The process of wind-blown dust emission is controlled by a number of environmental variables, including wind speed, soil texture, land use type, vegetation cover, and soil moisture. Dust deflation is favored by dry soil with low and sparse vegetation and constrained by high soil moisture. The dust emission scheme employed in the CMAQ is developed by Tong et al. (2016). The emission (vertical flux) of the dust F (g m-2 s-1) was estimated based on a modified Owen's equation (Owen, 1964): F=∑i=1M∑j=1NK×A×ρg×Si×SEP×u∗×(u∗2-u∗ti,j2)foru∗>u∗ti,j, where M is the erodible land use type, N is the soil texture type, and K is the ratio of vertical to horizontal flux calculated based on the amount of clay (clay %) within the soil: K=100.134×clay%-6,when:clay%<20%0.0002,when:clay%>20%.

The value of K (0.0002) for clay percentage > 20 % is used here following the recommendation of Marticorena and Bergametti (1995). A is a scaling factor (set as 32.0 in this study), ρ is air density, g is gravitational acceleration (9.8 m s-2), Si is dust source area for land type i, and SEP is the soil erodibility factor, which is calculated based on the fraction of clay, silt, and sand of the soil as SEP=0.08×clay+1.0×silt+0.12×sand. u∗ is the friction velocity, and u∗ti,j is the threshold friction velocity for soil type j and land use type i. More details of the dust emission algorithm have been given in Tong et al. (2016). Equation (1) is applied only when the model-calculated friction velocity exceeds the designated threshold value. Therefore, the value of threshold friction velocity is critical to determine the onset and magnitude of dust emission in the CMAQ model.

In the CMAQ dust module, the threshold friction velocity is dynamically calculated based on the presence of non-erodible elements and the change in soil moisture (Tong et al., 2016). The effect of non-erodible elements is represented by wind energy partitioning following Marticorena et al. (1997). The effect of soil moisture on dust emission is implemented following a two-step approach proposed by Fécan et al. (1999). First, the maximum water holding capacity (Wmax⁡) for each soil type is determined based on the amount of clay in the soil: Wmax⁡=(0.0014×clay+0.17)×clay.

In the case of soil moisture exceeding Wmax⁡, the threshold friction velocity is then adjusted (Marticorena et al., 1997) as u∗ti,j=u∗ci,j×Zi,j×fsoilmi,j, where u∗ci,j is the initial threshold friction velocity constant, and Zi,j is the surface roughness adjusting factor calculated with surface roughness length from the meteorology field, calculated as Z=C1C1×ln⁡Zruf-C2.

The constants C1=32 and C2=-5 used in this equation are derived from field measurement data from Gillette et al. (1980) and the relationship between roughness length and friction velocity described in Marticorena et al. (1997).

And fsoilmi,j is the moisture adjustment factor calculated using a revised Fécan formulation (Fécan et al., 1999): fsoilmi,j=1.0,forSm≤Wmax⁡1.0+1.21×Sm-Wmax⁡0.680.5,forWmax⁡≤Sm<Sl where Sm is soil moisture andSl is the saturation soil moisture limit determined by soil textures. The values of Sl are given by the North American Mesoscale (NAM) modeling system (Rogers et al., 2009), as summarized in Table 1.

Previously, the values of initial threshold friction velocity constant were taken from observed data from wind tunnel experiments conducted by Gillette and coworkers (Gillette et al., 1980, 1982). Fu et al. (2014) reported the u∗ci,j used in the CMAQ has an average value of 0.7 m s-1 among all soil types, which is too high to generate enough dust particles over East Asia. Fu et al. (2014) used a fixed value of 0.3 m s-1 based on a study of local measurements in a northern desert in China (Li et al., 2007). Although this smaller threshold helps generate a higher production of dust emission during the 6-day simulation episode from 1 to 6 May 2011, the arbitrarily designated threshold value for all land covers and soil categories prevents the model from reproducing spatial and temporal variations of dust emission. In this study, we have conducted a reanalysis of the field data (Gillette et al., 1980, 1982) and revised the threshold friction velocities through removing the double counting of soil moisture in the CMAQ dust emission modeling. In these field campaigns, some of these experiments were performed under rather dry conditions, but for other samples the effect of soil moisture cannot be ignored. Therefore, these values reported from field experiments are not always suitable for being used directly as the initial threshold friction velocity constant, which is assumed to represent extremely dry conditions. Meanwhile, in the CMAQ dust module, dynamic soil moisture data are used to adjust threshold friction velocity. Therefore, we need to convert the wet-condition data into threshold values under dry conditions. Otherwise, there will be the double counting of soil moisture in some cases.

