The study area is inland Dronning Maud Land around the Dome Fuji station (77.316∘ S, 39.701∘ E, 3810 m a.s.l.) with a radius of about 300 km (Fig. 1). Dome Fuji is the second-highest summit of the AIS and is located at the junction of four ice divides (Fig. 1a). Spatial and temporal variability of SMB around the ice divide near Dome Fuji is characterized as follows. The large-scale spatial pattern of the SMB depends on the surface elevation, distance from the moisture source (ocean) along the atmospheric pathway and large-scale surface topography (e.g., positions of ice divides) in combination with a dominant trajectory of air mass with moisture. Around Dome Fuji, SMB is larger on the windward (mostly northeastern) side of ice divides than on the leeward side (Suzuki et al., 2008; Fujita et al., 2011; Tsutaki et al., 2022). In addition, local variations in surface topography, which may be influenced by the bedrock topography and ice flow (Fujita et al., 2011; Tsutaki et al., 2022), also affect the precipitation patterns and snow redistribution and thus the spatial patterns of SMB (Furukawa et al., 1996; Van Liefferinge et al., 2021).

In this study, we use seven snow pit samples and 15 ice cores collected between 1993 and 2019 to reconstruct accumulation rates. The names and locations of each site are shown in Fig. 1 and Table 1. The locations of the sites are mostly distributed within 50 km from Dome Fuji (Fig. 1b), with additional sites further away (200–400 km), both on the windward and leeward sides of the main ice divides with respect to the origin and transport paths of moisture. At the Dome Fuji station, two deep ice cores have also been drilled in the 1990s and 2000s to the depths of 2503 m (340 ka BP) (DF1) and 3035 m (720 ka BP) (DF2), respectively (Watanabe et al., 1997a, b; Dome Fuji Ice Core Project Members, 2017; Motoyama et al., 2021). Two shallow cores (DF1993 and DF2001) are the shallow parts of the DF1 and DF2 deep ice cores. Because the NDF2018 core is the longest and oldest shallow ice core in this study (151 m long, covering ∼4800 years), we use the records of DF1 and DF2 deep cores to cover the same period. Four ice cores (DF1, DF2, NDF2018 and NDFN cores) cover more than 4000 years, three ice cores (DF1997, DFS2010 and DFS2011 cores) cover 1500 years and six ice cores (MD364, NDF2013, S79, S80, DFNW and DFSE cores) cover 800 years or less. The DF1997 core has poor ice core recovery and uncertain depth assignments for 0–20 m; thus, we use the data only below 20 m. The DF1999 core was measured for permittivity (for density estimation) but not for DEP for volcanic matching, and thus it was only used for estimating the average density profile at Dome Fuji.

We obtained density profiles of shallow cores by combining published density profiles, bulk densities measured at the field, and permittivity measured in a laboratory (Fig. 2, Table 2). We briefly describe each method.

Annual layer counting is not possible for our samples because some annual layers can be eroded due to the low accumulation rate (Kameda et al., 2008). Therefore, we used volcanic signals recorded as near-surface conductance of the direct current (ECM) (Hammer, 1980; Wolff, 2000), high-frequency-limit electrical conductivity (DEP) and non-sea-salt sulfate (nssSO42-), or tritium peaks from the past bomb tests to identify the age horizons. We used published data of the ECM of the DF1993 and DF1 cores (Fujita et al., 2015), nssSO42- concentrations of the DF2001 core (Motizuki et al., 2014), nssSO42- concentrations and tritium contents of the MD732 snow pits (Hoshina et al., 2014), respectively. For other ice cores and snow pit samples, we newly generated the data with the following methods.

