All the simulations analysed in this study use the low-resolution version of the latest UK physical climate model, HadGEM3-GC31-LL, hereafter HadGEM3 . HadGEM3 is a fully coupled climate model that uses the Unified Model (UM) for the representation of the atmosphere, the Joint UK Land Environment Simulator (JULES) for the representation of land surface processes , and the NEMO3.6 and the CICE5.1 models for the representation of the ocean and the sea ice, respectively.

In its low-resolution version (N96-ORCA1), HadGEM3 utilises a horizontal grid spacing of approximately 135 km on a regular latitude–longitude grid for the atmosphere. For the ocean, an orthogonal curvilinear grid with a grid spacing of approximately 1∘ is used. Note that the grid spacing for the ocean model decreases down to 0.33∘ between 15∘ N and 15∘ S of the Equator, as described by . For the vertical discretisation, the UM atmospheric model utilises 85 pressure levels (terrain-following hybrid height coordinates), while the NEMO ocean model uses 75 depth levels (rescaled-height coordinates).

The modifications and set-up of the applied sea-ice model CICE 5.1 (hereafter “CICE”) are described in . The standard elastic–viscous–plastic rheology (EVP) has been applied for ice dynamics with default CICE remapping advection algorithm and ridging schemes . Ice thermodynamics are based on with four ice and one snow layer. A semi-implicit coupling scheme between atmosphere and sea ice has been introduced to ensure the stability of the solver . The evolution of the sea ice is separately calculated for five ice thickness categories within each grid cell. For our study it is important to mention that the albedo calculation is based on the scheme used in the CCSM3 model but includes surface melt ponds by applying the explicit topographic melt pond model of and . Meltwater, formed as a result of snow melt, ice melt, and precipitation, runs downhill under the influence of gravity and collects on sea ice starting at the lowest surface height applying the sub-grid-scale ice thickness distribution. The evolution of pond fraction and depth as well as the formation of ice lids is calculated. The albedo of ponds of depth > 20 cm is 0.27 , significantly smaller than the albedo of ice and snow, which varies through the year in the range 0.5–0.9 and is calculated as described in . Above 70∘ N, during the months March–July, over the geographic regions we analyse, the albedo of open water is 0.07 ± 0.03 . In other sea-ice models without an explicit pond scheme, the ice albedo is reduced when the surface temperature approaches freezing temperature see e.g. to indirectly account for the impact of melt ponds. This adjustment has been removed here to not double count for the impact of ponds on albedo. For full details on model configuration, performance, and improved physics compared to older model versions, see .

The pre-industrial (PI) simulation used in this study was prepared and run by the UK Met Office as part of the sixth Coupled Model Intercomparison Project, CMIP6 . This simulation uses invariant solar, greenhouse gas (GHG), ozone, tropospheric aerosol, volcanic, and land-use forcing for the year 1850; see for details. The climate system took about 615 model years of spin-up to attain a steady state. These years are not used in our analysis. Of the subsequent 500 model years of production run , the first 200 are used here in our analysis.

The LIG simulation analysed in this study was first presented by ; it constitutes the UK's PMIP4 LIG contribution, as part of the wider CMIP6 project. The LIG experiment fully complies with the standard PMIP4 experimental protocol for Last Interglacial climate simulations, as described by . In more detail, this simulation is a time slice of the Earth's climate 127 000 years ago (i.e. 127 ka). The Last Interglacial climate was forced using 127 ka constant astronomical parameters based on and constant atmospheric trace GHG concentrations derived from ice core records (see , Table 1 for full details and values used). All other boundary conditions including ice sheets, topography, vegetation, aerosol, volcanic activity, and solar constant are identical to the PI simulation.

The LIG simulation was initialised from the end of the 615 years of PI spin-up. A further 350 model years of LIG spin-up were required for the atmosphere and the (upper-) ocean to reach quasi-equilibrium. See for details on how the LIG spin-up was evaluated and what metrics were used to assess the atmospheric and oceanic equilibria. After having attained quasi-equilibrium, the simulation was continued for a further 200 years of production run. This length of simulation has been shown to be long enough to capture model internal variability . The first 35 years of LIG spin-up and all 200 years of the LIG production run are used in our analysis.

