the background relative uncertainty (20%).
On the depth interval from 578 to 1086 m, the modeled scenario without any fitting to δ 15 N-inferred LID (blue curve, Fig. S7) is almost as effective as the one that was fitted (orange curve,  We thus conclude that we can keep the scenario combining δ 15 N-inferred LID and modeled LID in the construction of AICC2023.

Background uncertainties for LID, accumulation rate and thinning scenarios
Although there is no objective way to assign specific prior uncertainties, the values chosen by Bazin et al. (2013) seem unrealistic (i.e.80 % of uncertainty for the LID during some glacial periods at EDC whereas firn modeling and δ 15 N agree within a 20 %-margin at most).That is why we believe the prior uncertainties should be reduced in AICC2023 and implement the following major changes (blue plain line in Fig. S8 and S9): -The LID background relative uncertainty is reduced to values oscillating between 10 and 20 % at most, excluding values reaching 80 % used in AICC2012.The reason for this modification is that in 2012, the mismatch between firn model outputs and δ 15 N-inferred LID was not understood.In the meantime, much progresses have been made, confirming that the δ 15 N-inferred LID was correct and firn models or their forcing have been adapted (Parrenin et al., 2012;Bréant et al., 2017;Buizert et al., 2021).
-The thinning relative uncertainty is evolving linearly, rather than exponentially as it was done in AICC2012.The linear uncertainty permits to have a significant uncertainty at intermediate depth levels while with the exponential shape, the uncertainty was essentially located at lower depth levels, which was not realistic.
-The accumulation relative uncertainty is decreased to 20 %, as opposed to 60 % used in AICC2012.This choice is motivated by the study of Parrenin et al. (2007) who counted event duration in EDC and DF ice cores and found out an offset of 20 % on average.
We build different test chronologies by keeping the same age constraints and background scenarios as in AICC2023 but varying the background errors (Table S4).The largest age offset is observed between the test AICC2012 and the other test chronologies at around 650 ka BP.It reaches 400 years (see red arrow in Fig. S8), which is not significant considering the uncertainty associated with the test chronologies over this period (ranging from 1,800 to 3,400 years).Since varying the background uncertainties has no significant impact on the final age model and the background uncertainties of AICC2012 seem unrealistic, we reduce the background errors with respect to AICC2012 and we use the Test 5 configuration from Table S4 to construct AICC2023.S6).It is estimated from δ 15 N data or δ 40 Ar data (which also reflects evolution of the firn thickness) and corrected for thermal fractionation.The thermal fractionation term is estimated by the firn model running in the same configuration as for calculating the modeled LID at EDC (i.e.firn densification activation energy depending on the temperature and impurity concentration).The final LID scenario has been smoothed using a Savitzky-Golay algorithm (25 points), and then provided as an input file to Paleochrono (Fig. S11).

Background scenarios and relative errors for the construction of AICC2023
With respect to the AICC2012 chronology, the background LID scenarios for EDC, Vostok, EDML and TALDICE ice cores are revised in AICC2023 (Table S8).We also reduce the background relative uncertainties associated with the LID, thinning and accumulation functions at the five sites (see Sect. 3.2 in the Supplementary Material).
Table S8.Origin of the background scenarios of LID, thinning and accumulation for EDC, EDML, Vostok, TALDICE and NGRIP and associated relative errors used in AICC2023.The LID prior relative uncertainty is set between 0.1 or 0.2 whether δ 15 N data are available or not.The mention "AICC2012" means that the scenario is the same than in AICC2012 (Bazin et al., 2013).With respect to the AICC2012 chronology, new stratigraphic links between ice and gas series are used to constrain AICC2023 over the past 120 kyr.They include tie points between CH4 series from EDC, EDML, Vostok, TALDICE and NGRIP ice cores (Baumgartner et al., 2014) as well as volcanic matching points between EDC, EDML and NGRIP ice cores (Svensson et al., 2020) (Fig. S12).The gas stratigraphic links used to construct AICC2012 over the last glacial period come from matching CH4 and δ 18 Oatm variations between ice cores.Still, an offset of several centuries is observed between Antarctic and Greenland CH4 records during the rapid increases associated with Dansgaard-Oeschger (D-O) events in AICC2012 (Fig. S12).Baumgartner et al. (2014) substantially extended the NGRIP CH4 dataset and provided accurate tie points between NGRIP, EDML, EDC, Vostok and TALDICE CH4 records.By implementing these new gas stratigraphic links in AICC2023, we improve the alignment between the CH4 records by several centuries, up to 500 and 840 years for the North Atlantic abrupt warming associated with D-O 5 and 18 respectively.The background thinning function is the same for AICC2012 and AICC2023 (dark blue dotted line).