We took a three-step approach to calibrate the threshold friction velocity. First, the value of soil moisture was extracted for each sample from the raw field dataset (courtesy from Dale Gillette, retired). Next, these data were used to feed the Fécan formula (Fécan et al., 1999) to derive the dry-condition threshold velocity by casting soil moisture back to zero. We have reprocessed the data for the soil and land use types measured by Gillette (1980, 1982). In the case of missing data for certain soil types, we have chosen the values with the soil composition closest to a measured type following the USDA soil composition diagram (Fig. 1 of Gillette et al., 1980). Finally, the revised values of u∗ci,j are implemented in the CMAQ. To our knowledge the double-counting issue affects the dust emission module in CMAQ only because the microphysical parameterization is developed based on Gillette et al. (1980) data. It also explains the significant underestimations of dusty by CMAQ as reported in Appel et al. (2013) and Fu et al. (2014). The comparison of the default and revised constants is summarized in Fig. 1c. As the double-count of soil moisture has been corrected, the revised constants are lower than the default ones. The majority of land cover in the Gobi is categorized as shrubland, where the revised initial threshold friction velocity constants are significantly lower than the default values for all soil types as shown in Fig. 1c, indicating that the revised scheme is expected to produce higher dust emissions over the Gobi. The Taklamakan Desert is mainly configured as barren or sparsely vegetated land cover with a sandy soil type, which only shows a small drop in the threshold friction velocity constant from 0.28 to 0.23 m s-1, indicating that the changes in dust emission over the Taklamakan may not be substantial. The CMAQ distributes dust emission to four size bins: 0.1–1.0, 1.0–2.5, 2.5–5.0, and 5.0–10.0 µm, with the mass contribution as 3, 17, 41, and 39 % within each bin, respectively. The first two bins represent the fine-mode aerosol and the larger two represent the coarse mode. So the mass contribution is 20 % for fine-mode and 80 % for coarse-mode aerosol.

The emission of natural wind-blown dust particles is distributed to 19 aerosol species in the CMAQ, following the profile developed based on the EPA's SPECIATE database (Simon et al., 2010). As compared with other models that treat dust as a unique aerosol species, the CMAQ approach provides a more detailed description of dust mineralogy. This feature benefits CMAQ by enabling the model to maintain the consistency between dust and other types of existing aerosol in the modeling system. However, mass contributions of different chemical components may differ greatly among different source areas; thus, using a fixed profile for all dust sources may introduce uncertainty and lose the capacity to model the varieties of dust particles. The mass contribution of aluminum (Al) is around 5–8 % for pure minerals around the world, and the ratios between other trace metals and Al may vary substantially for different dust particles. Thus, the elemental mass ratio between calcium and aluminum (Ca / Al) is usually used to identify the source of dust samples (Huang et al., 2010; Sun et al., 2005). For example, the Ca / Al ratio for Saharan dust is around 0.9 and 1.0 for fine and coarse dust particles, respectively (Blanco et al., 2003; Formenti et al., 2003; Kandler et al., 2007; Reid et al., 2003); for Arabian dust, it is around 0.13 and 0.15 for coastal and inland dust, respectively (Krueger et al., 2004); for Taklamakan dust, it is about 1.5–1.9 (Huang et al., 2010); and for dust from the Gobi Desert, it is 0.4–1.1 (Arimoto et al., 2006; Zhang et al., 2003). To characterize the dust aerosols in the CMAQ better, source-dependent speciation profiles are developed in this study for the Gobi and Taklamakan deserts based on local measurement data collected by Huang et al. (2010). These two profiles are compared with the default one in the CMAQ as shown in Table 2. For the model species which are not measured in Huang et al. (2010), including primary organic carbons (APOC), non-carbon aerosols (APNCOM), elementary carbons (EC), silicon (ASI), and water (AH2O), their values for the Taklamakan and Gobi are kept the same as in the default profile. And for unspeciated (AOTHR) and non-anion dust (ASOIL), their values in the two new profiles are calculated based on the contributions of all other species, to keep the total mass contributions conservative. It is important to note that the model species refer to an anion or cation phase for sulfate (ASO4, SO42-), nitrate (ANO3, NO3-), chloride (ACL, Cl-), ammonium (ANH4, NH4+), sodium (ANA, Na+), Ca2+ (ACA), magnesium (AMG, Mg2+), and potassium (AK, K+) and an element phase for iron (AFE, Fe), Al(AAL), silicon (ASI, Si), titanium (ATI, Ti), and manganese (AMN, Mn). Mass contributions of different aerosols differ significantly among the profiles as shown in Table 2. For example, Ca2+ accounts for 7.94 % of the total fine-particle mass in the default profile, which is much higher than the Taklamakan (2.063 %) and the Gobi (1.788 %). For Mg2+, the default profile assumes 0 % of mass contribution, yet the values for the Taklamakan and Gobi are 0.165 and 0.799 %, respectively. The K+ contribution within the default profile is 3.77 %, while for the Taklamakan it is 0.153 % and for the Gobi it is 0.282 %. Si is one of the most abundant metals in the crust, yet the default speciation profile has an inappropriate assumption of zero Si content in coarse-mode dust particles. As no measurements are found for Si over the Taklamakan or Gobi, we use the element ratio of Al / Si as 8 % / 28 % to derive the mass contribution of Si in the coarse model, which is a conventional approach for trace metal analysis (Huang et al., 2010). Different configurations within the speciation profile will lead to significant differences in model predictions for these trace metals, demonstrated in more detail in Sect. 3.

