BICYCLE-SE is a carbon cycle box model consisting of 10 ocean boxes, a one box atmosphere and a 7 box terrestrial biosphere which also considers carbon exchange fluxes with the solid Earth by a sediment module, volcanic outgassing, weathering and coral reef growth. It is fully described in with the processes related to carbon isotopes being updated recently . The simulation scenarios shown here have already been discussed in . Scenario A0 is a run without abrupt changes in AMOC during Greenland stadials, while AMOC changes drastically in scenario A3. These AMOC changes are prescribed by the mean of two independent Iberian Margin sea surface temperature (SST) data sets which is then rescaled to obtain a minimal AMOC of 2 Sv during HS1, since this agrees best with 14C reconstructions in the deep Atlantic Ocean . See Fig. S1 in the Supplement for details on both the prescribed climate change and the simulated MRA with BICYCLE-SE.

LOVECLIM is an EMIC, which includes a quasi-geostrophic T21 atmospheric model, an OGCM (3°×3°, 20 vertical levels) coupled to a dynamic-thermodynamic sea-ice model, a vegetation model and a global carbon cycle model. The conditions at 54 kyrBP were obtained through a transient experiment starting at 140 kyrBP . The model is forced with the transient evolution of orbital parameters , atmospheric greenhouse gas concentration , and continental ice-sheet geometry and albedo. For the period 140–120 kyrBP, the evolution of the continental ice-sheet geometry and albedo are as described in the PMIP4 protocol of the penultimate deglaciation . Between 120 and 20 kyrBP, the continental ice-sheet geometry and albedo evolution are given by a 130 kyrBP off-line ice-sheet model simulation . Between 20 and 2 kyrBP, the model is forced by the continental ice-sheet geometry and albedo evolution from ICE4G . At 54 kyrBP, the Bering Strait is closed, and is gradually opened between 11 and 10 kyrBP.

Between 140 and 54 kyrBP, the atmospheric Δ14C is set constant at 0 ‰. The model is then re-equilibrated with atmospheric Δ14C at 54 kyrBP for 5000 years. From 54 kyrBP the model is forced by the transient evolution of atmospheric Δ14C using the IntCal20 data . To allow the ocean to fully equilibrate the oceanic Δ14C can be used in this context here from about 50 kyrBP onwards.

To simulate the millennial-scale variability of the last glacial period and the impact of deglacial ice-sheet disintegration (scenario fwf), meltwater is added into the North Atlantic (50–60° N, 10–60° W) with the timing of these events being based on Iberian Margin SST from . In detail, a meltwater flux of 0.3 Sv was added to achieve a collapsed AMOC during Heinrich stadials (HS) 1–4, while only 0.2 Sv was added during HS5. A triangular pulse of up to 0.3 Sv was also added between 13 and 12 kyr BP to simulate a reduced AMOC during the Younger Dryas. For comparison, a simulation without such meltwater fluxes and related millennial-scale changes is performed (scenario nofwf). See Fig. S2 for details on both the simulated climate change and the simulated MRA with LOVECLIM.

The Bern3D model is a coarse resolution (68×46×40 irregular spaced grid in longitude, latitude and depth) Earth system model (or EMIC) that couples a frictional geostrophic ocean to 2D energy-moisture balance atmosphere. In contrast to the detailed description of the model in , we here employ it without the dynamical ice-sheet component. Instead, ice-sheet evolution over the last 55 kyr is prescribed. For this, the Last Glacial Maximum (LGM) ICE-6G, and modern ice-sheet extents are linearly interpolated based on the benthic δ18O record of for the last 55 kyr. The biogeochemical cycle including the implementation of radiocarbon is described in and .

