Experiments in this study are performed using the atmosphere component (EAMv1) of the US Department of Energy (DOE) E3SMv1 model (Rasch et al., 2019). The model predicts the number and mass mixing ratios of seven aerosol species (i.e., mineral dust, black carbon – BC, primary organic aerosol, secondary organic aerosol, sulfate, sea salt, and marine organic aerosol – MOA) through a four-mode version of the modal aerosol module (MAM4) (Liu et al., 2016; Wang et al., 2020). The four aerosol modes are Aitken, accumulation, coarse, and primary carbon, while dust aerosols are carried in accumulation and coarse modes. Aerosol optical properties in each mode are parameterized following Ghan and Zaveri (2007). The dust optics used in this study are updated according to Albani et al. (2014).

EAMv1 includes a two-moment stratiform cloud microphysics scheme (MG2) (Gettelman and Morrison, 2015). We note that the WBF process rate in EAMv1 is tuned down by a factor of 10, which results in more prevalent supercooled liquid water clouds in high latitudes than observations and many other global climate models (Y. Zhang et al., 2019; Zhang et al., 2020). In addition, the Cloud Layers Unified By Binormals (CLUBB) parameterization (Bogenschutz et al., 2013; Golaz et al., 2002; Larson et al., 2002) is used to unify the treatments of planetary boundary layer turbulence, shallow convection, and cloud macrophysics. Deep convection is treated by the Zhang and McFarlane (1995) scheme.

In EAMv1, the heterogeneous ice nucleation in mixed-phase clouds follows the classical nucleation theory (CNT) (Hoose et al., 2010; Y. Wang et al., 2014). CNT holds the stochastic hypothesis, which treats the ice nucleation process as a function of time. Immersion, contact, and deposition nucleation on dust and BC are treated in the CNT scheme. More details about CNT parameterization are provided in Sect. S2.1 in the Supplement.

The experiments we conducted for this study are shown in Table 1. For the control experiment (hereafter CTRL), the EAMv1 was integrated from July 2006 to the end of 2011 at 1∘ horizontal resolution and 72 vertical layers. The first 6 months of the experiment were treated as model spin-up, and the last 5 years of results were used in analyses. The horizontal wind components were nudged to the Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) (Gelaro et al., 2017) meteorology with a relaxation timescale of 6 h (Zhang et al., 2014). In addition to CTRL, we conducted three sensitivity experiments to investigate the INP effect of dust from major source regions. In these sensitivity experiments, heterogeneous ice nucleation in the mixed-phase clouds by dust from local Arctic sources, North Africa, and East Asia is turned off (i.e., noArc, noNAf, and noEAs, respectively). The other settings of these three experiments are identical to CTRL. Analyses related to the sensitivity experiments are provided in Sect. 3.4.

Dust emission in the default EAMv1 is parameterized following Zender et al. (2003, Z03), which uses semi-empirical dust source functions to address the spatial variability in soil erodibility. The HLD emission is omitted in the Z03 scheme, since it was thought to be dubious (Zender et al., 2003). In this study, we replaced the Z03 scheme with another dust emission parameterization (Kok et al., 2014a, b, K14) that avoids using a source function (see more details about K14 in Sect. S1). The K14 scheme is able to produce the HLD emission over Iceland, the Greenland coast, Canada, Svalbard, and northern Eurasia (Fig. 1a). Furthermore, to address the overestimation of dust emission in clay size (< 2 µm diameter) (Kok et al., 2017), we changed the size distribution of emitted dust particles from Z03 to that based on the brittle fragmentation theory (Kok, 2011). 1.1 % of the total dust mass is emitted to the accumulation mode and 98.9 % of that is emitted to the coarse mode based on the brittle fragmentation theory, whereas the fractions are 3.2 % and 96.8 %, respectively, in Z03.

To quantify the source attribution of dust, we implemented a dust source-tagging technique in EAMv1. This modeling tool was previously applied to BC (H. Wang et al., 2014; Yang et al., 2017b), sulfate (Yang et al., 2017a), and primary organic aerosol (Yang et al., 2018) in the Community Atmosphere Model version 5 (CAM5). In this method, dust emission fluxes from different sources are assigned to separate tracers and transport independently so that dust originating from different sources can be tracked and tuned separately in a single model experiment. As shown in Fig. 1a, dust emissions from seven source regions are tagged: Arctic (Arc; above 60∘ N, HLD source), North America (NAm), North Africa (NAf), Central Asia (CAs), Middle East and South Asia (MSA), East Asia (EAs), and the rest of the world (RoW). The Arctic source is further divided into four sub-sources: Alaska (Ala), northern Canada (NCa), Greenland and Iceland (GrI), and northern Eurasia (NEu) (Fig. S1), which are used in the analysis of INP sources in Sect. 3.3. RoW represents the three major dust sources in the Southern Hemisphere (South America, South Africa, and Australia), along with very low emissions from Europe and the Antarctic.