The default heterogeneous chemistry scheme within the CMAQ considered the conversions from N2O5 to HNO3 and from NO2 to HONO and HNO3. These reactions play an important role in the nighttime production of nitrate aerosols (Dong et al., 2014; Pathak et al., 2011; Pun and Seigneur, 2001). Heterogeneous reactions are treated as irreversible in the model (Davis et al., 2008; Sarwar et al., 2008; Vogel et al., 2003). While dust particles serve as a platform for heterogeneous reaction, they also participate in some of the reactions to take up the gas-phase species. The uptake of gases onto the surface of dust particles is defined by a pseudo-first-order reaction rate K (Dentener et al., 1996; Heikes and Thompson, 1983) calculated as K=rpDg+4vgγg-1Ap, where rp is the radius of the particle, Dg is the diffusion coefficient of gas molecules, vg is the mean molecular velocity of gas, Ap is the surface area of the particle, and γg is the uptake coefficient for gas. Many research efforts have been devoted to quantifying the uptake coefficients. The reported values of the uptake coefficient may differ by more than 2–3 orders of magnitude, depending on the source of the dust samples and analytical methods (Cwiertny et al., 2008; Usher et al., 2003). While this work focuses on East Asia, the uptake coefficients are mainly collected from Zhu et al. (2010), which summarized the estimations for dust samples from deserts in China. The “best guess” regarding uptake coefficients was suggested based on the analysis of different measurement studies summarized in Zhu and Zhang (2010). But in this study both the lower and upper limits of uptake coefficients are examined to evaluate their impacts. Table 3 lists the 13 dust heterogeneous reactions implemented in the CMAQ in this study, with the lower and upper boundaries of uptake coefficients.

The CMAQ model simulation uses version 5.0.1. In this study, the CMAQ is configured with the updated 2005 carbon bond gas-phase mechanism (CB05), aerosol module AE6, in-line photolysis calculation and NO emission from lightning, and the Euler backward iterative (EBI) solver. The modeling domain covers East Asia and peninsular Southeast Asia as shown in Fig. 2. The CMAQ simulation is performed with a 36 km horizontal grid spacing and 34 vertical layers with a model top at 50 hPa and finer resolution at the near-surface layers to represent the planetary boundary layer clearly. The simulation period covers March and April from 2006 to 2010 to represent the spring dust episode of East Asia.

The meteorology field is simulated with the Weather Research and Forecasting model (WRFv3.4, Skamarock et al., 2008). Initial and boundary conditions are downscaled from the GEOS-Chem global model simulation for 2006–2010 by following the routines described in Lam and Fu (2009). Biogenic emission is from MEGAN2.1 (Guenther et al., 2006; Müller et al., 2008), biomass burning emission is from FLAMBE (Reid et al., 2009), and anthropogenic emission is from Zhao et al. (2013) over China and INTEX-B over other countries within the domain. More details about meteorology and emission datasets are given in Dong and Fu (2015a, b).

To examine the performance of the CMAQ model development with revised parameterization and dust heterogeneous reactions, a total of six scenarios is conducted as listed in Table 4. The simulations Dust_Off and Dust_Default are designed to investigate the performance of the CMAQ without dust emission and with the default dust plume rise scheme; Dust_Revised is designed to investigate the performance of applying the new parameterization of initial friction velocity threshold constants; Dust_Profile is designed to examine the improvement by applying source-dependent dust composition profiles; and Dust_Chem and Dust_ChemHigh are designed to examine the impacts of implementing heterogeneous chemistry with the lower and upper estimations of uptake coefficients, respectively. The scenario Dust_Profile, Dust_Chem, and Dust_ChemHigh are all performed based on Dust_Revised.