In addition to changes in orbital configuration , greenhouse gases (CO2, CH4, and N2O, ), and ice-sheets, aerosol radiative forcing due to the changing atmospheric dust load is prescribed as well for the transient simulations. The temporal evolution of the aerosol radiative forcing follow the EPICA Dome C dust record , which has been normalised to values between 0 during the late Holocene and -2 Wm-2, which corresponds to an average radiative forcing of about -1 Wm-2 at the LGM see. In scenario PallSTD, additional freshwater fluxes are applied into a box in the North Atlantic (45–70° N, 66° W–14° E) during all stadials, with a maximum freshwater flux of 0.4 Sv during HS and 0.2 Sv during non-Heinrich stadials. The timing of stadials is, similarly as in BICYCLE-SE, based on the Iberian Margin SST as published in .

The model was first spun up to PI conditions, followed by a 10 kyr spinup to the conditions of 65 kyrBP. The model was then run transiently from 65 kyrBP to present day, with the time from 65 to 55 kyrBP, serving as an additional transient spinup period (with atmospheric Δ14C fixed at its IntCal20 value for 55 kyrBP) to reach a realistic radiocarbon content in the deep ocean. The additional scenario Pnofwf is in all but the missing freshwater fluxes during stadials identical to PallSTD. Note that the simulations shown here have different spin-ups – and contain no flux corrections from the North Atlantic to the North Pacific (45–70° N) – and so provide different results to the runs published in which considered multi-proxies during Termination 1. See Fig. S3 for details on both the simulated climate change and the simulated MRA with Bern3D.

Output from the LSG OGCM has already been used within IntCal20 . LSG has been used with a horizontal resolution of 3.5° and a vertical resolution of 22 unevenly spaced levels. Further setup details are contained in . No additional simulations (or outputs) have been performed (or generated). The model has been run with temporally constant ocean circulation which differed in three climate setups. They contain a scenario that mimics present-day climate background conditions approximating the Holocene and interstadials. One glacial scenario aims at representing the LGM, features a shallower and by about 30 % weaker AMOC compared to interglacials. A second glacial climate scenario mimics cold stadials with further AMOC weakening by about 60 %. As described in detail in these different climate setups have been generated by forcing the surface ocean with output from an atmospheric model, that itself was driven by different SST reconstructions. Each of these three climate setups has been performed for three different atmospheric Δ14C forcings ending in nine simulations. The three atmospheric Δ14C forcings contain the posterior mean estimate for the Hulu-based 14C atmospheric reconstruction and two time series that cover its 95 % confidence interval (CI) (mean ±2σ). Here, the median MRA from these nine simulations is shown, which is also what has been used within IntCal20.

For an initial evaluation of model-based surface MRA we compare simulations with data-based MRA from the Global Ocean Data Analysis Project (GLODAP) , which is a synthesis of ocean sampling expeditions carried out in the 1990s within the World Ocean Circulation Experiment (WOCE), the Joint Global Ocean Flux Study (JGOFS), and the Ocean Atmosphere Carbon Exchange Study (OACES). Here, we calculate 2MRA=8033⋅ln⁡Δ14Catmosphere1000+1pre-bomb DI14CabioticPI DICabiotic in units 14C yr based on the approach used for the ocean model intercomparison project . Abiotic PI dissolved inorganic carbon (DIC) has been derived from the atmospheric CO2 concentration of 284 ppm and pre-bomb abiotic dissolved inorganic 14C (DI14C) from Δ14Catm=-24 ‰. According to , the latter represents the mean value of the decade before the onset of the bomb-peak in mid-1955, which has been used as pre-bomb baseline to reconstruct marine 14C from potential alkalinity . This so-called “natural” GLODAP-based MRA will always be a compromise since 14C is only corrected for bomb-14C, but not for fossil fuels, the so-called 14C-Suess effect .

