Comparing the impacts of ozone-depleting substances and carbon dioxide on Arctic sea ice loss

The rapid decline of Arctic sea ice is widely believed to be a consequence of increasing atmospheric concentrations of greenhouse gases (GHGs). While carbon dioxide (CO2) is the dominant GHG contributor, recent work has highlighted a substantial role for ozone-depleting substances (ODS) in Arctic sea ice loss. However, a careful analysis of the mechanisms and relative impacts of CO2 versus ODS on Arctic sea ice loss has yet to be performed. This study performs this comparison over the period 1955–2005 when concentrations of ODS increased rapidly, by analyzing a suite of all-but-one-forcing ensembles of climate model integrations, designed to isolate the forced response to individual forcing agents in the context of internal climate variability. We show that ODS have played a significant role in year-round Arctic sea ice extent and volume trends over that period, accounting for 64% and 32% of extent and volume trends, respectively. These impacts represent 50% and 38% of the impact from CO2 forcing, respectively. We find that ODS act via similar physical processes to CO2, causing sea ice loss via increased summer melt, and not sea ice dynamics changes. These findings imply that the future trajectory of ODS emissions will play an important role in future Arctic sea ice evolution.


Introduction
Observations of Arctic sea ice extent (SIE) and volume (SIV) have documented a precipitous sea ice decline over the satellite era Notz 2018, Meredith et al 2019). While internal climate variability has played a non-negligible role in observed Arctic sea ice loss (e.g. Kay et al 2011, Stroeve et al 2012, Notz 2015, Ding et al 2019, England et al 2019, the recent IPCC AR6 concluded that it is very likely that anthropogenic forcing due mainly to greenhouse gases (GHGs) is responsible for more than half of the observed summer sea ice loss (IPCC 2023). Carbon dioxide (CO 2 ) is the dominant source of anthropogenic GHG radiative forcing, and thus much of the sea ice community's work has focused on the sea ice response to CO 2 emissions (e.g. SIMIP Community 2020). However, recent work has highlighted that other anthropogenic GHG forcing agents-particularly ozone-depleting substances (ODS)-also played a substantial role in Arctic sea ice loss (Polvani et al 2020, Sigmond et al 2023. ODS are halogenated carbon compounds (e.g. chlorofluorocarbons), which were developed for use as refrigerants, aerosol propellants, and solvents. The substantial emission of ODS into the atmosphere began in the 1950s and continued to increase unabated until the signing of the Montreal Protocol in 1987, which strongly regulated their use (World Meteorological Organization 2018). ODS are well known as the primary driver of stratospheric ozone (O 3 ) depletion and the cause of the Antarctic ozone hole (Solomon 1999). Less well known are the direct impacts of ODS on the climate system via radiative forcing. ODS are long-lived gases with powerful radiative efficiencies: For example, CFC-11 and CFC-12 have 20-year global warming potentials roughly 10 000 times that of CO 2 (World Meteorological Organization 2018). Thus, despite their relatively low concentrations compared to CO 2 , ODS have a notable radiative forcing (Ramanathan 1975, Montzka et al 2011. Specifically, over the period  where ODS concentrations increased rapidly, the radiative forcing from ODS was 0.31 W m −2 (Meinshausen et al 2011). This represents 30% of the CO 2 forcing over this period (1.02 W m −2 ) and a slightly larger forcing than methane (0.23 W m −2 ; based on abundance changes), making ODS an important contributor to GHG forcing over the second half of the 20th century.
A number of earlier studies have examined the so-called 'World Avoided' by the signing of the Montreal Protocol, showing that the reductions in ODS emissions since 1987 have resulted in a substantial decrease in surface climate warming (Hansen et al 1989, Velders et al 2007, Morgenstern et al 2008, Garcia et al 2012, Polvani et al 2016, Goyal et al 2019 and Arctic sea ice loss . Recently, Polvani et al (2020) investigated a related but distinct question: how much of the historical climate change that has already been observed can be attributed to historical forcing from ODS? They found that when ODS are kept fixed at 1955 levels in the CESM1 climate model, Arctic September SIE loss and annual-mean Arctic surface air temperature (SAT) warming are only one half of those in the historical integrations (in which ODS concentrations increase markedly over the second half of the 20th century). In recent analogous experiments performed with the CanESM5 model, Sigmond et al (2023) also found that September SIE loss and Arctic warming are roughly one half of those in historical integrations.