Both ground-based measurements and satellite observations are used in this study to help examine the uncertainty and evaluate the performance of the model. The Air Pollution Index (API, http://datacenter.mep.gov.cn) reported by the Chinese Ministry of Environmental Protection (MEP) is used to evaluate the PM10 predictions from the CMAQ. The API is reported on a daily basis with a national coverage of 86 middle size or larger cities in China and has been applied in many modeling studies for evaluation purposes (Zhao et al., 2013). To investigate the transport of dust particles over downwind areas simulated by the model, we also used surface observations from the Acid Deposition Monitoring Network in East Asia (EANET; EANET, 2007) over Japan, and observations from the Taiwan Air Quality Monitoring Network (TAQMN; http://taqm.epa.gov.tw/taqm/en/default.aspx). EANET provides hourly or daily records of PM10, O3, NOx, and SO2 and also biweekly (or longer-interval) records of HNO3, SO42-, and NO3-. Data from EANET is also employed in this study for examining model performance with dust heterogeneous chemistry. The TAQMN provides observations of most criteria air pollutants, yet only PM10 observations at the Xinzhuang site are used in this study to focus on dust storm impact. These variables are used to examine the simulation responses by implementing dust heterogeneous chemistry. Ground-based measurements of K+, Mg2+, Ca2+, and PM2.5 at Duolun and Yunlin from Fudan University's observation network (Huang et al., 2010) are used to investigate the model's capability to simulate tracer metals by applying the source-dependent speciation profile for suspended dust particles. AOD from the AErosol RObotic NETwork (AERONET; http://aeronet.gsfc.nasa.gov/) operated by NASA is also collected to evaluate the CMAQ predictions. To examine the spatial distribution and column density of model-simulated dust particles, we also used AOD products from the Moderate Resolution Imaging Spectroradiometer (MODIS; http://modis.gsfc.nasa.gov/). All observations are collected for the period of March and April from 2006 to 2010, except for the Fudan University data, which only cover March and April in 2007. The locations of API, EANET, TAQMN, and AERONET observational stations are indicated in Fig. 2, along with the locations of Duolun and Yunlin. Table 5 summarizes detailed information on each observation network.

To examine model improvement by implementing new initial threshold friction velocity constants, simulation results from Dust_Default and Dust_Revised are compared for the spatial distribution of PM10 and evaluation bias against API. Dust_Default – Dust_Off and Dust_Revised – Dust_Off represent the PM10 from the default dust scheme and the revised dust scheme, respectively, as shown in Fig. 3a and b. PM10 concentrations generated from dust are averaged for March and April from 2006 to 2010. Figure 3a indicates that the default dust scheme produces a very limited amount of PM10 only over the Gobi Desert (less than 70 µg m-3), which can hardly represent the East Asian dust storms. The revised scheme produces higher PM10 concentrations of more than 400 µg m-3 in the Gobi source region, as shown in Fig. 3b. Dust plumes are generated by the revised model over the Gobi and Taklamakan deserts and also in sparse grassland over the northwest region of the Tibetan Plateau. Particles from dust plumes are transported southeastward and contribute 50–100 µg m-3 of PM10 over northern and eastern China and less than 50 µg m-3 over southern China, South Korea, and Japan. Huang et al. (2010) demonstrated that there are two transport pathways for Asian dust: plumes from the Gobi and Taklamakan are either pushed by prevailing winds eastward towards South Korea and Japan or southeastward down towards southern China and Taiwan. With the revised dust scheme, the CMAQ generally reproduces the spatial distribution of Asian dust as shown in Fig. 3b, which indicates the most significant impact over northern and eastern China and relatively weak impacts over downwind areas along the transport pathway towards the southeast. Figure 3c and d demonstrate the evaluation bias at the API cities over China for the Dust_Default and Dust_Revised scenarios, respectively. With the default dust scheme, the CMAQ shows a large negative bias for the entire domain. The most serious underestimation is found over northern and western China, with a negative bias of more than -80 µg m-3. Figure 3c also suggests that cities closer to the Taklamakan and Gobi deserts have a larger negative bias, indicating that the default scheme cannot generate sufficient dust emission to reproduce the observed PM10. With the revised scheme as shown in Fig. 3d, simulation biases for most of the cities are reduced down to ±20 µg m-3. The largest overestimation is found at Hohhot (39 µg m-3), and the largest underestimation is found at Xining and Lanzhou (-60 µg m-3). Figure 3d also indicates that the model tends to have a larger simulation discrepancy in cities close to the desert.