To evaluate this natural GLODAP-based MRA we calculate local anomalies in MRA, the so-called ΔR, by subtracting the Marine20-based global MRA for 1950 CE (407 14C yr), and compare the resulting map of ΔR with the entire content of the ΔR data base found at http://calib.org/marine/ (last access: 2 October 2025) . We take the 2000 entries of this data base which were available with stated uncertainties on 2nd October 2025. The data were collected mostly between the years 1729 and 1959 CE with three entries from the 17th century, one from 1512 CE and one from before CE. ΔR is calculated from the difference of the reverse-calibrated collection year using Marine20 and the measured 14C age and is assumed to stay constant in time. We average the data in a spatial resolution of 2° in both latitude and longitude ending with 609 values. Here, we use (as in the application of the online scripts of the data base) weighted means for averaging with the mean of ΔR (ΔR‾) given by 3ΔR‾=∑iΔRiσi2∑i1σi2 where σi is the uncertainty in ΔRi and the reported uncertainty σ of ΔR‾ is the maximum of the SD of ΔR‾ and the weighted uncertainty in ΔR‾ (see http://calib.org/marine/AverageDeltaR.html, last access: 2 October 2025, or , for details). We plot not only ΔR‾, but also a minimum and a maximum version with ΔR‾±1σ (Fig. S4). Independent of which version of the data-based ΔR we take we find in general a good agreement between them and natural GLODAP (differences of typically up to 100 14C yr) with the exception of single data points and the entire the west coast of North America, where values in the data base are more than 100 14C yr older than in GLODAP, potentially caused by coastal effects.

In order to compare model outputs with observations, we make use of compiled deglacial marine radiocarbon data from and regional time-series splines based on planktic foraminifera from . The latter include regional time series from 25 sites for the surface northeast Atlantic (>52° N and east of 24° W) and the Iberian Margin (∼38° N, 10° W). We use the deep Atlantic MRA (i.e. B-Atm, deeper than 2 km water depth) based on 34 sites that contributed to the baseline selection of in the realisation of . We do not compare model results to , since they used age models based on IntCal09 and the resulting 42 kyr-long time series of surface MRA cover 0–65° N in the Atlantic, a value which cannot be extracted from the box model simulations.

Additionally, point-wise surface MRA for the time slices of the LGM (19–21.8 kyrBP) and the HS1 (15–17.5 kyrBP) from the baseline selection of the data base of are used for further model evaluation. Here, multiple MRA entries for the same sites are averaged using weighted means (similar to Eq. ) reducing 67 entries for the LGM to 19 values (89 entries for the HS1 to 22 values), from which 13 exist for the same location in both periods making the calculation of HS1–LGM differences in MRA possible. Here, one entry for the LGM was rejected as outlier since its MRA was >1000 14C yr offset from four alternative MRAs for the same site. Most surface MRA data are from the northeast Atlantic (between Iberian Margin and Iceland) which are complemented with data from single sites in the Southern Ocean, the tropical East- and West Pacific and the tropical Atlantic.

Estimates of changes in surface MRA are necessary for 14C dating of marine carbon-containing material. For radiocarbon dating of marine organisms, such as foraminifera, it is important to take into account the organism’s seasonal/depth habitat, e.g. for the estimation of appropriate local variations in reservoir age, that is ΔR , or for comparison with model-based MRA estimates. Planktonic foraminifer habitats are species-specific, and often related to the mixed layer depth, but sometimes extend to greater depths . The mixed layer depth has a seasonal cycle, but for modern day latitudes <50°, it is typically less than 200 m , which in our two EMICs is indeed the case for the annual mean mixed layer depths during the PI (Fig. S5). Note that due to a lack of more recent results from LOVECLIM, results for 2 kyrBP are chosen to stand in for the PI reference. The mean surface MRA calculated for the top 50 m within IntCal20 from LSG output might be relevant to foraminifer data from the lower end of the mixed layer depth range, while the 100 m thick non-polar surface ocean boxes of the BICYCLE model used within Marine20 were probably in the middle of the relevant mixed layer depth range. From the GLODAP data we calculate a mean surface MRA for latitudes <50° of 291, 318 and 356 14C yr, when considering data from the top 50, 100, and 200 m water depth, respectively, illustrating the centennial-scale of the uncertainty related to the assumed habitat depth range. Furthermore, calculated surface MRA for different depth ranges from the transient simulations provide insights into how this depth-dependency might vary with time (Figs. S2 and S3). Here, the depth ranges over which surface MRA have been calculated differ in detail (Table ), depending on model grid, since we only consider full vertical layers. Differences between MRA based on roughly the top 50 m or 100 m are similar, and in agreement with GLODAP, while those based on roughly the top 200 m are about 50 and 100 14C yr higher than MRA of top 50 m in Bern3D and LOVECLIM, respectively.