The impacts from ODS on sea ice loss found by Polvani et al (2020) are remarkably large. Their surprising result, however, is in part related to the large cancellation that occurs between the warming effects of CO 2 and the cooling effects of aerosols over the second half of the twentieth century (e.g. Fyfe et al 2013, Kong andLiu 2023), which tends to highlight the impacts of ODS compared to this residual. Therefore, in order to better understand the impact of ODS on Arctic sea ice loss, it is necessary to (i) directly compare the impacts of ODS and CO 2 ; and (ii) understand the physical mechanisms contributing to the differences in sea ice trends. In this work, we address these topics by analyzing all-but-one-forcing climate model ensembles in which ODS and CO 2 are respectively held fixed at their 1955 levels and compared to a historical forcing ensemble. We use large ensembles for this study in order to robustly evaluate the forced response of sea ice to different forcing agents in the context of internal climate variability. We first quantify the monthly impacts of ODS on a range of sea ice variables and directly compare these with the impacts of CO 2 forcing. We then consider the physical processes contributing to the sea ice trend differences, analyzing sea ice mass budgets, Arctic surface energy budgets, and sea ice sensitivity across the experiments. Finally, we discuss the implications of these findings for past and future simulations of Arctic climate change.

Model experiments
This study is based on a multi-forcing suite of single-model initial condition large ensembles (SMILEs) run using the Community Earth System Model version 1 with the Community Atmosphere Model, version 5 (CESM1(CAM5); Hurrell et al 2013). Our baseline SMILE is the widely-used CESM1-LE (Kay et al 2015), which is a 40-member ensemble that uses CMIP5 historical radiative forcings from 1920-2005. The initial conditions for the CESM1-LE are nearly identical, differing only by roundoff-level perturbations in their atmospheric initial states. We select 10 members from this ensemble from which forcing perturbation experiments are branched in 1955. We refer to this 10-member ensemble of historical simulations as the 'All' forcing ensemble.
In addition, we here analyze three 10-member ensembles spanning 1955-2005, in which certain forcing agents are held fixed at 1955 levels. Specifically, we run 'FixCO 2 ' , 'FixODS' , and 'FixODSO 3 ' ensembles, which respectively fix the CO 2 , ODS, and ODS and O 3 concentrations at their 1955 levels. In these fixed-forcing experiments, all other forcing agents are specified to evolve according to their time-varying historical values. This 'all-but-one-forcing' approach allows one to directly quantify the impact of given forcings via comparison to the historical runs, and assuming a linear response. It is widely used in climate science (e.g. Polvani et al 2011, Deser et al 2020. For readers unfamiliar with ODS, we briefly note that these compounds include chlorofluorocarbons, hydrochlorofluorocarbons, halons, carbon tetrachloride, methyl chloride, methyl chloroform, and others. CAM5 specifies the concentrations of two ODS forcing agents, CFC-11 * and CFC-12, which are time-varying and 'well-mixed' , i.e. with the same value for all latitudes, longitudes, and vertical levels. CFC-11 * is a linear combination of CFC-11 and 25 other halocarbons, weighted by their radiative efficiency relative to CFC-11. For the FixODS experiments, CFCs and related halocarbons are held fixed at their 1955 values and for the FixODSO 3 experiments both ODSs and stratospheric O 3 are held fixed at their 1955 values. For simplicity, following Prather et al (2011) and Young et al (2013), stratospheric ozone is defined as concentrations in excess of 150 ppbv; this approach avoids having to specify a tropopause height at each horizontal grid point in the model. Also, we recall that the output from these same four CESM1(CAM5) model ensembles have previously been analyzed by Liang et al (2022), who evaluated their SAT and September sea ice responses and reported how different climate feedbacks in the Arctic result in a larger Arctic Amplification from ODS than from CO 2 forcing in this model. The present paper, instead, focuses on sea ice itself: we here examine both the extent and volume trends, spatial patterns of the forced response, and perform a mass and energy budget of the sea ice field to determine whether ODS and CO 2 impact sea ice via different mechanisms. A careful analysis of the Arctic sea ice response to ODS, and its comparison to the response to CO 2 , has not been performed to date.

Comparison methodology
In order to isolate the forced response, we focus our analysis on ensemble-mean trends across the multi-forcing ensemble. Ensemble-mean trends are obtained by first computing the ensemble mean of a given quantity (e.g. Pan-Arctic SIE) and then computing the linear trend of this ensemble mean. We estimate a 95% confidence interval for the ensemble-mean trends using a bootstrapping approach in which the ensemble is repeatedly resampled (with replacement) in order to produce an empirical distribution of ensemble-mean trends based on 1000 realizations.