Figure 4a to c summarize the evaluation statistics for simulated PM10 against observations from the API. Daily data pairs from observation and simulation are used here to calculate the statistics. The Dust_Off scenario underestimates PM10 concentration by -56.74 %. Since dust storms from the Taklamakan and Gobi substantially contribute to suspended particle concentrations over East Asia during spring, a simulation with no dust emission should be responsible for the large underestimation. With the default dust emission module, model performance for the scenario Dust_Default is only slightly improved and PM10 is still underestimated by -55.42 %, as shown by Fig. 4b. The comparison between Dust_Off and Dust_Default suggests that the default dust module is not able to generate sufficient elevated particles to match the observed PM10 levels from the API. On the other hand, the Dust_Revised scenario shows a much better performance with an normalized mean bias (NMB) value of -16.06 %, as shown in Fig. 4c. Simulation results are also evaluated against AOD observations from the AERONET, with the statistics shown in Fig. 4d to f. Statistics for AOD evaluation against the AERONET suggest similar model performances as for PM10. The CMAQ simulation without dust rise underestimates AOD by -31.34 % at AERONET stations, and the default dust module offers almost no improvement for AOD prediction. The Dust_Revised scenario underestimates AOD by -22.1 %, indicating that the revised scheme is also able to improve the model performance for simulating fine-mode dust.

The speciation of dust particles determines the contributions of crust species and trace metal concentrations predicted by the model. As described in Sect. 2.2, we modified the fixed speciation profile within the CMAQ to be source-dependent for the Gobi and Taklamakan, and in this section ground-based observations collected at Duolun and Yulin are used to investigate the impacts by applying different speciation profiles. The model simulations from the Dust_Revised and Dust_Profile scenarios are compared with observations for K+, Mg2+, and Ca2+ with data pairs, as shown in Fig. 5 (simulations and observations for K+ and Mg2+ are upscaled by 5 and 10 time, respectively, to make them comparable with Ca2+ in the same figure), and evaluation statistics are summarized in Table 6. The two cities are close to the Gobi Desert, as shown by Fig. 2. Huang et al. (2010) used back trajectory and Ca / Al ratio analysis to demonstrate that the Gobi Desert is the main contributor of dust particles at Duolun and Yulin. Figure 5a indicates that with the default speciation profile, the CMAQ overestimates K+and Ca2+ by 208.9 and 36.69 %, respectively. Predicted Mg2+ concentration is almost zero because there is no emission of Mg2+ in the anthropogenic emission inventories, and the default profile indicated a 0 % mass contribution of Mg2+ from dust emission. With the source-dependent speciation profile, the CMAQ simulation of Mg2+ is increased significantly as shown in Fig. 5b. The revised CMAQ underestimates K+ and Ca2+ by -47.83 and -53.12 %, respectively. A consistent negative bias for trace metals should be due to the underestimation of total fine-mode aerosol within dust. Figure 5c shows the comparison between observed and simulated PM2.5 concentrations at Duolun and Yulin. The Dust_Revised and Dust_Profile scenarios only differ in terms of their speciation profile for particles. The predicted PM2.5 from the two scenarios are almost identical, so only Dust_Profile is shown in Fig. 5c. The concentration of PM2.5 is underestimated by -45.59 %, which is consistent with the underestimations of Mg2+, K+, and Ca2+. The evaluation statistics suggest that total PM2.5 is underestimated at Duolun and Yunlin, but Ca2+ is overestimated under the default profile. So although the NMB value changes from 36.69 % by the default profile to -53.12 % for Ca2+ by the revised profile, it is highly possible that the revised profile provides a better estimation of the calcium mass contribution since the evaluation statistics for trace metals should be consistent with that for total PM2.5. The overall underestimation of PM2.5 and trace metals is probably due to the underestimation of fine-mode aerosol mass contribution (20 %) to total dust emission. On the other hand, the simulation bias shown in Fig. 3d suggests a slight overestimation of PM10 at cities close to the Gobi Desert. So it is highly possible that the fine-particle mass contribution configured within CMAQ may be underestimated. Unfortunately, measurements made by Huang et al. (2010) only collected data for total suspended particles (TSPs) and no PM10 observation is found for Duolun or Yulin. So no reliable conclusion is drawn at this point because the PM10 data are not available at Duolun or Yulin, although the TSPs are measured and indicate that the fine-mode aerosol should have a larger mass contribution (about 40 % in TSP).