Maps of surface MRA for 2 kyrBP (our PI reference) compare well with MRA based on natural 14C in GLODAP (Fig. ). The area-weighted root mean square errors of the residuals of the model-based differences to natural 14C in GLODAP are 202, 133, 86 and 265 14C yr for BICYCLE-SE, LOVECLIM, Bern3D and LSG, respectively, which are all substantially smaller than the 399 14C yr of the area-weighted root mean square of the natural 14C in GLODAP, evidencing the explanatory power of all the four models. Note that atmospheric Δ14C at 2 kyrBP was with -16 ‰ only slightly different from the pre-bomb value of -25 ‰ in 1950 CE (0 kyrBP). However, be aware that this comparison has certain weaknesses due to the compromises in the GLODAP data (not free of the 14C-Suess effect, see Methods section for details) and our choice of using simulation results at 2 kyrBP. For latitudes <50° the simulated surface MRAs are 461, 401, 344 and 443 14C yr for BICYCLE-SE, LOVECLIM, Bern3D and LSG, respectively and the differences from GLODAP are for all models typically less than 300 14C yr with BICYCLE-SE and LSG showing predominantly larger values, while the two EMICs (especially Bern3D) seem to be closer to GLODAP. The surface MRA in LOVECLIM in the North Atlantic seemed be an exception here, showing a difference of more than 200 14C yr, a larger offset to GLODAP than the other models. LSG contains remarkably large positive offsets from GLODAP of more than 600 14C yr in high latitudes probably related to the simulated sea ice. BICYCLE-SE shows a negative offset from GLODAP in the Southern Ocean of around 500 14C yr. This offset might be caused by BICYCLE-SE's 1000 m deep surface ocean boxes in the polar regions. However, since a similar offset to GLODAP is missing in the northern high latitudinal boxes, which are also 1000 m deep, some counteracting processes might be at work here.

As a metric of the simulated climate variations in the various models we here show changes in global mean sea surface temperature (GMSST). The simulated GMSST dropped during the LGM, depending on model, between -1.5 and -2.5 K with respect to PI (Fig. a). While the -2.5 K cooling (BICYCLE-SE) is in the range covered by data-based studies e.g. the upper end of the range of -1.5 to -2.0 K (Bern3D, LOVECLIM) is probably under-estimating climate change during full glacial conditions.

The timing of AMOC reductions in the model simulations was prescribed to coincide with the timing of Greenland stadials expressed in the ice-core record of δ18O e.g., and was also captured in the SST reconstruction at the Iberian Margin . Here, the precise timing in BICYCLE-SE and Bern3D is based on , while in LOVECLIM data from have been used, giving similar results. In Bern3D the applied freshwater fluxes were adjusted according to its known freshwater sensitivity to achieve a virtually collapsed AMOC during Heinrich stadials and a strongly reduced AMOC during non-Heinrich stadials, while in LOVECLIM AMOC collapsed due to the applied freshwater forcing only in HS1–4, but was slightly reduced in HS5 and the Younger Dryas.

In BICYCLE-SE sensitivity tests and a comparison of simulated deep ocean MRA with reconstructions has been used to choose the applied AMOC changes . The resulting changes in simulated AMOC strength across models (Fig. b) varied widely, e.g. 16–26 Sv for PI, 10–20 Sv for the LGM and 2–5 Sv for HS1 and the Younger Dryas, summarizing results for runs with abrupt AMOC shutdown. Note that the AMOC strength was calculated differently in each model (see caption to Fig.  for details). For earlier HS, most models show smaller AMOC reductions than for HS1 and LOVECLIM also contains a strong but short reduction in AMOC during the 8.2 kyr event not covered by the other models. Notably, the AMOC at the LGM in Bern3D, which has been evaluated in a multi-proxy approach to about 11 Sv is now ∼15 Sv, since flux corrections applied in the other study have been neglected here.