The depth of a volcanic layer is generally defined at the maximum value of high-resolution data of DEP, ECM or nssSO42- (Hofstede et al., 2004), with depth uncertainty defined by the data resolution. The volcanic signals of low-latitude eruptions are recorded in Antarctic inland snow with a typical lag of 1 to 2 years due to the long-distance transport (Hammer et al., 1980; Cole-Dai and Mosley-Thompson, 1999; Cole-Dai et al., 2000, 2009; Gao et al., 2006). This depositional lag is not constant and may generate additional uncertainty for the dating. To avoid the uncertainty of depositional lag, we take advantage of published sulfur data of the WAIS Divide ice core (WDC) with accurate layer-counting chronology (Sigl et al., 2014, 2016) which we can use to precisely synchronize the volcanic signals in our shallow cores and pit samples. We assumed no depositional lag between WAIS Divide and Dome Fuji. Based on the synchronization of DF1993 and DF1 cores to WDC by Oyabu et al. (2022a), we identified the eruptions in the other shallow cores by inter-comparisons of their volcanic signals. We only employed the peaks found in two or more of our ice cores. In total, we identified 59 volcanic layers for the dating between 1992 CE and 3151 BCE (Figs. 3 and B1). We also confirmed the consistency of our dating with previous studies that used the same volcanic layers (back to 1260 CE for the DF2001 core by Igarashi et al., 2011; back to ∼100 CE, for the DFS2010 core by Motizuki et al., 2014). We found several common peaks in our ice cores that are not recorded in the WDC, which were not used for the dating because their depositional age could not be accurately known. We note that a common 1 cm thick tephra layer was found in the DF2, NDF and NDFN cores at 133.51, 118.66 and 122.18 m, corresponding to 1613±5, 1613±5 and 1611±4 BCE, respectively, according to their chronologies (Sect. 2.5). The tephra layer is not found in WDC, but it is probably the same as those found in the Dome C, Vostok and South Pole data at 132.6, 103.14 and 303.44 m, respectively (Palais et al., 1987; Narcisi et al., 2005), dated around 1550–1650 BCE (on Vostok GT4 and EDC2 chronologies, Narcisi et al., 2005). While the synchronization using tephra layers is generally difficult because of their sparseness and large signals from local eruptions, this tephra layer is easily visible in the field, and the age of the layer determined here (1613 BCE or 3563 yr BP) may be useful as an age constraint for other shallow ice cores from EAP.

The uncertainty in duration between two neighboring volcanic tie points, which is required for the error estimate of SMB, consists of the following components: (1) depth of ECM or DEP measurement (±0.01 m as 1σ for each volcanic layer, corresponding to 0.15–0.35 years), (2) identification of peak positions (±0.01 m as 1σ for each volcanic layer, corresponding to 0.15–0.35 years), and (3) duration between the two horizons originating from the age uncertainty of WD2014 (<0.75 years as 1σ) (Sigl et al., 2016). The overall uncertainties for the shallow cores are smaller than 1 year (1σ).

For the relatively long interval between 700 and 1000 CE without volcanic tie points, we employed three published age tie points for the DF2001 core based on the matching between the ice core 10Be and tree ring 14C (Horiuchi et al., 2008; Miyake et al., 2015, 2019; Oyabu et al., 2022a). The error in the durations between the tie points are also shorter than 1 year (1σ).

For the snow pit data, we used the sulfate peaks for the Pinatubo eruption (deposition in 1992) for all snow pits except for Dome Fuji; sulfate peaks for the Agung eruption (deposition in 1964); and Tritium peaks (deposition in 1966, Jouzel et al., 1979; Fourré et al., 2006) for MD732, Dome Fuji, and NDF2018 snow pits (Fig. 4). We do not use the Pinatubo eruption for the age tie point in the Dome Fuji pit data because there are three possible nssSO42- peaks for the eruption within ∼0.5 m interval (around 1.0 m), making the precise depth–age assignment impossible. Unidentified regional eruptions may have left impurity peaks near the Pinatubo depth, as have been found elsewhere in East Antarctica (Cole-Dai et al., 1997; Plummer et al., 2012; Crockart et al., 2021). The error of the duration between the tie points is estimated to be 0.5 year and that for the peak depth is estimated to be 0.1 m (corresponding to ∼1 year); thus, the overall uncertainty is ∼1.1 year (1σ).

Mean accumulation rates between the dated horizons are calculated by dividing the depth difference between the age markers, corrected for firn density and layer thinning (due to ice flow), by the time span. For the thinning function, we used a published model result for the Dome Fuji core (Parrenin et al., 2007) and applied it to other cores using normalized ice-equivalent depth between the surface and bottom of the ice sheet (Fig. 5). We assumed that the uncertainty of the thinning function linearly increases with depth at a rate of 0.01 % m-1 (1σ), giving reasonable uncertainty including the assumption of linear scaling of the Dome Fuji thinning function to other sites (note that the difference between the thinning function between DF and NDF is about 1.5 % at 150 m). The effect of the thinning correction on the resulting accumulation rate is smaller than the overall uncertainty (see below) for the depths shallower than 40 m, roughly corresponding to the last millennium, while it is highly important for the deepest part (4–5 ka BP) of the DF, NDF and NDFN cores (about 3 times the overall uncertainty) to discuss the long-term variations of SMB. The overall uncertainty of SMB for a given core and interval was estimated by using a Monte Carlo approach, in which the density profile, age of horizons and thinning functions were randomly varied 1000 times according to the respective error estimates to calculate the probability distribution of accumulation rate.

