In the framework of the NEEM project, a 2540 m deep ice core was drilled in northwestern Greenland (77∘45′ N, 51∘06′ W) at 2479 m a.s.l. The site is characterized by an average annual temperature of -29 ∘C and a modern accumulation of 22 cm ice equivalent per year. According to the GICC05modelext-NEEM-1 timescale, the ice core covers the last 128 kyr (Rasmussen et al., 2013). The ice cores were cut to obtain ice sticks with a square cross section of 36 mm × 36 mm. They were continuously melted on a continuous flow analysis (CFA) system with a typical melt speed of 3.5 cm min-1 (Schüpbach et al., 2018). The CFA system provides meltwater from the inner and least likely to be contaminated part of the core; thus, we did not adopt any further decontamination procedure. The inductively coupled plasma mass spectrometry (ICP-MS) samples were manually collected at a low resolution (110 cm). The temporal resolution depends on the accumulation rate and decreases with depth because of the ice thinning. According to the available timescale (Rasmussen et al., 2013) and considering the 110 cm sampling resolution, the temporal resolution varies from decadal to millennial (Table 1).

Samples were collected in vials that had been previously cleaned as follows: stored in HNO3 5 % for 7 d (Suprapure, ROMIL, UK), rinsed three times with ultrapure water (UPW, ELGA, UK), stored in HNO3 2 % for 7 d (Suprapure, ROMIL, UK), rinsed three times with UPW and then stored in HNO3 1 % (Ultrapure, ROMIL, UK) until the day before the sample collection, when they were rinsed three times with UPW and dried overnight under a laminar flow hood (Class 100). The samples were kept frozen and shipped to Italy for analysis. Once melted, the samples were acidified to pH 1 using HNO3 (Suprapure, ROMIL, UK). To ensure an effective dissolution of Fe particles, samples were stored at room temperature and analyzed 30 d after the acidification without any additional filtration step. We adopted this approach because analysis immediately after the acidification step might have led to uncertainties attributable to the Fe dissolution kinetics (Edwards, 1999; Koffman et al., 2014). Our choice was consistent with other studies that have indicated that samples to be used for the calculation of atmospheric fluxes must be acidified for at least 1 month prior to analysis in order to avoid any possible misinterpretation of the trace element data (Koffman et al., 2014). We will refer to this fraction as total dissolvable Fe (TDFe); total dissolvable Fe includes both the most labile fraction (dissolved iron, DFe), which is rendered soluble under mildly acidic conditions (Hiscock et al., 2013), and the fraction enclosed in iron-bearing mineral particles. TDFe does not directly represent the actual bioavailable Fe that can be dissolved into seawater at pH 8, but, considering that TDFe and DFe are significantly correlated (Du et al., 2020; Xiao et al., 2020), it represents an upper limit of the eolian Fe potentially available for the phytoplankton (Edwards et al., 2006).

The ice samples were analyzed with an Inductively Coupled Plasma Single Quadrupole Mass Spectrometer (ICP-qMS, Agilent 7500 series, USA) equipped with a quartz Scott spray chamber for the determination of Ca, Na and Fe. To minimize any kind of contamination, all of the instrument tubes were flushed before the analysis for 2 h with 2 % HNO3 (Suprapure, ROMIL, UK). A 120 s rinsing step with 2 % HNO3 (Suprapure, ROMIL, UK) occurred after each sample analysis to reduce any possible memory effect. The vials used for the standard preparation were cleaned following the same procedure adopted for the ice samples. Considering the isobaric and polyatomic interferences affecting Fe, this element was quantified using the interference-free isotope 57Fe. External calibration curves with acidified standards (2 % HNO3, Suprapure, ROMIL, UK) were prepared for Ca, Na and Fe from the dilution of a certified single-element 1000 ppm ± 1 % standard solution (Fisher Chemical, USA). The resulting R2 for the external calibration curves was 0.999 for all of the elements. The limit of detection (LoD) for 57Fe, calculated as 3 times the standard deviation of the blank, was 0.8 µg L-1. To assess accuracy for Fe, the TM-RAIN04 certified reference material (National Research Council of Canada) was measured every 50 samples. The accuracy was determined as a recovery percentage calculated as O/T %, where O is the determined value, and T is the certified value. For Fe, the accuracy was 104 %, whereas precision, calculated as relative standard deviation (RSD %) of selected samples read multiple times (n = 5) during the analysis, was on average 5 % (7 % for samples (n = 3) from the interglacial period, and 4 % for samples (n = 3) from the last glacial period). For Ca and Na, the LoD was 1 and 3 µg L-1, respectively. In the absence of a certified reference material, Ca and Na accuracy was calculated using a quality control (QC) sample prepared at 10 µg L-1 and measured every 50 samples. Accuracy for Ca and Na, calculated as described above, was 94 % and 108 %, respectively, whereas precision (RSD %) was on average 6 % (4 % for samples from the interglacial period, and 7 % for samples from the last glacial period) and 2 % (for both periods), respectively.