The simulated non-polar MRA (<50°) contains remarkable model-specific differences. From the absolute values we can identify the model-specific offsets with Bern3D simulating the lowest PI MRA, while BICYCLE-SE simulates the highest PI MRA (Fig. d). On the other hand, at the LGM (20 kyrBP), the smallest MRA is simulated by LSG, followed by Bern3D, LOVECLIM, and BICYCLE-SE. Across the last glacial period, Bern3D simulates the largest changes in MRA, with a 1000 14C yr anomaly with respect to PI during the Laschamps geomagnetic excursion around 42 kyrBP , while LOVECLIM and LSG display the smallest MRA anomalies with respect to PI (Fig. e). The differences to Marine20 – when plotted with respect to PI – show to a large degree model-specific responses which nearly all fall in their sizes within the 95 % CI of Marine20 (Fig. f). One more general pattern is the decrease of this difference in results from BICYCLE-SE and Bern3D during Greenland stadials not similarly seen in LOVECLIM.

Only when we calculate differences from simulations with and without abrupt AMOC shutdown during stadials for the same model do we find the tendency of smaller non-polar surface MRA during stadials (Fig. g). However, a notable exception here is HS1 in LOVECLIM which shows a more complex (rather opposite) dynamic. showed that ΔMRA remains small (∼100 14C yr) when changing the eddy diffusivity in the box-diffusion model that has been widely used to convert atmospheric Δ14C changes in terms of oceanic changes ( and subsequent IntCal calibration iteration until 2013). The main point outlined by is that the global average surface MRA and mean deep ocean 14C age vary in opposite directions when eddy diffusivity changes, mimicking global overturning variations (all else being equal). Interestingly, scaling the eddy diffusivity to our AMOC proxy curve SST record by leads to a similar ΔMRA pattern in the box-diffusion model as in the other models when they are averaged in the low and mid-latitudes (Fig. g). The linear scaling of the eddy diffusivity assumes a modern value of 4000 m2yr-1 and a minimum value of 500 m2yr-1 at ca. 40 kyrBP during the coldest interval of HS4 (ΔSST of -10 K, Fig. d). Instantaneous steady state MRA values are then calculated with the analytical Eq. (4) derived by . In the box-diffusion model, the 14C production is mainly balanced by radioactive decay in the deep ocean reservoir. Reducing or stopping the exchange with the mixed layer implies that 14C remains confined to the atmosphere and the mixed layer. These two boxes, which contain comparable total carbon inventories, would tend to homogenize, thus reducing the MRA.

When plotting changes in MRA reconstructions for the LGM together with model outputs we find that most models underestimate for most locations the changes in surface MRA (Fig. ). The best agreement, especially at high latitudes, is obtained for LSG. This comparison is rather limited, due to the availability of only 19 data points. It has as additional caveat that simulations are focused on 20 kyrBP, while the data cover the wider time window of 19–21.8 kyrBP. Changes in models and data only roughly agree in a few areas (Iberian Margin, Caribbean) but elsewhere differ widely, with models underestimating the MRA derived from proxy records. Note that the about 10 data point in the northeast Atlantic show partly very different changes, ranging from a reduction in MRA to a rise by more than 1000 14C yr. This local diversity of the reconstructions challenges our model-data comparison since small-scale localised effects seen in the data might not be contained in the coarsely resolved models. Notably, there is also significant disagreement between models in particular regions. One reason for the data-model offsets might be the habitat depth of the planktonic foraminifera, even though the EMICs give surface MRA that differ by less than 50 14C yr when based on roughly the top 200 m instead of the top 50 m used so far (Fig. S6). Interestingly, annual mean mixed layer depth at 20 kyrBP in the two EMICs differ in the non-polar regime only very little from the PI control (Fig. S5). Another reason may be that the models are all missing key processes that might influence vertical diffusivity in the upper ocean or air–sea gas exchange at the surface, particularly at high latitudes, or that they tend to simulate a different radiocarbon distribution in the interior ocean than prevailed in the past. Going here into more detail, e.g. by an extended deep ocean 14C model-data comparison is beyond the scope of the present study, but a potential promising avenue for future studies.