For DF1, DF2, DFS2011, DFS2010, NDFN and NDF2018 cores covering more than 2 kyr, we calculated the average accumulation rates over the periods of 1–1850 CE and 2000 BCE–1850 CE. Because there are no volcanic tie points exactly at 1 CE and 2000 BCE, we interpolated the volcanic depth–age relationships for the DF1, DF2, NDF, NDFN, DFS2010 and DFS2011 cores using a probabilistic dating model Paleochrono (Parrenin et al., 2021) to derive the depths for the exact years in the cores. The model estimates the age scale of a core by optimizing the prior estimates of accumulation rate (A) and thinning (τ) as functions of depth to reproduce the chronological constraints. We used the accumulation rate and thinning function and their uncertainties from the method described above as the prior A and τ. For the age constraints, we used the age differences between volcanic layers (ice age intervals) constructed from volcanic synchronization. The uncertainty of the ice age interval is estimated to be 0.1–1.2 years (1 standard deviation) from the following components: (1) uncertainty in the layer-counted WD2014 age scale (0.1–0.9 years for the intervals, larger for the deeper depth) and (2) depth uncertainty in matching our cores to WDC (<1 year, associated with data resolution).

Figure 6 shows the time series of accumulation rates of all 13 ice cores from nine sites. The longest record covers 5152±10 years (2001 CE–3151 BCE, DF2 core). The mean accumulation rate over the last 5 kyr is 26.2±1.0 kg m-2 yr-1 at the Dome Fuji station (mean of DF1 and DF2). A previously published estimate of the accumulation rate at the Dome Fuji station is 25 and 25.5 kg m-2 yr-1 over the period 1260–1993 CE and 1260–2001 CE, based on the DF1993 (shallow part of the DF1 core, Watanabe et al., 1997a) and DF2001 (shallow part of the DF2 core, Igarashi et al., 2011), respectively, with fewer age control points. They agree with our result for the same core and period (24.7±0.1 and 25.5±0.1 kg m-2 yr-1 for DF1993 and DF2001, respectively).

For the sites within 100 km of the Dome Fuji station, the accumulation rates generally range from ∼20 to 30 kg m-2 yr-1. We note that the variability (periodicity and amplitude) of each record is determined not only by the natural variability but also by the number of age tie points. For example, the DF2 data show higher centennial-scale variability than the DF1 data because of more tie points. We also note that at around 1000 BCE the DF1 tie points are scarce and the DF1 accumulation rate is larger than the DF2 value on average over ∼600 years. The reason is not clear, but it might suggest possible errors in depth assignments of the ice cores (we note that it is just below the firn–ice transition where the shallow ice core data are connected to the deep ice-core data for the DF cores). The discrepancy for this part does not affect the following discussion and conclusion about the long-term trends and differences between the sites. In general, the accumulation rate is larger to the north and smaller to the south of the Dome Fuji station. The mean accumulation rate and its variability at MD364 (range: ∼20–60 kg m-2 yr-1) are much larger than those in the Dome Fuji area. The site is located at the altitude of ∼3350 m, lower by ∼450 m than at Dome Fuji station, on the north side of the topographic ridge, and thus the mean precipitation is expected to be larger than at Dome Fuji. The larger variability in the depositional environment may be expected from the horizontal ice flow (4.1 m yr-1) over rough bedrock topography, creating temporal variations in local surface topography (e.g., surface slope and convex or concave curvatures) (Fujita et al., 2002; Kahle et al., 2021) or from the variable wind direction for deposition in combination with surface topography (Urbini et al., 2008).