The global dust emission for CTRL is 5640 Tg yr-1, which is tuned so that the global average dust aerosol optical depth (DOD) is 0.031. This is within the range of the observational estimate (0.030 ± 0.005) by Ridley et al. (2016). To maintain the magnitude of the global averaged DOD, our tuned global dust emission exceeds the range of the AeroCom (Aerosol Comparisons between Observations and Models) models (500 to 4400 Tg yr-1; Huneeus et al., 2011), likely due to a short lifetime caused by dust dry deposition that is too strong at the bottom layer near the dust source regions in EAMv1 (Wu et al., 2020). It is also about 2000 Tg yr-1 higher than the previous EAMv1 studies (Shi and Liu, 2019; Wu et al., 2020) because we distribute less dust mass into the accumulation mode and more dust mass into the coarse mode based on Kok (2011). The HLD emission is further tuned up by 10 times so that it accounts for 2.6 % (144 Tg yr-1) of the global dust emission (Fig. 1b), which is comparable with the recent estimates of 2 %–3 % above 50∘ N by Bullard et al. (2016) and of 3 % above 60∘ N by Groot Zwaaftink et al. (2016). The majority of global dust emission is contributed from North Africa (51.9 %, 2929 Tg yr-1) and Asia (37.7 %, 2124 Tg yr-1), with Asian emissions composed of MSA (20.2 %, 1140 Tg yr-1), EAs (10.9 %, 613 Tg yr-1), and CAs (6.6 %, 371 Tg yr-1). NAm has a weak dust emission of 33.4 Tg yr-1 that only contributes 0.6 % to the global emission, while the RoW has a combined contribution of 7.3 % (410 Tg yr-1). In addition, the seasonal variations between HLD and LLD emissions are different – the HLD (Arctic) source is more active in late summer and autumn, while the LLD sources (e.g., NAf, MSA, EAs) peak in spring and early summer (Fig. 1c).

To evaluate the model performance in simulating the dust cycle, we compare the model predictions with measured aerosol optical depth (AOD), dust surface concentrations, and dust deposition fluxes from global observation networks (Fig. 2). We select and process the level 2.0 AOD data (2007–2011) at 40 “dust-dominated” AErosol RObotic NETwork (AERONET; Holben et al., 1998) stations following Kok et al. (2014b). We note that the AERONET AOD measurements are biased towards clear-sky conditions due to the cloud-screening procedure (Smirnov et al., 2000). For dust surface concentrations, we use the same measurements at 22 sites, which Huneeus et al. (2011) used for the AeroCom comparison, and further extend the dataset with measurements at three high-latitude stations: Heimaey (Prospero et al., 2012), Alert (Sirois and Barrie, 1999), and Trapper Creek (Interagency Monitoring of Protected Visual Environments, IMPROVE; Malm et al., 1994). It is noted that the measurements at Trapper Creek only include dust particles smaller than 2.5 µm and are only compared with simulated dust concentrations at the same size range. All other concentration measurements capture dust particles below 40 µm and are compared with simulated dust over the whole size range (< 10 µm). The dust deposition flux dataset, which includes 84 stations, is also the same as Huneeus et al. (2011). The locations of the observation network are shown in Fig. 2d, with the AOD data taken close to source regions and the dust surface concentrations and deposition fluxes measured at relatively remote regions. The Pearson correlation coefficient (r) is provided for each comparison. We note that the comparisons are subject to representative biases caused by comparing an observational station with a global model grid point (with a horizontal resolution of ∼ 100 km). The comparisons of dust concentration and deposition flux also have systematic errors because the measurements were for a different time period than that of the model simulation.