To further elucidate the model-specific responses for HS1, we compare simulations with reconstructions for that time interval (Fig. ). Here, we focus on the HS1–LGM difference. Since the simulations with the LSG OGCM were not containing abrupt AMOC changes during Greenland stadials, the output from this model is not included here. For this comparison, 13 data points exist in the data base of . Both EMICs show a general MRA decrease at mid and low latitudes notably in the Indo-Pacific. This almost global MRA reduction is compatible with that obtained with the BICYCLE-SE model. In addition to this common pattern, the models exhibit different dynamical behaviour (e.g. a different ocean overturning or a different air–sea gas exchange of 14C) that is responsible for an increase of MRA in the northern North Atlantic, with varying intensity and spatial extent: confined to the northern North Atlantic (+Arctic) in BICYCLE-SE, intermediate in Bern3D, and widespread to most of the Atlantic basin in LOVECLIM. Thus, the different response in the Atlantic in HS1 readily explains LOVECLIM's different dynamics when comparing non-polar averages of runs with and without abrupt AMOC shutdown (Fig. g). There are areas where models disagree with available data. In particular, the Iberian margin data, e.g. shown previously in , exhibit a clear MRA increase during HS1, as already acknowledged in , which is covered in LOVECLIM, but not in Bern3D. Both EMICs show the existence of a radiocarbon bipolar seesaw pattern, with widespread decrease in the Southern Ocean, not restricted to the Atlantic sector. This radiocarbon bipolar seesaw pattern with MRA increases in the northern Atlantic and MRA reductions in the Southern Ocean is very reminiscent of the thermal bipolar seesaw available from paleo-SST data e.g. and from numerous model experiments e.g.. Indeed, this seesaw pattern in MRA variability is directly consistent with what has been previously observed in direct MRA reconstructions , where it has been hypothesized to relate to sea-ice variability in each hemisphere. Interestingly, the mixed layer depth is reduced in both the North Atlantic and the Southern Ocean between HS1 and LGM, especially in Bern3D (Fig. S5, right column). This in-phase change of the mixed layer depth with the MRA in the north together with their anti-phase change in the south suggests that other processes are more important for explaining surface MRA in connection with AMOC weakening.

In some of the regions where the two EMICs disagree, reconstruction-based evidence also exists – in particular for the Iberian Margin. However, in the equatorial Atlantic, where the models show opposite changes in MRA, such direct measurements are not available. This prevents a data-based evaluation of the two diverging simulations.

An alternative to calculating the MRA anomaly during HS1 is based on the different surface MRA at 16 kyrBP from simulations with and without abrupt AMOC shutdowns (Fig. S7, right column). However, such a model-based anomaly cannot directly be compared with reconstructions. Here, results from all contributing models (BICYCLE-SE, LOVECLIM, Bern3D) show more positive ΔMRA than in the calculations based on the HS1–LGM difference, but contain similar patterns. Increases in surface MRA in the non-polar Atlantic basin are now also contained in BICYCLE-SE.