For a given ice-covered cell (defined as a cell with sea-ice concentration > 0.15), the area of the cell at a sub-grid scale of (i) bare ice and snow, (ii) melt ponds on ice, and (iii) ocean that is exposed to the atmosphere and not covered by ice (hereafter the open-water area) may be computed as follows. The area of the cell covered by ice and snow and (ii) are returned as model output variables. (iii) is then computed as the grid-cell area not covered by ice and snow, and (i) is the ice area not covered by ponds.

The mean surface albedo over a grid cell is computed as 1 - (absorbed surface short-wave flux/downwelling surface short-wave flux), where absorbed and downwelling short-wave flux are model output variables. Unless otherwise indicated in the figure caption, figures showing the daily climatology of a variable are calculated using only ice-covered grid cells above 70∘ N, and the 200-year mean for each day is shown with the error as plus or minus twice the standard deviation (reflecting the inter-annual variability). Maps showing the Northern Hemisphere for each month from April–September are computed using the 200-year time average from monthly model output.

Here, we examine the HADGEM3-simulated LIG sea ice and upper ocean, in order to identify factors that contributed to LIG summer sea-ice loss. We first compare LIG and PI sea-ice area, as well as sea surface temperatures and salinities. We then consider the LIG production run and spin-up to determine the primary drivers of LIG ice loss.

Figure  compares the annual cycle of the Arctic sea-ice area between the PI and LIG period from the HadGEM3 simulations. The LIG winter ice area was slightly lower than the PI, with the smallest difference in area at 0.61 ± 0.56 million square kilometres in early April (Fig. ). Compared to the PI, an enhanced rate of LIG sea-ice melt from early May until late June led to a consistently ice-free LIG Arctic from early August until early October, as shown in Fig. . Maps of monthly sea-ice concentration are shown in Figs. S1 and S2. The minimum LIG ice area in September was 0.09 ± 0.11 million square kilometres, while the PI minimum ice area was 5.54 ± 0.98 million square kilometres. Long-term mean ice thickness during the LIG was also thinner than during the PI in all months, as seen in Fig. S3.