The statistical significance of trend differences between two ensembles are assessed using the bootstrapped distributions of trends. The standard errors in ensemble-mean trends are added in quadrature to obtain a standard error for the trend difference. If the trend difference exceeds 1.96 standard errors, the difference is considered statistically significant at the 95% confidence level.
We assess the fidelity of simulated trends by comparing to monthly-averaged passive microwave satellite SIC observations from the National Snow and Ice Data Center (NSIDC) processed using the NASA Team retrieval algorithm (data set ID: NSIDC-0051, Cavalieri et al 1996). We focus our analyses on the time period 1979-2005, since this allows for direct comparison to observations and also is the time period that shows the largest differences between experiments. Computing trends over the period 1955-2005 does not change the qualitative findings of this study. For sea ice thickness and volume, we do not compare to an observational estimate due to the lack of a continuous thickness observational record over this time period.

Impact of ODS and CO 2 on Arctic sea ice trends
We first consider the relative impacts of ODS and CO 2 on Arctic sea ice trends. Figure 1 shows simulated ensemble-mean SIE timeseries across the multi-forcing ensemble. For both March and September SIE, the ensemble-mean trends are relatively indistinguishable over the period , and begin to show differences over the period 1975-2005. Consistent with earlier work (e.g. Swart et al 2015), we find that the All forcing ensemble simulates ensemble-mean SIE trends that are in reasonably good agreement with observations. The observed trends are generally lower than the ensemble-mean trends, but fall within the distribution of simulated trends in most non-winter months of the year (not shown). The modeled winter SIE trends are higher than observed, potentially related to underestimated trends in ocean heat transport into the Arctic Ocean (Li et al 2017).
Fixing CO 2 at 1955 levels has a striking impact on simulated SIE, as expected from its dominant contribution to GHG forcing (Meinshausen et al 2011). For both summer and winter SIE, we find that the fixed CO 2 simulations exhibit relatively flat SIE evolutions over the simulation period. This suggests that in CESM1(CAM5) the cooling effects of aerosol forcing and land use roughly balance the warming effects of non-CO 2 GHGs over this period. ODS and O 3 are two of these non-CO 2 GHGs: note, however, that the former are well mixed, but not the latter. For both September and March SIE, we find that the FixODS run lies roughly halfway between the All and FixCO 2 runs, indicating that ODS have made a substantial contribution to Arctic sea ice loss over the late 20th century. With the exception of a winter SIE anomaly in the 1970s, the FixODSO 3 runs show a similar SIE evolution to FixODS, indicating that the contribution of stratospheric O 3 to SIE trends is minimal relative to ODS.
A comparison of monthly ensemble-mean trends in SIE and SIV is shown in figure 2. We find that fixing ODS forcing causes a substantial reduction in simulated SIE trends in all months of the year, with the All forcing trends being reduced by 64% in the annual mean (compare blue and green bars; monthly SIE trends are reduced by 40%-94%). The SIE trend differences between the FixODS and All forcing ensembles are statistically significant in 8 months of the year (insignificant differences in Jan, Apr, May, June). We find that additionally fixing O 3 has only a small impact on SIE trends (compare red and blue bars), and that the SIE trend differences between the FixODSO 3 and FixODS ensembles are not statistically significantly different in any month of the year. As one would expect, fixing CO 2 has the largest impact on SIE trends, eliminating or even reversing the trends present in the All forcing run (compare magenta and green bars). The average trend reduction from fixing CO 2 is 127% of the All forcing trends (monthly trends are reduced by 88%-170%), and these trend differences are statistically significant in all months of the year.
ODS and CO 2 also have clear impacts on SIV trends ( figure 2(b)), however the quantitative trend reductions differ from the SIE trends. In particular, fixing ODS results in an average 32% reduction in SIV trends relative to the All forcing run, whereas fixing CO 2 results in an average 84% reduction. SIV trends also show a more clear separation between the FixODSO 3 and FixODS ensembles, however these differences are not statistically significant in any month of the year. The FixCO 2 and FixODSO 3 SIV trends have statistically significant differences with the All forcing ensemble in all months of the year, whereas the FixODS differences are only significant in 5 months of the year (insignificant differences in January-July).