Figure S2 .
Figure S2.Evolution of EDC δO2/N2 record between 260 and 100 ka BP and between 560 and 300 ka BP.(a) EDC raw δO2/N2 old data between 800 and 100 ka BP (black circles for data of Extier et al., 2018b; and purple squares for data of Landais et al., 2012), outliers (grey crosses) and low-pass filtered signal (black and purple lines).EDC raw δO2/N2 new data (blue triangles, this study) and low-pass filtered signals (blue line).(b) Compilation of the two datasets and low-pass filtered (blue line) or band-pass filtered (red line) compiled signal.(c) 21 st December insolation at 75° S on a reversed axis.

Figure S3 .
Figure S3.Identification of peaks position in filtered or unfiltered δO2/N2 record between 260 and 180 ka BP.(a) EDC δD (Jouzel et al. 2007).(b) EDC δO2/N2 (blue dashed curve) and low-pass filtered EDC δO2/N2 (red curve).Peaks position in the δO2/N2 record is identified as per methods a or b.Following the method a, the maximum position age max(i),a (on the bottom horizontal axis) is the middle of the age interval [x 1 ; x 2 ] (blue vertical rectangles) in which δO2/N2 values are superior to 95 % of the difference D i (vertical blue bars).The other peaks position is indicated in a similar way on the bottom horizontal axis.Following the method b, the extremum position is given by a 0 value in the temporal derivative of the filtered δO2/N2 record.The peak positions obtained with the method b (age max(i),b , age min(i),b ) are indicated by red vertical bars and displayed on the top horizontal axis.

Figure S4 .
Figure S4.Alignment of EDC δ 18 Oatm and climatic precession and impact on the chronology between 810 and 590 ka BP.(a) EDC ice age difference between AICC2012 and three test chronologies (1) test chronology 1 (grey dotted line), (2) test chronology 2 (black dashed line), (3) test chronology 3 (purple plain line).AICC2023 ice age 1σ uncertainty is shown by the orange area.The largest age difference between chronology 1 and 2 is indicated by the red arrow at 712.0 ± 2.6 ka BP.(b) Compiled EDC δ 18 Oatm (purple circles).(c) Precession delayed by 5 kyr (grey dotted line) and not delayed (black dashed line) (Laskar et al. 2004).(d) Temporal derivative of precession (black dashed line), delayed precession (grey dotted line) and of the compiled δ 18 Oatm record (purple plain line).(e) IRD (red by McManus et al. 1999; blue by Barker et al. 2019, 2021).The gray squares indicate periods where IRD counts are superior to the 10 counts g -1 threshold shown by the blue dotted horizontal line.Grey vertical bars illustrate new tie points between EDC δ 18 Oatm and delayed precession mid-slopes (i.e.derivative extrema) when IRD counts are superior to the threshold.Black vertical bars illustrate new tie points between EDC δ 18 Oatm and precession mid-slopes (i.e.derivative extrema) when no Heinrich-like events is shown by IRD record.The 12 kyr 2σ-uncertainty attached to the tie points is shown by the horizontal error-bars in panel b.

(
2022) identified a large peak (up to 61%) in the IRD record ofMcManus et al. (1999)  (red plain line in panel e) corresponding to HE 11 between 131 and 125 ka BP.Yet, if we consider the IRD record ofBarker et al. (2019Barker et al. ( , 2021) ) used in our study because it covers the last 800 kyr (blue plain line in panel e), we observe another large peak (up to 56 counts g -1 ) at around 150-156 ka BP.Because of this presence of IRD, to establish the test chronology 3, we tuned δ 18 Oatm to the 5-kyr delayed precession over the whole period stretching from 155 to 124 ka BP (gray frame), which is larger than the duration covering only HE 11.

Figure S5 .
Figure S5.EDC ice age difference between test chronology and AICC2023 between 300 and 100 ka BP.(a) EDC ice age difference between AICC2023 and 4 tests chronologies: (1) test chronology 1 (grey dotted line), (2) test chronology 2 (black dashed line), (3) test chronology 3 (purple plain line) and (4) test chronology 4 derived using only δ 18 Oatm-δ 18 Ocalcite matching (red plain line).AICC2023 ice age 1σ uncertainty is shown by the red area.(b) δ 18 Oatm data from EDC (purple circles) and Vostok (blue circles).(c) Precession delayed by 5 kyr (grey dotted line) and not delayed (black dashed line) (Laskar et al. 2004).(d) Temporal derivative of precession (black dashed line), delayed precession (grey dotted line) and of the compiled δ 18 Oatm record (purple plain line).(e) IRD (red by McManus et al. 1999; blue by Barker et al. 2019, 2021).The gray squares indicate periods where IRD counts are superior to the 10 counts g -1 threshold shown by the blue dotted horizontal line.Grey vertical bars illustrate new tie points between EDC δ 18 Oatm and delayed precession when IRD counts are superior to the threshold.Black vertical bars illustrate new tie points between EDC δ 18 Oatm and precession when no Heinrich-like event is shown by IRD record.The 12 kyr 2σ-uncertainty attached to the tie points is shown by the horizontal error-bars in panel b.
Figure S6.Mismatch Δ between background and analyzed LID for EDC over the 100-3200 m depth interval.(a) Experimental LID (orange) and modeled LID scenarios as per configuration 1 (with impurities, blue dots) and configuration 2 (without impurities, red dots).(b) Composite background LID as per tests A (black), B (blue) and C (red).(c) Analyzed LID scenarios given by Paleochrono.(d) Three values of the misfit Δ are calculated for the three composite LID: ∆  , averaged over the two depth intervals where δ 15 N data are not available (either between 578 and 1086 m or between 1169 and 1386 m, see intervals shown by grey rectangles), and ∆  , averaged over the whole 3200 m.