For all the individual SMB time series of six cores (DF1, DF2, DFS2010, DFS2011, NDF2018 and NDFN) covering more than 2000 years, we find clear decreasing trends at the rates from -0.05 to -0.09 kg m-2 per century (Fig. 6, Table 3). In addition, there appears to be a significant increasing trend for the last 200 years for most cores. The increasing trend in the last 200 years is not observed at DF1 (drilled in 1993), possibly due to the lack the record after 1993, when the accumulation rate might have significantly increased, or the depth assignment might be imprecise for the shallowest part of the core. The increasing trend is also not observed for MD364 with large temporal variabilities (see above).

To examine the spatial differences in accumulation rate in the inland plateau region around Dome Fuji, mean accumulation rates over fixed time intervals were calculated for each core. For the averaging periods, we chose (1) 1461–1816 CE (all cores), (2) 1–1850 CE (DF1, DF2, DFS2010, DFS2011, NDFN and NDF2018 cores), and (3) 2000 BCE–1 CE (DF1, DF2, NDFN and NDF2018 cores) (Fig. 7) for the following climatological reasons.

We chose 1461–1816 CE because it is the most accurately constrained period by the volcanic years common for all the studied ice cores and also a good representation of the preindustrial late Holocene for averaging SMB without decadal noise.

For 1–1850 CE (Fig. 7b), the order of SMB of the DF, DFS, NDFN and NDF sites (from the largest to smallest) is consistent with those for 1461–1816 CE (Fig. 7c). The ratios of SMB at NDFN and NDF sites to that at DF are also similar for all three periods (Fig. 7a–c), suggesting that the relationships of SMB between the sites on multi-centennial or longer timescales have been stable. The mean SMB at the DF, NDFN and NDF sites for the period 2000 BCE–1 CE (Fig. 7a) are slightly larger than those for the period 1–1850 CE and significantly larger than those for the period 1461–1816 CE.

The recent accumulation rates estimated from snow pit observations are shown in Table 4. The average accumulation rate of all pit data since 1992 CE is 24.8±1.0 kg m-2 yr-1, which is in general agreement with previously reported values over similar periods from a pit study (25.6 kg m-2 yr-1 for 1992–2007 CE, Hoshina et al., 2014) and a snow stake study (27.3±1.5 kg m-2 yr-1 for 1995–2006 CE, Kameda et al., 2008). However, there are no consistent spatial gradients relative to the topographic ridge (i.e., less accumulation on the inland side of the ridge) for the average values since 1992 CE as found for the long-term reconstructions from the ice cores. The large variability in density near the surface and redistribution of surface snow (intermittent removal, Kameda et al., 2008) may introduce relatively large uncertainty in the estimates. On the other hand, the mean accumulation rates at MD732 (close to the DF station) since 1966 CE and 1964 CE are significantly larger than those at NDF, which is consistent with the spatial gradient found from the ice cores. The accumulation rates at MD732 and NDF since 1964 CE are significantly larger than the average values over the last 4 kyr and comparable to the largest values about 5 kyr ago. Although precise comparisons may be difficult because of the different averaging periods (pit and ice core records average about ∼50 years and a few centuries, respectively), the accumulation rate around Dome Fuji in the last few decades appears to be high in the long-term perspective.

To robustly estimate the trends and variability of SMB in the Dome Fuji area, we stack the reconstructed SMB from the individual cores (Fig. 8). We find that the modern SMB values at the study sites are within ±10 % of that at Dome Fuji, and it is not known if the variability of SMB is a function of its mean value. Thus, we stacked the records in two methods, with or without normalizing the variability. We excluded the DF1997 core from the stacking because it lacks reliable SMB reconstruction for the last ∼550 years.

In the first stacking method (Stack 1), each of the 12 SMB time series from nine sites is normalized to its average and 1 standard deviation over the period 1461–1816 CE, and 12 records are simply averaged. The stacked record is smoothed with a 71-year moving average to reduce short-term noise. In the second stacking method (Stack 2), each of the nine SMB time series from within 100 km of the Dome Fuji are normalized to their average between 1461 CE and 1816 CE and smoothed with a 31-year moving average. The smoothed records are averaged, added by the average SMB of the nine sites (23.90 kg m-2 yr-1 for 1461–1816 CE) and smoothed with a 71-year moving average. We exclude S79, S80 and MD364 records, which are located more than 100 km away from the Dome Fuji station. The advantage of this method is to permit a quantitative reconstruction, and it is necessary to exclude MD364 because its average and variability are significantly larger than those near Dome Fuji. The S79 and S80 have similar accumulation rates to Dome Fuji, but their locations are clustered in one direction relative to the dome (Fig. 1). The uncertainties of Stack 1 and Stack 2 were estimated by a Monte Carlo approach, in which the SMB value of each core in each age segment (between the age control points) was randomly varied 1000 times according to the error estimate and stacked. The uncertainties in the long-term trends of Stack 1 and Stack 2 were also estimated as the standard deviation of the slope of regression lines to the same 1000 stacked records by the Monte Carlo method.