In general, the three comparisons indicate that our CTRL simulation is capable of capturing the global dust cycle in both near-source and remote regions. As shown in Fig. 2a, the modeled AOD is within a factor of 2 of the observations over most of the stations. The correlation of the AOD comparison is 0.73, which is comparable to the best-performing simulation (r = 0.72) in Kok et al. (2014b). Our model also does a fairly good job in simulating the dust surface concentrations (Fig. 2b) and produces a correlation coefficient of 0.84. For the three high-latitude sites, the model shows moderate underestimation at Heimaey and Trapper Creek and large positive bias at Alert (see discussion below). The correlation coefficient for simulated dust deposition fluxes (r = 0.48) is also within the range of the AeroCom comparisons (0.08 to 0.84) in Huneeus et al. (2011). The model results over most of the sites are within 1 order of magnitude difference, except at the polar regions. In particular, the model overestimates the dust deposition flux in Greenland (red triangles in Fig. 2c and d) by around 2 orders of magnitude, likely due to local emissions simulated near the coast of Greenland that are too strong (Fig. 1a).

The seasonal cycle of dust surface concentrations at the three Arctic stations (Heimaey, Alert, and Trapper Creek) is shown in Fig. 3, along with the contribution from seven tagged sources. The simulated dust concentrations at Heimaey are dominated by HLD and agree well with the observation in late summer and autumn (Fig. 3a). Its annually averaged low bias shown in Fig. 2b mainly comes from the springtime, when Prospero et al. (2012) found the observed dust to be related to dust storms in Iceland, indicating a possible underestimation in the simulated Icelandic dust during this time. The HLD also dominates the surface dust concentrations at Alert (Fig. 3b), leading to a large overestimation from June to August in our simulation, which possibly implies a high bias and wrong seasonal cycle of HLD emission over Greenland and northern Canada. The Trapper Creek station is instead dominated by LLD from East Asia and shows an underestimation for most of the year. It is noted that we only include fine dust (diameter < 2.5 µm) for the comparison at Trapper Creek. A larger size range is likely to be more influenced by HLD sources. The low bias here, especially that during the autumn, can be related to missing local emissions from the coast of southern Alaska (Fig. 1a) that occurs most frequently in autumn (Crusius et al., 2011). An underestimation of the transport from Saharan dust may also contribute slightly, as the influence from Saharan dust is found during mid-May at Trapper Creek (Breider et al., 2014).

The simulated Arctic dust vertical profiles are also compared with the measured dust concentrations during the Arctic Research of the Composition of the Troposphere from Aircraft and Satellites (ARCTAS) flight campaign (Fig. 4) (Jacob et al., 2010). The ARCTAS campaign was conducted over the North American Arctic in April and July 2008. The simulated profiles are averaged over the regions where the aircraft flew, in accordance with Groot Zwaaftink et al. (2016). In April, the model does a good job in capturing the Arctic dust vertical profiles (Fig. 4a). However, in July, the model underestimates dust by a factor of 2 to 5 between 3 and 10 km (Fig. 4b). It also shows an overestimation near the surface in July, which agrees with the surface concentration comparison at Alert station (Fig. 3b). The underestimation in the upper troposphere and overestimation near the surface likely imply a vertical transport of HLD that is too weak in the North American Arctic in summertime. The high bias in the upper troposphere may also be related to an underrepresentation of LLD transport.

Finally, we evaluate the simulated dust extinction against the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) retrieval (Luo et al., 2015a, b; Yang et al., 2022), which includes nighttime dust extinction for the period of 2007 to 2009. This dataset has improvements in dust separation from other aerosol types and thin dust layer detection in the Arctic compared to the standard Cloud–Aerosol Lidar with Orthogonal Polarization (CALIOP) level 2 product (Winker et al., 2013). To make an apples-to-apples comparison, the modeled dust extinction is sampled along the CALIPSO tracks and screened by cloud fraction (Wu et al., 2020). For this comparison, we only use the first 3 years (2007 to 2009) of the CTRL simulation to be consistent with the observation period. Overall, the model does a good job in capturing the Arctic dust extinction vertical profiles (Fig. 5). We notice that the simulated dust extinction is lower than CALIPSO retrievals at the upper troposphere in summer, which agrees with the ARCTAS comparisons. The simulated dust extinction also shows a consistent underestimation in springtime (MAM) and a near-surface underestimation in wintertime (DJF). Since the Arctic is mostly covered by ice and snow in these two seasons, the impacts of HLD are expected to be limited and the low biases are most likely due to the underprediction of LLD transport. The near-surface underestimation in DJF may indicate an LLD transport that is too weak in the lower troposphere (e.g., the transport of dust emitted from Central Asia; see Fig. 7 and the corresponding discussions in Sect. 3.2). Moreover, the HLD has a large contribution in the lower troposphere in boreal summer and autumn, which is consistent with its strong emission at that time. In contrast, LLD plays a more dominant role in the upper troposphere, where African dust contributes the most in the springtime and East Asian dust has a larger contribution in the other seasons.