Figure S7 .
Figure S7.Mismatch Δ between background and analyzed LID for EDC over the 100-1500 m depth interval.(a) Background LID with and without adjusting the modeled LID to experimental LID values (orange and blue curves respectively).(b) Analyzed LID.(c) The averaged value of the misfit, Δ, is calculated for the two LID over the two depth intervals where δ 15 N data are not available (either between 578 and 1086 m or between 1169 and 1386 m, see intervals shown by grey rectangles).
BP (Fig.S10).To do so, the Vostok δ 18 Oatm record and the Chinese δ 18 Ocalcite signal are linearly interpolated every 100 years, smoothed (25 points Savitzky-Golay) and extrema in their temporal derivative are aligned.35 new tie points are identified and attached to a 1σ-uncertainty between 2.3 and 3.5 kyr.They replace the 35 age constraints obtained by aligning δ 18 Oatm and delayed precession, associated with a 6 kyr 1σ-uncertainty and used to construct AICC2012.
EDML using δ 15 N data and firn model estimatesWhen δ 15 N measurements are not available,Bazin et al. (2013) used a synthetic δ 15 N signal based on the correlation between δ 15 N and δD to estimate the background LID scenario at EDML and to constrain the AICC2012 timescale.In this work, the background LID scenario at EDML is estimated from δ 15 N data (when available), which is corrected for thermal fractionation.The thermal fractionation term is estimated by the firn model.Otherwise, the background LID is calculated by the firn model running in the same configuration as for calculating the modeled LID at EDC (i.e.firn densification activation energy depending on the temperature and impurity concentration).The final LID scenario has been smoothed using a Savitzky-Golay algorithm (25 points), and then provided as an input file to Paleochrono (Fig.S11).TableS7.Method of determination of LID background scenario according to EDML depth range.The thermal fractionation term is estimated by the firn model running in the same configuration as for calculating the modeled LID, i.e. firn densification activation energy depending on the temperature and impurity concentration.

Figure S12 .
Figure S12.CH4 records from Antarctic and NGRIP sites over the last 122 kyr.CH4 from EDML, TALDICE, NGRIP and EDC ice cores on the AICC2012 gas timescale (top panel).CH4 from EDML, TALDICE, NGRIP and EDC ice cores on the AICC2023 gas timescale (bottom panel).Stratigraphic links between CH4 series from EDC, EDML, Vostok, TALDICE and NGRIP ice cores (blue triangles and black squares, Baumgartner et al., 2014) and between volcanic sulfate patterns from EDC, EDML and NGRIP ice cores (vertical bars, Svensson et al., 2020) are used to constrain AICC2023 over the last 122 kyr.Abrupt D-O events are shown by grey rectangles and numbered from the youngest to the oldest (1-25)(Barbante et al., 2006).

Figure S14 .
Figure S14.Analyzed accumulation and thinning functions for EDC over the last 800 kyr.They are provided by AICC2012 and AICC2023 (black and blue plain lines respectively) along with their absolute uncertainties (gray and yellow respectively).

Table S4 . The different prior relative uncertainties tested for LID, thinning and accumulation
. The LID prior relative uncertainty is set between 0.1 or 0.2 whether δ 15 N data are available or not.

EDC ice age difference between each test chronology and AICC2012 timescale between 800 and 0 ka
BP.The ice age uncertainty (1σ) obtained for each test is shown by the dotted lines.The red arrow indicates the largest age mismatch between the test chronologies.Figure S9.

EDC ice age difference between each test chronology and AICC2012 timescale between 170 and 50 ka BP.
The ice age uncertainty (1σ) obtained for each test is shown by dotted lines.

Table S6 . Method of determination of LID background scenario according to Vostok depth range.
The thermal fractionation term is estimated by the firn model running in configuration 1: Firn densification activation energy depending on the temperature and impurity concentration.