The mean SMB of Stack 2 is 25.2±0.8 kg m-2 yr-1 over the entire period. The centennial-scale variabilities in both stacks are very similar to each other (Fig. 8a and b), although the amounts of stacked data are different between the two methods for the last 1500 years. There are general long-term declining trends in both stacked records over the entire period (∼5000 years), with a minimum accumulation rate around 1700–1800 CE and the greatest increase during the last 150 years.

Both stacked records show statistically significant decreasing trends over the most recent ∼5000 years (1850 CE–3151 BCE), with a slope of -0.037±0.005 kg m-2 per century for Stack 2. We analyze the robustness of the long-term trend following the method of McGregor et al. (2015) by calculating the average accumulation rate over 200-year intervals (bins) for each core, taking the median value in each bin, making the composite record by connecting the median values and drawing the regression line (Fig. 9). We also shift the position of the bin in 50-year increments to assess its influence on the regression slope (McGregor et al., 2015). All binned records show similar decreasing trends over the 5000 years, indicating that the decreasing accumulation rate is a robust feature around Dome Fuji. The slopes from the binned reconstructions are smaller than that of Stack 1, probably because the binning method reconstructs a lower accumulation rate from ∼1000 BCE to ∼1000 CE than Stack 1 by taking the median values rather than the average values. The long-term trends for the periods of 1 to 1850 CE (Table 3) for Stack 1 and Stack 2, as well as all the binned records, also show negative trends (Figs. 8 and 9).

We also made three local stacks using the subsets of data from sites close to each other as follows: “DF stack” with DF1, DF2, and DF1997 cores; “DFS stack” with DFS2010 and DFS2011 cores; and “NDF-NDFN stack” with NDFN, NDF2013, and NDF2018 cores (Fig. 10). For the last 2700 years, the local SMB stacks show systematic differences (DF > DFS > NDF-NDFN) with similar centennial-scale variations. We also find significant decreasing trends in all local stacks, with the slope of the DFS stack being slightly steeper than the other two stacks. The difference in the trends might be related to possible long-term migration of the dome summit position in response to glacial–interglacial changes in SMB and grounding line position (Saito, 2002; Parrenin et al., 2016). The dome migration could also possibly be a consequence of spatial and temporal variations in SMB, which may be influenced by meridional midlatitude atmospheric variability and precipitation events (Massom et al., 2004; Scarchilli et al., 2011; Turner et al., 2019), as has been discussed for Dome C and Talos Dome (Urbini et al., 2008).

We first discuss the robustness of the long-term trends. The long-term decreasing trend in the DF area is largely determined by the six ice cores going back by more than 2 kyr (DF1, DF2, DFS2011, DFS2010, NDFN and NDF2018). As the depths of these cores are more than 100 m, the corrections for layer thinning due to ice flow impact the reconstructed SMB records. Because the thinning is weakly constrained for the shallow cores (except for the DF core with good age control from the deeper depths), we investigate the possibility of reversing the sign of trend due to errors in thinning. In particular, we assess whether the negative accumulation rate trend is possibly an artifact of thinning correction (Sect. 2.5) because the correction increases the estimated mass (hence SMB) at deeper depths. Here, we calculate the SMB under the (unrealistic) assumption of no layer thinning for the six cores. As shown in Fig. 11, all cores exhibit a significant negative accumulation rate trend even without the thinning correction. Therefore, we conclude that the multi-millennial decreasing trend in accumulation is a robust feature in the Dome Fuji area.