The non-sea-salt Ca concentration is commonly used as proxy for terrestrial inputs in polar regions, and it is calculated as nssCa = [Ca] - ([Ca] / [Na])sw ⋅ [Na], where “sw” indicates seawater.

Fe and nssCa concentrations and fluxes were calculated as F=C⋅A, where F is the Fe flux (in mg m-2 yr-1), C is the Fe or nssCa concentration (in ng g-1) and A the accumulation (in m yr-1 ice equivalent) previously determined by Rasmussen et al. (2013). A pattern of higher dust (expressed as nssCa2+) and Fe fluxes during colder climate periods and lower dust and Fe fluxes during warmer climate periods is clearly recognizable (Fig. 2, Table 2).

The Holocene (0.042–11.7 ka) was characterized by average Fe fluxes of 0.5 mg m-2 yr-1 that varied between 0.01 and 5.3 mg m-2 yr-1 (Fig. 2). The coefficient of variability (CV), calculated as the ratio between the standard deviation and the mean value, was 1.2. The more recent 4000 years are characterized by the highest average Fe fluxes (0.6 ± 0.4 mg m-2 yr-1). The lowest Fe fluxes were recorded between 4000 and 8000 years b2k (0.3 ± 0.2 mg m-2 yr-1). During the Younger Dryas (YD; 11.7–12.9 ka), an abrupt cooling was observed with a drop in the δ18O value from -36.9 ‰ to -43.1 ‰. Coincidently, the recorded average Fe fluxes rose to 1.2 ± 0.4 mg m-2 yr-1, higher than both the 12.9–13.9 ka (0.5 ± 0.3 mg m-2 yr-1) and the 10.7–11.7 ka (0.3 ± 0.2 mg m-2 yr-1) periods.

The last glacial period (11.7–108 ka) showed Fe fluxes 4 times higher (2.0 ± 2.2 mg m-2 yr-1) than the Holocene, spanning from 0.05 to 16.5 mg m-2 yr-1 (Fig. 2). However, a significant variability during the last glacial period was detected. During the LGM and MIS 4, average Fe fluxes were 7 times (3.6 ± 2.3 mg m-2 yr-1) and 10 times (5.8 ± 2.8 mg m-2 yr-1) greater than the Holocene average. Fe fluxes also increased during the MIS 5c–MIS5b transition (87 ka), when a concurrent decrease in δ18O values was observed. During MIS 5c and MIS 5d, Fe fluxes were comparable with those detected during the Holocene. The high frequency of the Dansgaard–Oeschger (D–O) events during MIS 3 was mirrored by the high variability in both nssCa and Fe fluxes. Each stadial period corresponded to an increase in both Fe and nssCa. However, their variability was significantly different. During MIS 3, Fe fluxes showed maximum values greater than 5 mg m-2 yr-1 during D–O 4, 9, 12 and 15 (8.5, 6.5, 7.5 and 6.6 mg m-2 yr-1 respectively), and lower than 5 mg m-2 yr-1 during D–O 6, 7, 8, 10, 11 and 13 (3.9, 2.6, 4.1, 2.6, 2.7 and 3.2 mg m-2 yr-1 respectively). This variability was significantly higher than that recorded for nssCa, which showed maximum values close to 20 mg m-2 yr-1 for all of the D–O events.