Table 2 summarizes the relative contributions from individual sources to the total Arctic dust burden. The transport pathways can be identified from the dust burden spatial distribution for each source in Fig. 6, while the relative contribution of each source to the total dust burden is shown in Fig. S2. We also calculate the regional burden efficiency for each source (Table S1), which is defined as the mean contribution to the Arctic dust column burden divided by the corresponding dust emission (H. Wang et al., 2014). This metric represents the sensitivity of Arctic dust loading to per-unit changes in dust emission from each source (i.e., the poleward transport efficiency of each source).

Our model results suggest that the HLD (Arc) is the largest contributor (30.7 %) to the annual mean Arctic dust burden among all the tagged sources. As shown in Figs. 6a and S2a, the local dust is confined within the high latitudes, with higher amounts and higher contributions to the total dust burden near the sources in northern Canada as well as Iceland and the coast of Greenland. The interior of the Greenland ice sheet, with its higher elevations, is more influenced by LLD from North Africa and East Asia than HLD (Fig. S2c and f). This is due to the weak vertical transport of local emissions in the Arctic (see more discussion below).

On the other hand, all LLD sources are responsible for 69.3 % of the dust loading in the Arctic, with considerable contributions from North Africa (24.2 %) and Asia (in total 44.2 %; EAs: 19.9 %, CAs: 12.8 %, MSA: 11.5 %), as well as minor contributions from NAm (0.9 %) and RoW (nearly 0). The North African dust is primarily transported westward to the Atlantic and southward to the Sahel, with a smaller fraction transported directly northward or northeastward across Eurasia to the Arctic (Fig. 6c; Shao et al., 2011). The westward trajectory can also bring dust to the Arctic through the Azores high (e.g., VanCuren et al., 2012), but this pathway is not clearly seen in Fig. 6c, likely due to the strong wet removal process over the North Atlantic. As evident by the low transport efficiency in Table S1, the significant contribution of the North African dust to the Arctic dust burden is mainly due to its massive emission. However, this is not the case for EAs. The East Asian dust is first lifted vertically by topography and convection (Shao et al., 2011) and is widely spread over the Northern Hemisphere midlatitude and high-latitude regions through the westerly flow in the upper troposphere (Fig. 6f). The high elevation of East Asian dust plumes results in weaker removal processes and thus an efficient poleward transport. As shown in Table S1, the annual transport efficiency of the East Asian dust is relatively high among the LLD sources, which is nearly 3 times larger than that of the North African dust. The poleward transport of dust from CAs and MSA takes the pathway across Siberia (Fig. 6d and e). The transport efficiency of the CAs dust is 2 times higher than that of the MSA dust (Table S1). This is attributed to CAs being closer to the Arctic and having less southward dust transport than MSA. Overall, the LLD from North Africa and Asia contributes more to Eurasia and the Pacific sector of the Arctic (Fig. S2c to f). The impact of NAm dust is limited by its weak emission (Fig. 6b), while dust emitted in the Southern Hemisphere (RoW) can hardly pass the Equator (Fig. 6g).

Earlier modeling studies (Breider et al., 2014; Groot Zwaaftink et al., 2016; Luo et al., 2003; Tanaka and Chiba, 2006) also quantify the relative contributions of dust from various regions to the Arctic dust loading. Among these studies, only Groot Zwaaftink et al. (2016) include HLD. Our estimate about the HLD percent contribution is close to that from their study (27 %). For LLD, our conclusion about the dominant role of African and Asian dust in the Arctic dust burden is also corroborated by these previous studies. However, the relative importance of African and Asian dust is uncertain. Based on our results, the Asian dust is responsible for 65 % of the LLD transport to the Arctic, while the African dust only contributes 35 %. Other studies find that 50 % (Groot Zwaaftink et al., 2016; Luo et al., 2003; Tanaka and Chiba, 2006) to as much as 65 % (Breider et al., 2014) of the LLD in the Arctic is attributed to North Africa. These discrepancies may be explained by the different dust emission and scavenging, dust size distribution, meteorological fields, and/or time periods for the model simulation. For example, the wet removal process is expected to have large discrepancies among different models because of the large uncertainties in the model representation of clouds and precipitation. The different spatial distributions of dust emission due to the use of different emission parameterizations may also contribute to the discrepancies (e.g., North Africa dust in our study contributes slightly less at 51.9 % to the global dust emission than the other studies: from 57 % to 67 %). Isotopic analysis (Bory et al., 2002, 2003) and case studies (Huang et al., 2015; Stone et al., 2005; VanCuren et al., 2012) have proven that both Asian and African dust can be transported to the Arctic. However, it remains unclear which of them contributes more to the Arctic dust loading due to limited observational constraints.