The earlier sea-ice retreat resulted in a warmer and less salty ocean surface during July, August, and September during the LIG (Figs.  and ). The differences reached more than 5 ∘C (LIG–PI SST anomaly) and around 1–2 ppt for temperature and salinity respectively. The interaction of ocean surface conditions and sea-ice extent in the model simulations potentially allow sea surface temperature (SST) observations from proxy records to be used as a signature of ice conditions during the LIG. In particular it may be informative for the marine core community to search for large summer SST differences from the PI, of >4 ∘ C, at latitudes near the expected LIG ice edge (e.g. 70∘ N; east of Greenland) and smaller summer SST differences from the PI, of > 2 ∘C, in the central Arctic/Beaufort sea. This could help identify an sea-ice-free summer LIG Arctic in marine observations. However, we note that a partially-ice-covered summer Arctic, with thin LIG sea ice, might yield a similar signature to this but with lower magnitude temperature anomalies everywhere, as very thin summer ice cover has a small insulating effect . Thus it is additionally useful to consider sea surface salinity (SSS) changes (Fig. ). These have a clearer signature than SST. While the salinity patterns were similar in May for both simulations, the stronger melt during the LIG caused the region with LIG winter ice cover to become significantly fresher than the PI, by around 1–2 ppt in June and July. Additionally, a difference of 0.5–1.5 ppt was retained at least from March–September. Beyond helping identify the observational signature of a sea-ice-free Arctic, we aim to understand the following questions: what caused the large differences in summer sea ice in this model? Particularly, why did the spring melt increase so significantly in spite of the lower atmospheric CO2 concentration in the LIG? Processes with a significant impact on summer sea ice include thermodynamic processes such as ice–albedo feedbacks, ocean heat transport, and cloud cover feedbacks. Ice–albedo feedbacks may be significantly impacted by the presence of melt ponds, as discussed in Sect. . The preconditioning of winter sea ice may also lead to a reduced sea-ice extent the following summer . In addition, summer sea-ice extent may be affected by changes to sea-ice distribution caused by changes to wind patterns and ocean circulation further discussed in Sect. . It is currently unknown which process was most significant for the enhanced LIG sea-ice loss. We first look at the spin-up simulation for the LIG, with a particular focus on the first year of the spin-up (Figs.  and ). Fig.  shows that the winter Arctic sea ice retained a similar area to the PI control period over the first 15 years of the spin-up period, beyond which it decreased only slightly (< 1 million square kilometres) from the PI control. However, an ice-free summer state was reached after only 4 years ; we find August sea-ice area halved during the first year of the spin-up run from nearly 6 million to around 3 million square kilometres. Differences between PI and LIG developed rapidly during this first summer, within a few months of the January switch from PI to LIG forcings: this suggests that winter preconditioning did not play the dominant role in the enhanced melting of sea ice during the LIG spin-up. Similarly, this rapid loss of sea ice within the first few spin-up years cannot be explained by a change to ocean heat transport, which is often linked to reduced sea ice : upper ocean heat transport takes decades to equilibrate in numerical models after a significant perturbation, and deeper ocean heat transport takes centuries to millennia . Therefore, as can be seen from Fig. , changes to ocean heat transport could not have been the first-order driver of the rapid LIG sea-ice loss observed. Other factors that were present by the summer of LIG spin-up year 1 must be key to the enhanced melt throughout the simulated LIG period. In order to deduce the importance of two of the remaining processes, ice–albedo and cloud cover feedbacks, we consider the surface energy budget (Fig. ). This is because the Arctic Ocean's surface heat balance is closely linked to its ice mass balance : at the surface, short-wave (SW) flux in particular is known to play a dominant role in summer sea-ice melt and is linked to ice–albedo feedbacks ; longwave flux is related to the longwave cloud radiative forcing and cloud cover feedbacks . Summer positive TOA radiation led to a positive net absorbed SW flux anomaly, of up to 75 Wm-2 in June, that contributed nearly all of the net absorbed surface heat flux anomaly (Fig. a). Unlike changes related to preconditioning and ocean heat transport, this net SW flux anomaly was already present in the first year of LIG spin-up run, reaching 55 Wm-2 in June (Fig. b). This immediate response, coupled with the immediate halving of August sea-ice area, suggests the absorbed short-wave radiation played a key role in summer ice loss and thus that ice–albedo feedbacks may also have been important. We note also that the longwave anomaly accounted for < 5 Wm-2 of the total surface heat flux anomaly in Fig. b, so longwave forcings and feedbacks related to cloud cover were likely not dominant contributors to the enhanced LIG sea-ice loss, as confirmed by . Other differences between the LIG and PI surface heat budget contributed < 20 Wm-2 monthly to the surface energy budget (Fig. ). This suggests that thermodynamic processes that led to the enhanced LIG summer sea-ice melt must predominantly have resulted from this surface SW anomaly.

Therefore, by elimination, the dominant contributors to the enhanced LIG summer sea-ice loss were (i) thermodynamic processes, driven by this surface SW anomaly and likely related to ice–albedo feedbacks, and/or (ii) changes to the summer ice distribution. In the next section, we investigate the relative importance of these processes.

Today's diminishing Arctic sea ice has led to a new focus on seasonal forecasting of Arctic sea-ice conditions and especially predicting the minimum sea-ice area each year . Spring melt ponds are a good predictor for summer sea-ice conditions . It is of interest to see how the predictability of spring melt pond area and August–October sea-ice area varied between the PI and LIG, since this may yield insight into how predictability may change in future under conditions of reduced Arctic sea ice. Predictability is investigated here by considering interannual variability within each of the LIG and PI periods. The interannual variability in the radiative forcing at the surface was much smaller than the difference in the radiative forcing between the PI and LIG. Thus, investigating the relation between spring melt pond area and August–October sea-ice area within each of these two periods gives insight into the impact of melt ponds on the summer sea-ice area for similar TOA radiative forcing and winter ice conditions each year.