The NEEM Fe ice core record allows for the first comparison of Fe concentrations and fluxes between the Arctic and Antarctica (Fig. 3, Table 3). The only Antarctic Fe records that can reach at least the LGM are from Talos Dome (TD) (Spolaor et al., 2013; Vallelonga et al., 2013), Law Dome (LD) (Edwards et al., 2006; Edwards et al., 1998) and EPICA Dome C (EDC) (Wolff et al., 2006). However, we point out that both the samples from Dome C and Talos Dome were acidified for at least 24 h, leading to a possible underestimation of the actual TDFe concentration. This implies that the general trends and features can be comparable with the NEEM record, whereas absolute concentrations might differ due to the different acidification procedure used (Koffman et al., 2014).

In Antarctica, the average Fe flux and concentration values varied significantly among the different sites during the Holocene with similar values recorded at the coastal sites (TD) and lower values recorded in the internal Antarctic Plateau (EDC) (Table 3). For TD, this was explained both through changes in atmospheric transport patterns across Antarctica and through an additional local input of dust from proximal Antarctic ice-free zones that affected coastal sites more than the central plateau, which was exclusively exposed to remote sources such as southern South America (Albani et al., 2012; Delmonte et al., 2010b; Vallelonga et al., 2013).

During the LGM, both TD and EDC shared a similar dust flux loading, comprised between 10 and 15 mg m-2 yr-1 (Baccolo et al., 2018), and the same dust source region, as confirmed by the Sr–Nd isotopes (Delmonte et al., 2010a). Compared with the Holocene, the atmospheric dust fluxes in TD increased by a factor 6, whereas the increase was by approximately a factor 25 in EDC (Delmonte et al., 2010b). This is mirrored by a similar average Fe flux enhancement compared to the Holocene with values that were up to 4- and 21-fold higher, respectively (Vallelonga et al., 2013; Wolff et al., 2006). These discrepancies between the two sites are likely due to the higher Holocene dust flux observed in TD compared with EDC, as a consequence of a relevant local dust contribution at TD (Baccolo et al., 2018; Delmonte et al., 2010b).

During the last glacial period, the most relevant dust source was southern South America for both TD and EDC (Basile et al., 1997; Delmonte et al., 2010b; Lambert et al., 2008). Dust fluxes peaked during MIS 4 where both sites recorded maximum values around 10 mg m-2 yr-1 (Lambert et al., 2008; Vallelonga et al., 2013) and comparable Fe fluxes (0.17 ± 0.07 mg m-2 yr-1 at TD and 0.12 ± 0.07 mg m-2 yr-1 at EDC) (Vallelonga et al., 2013; Wolff et al., 2006).

The LD record, due to the different analytical preparation of the samples, is not directly comparable with TD and EDC. Nevertheless, we can still evaluate and discuss the Fe flux ratio between the Holocene and the LGM. Unfortunately, for the LD record, there is no dust profile available, meaning that it is not possible to define the main dust and Fe sources for this location, although the Australian continent has been an important source of mineral dust in the recent past (Edwards et al., 2006; Vallelonga et al., 2002). During the LGM, Fe fluxes increased 10-fold compared with the Holocene period, 2.5 times more than what was observed in TD. Similarly to what was observed in the EDC record, this difference might be explained by the absence of local dust sources that affected LD during the Holocene or by the lower sampling frequency for the LD record (n = 27) compared with TD (n = 801).

Despite the different acidification times, the overall picture during the Holocene is that the average Fe fluxes in NEEM (0.5 mg m-2 yr-1, CV = 1.2) were higher than in Antarctica. Among the Antarctic Fe fluxes, TD (0.09 mg m-2 yr-1, CV = 1.2) and LD (0.04 mg m-2 yr-1, CV = 0.5) were higher than those recorded at EDC (0.007 mg m-2 yr-1, CV = 0.2).

In NEEM, the LGM (19–26.5 ka) was characterized by a 10-fold and 7-fold enhancement in dust (expressed as nssCa) and Fe fluxes, respectively. A similar behavior was observed in the Antarctic cores as described above (Table 3). Considering that the atmospheric CO2 concentration dropped down to 180 ppm (Köhler et al., 2017), the global Fe flux enhancement likely contributed to part of this decrease, promoting marine productivity in some HNLC regions (Amo and Minagawa, 2003; Kawahata et al., 2000; Martínez-Garcia et al., 2011).