The fixed ODS and CO 2 ensembles allow us to perform a direct 'head-to-head' comparison of their impacts on Arctic sea ice loss. The global radiative forcing from ODS over the period 1955-2005 is 0.31 W m −2 , which is 30% of the radiative forcing from CO 2 over the same period (1.02 W m −2 ; Meinshausen et al 2011). Thus, a reasonable null hypothesis would be for the impact of ODS on sea ice loss to be 30% that of CO 2 . However, we find that ODS have an outsized impact on Arctic sea ice loss relative to this expectation, with impacts on SIE and SIV trends that are 50% and 38% of the CO 2 impact, respectively. This finding is consistent with the results of Liang et al (2022), who reported that ODS also produce larger Arctic amplification than CO 2 .
To provide a more complete picture of the response of sea ice to ODS and CO 2 , in figures 3 and S1 we plot spatial trends across the experiments, which show coherent behavior across a range of sea ice variables. We show summer spatial trends in figure 3, as these are the months with the largest impacts from ODS, and show winter spatial trends in the supplementary material (figure S1). Relative to the FixCO 2 and FixODS simulations, the All forcing run shows trends towards less extensive (row (a)), thinner (row (b)), and younger (row (d)) sea ice in both the summer and winter seasons. The All forcing runs also shows trends towards less snow on sea ice (row (c)) and warmer SAT (row (e)) in both summer and winter. The spatial trend patterns are broadly consistent across the experiments, with the trends becoming progressively more muted as ODS, ODSO 3 , and CO 2 are respectively held fixed. In both summer and winter, the SIC trend differences occur near the ice-edge location (row (a)), whereas the sea ice thickness (SIT; row (b)), snow depth (SND; row (c)), and SAT (row (e)) trend differences occur throughout the Arctic Ocean domain, suggesting that a common broad-scale mechanism is driving the trend differences. The sea ice age differences (row (d)) also occur throughout the Arctic basin and are intensified along the pathway of the transpolar drift stream. The summer SIC differences are relatively zonally symmetric around the summer sea ice edge location, whereas the winter SIC differences are largest in the Barents and Greenland Seas and the Sea of Okhotsk.

Physical processes contributing to sea ice trend differences
Next, we consider the physical processes responsible for the trend differences identified in the fixed CO 2 and ODS experiments. We first analyze trends in sea ice mass budget tendency terms. The sea ice mass budget consists of a dynamic tendency term (ice mass transport convergence) and a thermodynamic tendency term (ice melt and growth). All terms use a sign convention with positive values corresponding to mass gain and negative values correspond to mass loss. In figure 4, we compute mass budgets over a Central Arctic basin domain consisting of the Central Arctic, Canadian Arctic Archipelago, Kara, Laptev, East Siberian, Chukchi, and Beaufort Seas, following the regional definitions of Bushuk et al (2017) (see inset in figure 4(e)). This analysis domain spans the region of thickest Arctic sea ice and its boundaries encompass all flux gates through which ice passes into the North Atlantic and North Pacific sectors. The mass budget terms are computed as areal integrals over the analysis domain, following the approach used in the mass budget intercomparison study of Keen et al (2021). We also plot summer and winter spatial trends of mass budget terms in figures 5 and S2, respectively.
We find that the sea ice trend differences across experiments are dominated by differences in thermodynamic, rather than dynamic, sea ice processes (figures 4(a) and (b); figures 5(a) and (b)). In particular, the All forcing run has the strongest trend towards increased summer melt and these melt trends are correspondingly reduced in the runs with Fixed ODS and CO 2 . The melt changes due to ODS represent 57% of the melt changes due to CO 2 . The melt trend differences are due to both top melt and basal melt, with these terms making roughly equal contributions to the summer melt differences across experiments The summer melt differences are broad-scale, occurring across the Central Arctic (figures 5(b)-(d)), and closely mirroring the SIT differences in figure 3(b). Note that the ring of positive melt trends at the sea ice periphery occurs due to the negative trends in SIT-there is simply less sea ice mass available to melt at those locations. The spatial changes associated with ice dynamics are noisier ( figure 5(a)) and essentially cancel when integrated over the Central Arctic basin domain. There is an increasingly positive trend from ice dynamics in the Central Arctic as one moves across figure 5(a) from the fixed CO 2 to the All forcing run. This is possibly due to the fact that ice strength decreases as ice thins, making it easier to import ice into the Central Arctic, which is typically a region of ice convergence.