Dust heterogeneous chemistry involves the uptake of gas-phase species and the production of secondary inorganic aerosols. In this section we investigate the impacts of implementing dust chemistry in the CMAQ by examining the simulation difference between Dust_Chem and Dust_Profile and the difference between Dust_ChemHigh and Dust_Profile. Figure 6 shows the concentration changes (color contours represent the absolute concentration changes, and dashed lines represent the percentage changes) under heterogeneous chemistry with lower (left column) and upper (right column) limits of uptake coefficients for O3 (first row), SO2 (second row), SO42- (third row), HNO3 (fourth row), NOx (fifth row), and NO3- (sixth row). All variables are averaged for March and April from 2006 to 2010. The spatial distributions shown in Fig. 6 suggest that impacts of dust chemistry are more pronounced in eastern China over the downwind areas than in the dust source area near the deserts. This is because eastern China has intense anthropogenic emissions that help to accelerate the dust chemistry, while the desert areas in northern and western China have much lower concentrations of the reactive gases. Dong and Fu (2015a) reported that in spring, O3 concentration is around 30 ppbv and less than 5 ppbv for NO2 and SO2 over the Gobi, while the concentrations over eastern China are 50–60 ppbv for O3 and 10–40 ppbv for NO2 and SO2. A greater proportion of these reactive gases participates in the heterogeneous reactions on the surface of dust particles transported from deserts and thus lead to a more significant impact of dust chemistry over the downwind area instead of over the deserts. Li et al. (2012) also reported that the impact of dust chemistry for O3, SO2, and NO2 is less than 5 % over the Gobi but up to 30–40 % in eastern China and even higher over the western Pacific. O3concentration is reduced by 3–6 ppbv (2–10 %) and 5–11 ppbv (4–20 %) with the lower and upper limits of uptake coefficients, respectively, due to irreversible Reaction (R1) as listed in Table 3, which agrees well with the values reported by Tang et al. (2004; 20 % and Li et al. (2012; 5–20 %. Wang et al. (2012) reported lower O3 reduction due to dust chemistry by 3.8 and 7.3 ppbv with lower and upper uptake coefficients, respectively, which could be because of using a different simulation year (2001) with a lower baseline O3 over East Asia. The concentration of SO2 is reduced by ∼ 2 ppbv (10 %) and ∼ 6 ppbv (30 %) with lower and upper limits of uptake coefficients, respectively, due to the consumption in reaction R13, which also leads to the increase in SO42- concentration by ∼ 3 µg m-3 (8 %) and more than 5 µg m-3 (15 %) under Dust_Chem and Dust_ChemHigh, respectively. Impacts on SO2 reported by other studies differ moderately by a factor of 2 or more, varying from 55 % by Tang et al. (2004) as the highest, 10–20 % as the medium (Li et al., 2012), and 5–8 % as the lowest (Wang et al., 2012) impact. Different impacts caused by heterogeneous chemistry reported by different studies should mainly be attributed to the different simulation episodes. Tang et al. (2004) focused on dust episodes only in 2001 with lower baseline pollutants from anthropogenic emission; Li et al. (2012) also simulated dust episodes but in 2010; and Wang et al. (2012) simulated the year 2001 but focused on the entirety of April, when the monthly averages of particles are apparently smaller than the values from dust episodes only. Reaction (R7) indicates the consumption and Reaction (R11) indicates the production of HNO3, while the net effect of dust chemistry is found to decrease HNO3 concentration by 0.2–0.8 ppbv (8–30 %) as shown in Fig. 6e and f. Our result is comparable with the values reported by Li et al. (2012) as 5–40 % but smaller than that reported by Tang et al. (2004) as 30–70 %. Although the reaction R9 indicates the uptake of NOx by dust particles, simulation results suggest that NOx concentration is increased by 0.2–1 ppbv over eastern China and the west Pacific. The elevation of NOx concentration should be attributed to the depletion of O3 and the conversion of gas-phase HNO3 back to NOx (Yarwood et al., 2005). As a result of excessive SO42- production from dust chemistry, the concentration of NO3- is decreased under the Dust_Chem scenario due to the thermal–dynamic equilibrium between SO42--NH4+-NO3-. The equilibrium drives the inorganic aerosols to convert from NH4NO3 to (NH4)2SO4 over eastern China, with intensive anthropogenic SO2 and NOx emissions from industry and power sector but insufficient NH3 to neutralize all the acid anions. On the other hand, over the western Pacific and Japan, where SO2 and NOx emissions are less intensive, the concentration of NO3- is increased slightly due to the production indicated by Reactions (R8) and (R10). Consequently, over eastern and central China, removed NO3- evaporates back to HNO3, which again pushes the gas-phase equilibrium towards the production of NOx and thus leads to the increase in NOx but a decrease in HNO3 and NO3-. Meanwhile, with the upper limit of uptake coefficients, the production rate of HNO3 catches up with the removal rate of NO3-, which helps to slow down the decrease in NO3- over China and accelerate the increase in NO3- over the western Pacific and Japan. Our result is consistent with the findings from other studies. Wang et al. (2012) also reported the increase in NOx and decrease in HNO3 and NO3- concentrations due to dust chemistry over East Asia. Li et al. (2012) reported that NO3- concentration with a lower uptake coefficient is about 5 µg m-3 (30 %) lower than the base case simulation (with a “best guess” uptake coefficient suggested in Zhu et al., 2010) and NO3- predicted by high uptake coefficient is about 12 µg m-3 (100 %) higher than the base case at Shanghai and Xiamen.