HLD and LLD source regions also have very distinct vertical distributions in the Arctic. Figure 7a and b show the annual mean vertical profiles of Arctic dust concentrations from various sources and their percentage contributions, respectively. The Arctic dust in the lower atmosphere is dominated by the local source. HLD accounts for more than 30 % of the Arctic dust concentrations below 800 hPa, with up to 85 % contribution near the surface. However, the HLD contribution decreases rapidly with height and is less than 10 % above 700 hPa. This is because the lower troposphere of the Arctic is more stratified than the middle and low latitudes, which suppresses the vertical transport of HLD. The lower tropospheric stability (LTS) from the CTRL simulation and comparison with the MERRA-2 reanalysis data are shown in Fig. S3. The weak HLD vertical transport in the Arctic is also reported by previous studies (Groot Zwaaftink et al., 2016; Baddock et al., 2017; Bullard, 2017). Moreover, the LTS over the Arctic sea ice is much larger than that over open-ocean surface (Schweiger et al., 2008), which may lead to a stronger vertical transport of HLD over open waters. This suggests that the vertical transport of HLD may change with the sea ice reduction in a warming future.

In contrast, LLD has a higher contribution in the middle and upper troposphere than near the surface. Such a vertical distribution of LLD is consistent with Stohl (2006) and Groot Zwaaftink et al. (2016). As Stohl (2006) found, aerosols originating from the warm subtropics are transported poleward following the uplifted isentropes and the Arctic lower atmosphere is dominated by the nearly impenetrable cold polar dome. Therefore, there is a slantwise lifting of low-latitude aerosols during their poleward transport. NAf and EAs are the two key contributors to the Arctic dust vertical concentrations, each of which contributes up to one-third of the total dust concentrations above 700 hPa. Dust emission from MSA also has a moderate contribution (15 %–20 %) that increases gradually with height, while the contribution from CAs peaks at 700 to 800 hPa, indicating a lower-altitude transport pathway than the EAs and MSA dust.

In addition, the Arctic dust undergoes a strong seasonal cycle (Table 2 and Fig. 7c–j). Because of the strong local emissions (Fig. 1c), about half of the Arctic dust burden in summer and autumn comes from HLD, with more than 50 % contribution of Arctic dust concentrations below 850 hPa in these two seasons. In contrast, LLD plays a dominant role in spring and winter. The North African dust has the largest contribution in spring, which accounts for about 45 % of the total dust concentrations above 700 hPa. The East Asian dust is more important in the other three seasons. Due to its high emission height, the relative contribution from EAs tends to increase with height and reaches 30 % to 50 % of the total dust concentration above 500 hPa in summer, spring, and winter.

We are particularly interested in the contribution of various dust sources to the Arctic INP populations. Therefore, we compare the simulated INP concentrations with nine Arctic field measurements, which are summarized in Table 3. The modeled dust INP concentrations are diagnosed from monthly averaged aerosol properties using the default CNT scheme and two empirical ice nucleation parameterizations: DeMott et al. (2015; hereafter D15) and Sanchez-Marroquin et al. (2020; hereafter SM20). The D15 parameterization, which is representative of Saharan and Asian desert dust, relates dust INP number concentrations to the number concentration of dust particles larger than 0.5 µm diameter and is found to produce the most reasonable LLD INP concentrations in EAMv1 (Shi and Liu, 2019). CNT and D15 are applied to LLD only and all the dust aerosols (LLD and HLD) in Fig. 8a–b and d–e, respectively. The SM20 parameterization, which is derived for the HLD Icelandic dust, describes the dust INP number concentrations as a function of surface-active site density and total dust surface area. Considering the possibly different ice nucleation ability between HLD and LLD, we only applied the SM20 parameterization to HLD, and the CNT and D15 parameterizations are still applied to LLD in Fig. 8c and f, respectively. To account for the contributions from other aerosol types, we also calculate the INP concentrations from BC (Fig. 8g) and sea spray aerosol (SSA; includes MOA and sea salt) (Fig. 8h) following Schill et al. (2020; hereafter Sc20) and McCluskey et al. (2018; hereafter M18), respectively. More details about the ice nucleation parameterizations are provided in Sect. S2. We discuss the choice of dust ice nucleation schemes in Sect. S2.6 in the Supplement.