The trends towards greater summer melt are partially, but not completely, offset by trends towards greater sea ice growth in autumn and winter months (November-March; figure 4(b)). This increased winter ice growth is consistent with a stronger negative ice thickness-growth feedback, associated with the thinning trends in these simulations (Bitz andRoe 2004, Petty et al 2018). As expected from this negative feedback, we find the largest winter ice growth increase in the All forcing simulations, followed by the FixODS and FixCO 2 simulations, respectively. The growth differences have a broad scale pattern that closely resembles the winter thickness differences between simulations (compare figure S2(b) to figure S1(b)). We also find that snow melt on sea ice increases in the month of June and decreases in the month of July in all experiments, indicating a shift of snow melt to earlier in the summer season ( figure 4(f)). This snow melt shift is largest in the All forcing experiments, and is progressively smaller in the FixODS and FixCO 2 simulations, respectively. Note that in these simulations, all snow on sea ice melts eventually over the summer. This explains the decreases in July snow melt-there is less snow mass available to melt in July due to the increased snow melt that occurred in June. The key point revealed by this sea ice mass budget analysis is that the impact of ODS on sea ice is similar to that of CO 2 , only smaller in magnitude owing to its smaller radiative forcing. Figure 5. Ensemble-mean summer (July-August-September) trends in SIT tendency trends due to dynamic processes (row (a)), thermodynamic processes (row (b)), top melt (row (c)), and basal melt (row (d)) in the CESM FixCO2, FixODSO3, FixODS, and All forcing simulations. All terms have units of cm day −1 decade −1 , with positive values corresponding to mass gain and negative values corresponding to mass loss.
We have attributed the stronger trends in the All forcing simulations to increased summer melt relative to the fixed ODS and CO 2 runs. Next, we consider the drivers of this increased summer melt. In figure 6(a), we plot annual-mean trends of surface energy budget terms averaged over an Arctic domain north of 60 • N. All terms are defined as positive downwards, with positive fluxes contributing to surface warming and sea ice melt. We find consistent mechanisms of Arctic warming present across the experimental suite, with surface energy budget trends that are ordered according to their Arctic SAT trends. These SAT trends are largest in the All forcing run, followed by the FixODS, FixODSO 3 , and FixCO 2 runs, respectively.
We find that the two key drivers of surface Arctic warming are increased downwelling longwave radiation and decreased upwelling shortwave radiation, consistent with earlier work (e.g. Boeke and Taylor 2018). Relative to the All forcing run, the fixed ODS and CO 2 runs have reduced downwelling longwave radiation due to their reduced GHG-based radiative forcing, with the ODS impact representing 35% of the CO 2 impact. The fixed ODS and CO 2 runs also have notable reductions in upwelling shortwave radiation. These reductions are consistent with the increased sea ice cover in these runs, which increases the surface albedo and produces more reflected shortwave radiation. In particular, the ODS impact on upwelling shortwave is 44% that of CO 2 , which closely reflects their relative impacts on surface albedo and SIE trends, which are 43% and 50%, respectively.
The upwelling longwave and downwelling shortwave heat fluxes each provide notable negative feedbacks which oppose this Arctic warming. The upwelling longwave trends act to cool the surface via the Planck feedback, and closely resemble the negative of the Arctic SAT trends. Consistent with this, the relative impact of ODS on upwelling longwave radiation and SAT are 35% and 35% of the CO 2 impacts, respectively. We also find systematic increases in downwelling shortwave radiation in the fixed forcing experiments. These Arctic SIE sensitivity to global-mean and Arctic-mean surface temperatures (defined as the ratio of annual-mean SIE trends to annual-mean temperature trends). Error bars indicate the 95% confidence range for the ensemble-mean trends and sensitivities, obtained by bootstrapping the ensemble.
increases are consistent with increased multiple reflections off clouds due to the increased sea ice cover and surface albedo in these runs (Wendler et al 1981). Indeed, we find that the fixed ODS runs provide 50% the impact of CO 2 on downwelling shortwave radiation, which generally reflect their relative impacts on surface albedo and SIE (43% and 50%, respectively). Compared to the radiative flux terms, the contributions from latent and sensible heat fluxes are relatively modest. But, again, the key point is that ODS impact the same surface energy budget terms as CO 2 , only with a smaller amplitude.