We discuss possible causes for the decreasing trend in accumulation rate around Dome Fuji over the last ∼5000 years in the preindustrial period. Note that the main objective of this study is a reliable reconstruction of SMB over long timescales. The thorough investigation of the mechanisms would require reliable SMB reconstructions in other parts of Antarctica and climate modeling works that are beyond the scope of this study. The first obvious candidate for the long-term change in accumulation rate is the secular changes in orbital forcings. Obliquity (the tilt of Earth's rotation axis) becomes smaller over the last 5000 years (Fig. 12c), which gradually decreases the annual mean insolation at high latitudes. The declining annual mean insolation may cool the ocean surface and atmosphere over the Southern Ocean and Antarctica, leading to reduced evaporation and atmospheric moisture content, and possibly the general decrease in the snow accumulation over the Antarctic inland. On the other hand, the radiative forcing of greenhouse gases (GHGs) increases over the same period (Fig. 12d), which might counteract other forcings for reducing snow accumulation. Other potentially important forcings are the changes in solar irradiance (Fig. 12e) and volcanic forcing (Fig. 12f), although they have significant uncertainties in reconstructions.

For the AIS SMB, sensitivity experiments with a regional climate model have suggested the roles of Southern Ocean sea surface temperature (SST) and sea ice concentration for the SMB (Kittel et al., 2018). Their results showed that increases in sea ice concentration led to a decrease in the SMB over most of the AIS, but lower SST led to contrasting changes in the coastal and inland areas of AIS, i.e., a decrease in coastal areas and an increase in inland areas, because of the relationship between atmospheric vapor content and saturation pressure along the air mass trajectory towards the inland. On the other hand, Vannitsem et al. (2019) analyzed the relationship between reconstructed AIS SMB (Thomas et al., 2017) and the results of global climate models for 850–2005 CE and found that the SMB over the Antarctic Plateau is mostly influenced by the surface air temperature and sea ice concentration and not by large-scale atmospheric modes such as El Niño or the Southern Annular Mode (SAM). Therefore, to provide data-based clues for the long-term relationships between the surface temperature, sea ice and SMB, we compare our SMB trend with the published temperature and sea ice reconstructions over the latter half of the Holocene.

Long-term coolings from the middle to late Holocene at high latitudes in the Southern Hemisphere have been suggested by previous studies, although the relevant climatic forcings and mechanisms (e.g., in terms of orbital configurations and GHGs) are not clear. Recent compilations of surface temperature reconstructions exhibit cooling trends for the middle to high latitudes of the Southern Hemisphere (Marcott et al., 2013; Kaufman et al., 2020a, b), and stable water isotope records on the EAP (EDML, Dome Fuji, Vostok, Dome C, TALDICE) commonly show decreasing trend from the middle to late Holocene (Masson-Delmotte et al., 2011). On the other hand, individual reconstructions of sea surface temperature (SST) show a variety of trends from clear cooling to little trend over the last 5000 years (e.g., Hodell et al., 2001; Nielsen et al., 2004; Anderson et al., 2009; Divine et al., 2010; Lamy et al., 2010; Shevenell et al., 2011; Etourneau et al., 2013; Xiao et al., 2016), which may be partly because local SST may be sensitive to the positions relative to ocean currents. Cooling on land in the latter half of the Holocene is also suggested by the advances of glaciers in the Southern Hemisphere (Solomina et al., 2015).

Reconstructed sea ice extents from multiple proxy records for different parts of the Southern Ocean show general advancements for the last 5000 years. The sea ice advance around 5000 to 4000 years ago was found from TN057-13 in the southeastern Atlantic (Hoddel et al., 2001), from JPC24 in Prydz Bay in the Atlantic Ocean sector (Denis et al., 2010), and from MD03-2601 (Crosta et al., 2008; Denis et al., 2010) and U1357 (Ashley et al., 2021) offshore Adélie Land in the Indian Ocean sector. Denis et al. (2010) suggested that increasing sea ice cover over the late Holocene is a common feature in the coastal Antarctic. From the Antarctic ice cores, secular increases in sodium, which is a proxy for sea ice extent in the Southern Ocean (Wolff et al., 2003), during the Holocene have been widely observed (Winski et al., 2021). From ice cores in the EAP, the increases in sodium fluxes and concentrations since the middle Holocene were observed at EDML (Fischer et al., 2007), Dome Fuji (Iizuka et al., 2008), Dome C (Fischer et al., 2007) and TALDICE (Mezgec et al., 2017). Thus, sea ice extent was probably increased in the areas of the Southern Ocean that supply sodium to the EAP.