For the 200 years of simulation output for both LIG and PI, Pearson’s correlation coefficient was calculated between the mean August to October sea-ice area and each of the following four variables in each of April, May and June:

the mean monthly melt pond fraction;

Results are shown in Fig. 12. We note that, for both time periods, the April through June thin-ice fraction was a statistically significant (p<10-5) predictor of summer ice area, as might be expected . However, of more interest here is the significance of the pond-related correlations through the spring. Whilst not significant in April for either period, by May, there was a significant negative correlation between melt pond formation and summer sea-ice area of -0.34 for the PI and -0.42 for the LIG. This is only a slightly weaker correlation than that of the absorbed short-wave fraction of -0.37 for the PI and -0.46 for the LIG (Fig. ). By June both of these correlations were much stronger than the thin-ice fraction correlation: the melt pond correlation was -0.55 for the PI and -0.73 for the LIG, and the short-wave correlation was -0.55 for the PI and -0.76 for the LIG. These strong correlations show that spring melt pond formation alone was nearly as reliable a predictor of September ice cover as the presence of any water (including both ponds and open water) over ice-covered grid cells.

Compared to the PI, the LIG spring sea-ice area above 70∘ N was on the order of a million square kilometres lower, and the mean sea-ice thickness is about 1 m less. Therefore, the correlations shown in Fig.  imply that May–June pond-related quantities can be used to make more reliable seasonal predictions of summer sea ice for thinner (LIG) than thicker (PI) sea ice, as thinner ice is more sensitive to pond formation and albedo changes. This is of importance for future seasonal predictions: as the climate warms and Arctic sea ice continues to thin, the spring melt pond area is likely to become an increasingly reliable predictor of September sea-ice area.

Despite their significant impact on sea-ice melt, melt ponds have only recently been explicitly included in global climate models e.g.. Prior to this, models without an explicit pond scheme had the ice albedo reduced when conditions approached near-melting . Until now, it has not been clear whether this melt pond parameterisation approach adequately accounts for the impact of ponds during warmer-than-present-day conditions .

found that HadGEM3 simulated an ice-free summer LIG Arctic and suggested this was linked to HadGEM3's realistic representation of melt pond physics. Here, we have quantified the importance of melt ponds during the LIG in depth, showing that greater melt pond formation, earlier in the year, directly led to the large surface energy budget difference between LIG and PI that was first highlighted in . We have demonstrated the impact of melt ponds on thermodynamic processes during the LIG and their role in determining LIG spring albedo over ice-covered regions, leading to albedo feedbacks and ultimately the extreme ice loss first presented in . Additionally, our analysis of the HadGEM3 PMIP4 LIG spin-up simulation years, showing that changes to longwave radiation and ocean heat transport were not primary drivers of the observed enhanced LIG ice loss, was supported by , who used the LIG production run to show negligible differences in these factors between the simulated LIG and PI.

Of the CMIP6 models that have simulated the LIG according to CMIP6-PMIP4 experimental protocol, the two to include an explicit melt pond scheme are the two with the lowest minimum monthly SIA: HadGEM3, 0.2 million square kilometres, and CESM2, 1.2 million square kilometres . The only other model that simulated a minimum LIG SIA less than 2.0 million square kilometres was NESM3, at 1.3 million square kilometres. NESM3 has an unrealistic sea-ice representation for the PI period, with amplitude of the seasonal cycle and winter sea ice significantly overestimated ; thus, we disregard this model in our analysis. For the remaining models which use parameterised rather than explicitly modelled ponds, the mean LIG minimum SIA was 3.5 ± 1.2 million square kilometres. Since only 16 models have run this simulation, it is not possible to say that implementing an explicit melt pond scheme will tend to result in lower simulated LIG summer ice area; other aspects of the model simulation set-up may also be significant in the simulation of Arctic sea ice. For example, the IPSL-CM6L model simulates some compensating ocean circulation and cloud cover changes that contribute to preserving LIG summer sea ice . However, despite these caveats, it is striking that the only two models to include explicitly modelled melt ponds also simulated minimum LIG SIA at least 2.3 million square kilometres less than (i.e. 1.9 standard deviations from) the mean of the models which use parameterised ponds, whilst simulating a realistic PI SIA annual cycle, as well as maximum LIG SIA similar to (i.e. 0.5 standard deviations from) the mean of the models which use parameterised ponds.