The impact of dust chemistry shown in Fig. 6 suggests comparable results to other modeling assessments, but very few previous studies incorporated observation data to validate further the impact indicated by the model. To evaluate the model's capability to represent dust chemistry and also determine the best-fit uptake coefficients, observations from EANET are used to compared with simulations from Dust_Profile, Dust_Chem, and Dust_ChemHigh. Evaluation statistics are summarized in Table 7. Simulation results from different scenarios for O3 all agree well with observations, as indicated by the statistics. O3 is overpredicted by 1.26 % without dust chemistry in the CMAQ and underpredicted by -1.97 and -4.43 % with lower and upper uptake coefficients, respectively. SO2 is overpredicted in all scenarios, but the NMB value is reduced from 90.7 % without dust chemistry to 69.8 and 63.7 % with lower and upper uptake coefficients, respectively. Evaluation statistics for HNO3 and NOx show a similar response to SO2, where heterogeneous chemistry helps to reduce the large overestimations from 109.03 % without dust chemistry to 85.17 and 81.24 % with lower and upper limits of uptake coefficients, respectively. The positive bias for SO2 and HNO3 should be attributed to the overestimated anthropogenic emissions (Dong and Fu, 2015a; Wang et al., 2011). For NOx evaluation, however, model overestimation is increased from 35.61 % under the Dust_Profile scenario to 37.79 and 38.21 % under the Dust_Chem and Dust_ChemHigh scenarios, respectively. The overestimation of NOx emission should be responsible for the positive bias from the CMAQ as indicated by previous studies (Dong and Fu, 2015a), but implementing dust chemistry in the model leads to a larger overprediction of NOx. The concentration of SO42- is underpredicted by 16.28 % without dust chemistry, yet the simulation overpredicts SO42- by 13.74 and 29.43 % under the Dust_Chem and Dust_ChemHigh scenarios, respectively. For NO3- predictions, dust chemistry helps to reduce the underprediction from 13.07 % under the Dust_Profile scenario to -1.97 % under Dust_Chem scenario. But the simulation is boosted too much with the upper limit of coefficients and it overpredicts NO3- concentration by 24.09 % under the Dust_ChemHigh scenario. Note that EANET data are collected from Japanese sites so that Dust_Chem and Dust_ChemHigh show consistent increases in NO3-, as shown in Fig. 6. Statistics shown in Table 7 suggest that implementing heterogeneous chemistry seems able to improve the CMAQ performance for most of the species except O3 and NOx, but the lower limit of uptake coefficients favors the prediction of SO42- and NO3-, and the upper limit of uptake coefficients has a better prediction for SO42- and HNO3. Although these statistics show competitive performance between Dust_Chem and Dust_ChemHigh, the lower limit of the uptake coefficient may be more appropriate if we consider the uncertainty within the baseline anthropogenic emissions. With both surface observations and satellite retrievals, Dong and Fu (2015a) demonstrated that the CMAQ overpredicted NOx and SO2 over East Asia between 2006 and 2010 by around 30 and 20 %, respectively, due to overestimation in anthropogenic emissions. Wang et al. (2011) also report the overestimation of SO2 by 14 % over China. Implementing dust chemistry helps to reduce simulated concentrations of SO2, NOx, and HNO3, so it can balance part of the positive bias caused by anthropogenic emissions, but the statistics for SO42- and NO3- indicate that the counter effect caused by using the upper limit of uptake coefficients may be too excessive and push the balance towards an overestimation of aerosols as a side effect. Consequently, without explicitly excluding the bias within anthropogenic emissions, no reliable conclusion could be achieved regarding the preference of uptake coefficients.

In this section we probe the capability of the CMAQ to reproduce a severe dust event. Many studies have reported that spring 2010 had the most severe dust storms in recent decades (Bian et al., 2011; Li et al., 2012) due to a nation-wide drought in China. PM10 observations were more than 1000 µg m-3 at Beijing (Han et al., 2012), 1600 µg m-3 at Seoul (Tatarov et al., 2012), and 1200 µg m-3 at Taiwan (Tsai et al., 2013). These studies mainly focused on the impact of dust storms on a local scale and the understanding of the emission and transport of the dust event on a regional scale is not well-developed. Here we examined this severe dust event with model simulations, satellite observations, and also surface measurements from multiple networks. Figure 7 displays the MODIS AOD and simulated AOD from the CMAQ Dust_Chem scenario during the severe dust storm episode from 19 to 21 March 2010. Simulated AOD is derived by following the approach described in Huang et al. (2013) for 11:00 LT only to make it consistent with the nadir view time by the MODIS. Spatial distributions of the satellite data agree well with the simulation on a daily scale, indicating that the model can generally reproduce the column density and long-range transport of dust particles. As shown in Fig. 7b, the CMAQ simulation suggests that a large amount of dust emissions were uplifted on 19 March from the Gobi and increased AOD values over the desert and northern China. The heavy dust emission on 19 March has been identified with an ozone monitoring instrument (OMI) by Li et al. (2011); Fig. 8a also indicates consistently high AOD values around northern China. As the dust plume moved eastwards, both the MODIS and the CMAQ suggest that AOD in the eastern coastal area of China increased from about 0.8 on 19 March to more than 2.0 on 20 March. On 21 March, the majority of the dust plume was pushed eastward and started to build up AOD over the west Pacific and Japan.