From the above evidence, the decreasing trend in the Dome Fuji SMB over the last 5000 years appears to be associated with surface cooling and sea ice expansion in the potential water vapor source areas for the Dome Fuji region, namely the Atlantic Ocean and Indian Ocean sectors of Southern Ocean (Suzuki et al., 2008). The cooling trend is also seen over the EAP, possibly suggesting a general cooling trend of the atmosphere from the vapor source regions to the EAP. Our observation is consistent with the analyses of Vannitsem et al. (2019) that both surface air temperature and sea ice trends could have contributed to the decreasing Dome Fuji SMB. On the contrary, our observation may be inconsistent with the sensitivity study of the regional climate model (Kittel et al., 2018) that shows that the decrease in accumulation rate in the EAP requires an increase in SST. A thorough investigation of the inconsistency is beyond the scope of this study, but it might be related to the differences in the background climate states between the model runs (1979–2015) and the past 5000 years.

We compare the centennial-scale variabilities of SMB records at Dome Fuji and other parts of Antarctica for the last 1000 years. For Dome Fuji, the deviation of SMB from the long-term trend was calculated by subtracting a linear function fitted through 1236–1850 CE (the oldest year for the fitting was adopted from the other records with a shorter duration). For the comparison, we use the SMB reconstruction for the Atlantic sector of DML based on the analyses of six firn cores (Hofstede et al., 2004), as well as the composite SMB records for the EAP, all of East Antarctica (EA; including coastal area of DML, Wilkes Land and the East Antarctic Plateau) and parts of West Antarctica (WA; the Antarctic Peninsula is not included) based on the compilations of multiple ice cores (Thomas et al., 2017). The published records by Thomas et al. (2017) were smoothed by a 71-year moving average to match the resolution with that of the Dome Fuji data, and the deviations from their linear trends were calculated in the same manner as the Dome Fuji record. Figure 13 shows the SMB anomaly for Dome Fuji and the Atlantic sector of DML, EAP, EA and AIS. The SMB anomaly around Dome Fuji has four distinct periods: mostly negative before 1300 CE, slightly positive for 1300–1450 CE, slightly negative for 1450–1850 CE, and positive after 1850 CE (Fig. 13a). Before 1460 CE, the Dome Fuji, DML, EAP and EA records commonly exhibit a peak around 1300–1450 CE, suggesting this anomaly is a real climatic signal and extended to much of East Antarctica. For the middle part (1460–1850 CE), the DML, EAP and EA records are also characterized by a broad minimum around 1500 CE and a relatively short maximum around 1600 CE. The Dome Fuji record shows consistently negative values in this period, but it also has a broad minimum around 1500 CE and small maximum around 1600 CE. The WA composite also shows a positive anomaly around 1600 CE. After ∼1650 CE, Dome Fuji, DML, EAP, EA and WA records all show a minimum around 1700 CE, although the durations of the negative anomalies are different. The Dome Fuji and DML records show minima in more recent periods (around 1850 CE in the Dome Fuji record and around 1900 CE in the DML record), which are not seen in the other records. For the most recent part, all records show increases with different durations (over 100–300 years).

We further compare our Dome Fuji record with other published reconstructions. For Princess Elizabeth Land, the SMB anomaly is relatively high for ∼1207–1450 CE, low for ∼1450–1850 CE and significantly increased afterward (until 1996 CE) (Li et al., 2009). In the South Pole record (Ferris et al., 2011), the SMB anomaly is relatively large for ∼1200–1400 CE, reduced during the period of 1500–1900 CE, and then increased afterward. On the continental scale, a stacked SMB record by Frezzotti et al. (2013) for the last 800 years, which includes more data from the Antarctic Plateau than the composite of Thomas et al. (2017), identified three periods of low accumulation rate (1250–1300, 1420–1550 and 1660–1790 CE). In summary, the reconstructions for the EAP commonly show a higher accumulation rate for the periods of ∼1250–1400 and ∼1550–1650 CE and a lower accumulation rate around 1500 and 1700 CE. The increases in the snow accumulation over the 20th century are also observed, although the timing of the onsets is different among the records.