Overall, only including LLD as INPs results in up to 4 orders of magnitude underprediction compared to observations (Fig. 8a and d), while taking into account the contribution from HLD greatly improves the model performance by increasing the simulated dust INP concentrations (Fig. 8b, c, e, and f). The CNT parameterization produces 5 to 10 times more INP concentrations than the other two schemes at moderately cold temperatures (-22 to -28∘), while it has a significant underestimation of observed INP concentrations at warm temperatures (T > -18∘) (also see Fig. S4). D15 and SM20 agree well with each other in simulating HLD INPs, with SM20 producing slightly higher results than D15. Our modeling results also indicate that BC and SSA have much less of a contribution to INP than dust in all nine field campaigns (Fig. 8g and h).

A detailed analysis of sources of the INPs for the nine datasets based on modeling analyses and the corresponding observations in the literature is provided in Table 3. Modeling results indicate that HLD has larger contributions to the INPs for the campaigns conducted in summer and autumn than spring, in agreement with the observations. Also, ground-based measurements are more influenced by the nearby HLD sources, while LLD from EAs and NAf contributes more to the aircraft measurements.

Our modeling analyses about the INP sources agree well with the observational studies at Alert in spring 2016 and near Iceland in autumn 2014 (symbols C and I in Fig. 8, respectively), while the model underestimates the observed INP concentrations in both cases. The low bias in dataset C indicates an underprediction in the long-range transport of Asian dust to the Arctic surface in springtime. The underestimation in dataset I is more likely due to the fact that some of the aircraft measurements were taken inside the Icelandic dust plumes (Sanchez-Marroquin et al., 2020), which cannot be resolved by the monthly mean model output and the coarse model horizontal resolution (1∘). Such uncertainties exist in all the model–observation comparisons.

Some other comparisons in INP sources reveal the lack of marine and carbonaceous INPs in the model. The model results show a dominance of dust INPs in spring 2017 at Zeppelin and Oliktok Point (symbols D and E in Fig. 8) and in autumn 2004 at Utqiaġvik (symbol H in Fig. 8), while the observational studies suggested the importance of marine sources at the first two locations and of carbonaceous aerosols at Utqiaġvik. Therefore, it is likely that the model underestimates the contribution of MOA (Wilson et al., 2015; Zhao et al., 2021a) and does not account for terrestrial biogenic INPs (Creamean et al., 2020) due to the lack of treatments in the model. In addition, both D15 and SM20 schemes cannot represent the high ice-nucleating ability of HLD at warm temperatures at Zeppelin in summer 2016 (symbol G in Fig. 8), which is attributed to soil organic matter by Tobo et al. (2019). When these organics are taken into account in the model, model overestimation for site G will get even worse, implying an overestimation of surface dust concentrations and/or HLD emission at Svalbard in the summertime. In summary, the model's INP biases in the Arctic are likely due to biases in the simulated aerosol fields (e.g., dust, MOA, and BC) and uncertainties in current ice nucleation parameterizations or missing representations of other INP sources (e.g., terrestrial biogenic aerosols).

In addition, we do not explicitly represent the potential ice nucleation ability differences in freshly emitted HLD and long-range-transported LLD caused by aging and the coatings of pollutants (Kulkarni et al., 2014; Boose et al., 2016). However, D15 and SM20 may already take the aging effect into account implicitly. D15 is based on the Saharan and Asian dust data collected over the Pacific Ocean basin and US Virgin Islands, respectively, which are far away from the corresponding LLD sources, while SM20 is derived from the freshly emitted Icelandic HLD, which is subjected to less of an aging effect.

The comparisons above are based on INP concentrations at a given temperature set by the INP instruments, which reflects the potential INP populations under ambient aerosol conditions. Next, we examine the immersion freezing rate of dust originating from the seven tagged sources (Fig. 9) to evaluate the influences of HLD and LLD on ice nucleation processes in mixed-phase clouds. It is noted that the immersion freezing rate here is calculated online in the model using the ambient temperature and the default CNT ice nucleation parameterization.