During MIS 4 (60–71 ka), NEEM Fe fluxes were higher than all of the other investigated records. Compared with the LGM average, dust (Ruth, 2007), nssCa and Fe fluxes (this work) in the Arctic during MIS 4 exhibited a ≈ 1.5-fold increase (Table 3), while they were lower both in TD and EDC. To explain this behavior we advance some hypotheses. The first is that the increase in dust and Fe fluxes can be attributable to changes in the atmospheric circulation, likely due to the topographic influence of the Laurentide Ice Sheet (LIS). Indeed, during the LGM, the LIS was nearly 2 times larger than at MIS 4 (Löfverström et al., 2014; Tulenko et al., 2020), and it might have caused a stronger meridional splitting of the westerlies (Löfverström et al., 2014) and a southward migration of their mean position (Kang et al., 2015; Manabe and Broccoli, 1985). The southward shift during the LGM might have produced a reduction in strong winds passing over the source areas (i.e., Taklimakan and Gobi deserts) (Kang et al., 2015) and/or a stronger southward Fe and dust deposition over the Chinese Loess Plateau (Zhang et al., 2014) and the midlatitude North Pacific (Sun et al., 2018). In contrast, during MIS 4, the westerlies might have been located northward (i.e., over the Taklimakan and Gobi deserts) and characterized by a less marked meridional splitting (Löfverström et al., 2014), conveying a larger amount of dust to Greenland. We also propose two alternative hypotheses that rely on (1) the possibility that additional dust sources (e.g., Saharan dust) might have reached Greenland during MIS 4 and on the fact that (2) during MIS 4, the Asian monsoon system was stronger in winter than in summer, producing drier conditions that caused an enhanced dust production and transport to Greenland (Xiao et al., 1999). However, to better address this point, a more comprehensive investigation that involves a large set of paleorecords and atmospheric modeling is required, although this is beyond the scope of this paper.

A parallel study that reported the Fe concentration from the NEEM ice core was recently published (Xiao et al., 2020). It reports the TDFe and DFe concentrations and fluxes with a lower temporal resolution (n = 166) than the current investigation (n = 1596). Moreover, the analytical approach was different, as the melted ice samples were filtered at 0.45 µm and acidified for 6 weeks before the analysis. Even though the overall pattern between the two records is similar, we observe several differences between Xiao et al. (2020) and our study: (a) the average Fe concentration over the entire record is 4-fold higher than that found in our investigation (101.4 ng g-1 vs. 20.4 ng g-1); (b) the Fe concentration range is wider (1.5–1194.5 ng g-1 vs. > LoD – 457.6 ng g-1) compared with the data presented in this paper; (c) average Fe fluxes are 2.4-fold higher during the Holocene (1.2 mg m-2 yr-1 vs. 0.5 mg m-2 yr-1) and 3.5-fold higher during the LGM (12.5 mg m-2 yr-1 vs. 3.6 mg m-2 yr-1) compared with those recorded in this study; (d) the LGM Fe flux showed a 10-fold increase during the Holocene, compared with the 7-fold enhancement that we observed; (e) the TDFe fluxes and concentrations were higher during the LGM than during MIS 4, whereas we found higher fluxes during MIS 4, consistent with a similar enhancement of nssCa and dust (Ruth, 2007).

Possible reasons for these differences might stem from the different temporal resolution and the discrepancies between the adopted analytical approaches; this highlights the need to standardize the analytical procedures when trace elements are analyzed in ice and snow samples in order to have a more reliable comparison among both different and identical locations.

Considering the biological relevance of Fe and taking advantage of the Fe flux record retrieved from the NEEM ice core, one important question remains regarding whether its flux increase during the last glacial period triggered the marine productivity in the HNLC region of the North Pacific (Olgun et al., 2011).