Multidecadal to centennial SMB variabilities may be driven by external forcings such as solar irradiance and volcanic activity, as well as by internal variabilities of the atmosphere–ocean system (Goosse et al., 2012; Frezzotti et al., 2013; PAGES 2k Consortium, 2013, 2019; Medley and Thomas, 2019; Mann et al., 2021). Previous studies have suggested the correspondence between the periods of reduced SMB and strong volcanic forcings or solar minima over the last millennium (Bertler et al., 2011; Frezzotti et al., 2013; Osipov et al., 2014; Thomas et al., 2017), possibly through the reduction of incoming solar radiation and associated changes in the climate system (e.g., surface cooling and circulation change). In the detrended SMB record at Dome Fuji over the last 1000 years, the intervals of small SMB anomalies appear to correspond to those with large volcanic eruptions (e.g., Samalas in 1257 CE, an unknown eruption in 1458 CE, Tambora in 1815 CE and Krakatau in 1883 CE), as well as the Spörer, Maunder and Dalton grand solar minima (Fig. 13). Note that the eruption in 1458 was previously estimated as being from Kuwae, but it is now interpreted as an unknown event (perhaps in the Southern Hemisphere) (Hartman et al., 2019). On the other hand, the Oort and Wolf grand solar minima do not correspond to low SMB anomaly periods at Dome Fuji, but they coincide with the low SMB anomaly periods in the EA composite. Thus, our data are partly consistent with the previous suggestions for the last 1000 years. As discussed above, the Dome Fuji stacked record shows a negative SMB anomaly for the 15th to 19th centuries. During this period, in addition to strong volcanic forcing and weak solar activity, the radiative forcing of GHGs becomes the smallest in the last 2 kyr, which may have contributed to the lower SMB.

For the older period, the detrended SMB for 1500 BCE–1000 CE shows variability similar to the last 1000 years, and it shows reduced variability before 1500 BCE (Fig. 12). In addition, the correspondence between the intervals of low SMB anomalies and strong volcanic or weak solar forcings is not clear in the older period. Before 1000 CE, the average solar activity was larger and the frequency of large volcanic eruptions was lower than during the last 1000 years. It might be possible that the clear relationship between the significant negative SMB anomalies and solar and volcanic forcings is only visible when the two forcing anomalies are strong and coincide, which seems rarer in the older period than in the last 1000 years. However, there is also a possibility that the SMB variability on the multidecadal to centennial scales is not captured well by our stack because of the lower number of ice cores and age constraints. The reconstructions of volcanic and solar forcings might also have larger uncertainties in the older part. In any case, it is desirable to obtain long-term and detailed reconstructions of the AIS SMB at multiple sites to examine their relationships with the climatic forcings.

The Dome Fuji SMB record reveals that the magnitude of the snow accumulation increase during the last 150 years is the greatest in the last 5000 years (+1.308±0.339 kg m-2 100 yr-1, Table 3). We speculate that the large increase may have been created by the combination of anomalously low accumulation in the 18–19th centuries (23.8±0.4 kg m-2 yr-1 in Stack 2) and a continuous increase in the industrial period. As previous studies have suggested, the significant increase in accumulation rate in the 20th century may be attributed to anthropogenic forcings, such as increased atmospheric GHG concentrations and stratospheric ozone depletion (e.g., Medley and Thomas, 2019). Climate models suggest that enhanced GHG radiative forcing would increase temperature, and thus the moisture content in the atmosphere over the Southern Ocean and Antarctica, leading to the increase in Antarctic precipitation (e.g., Previdi and Polvani, 2016). The ozone depletion is suggested to strengthen the mean zonal wind at middle to high latitudes in the lower troposphere, which is balanced by a stronger mean meridional wind (Chemke et al., 2020). In addition, ozone depletion would enhance barotropic instability, which also increases the poleward eddy moisture fluxes (Chemke et al., 2020). In the preindustrial period, the variation in the combined radiative forcing of solar activity, volcanic forcing and GHGs is smaller than 1 W m-2, while the anthropogenic radiative forcing alone is +2.7 W m-2 in the industrial period (in 2019 relative to 1750) (IPCC, 2021). Although the current accumulation rate is likely lower than the maximum value in the last 5000 years of the preindustrial period, we speculate that it could eventually exceed the natural range in the future as the anthropogenic forcings continue to change the atmosphere–ocean system around Antarctica.