Compared with its contribution to the dust burdens, the contribution of the HLD to the annual mean mixed-phase cloud immersion freezing rate is relatively small (∼ 10 % below 600 hPa) (Fig. 9a). This is because the HLD is mainly located in the lower troposphere and not a lot of HLD can reach the mixed-phase cloud levels (or the freezing level), especially in the case that the HLD tends to be more prevalent in the warm seasons (see more discussion below). Among the LLD sources, North African dust (Fig. 9c) and East Asian dust (Fig. 9f) are the two major contributors, both of which are responsible for more than 20 % of the annual mean immersion freezing rate in the mixed-phase clouds. Consistent with the vertical distribution of dust concentrations, the North African dust has its maximum contribution (30 %–40 %) at around 500 hPa, while the East Asian dust plays a more important role at higher altitudes (above 400 hPa). Dust from Central Asia also has a moderate contribution (∼ 20 %) to the immersion freezing rate in the Arctic (Fig. 9d).

Considering the different seasonality of HLD and LLD in the Arctic, we next investigate the seasonal variations of the immersion freezing rate in the Arctic mixed-phase clouds from HLD and two dominating LLD sources (NAf and EAs) (Fig. 10). HLD has the largest contribution to the Arctic immersion freezing rate in boreal autumn, with more than 30 % below 700 hPa and up to 50 % near the surface (Fig. 10c). It is related to the prevalence of HLD and relatively cold temperatures during this time in the Arctic. This is not the case for the summer, when the freezing level is relatively high. Although it is responsible for 50 % of the total Arctic dust burden in the boreal summer, HLD has a limited contribution to the immersion freezing rate in the clouds (Fig. 10b) because its weak vertical transport makes it hard to reach the freezing line. The contrasting results in summer and autumn suggest that the immersion freezing rate in the Arctic clouds is influenced by air temperature in addition to the aerosols. It also implies that the surface INP measurements may not reflect the complete picture of INP effects, and more aircraft INP measurements are needed in the future. The seasonal variations of the immersion freezing rate from NAf and EAs are weaker than that from HLD but are still subjected to the vertical temperature change with season. The North African dust contributes more in spring and winter, while the East Asian dust is more important in summer and autumn.

Dust INPs can freeze the supercooled liquid droplets, which impacts the cloud microphysical and macrophysical properties and modulates the Earth's radiative balance. To examine such impacts, we conduct three sensitivity experiments that turn off the heterogeneous ice nucleation in the mixed-phase clouds by dust from Arctic local sources, North Africa, and East Asia, respectively (i.e., noArc, noNAf, and noEAs in Table 1). The impacts of dust INPs from each source are determined by subtracting the respective sensitivity experiment from CTRL. Due to the feedbacks in dust emission and wet scavenging caused by changing cloud properties, the dust concentrations in the sensitivity experiments are not identical to CTRL, but the absolute differences are mostly within 5 % (Fig. S5 in the Supplement).

The cloud liquid and ice changes caused by dust INPs from each source are shown in Fig. 11. Due to the strengthening of heterogeneous ice nucleation processes, INPs from all three sources consistently reduce the total liquid mass mixing ratio (TLIQ) (Fig. 11, first column) and cloud liquid droplet number concentration (NUMLIQ) (Fig. 11, third column). The influence of HLD is mainly in the lower troposphere (Fig. 11, top row), and the influence of LLD extends to higher altitudes (Fig. 11, bottom two rows). Moreover, the cloud ice number concentration (NUMICE) decreases in the upper troposphere (Fig. 11, fourth column), likely due to fewer cloud droplets available for the homogeneous freezing in cirrus cloud after introducing dust INPs in the mixed-phase clouds. With fewer ice crystals falling from the cirrus clouds to the mixed-phase clouds, the WBF process in the mixed-phase clouds is inhibited (Fig. S6). Other ice-phase processes such as the accretion of cloud water by snow and the growth of ice crystals by vapor deposition also become less efficient, which decreases the total ice mass mixing ratio (TICE) above 600–700 hPa of altitude (Fig. 11, second column). TICE in the lower troposphere is increased because of immersion freezing and snow sedimentation from above.

Since liquid water path (LWP) is found to play a critical role in the Arctic radiative budget (e.g., Dong et al., 2010; Hofer et al., 2019; Shupe and Intrieri, 2004), we further investigate the seasonal variations of LWP changes caused by dust INPs from the three sources (Fig. 12). Corroborated by their large contribution to the immersion freezing rate during this time (Fig. 10c), HLD INPs produce the strongest LWP decrease (-1.3 g m-2) in boreal autumn (Fig. 12c), especially over northern Canada and Greenland. The influence of LLD INPs on LWP peaks in spring and winter. North African dust tends to have a larger impact on northern Eurasia, while East Asian dust impacts the western Arctic more.