Nowadays, a significant amount of Asian dust (250 Mt yr-1) is primarily deposited over the HNLC region of the subarctic Pacific (Serno et al., 2014; Zhang et al., 2003), and the marine productivity changes in this oceanic region might reflect potential Fe fertilization effects promoted by atmospheric Fe supply. During modern times, both increases in the eolian influx from Asia (Young et al., 1991) and sporadic Fe input from volcanic eruptions (Langmann et al., 2010) have resulted in an enhancement of MPP by more than 60 %. Moreover, recent Fe fertilization experiments performed south of the Gulf of Alaska (McDonald et al., 1999; Tsuda et al., 2003) have shown significant increases in the abundance of diatoms and in the chlorophyll-a concentration (Boyd et al., 1996), indicating that the North Pacific is possibly sensitive to atmospheric Fe inputs. However, no data are available to evaluate if the Fe sensitivity of the subarctic Pacific Ocean holds over even longer timescales or if an increase in the eolian Fe supply, observed during glacial periods, could explain the MPP variability in the subarctic Pacific Ocean. To address these points, we compared the NEEM Fe record with different marine sediment cores (Table 4).

For both interglacial and glacial periods, previous geochemical evidence indicates that the dust source influencing Greenland and the North Pacific mainly originated from the East Asian deserts (Schüpbach et al., 2018; Serno et al., 2014). However, considering that there are no eolian Fe flux records from the marine sediment cores, they might have received different amount of Fe compared with what is observed in the ice core record. Through a comparison between a marine sediment record from the western subarctic Pacific Ocean (SO202-07-6) and the NGRIP ice core, it has been shown that dust fluxes changed coherently and simultaneously during abrupt climate changes, although the amplitude of the change was different (Serno et al., 2015). The larger variability observed in NGRIP, as well as in NEEM, compared with marine sediments indicates changes in the atmospheric dust transport from the source areas to Greenland (e.g., rate of aerosol rainout and different wind strength).

Recently, it has been proposed that additional dust sources might have influenced Greenland in the last 31 kyr (Han et al., 2018; Lupker et al., 2010). Strontium and lead isotopes indicate that Saharan dust contributed to the overall NEEM dust budget during the Younger Dryas (12 %–73 %) and between 17 and 22 ka (16 %–70 %), while the Taklimakan and Gobi contribution (i.e., eastern Asia sources) was dominant (55 %–94 %) prior to 22 ka (Han et al., 2018). However, despite the Saharan dust source, we assume that the main dust source for the NEEM ice core during the last glacial period is represented by the Gobi and Taklimakan deserts (Svensson et al., 2000). This is coherent with the dust changes' synchronicity among Greenland, the Chinese loess (Ruth et al., 2007) and the northern Pacific sediment records located downwind of the Asian dust sources (Schüpbach et al., 2018; Serno et al., 2015). Nevertheless, additional investigations to assess the magnitude of the Saharan dust contribution prior to 31 ka and to identify other possible source regions are needed.

All variables considered (i.e., different dust amplitude and other potential dust sources), and observing that the overall pattern of higher dust deposition during the coldest periods is consistent between the ice and sediment core records, we assumed that the Fe flux changes observed in NEEM are representative for the eolian Fe supply to the subarctic Pacific Ocean.

To evaluate whether past marine productivity was influenced by the atmospheric Fe supply for the period ranging from the LGM to the Holocene, we compared the NEEM record with the high temporal resolution SO202-27-6 (from the Patton–Murray Rise plateau, eastern subarctic Pacific Ocean) and the SO202-07-6 (from the Detroit Seamount, western subarctic Pacific Ocean) productivity records (Méheust et al., 2018). For a long-term record, we relied on the ODP887 (McDonald et al., 1999) and the ODP882 (Haug et al., 1995) sediment cores, located close to SO202-27-6 and SO202-07-6, respectively. A comparison over the last 108 kyr between the NEEM record and the S-2 sediment core (from the Shatsky Rise, midlatitude North Pacific) was also performed (Amo and Minagawa, 2003) (Fig. 4, Table 4).

The past marine primary productivity reconstruction was performed relying on the Si/Al ratio, the percentage of biogenic silica and the brassicasterol concentration. The Si/Al ratio is used as a proxy for opal, or biogenic silica (diatoms), in the absence of directly measured opal concentrations. The normalization to Al removes any possible variable inputs of lithogenic detritus (McDonald et al., 1999). Brassicasterol is a sterol compound that has been used as a molecular indicator of the presence of diatoms (Sachs and Anderson, 2005). The brassicasterol concentration is also used, along with highly branched isoprenoid alkenes (IP25), for the phytoplankton IP25 index (PIP25) calculation, which is a proxy for the evaluation of past sea ice conditions (Méheust et al., 2018; Müller et al., 2011)