As a final step, we compare simulated and reconstructed time series of surface northern Atlantic (>50° N) and deep Atlantic MRA – all with respect to PI (Fig. ). The maximum lengths of reconstructed timeseries are limited to the past 25–30 kyr, but they provide useful insights for Termination I, including HS1. As expected, simulations with abrupt AMOC changes contain higher ΔMRA during HS1 than those without these abrupt changes. However, only LOVECLIM reached changes of 1300 14C yr in the surface northern Atlantic as found in the reconstructions, and only Bern3D reaches the HS1 peak of 1200 14C yr in the deep Atlantic, although some thousand years later. The model-data offset at the surface northern Atlantic might have contributed to aliasing of spatial patterns, whereby the compiled data are largely restricted to east of 24° W in the northern Atlantic >42° N whereas model results include the whole northern Atlantic and Arctic Ocean (>50° N). However, if this was the case, MRA variability in the western North Atlantic (and Arctic) should have been greatly reduced compared to the eastern part of the basin, which is opposite to what the model simulations tend to suggest (Fig. ). Note that in BICYCLE-SE, ΔMRA in both surface northern Atlantic and deep Atlantic are very similar probably linked to the box model geometry and prescribed fluxes which contain a direct connection and vigorous exchanges between both boxes. All models tend to agree on having peaks in both surface northern Atlantic and deep Atlantic ΔMRA connected with millennial-scale climate change. However, amplitudes are model-specific with Bern3D simulating the largest amplitudes which even appear, although in smaller size, in the simulation without abrupt AMOC changes. This indicates as already notified in that they are partly related to the variability contained in either atmospheric Δ14C or CO2 and not only to abrupt AMOC changes, though the impact of such atmospheric effects is likely limited to less than ∼300 14C yr equivalent for the global deep ocean and far less for the surface ocean . Possible explanations might be related to changes in the terrestrial carbon cycle e.g..

Finally, we calculate changes in surface MRA from a multi-model mean (MMM) based on low latitude (<50°) results of scenarios with abrupt AMOC changes in the two EMICs, which are then compared with Marine20 (Fig. a, b). While the full range of simulation results is nearly always within the 95 % CI of Marine20, it is on its lower edge since the LGM. Furthermore, the MMM is consistently about 100 14C yr smaller than Marine20. We tested if this offset might be based on the chosen depth of the analysed surface MRA. However, we only found a slightly better agreement between MMM and Marine20 when using results from the mean of the top 200 m instead of the top 50 m as done in the default case (Fig. a). This offset between MMM and Marine20 again illustrates that the pre-bomb non-polar MRA, which was about 407 14C yr for the year 1950 CE when using BICYCLE within Marine20 is model-specific.

In a very last step we additionally calculate changes in the surface MRA of the high latitudes (>50°) and of the global mean in order to understand how much they differ from the MRA of the non-polar areas (Fig. S8). The MRA in both high northern and high southern latitudes in LSG contain a jump by more than 1000 14C yr at 10.7 kyrBP; however, this jump is related to an arbitrary change to an ‘interglacial’ overturning scenario from the reduced “glacial” overturning scenario in the LSG model. As the much larger MRA at the LGM in high latitudes in LSG agreed better with reconstructions than the smaller values of the EMICs (Fig. ), this raises some questions: if these sparse LGM-based reconstructions are broadly representative of glacial conditions, one might ask if the EMICs are missing or inadequately resolving key processes in the high latitudes during glacial times. Alternatively, if the EMIC-based glacial polar MRA turn out to be more accurate than the LSG-based results any polar age calibration based on the latter are then potentially biased towards too old values. For the two EMICs, the global mean MRA and the non-polar MRA are very similar (Fig. S8) suggesting that the non-polar MRA might indeed be of global applicability. However, when looking to the details we find that the range and the MMM of the changes in the northern polar MRA (which is similar, but not identical to the MRA of the northern North Atlantic discussed earlier) is – apart from certain HSs with several centuries older MRA – indeed comparable with the 95 % CI of Marine20 (Fig. c). The changes in the southern polar MRA are much higher than in Marine20, both in range and in MMM (Fig. d), especially in Bern3D, although still by a few centuries younger than in LSG (Fig. S8). Thus, a global usage of the non-polar MRA would lead especially in the Southern Ocean but also at the Iberian Margin; to underestimations of MRA by at least one century, leading to an overestimate of calendar age in probes whose radiocarbon age calibration is based on them. This Southern Ocean age offset in the models is probably related to the higher glacial sea ice coverage (Fig. S9) and the larger glacial mixed layer depth (Fig. S5). Here, larger glacial summer sea ice in the Southern Ocean and larger glacial mixed layer depth in Bern3D compared to LOVECLIM potentially explain at least part of the model-specific ΔMRA in the Southern Ocean.