Dust INPs from the three sources consistently increase (decrease) the annual mean downwelling shortwave (longwave) radiative flux (FSDS and FLDS) at the surface (Fig. 13, left and middle columns). This is mainly due to the LWP decrease, which reduces the cloud albedo and longwave cloud emissivity. For HLD INPs, the FLDS reduction dominates over the FSDS increase and causes a net cooling effect at the Arctic surface (-0.24 W m-2) (Fig. 13c). In contrast, FSDS and FLDS changes related to the LLD INPs are comparable, which cancel each other and yield a small net radiative effect (0.08 W m-2 for NAf and -0.06 W m-2 for EAs) (Fig. 13, bottom two rows). These differences in the net radiative effect are associated with different seasonalities of HLD and LLD. The insolation in the Arctic is strong in spring and summer but very limited in autumn and winter. Since the HLD INPs have a much stronger influence on LWP in autumn and winter than spring and summer (Fig. 12), their contribution to the FSDS warming is weak, and the FLDS cooling in autumn and winter dominates the annual mean effect (Table 4a; also seen in Figs. S7 to S9). LLD INPs are also important in spring and summer, so their FSDS warming effect is comparable to, and compensates for, the FLDS cooling effect.

We also examined the dust INP effect on cloud radiative forcing (CRF) at the top of the atmosphere (TOA) (Table 4b). Dust INPs from the three sources induce a small net cooling (from -0.03 to -0.05 W m-2) in the Arctic, with SW warming and LW cooling effects. The net cooling persists throughout the year, except for the summertime when the sufficient insolation results in a strong SW warming and, consequently, a net warming effect. Shi and Liu (2019) also found that LLD can induce a general net cooling effect above 70∘ N (0.18 to -1.95 W m-2), but at a much higher magnitude than the sum of NAf and EAs dust INP effects (-0.15 W m-2 above 70∘ N, not shown in Table 4), which implies that the aerosol glaciation effect on mixed-phase clouds is highly nonlinear.

Finally, we evaluate the model performance in simulating the Arctic LWP and radiative fluxes against the Moderate Resolution Imaging Spectroradiometer (MODIS) LWP and the Clouds and the Earth's Radiant Energy System Energy Balanced and Filled Edition 4.1 (CERES-EBAF Ed4.1) products (Loeb et al., 2018; Kato et al., 2018), respectively (Fig. 14). Two MODIS datasets are used, including the standard Collection 6.1 product (Pincus et al., 2012 – P12; Khanal et al., 2020 – K20). The P12 product combines MODIS observations from Terra and Aqua and is designed for apples-to-apples comparisons with modeling results from the Cloud Feedback Model Intercomparison Project (CFMIP) Observation Simulator Package (COSP). The standard product has a well-known positive zonal bias near the poles that is strongly correlated with the solar zenith angle (SZA). The K20 product largely reduces this bias by utilizing the SZA and cloud heterogeneity index in the retrieval algorithm. The MODIS simulator is used to calculate the simulated LWP. According to Fig. 14, the simulated LWP from the four experiments is lower than P12 but higher than K20. All four experiments also underestimate FSDS with shortwave cloud radiative forcing (SWCRF) that is too strong and overestimate FLDS with longwave cloud radiative forcing (LWCRF) that is too strong, which likely points to the biases of modeled clouds (e.g., LWP that is too large compared to K20). The differences among the model experiments are very small compared to their discrepancies with observations. We notice that including dust INPs from the three sources decreases the simulated LWP (i.e., CTRL has smaller LWP than the other experiments) (Fig. 14a), which makes the model performance better if compared to K20. Moreover, it shows noticeable improvements in simulating both surface and TOA radiative fluxes after including dust INPs from each of the three sources (i.e., the results from CTRL are closer to the CERES results than the other three experiments) (Fig. 14b–e).

Overall, including HLD or LLD INPs does not contribute a lot to the reduction of biases in simulating the LWP and radiative fluxes in the AMPCs. However, the representation of AMPCs in global climate models is associated with multiple cloud macrophysical and microphysical processes, as well as large-scale dynamics (Morrison et al., 2012) (see more discussion in Sect. 4), which interact with one another nonlinearly. Therefore, even though including HLD or LLD INPs does not improve the representation of AMPCs significantly in our model, a good representation of dust INPs, especially including HLD INPs, could still be of great importance for parameterizing AMPCs in the model.