Finally, we consider whether ODS influence the sensitivity of Arctic sea ice to surface temperature warming, a quantity evaluated by many prior studies (e.g. Gregory et al 2002, Winton 2011, Mahlstein and Knutti 2012, Rosenblum and Eisenman 2017, SIMIP Community 2020, Bonan et al 2021. In figure 6(b), we plot the ratio of annual-mean Arctic SIE trends to annual-mean surface temperature trends across the experimental suite. The surface temperatures are averaged either globally or over an Arctic domain north of 60 • N, and ensemble means are computed prior to taking the ratio. Note that reliable sensitivities could not be computed for the FixCO 2 run because temperature trends are close to zero in this experiment, and thus the sensitivity metric is poorly conditioned. Therefore, we instead consider a 'CO 2 Only' sensitivity, based on trend differences between the All forcing and FixCO 2 runs. This difference has large trends in both SIE and temperature, allowing for a robust computation of sensitivity to CO 2 forcing alone.
We find that Arctic SIE sensitivity to global-mean surface temperature (GMST) is lower in the runs that fix ODS compared with the All forcing and CO 2 Only sensitivities. In both the FixODS and FixODSO 3 runs, the sensitivity differences are significant at the 95% level relative to both CO 2 Only and All forcing. This lower sensitivity to GMST in fixed ODS runs is consistent with the earlier results of Liang et al (2022) and Sigmond et al (2023), who found that ODS forcing produces greater Arctic Amplification than CO 2 forcing. The SIE sensitivity to Arctic temperatures is more consistent across experiments. The fixed ODS runs have slightly lower sensitivity than CO 2 Only and All forcing, but the differences are generally not significant at the 95% level, with the exception being the difference between the FixODS and All forcing runs. Our finding that sea ice sensitivity to GMST depends on the forcing agents considered emphasizes the fact that historical sea ice sensitivity needs to be evaluated using realistic historical forcings, consistent with the recent results of DeRepentigny et al (2022).

Discussion and conclusions
Using all-but-one-forcing ensembles, we have investigated the role that ODS have played in historical Arctic sea ice loss and have placed this ODS impact in the context of CO 2 forcing. We have found that ODS forcing played a significant role in year-round SIE and SIV trends over the second half of the 20th century, with SIE and SIV trends reduced by 64% and 32%, respectively, when the impacts of ODS are removed. These ODS trend contributions represent a substantial fraction (50% and 38%, respectively) of the sea ice loss from CO 2 forcing. We found that ODS, like CO 2 , affect sea ice by thermodynamic rather than dynamic sea ice processes, and that the losses in both cases can be attributed to summer sea ice melt. These melt differences are dominated by changes in downwelling longwave radiation associated with GHG radiative forcing and changes in upwelling shortwave radiation associated with sea ice and surface albedo changes.
We note that the impact of ODS on sea ice loss compared to CO 2 in this model appears to be larger than would be expected from global mean radiative forcing alone (Meinshausen et al 2011). Relatedly, we have also found that the sensitivity of Arctic sea ice to global-mean surface temperature change is larger when ODS forcing is included. These findings suggest that ODS might be more efficient than CO 2 at causing Arctic sea ice loss. However, since we do not have the actual radiative forcings over the Arctic for our model over the period of interest , we are unable to explore this question further at this time. We plan new dedicated runs to compute the separate surface and top-of-atmosphere radiative forcings over the Arctic to better quantify this potentially important effect.
Finally, we emphasize the remarkable fact that the Montreal Protocol, originally designed to halt the formation of the ozone hole over the Antarctic, has resulted in a substantial reduction of sea ice loss in the Arctic . Given the different time evolution of GHG forcing agents, it is important to move beyond sole consideration of CO 2 and also consider the influence of non-CO 2 GHG forcing agents in historical Arctic sea ice evolution. This may be especially important for ODS since (i) they exhibit a highly distinct and non-monotonic forcing history compared to CO 2 and (ii) they may have a high efficacy in causing Arctic sea ice loss, which would make their contribution larger than previously expected. Therefore, consideration of ODS forcing trajectories will likely be critical in understanding both past and future sea ice changes. It is also crucial to note that the results of this study are based upon a single model. The recent study of Sigmond et al (2023) found a similar response to ODS forcing in the CanESM5 model, but similar experiments with other climate models are required to build confidence in these findings, particularly using CMIP6 historical forcings which will allow for results to be extended to the end of 2014.

Data availability statement
The CESM-LE data are available for download from the Earth System Grid Federation data portal (www. cesm.ucar.edu/projects/community-projects/LENS/data-sets.html). The CESM all-but-one-forcing experiment data and analysis code are available via a Zenodo repository (https://zenodo.org/record/ 7469290). The NASA team sea ice concentration observations used in this study are available from the National Snow and Ice Data Center website (http://nsidc.org/data/NSIDC-0051/versions/1).