To further examine the dust event, a forward trajectory is analyzed to characterize the transport pathway of dust plumes with the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model from the NOAA/Air Resources Laboratory (Draxler and Rolph, 2015; Rolph, 2015). The movement of air mass was analyzed for 72 h, starting from 00:00 UTC (08:00 LT) on 19 March 2010 in the Gobi, with the forward trajectories shown in Fig. 8f. Air masses at 500 (red line), 1000 (blue line), and 2000 m (green line) moved southeastward until 20 March, when the higher plume turned east and moved across Japan and the west Pacific, while the lower plumes continued towards the eastern coastal area of China and finally reached Taiwan on 21 March. The HYSPLIT trajectory showed the same consistent transport of dust plumes as the MODIS and the CMAQ AOD analysis.

To understand the impact of dust storms along the transport pathway, we compare the simulated and observed surface level PM10 on a daily scale for all of March 2010 at selected stations along the transport pathway, as indicated by Fig. 8f. Simulations from the Dust_Chem (black lines in Fig. 8a–e and g–o) and the Dust_Off scenarios (blue lines in Fig. 8a–e and g–o) are analyzed to examine the model performance. Temporal variations of PM10 with observations from the API (red circles) are examined at Beijing, Lanzhou, Nanjing, Fuzhou, Lianyungang, and Shanghai as shown in Fig. 8a, b, c, d, e, and k, respectively. Observed PM10 concentrations increased rapidly from less than 300 µg m-3 on 18 March to more than 600 µg m-3 at Beijing and 480 µg m-3 at Lianyungang on 19 March. In central and eastern China, concentrations of PM10 peaked on 21 March at Nanjing and Shanghai, and the API reached the measurement ceiling value of 600 µg m-3. In southern China, PM10 was also elevated to 600 µg m-3 at Xiamen on 22 March. Temporal variation in these cities suggested that PM10 concentrations were elevated with the onset of the dust storm, which moved from the Gobi to southeastern China from 19 to 21 March. PM10 concentrations fell back under 300 µg m-3 after the event. Lanzhou reached a peak of PM10 concentration with 500 µg m-3 on 14 March, which should be attributed to the impact of the dust storm that originated from the Taklamakan. Ling et al. (2011) also reported an observed 507 µg m-3 PM10 on 14 March at Lanzhou. Observations from EANET and the TAQMN are also employed to examine the long-range transport of dust over the west Pacific and Taiwan. Temporal variations of PM10 at three EANET sites (green diamonds) including Oki, Ogasawara, and Hedo are shown in Fig. 8h, i, and j, respectively. PM10 concentrations at these sites all showed a consistent increase with the onset of dust on 21 or 22 March. At Xinzhuang, as shown in Fig. 8g, observations from the TAQMN (purple circles) demonstrated that local PM10 was increased from less than 100 µg m-3 on 20 March to more than 700 µg m-3 on 21 March due to the impact of the dust storm. Simulated PM10 from the Dust_Chem scenario agreed well with observations from different networks all over the domain. Predictions from the Dust_Chem and Dust_Off scenarios were almost the same at all stations during the non-dust period from 1 to 10 March, yet the Dust_Chem scenario was able to reproduce the rapid elevation of PM10 during the dust event. However, a noticeable discrepancy was also found between the Dust_Chem prediction and observations. In general, the CMAQ overpredicted PM10 slightly during the dust event at most of the API sites in China but failed to reproduce the high concentrations at Lanzhou before 15 March and after 20 March. To help understand the model performance in predicting fine particles from dust, daily variations of AOD from the AERONET observations and the CMAQ simulations are also examined at four stations, including Beijing, the Semi-Arid Climate Observatory Laboratory (SACOL) station at Lanzhou (Ling et al., 2011), Osaka, and EPA-NCU (Taiwan Environment Protection Agency station at National Central University), as shown in Fig. 9l, m, n, and o, respectively. Temporal variations of AOD were consistent with the daily changes in PM10 at these cities along the dust plume movement trajectory. The highest AOD was found on 14 March at the SACOL station, which was consistent with the rapid increase in PM10 concentrations at Lanzhou. Moderate underestimations of AOD were also found at Lanzhou and EPA-NCU stations during the dust events, indicating that fine-mode aerosols were also underestimated over this region. In general, comparisons between the CMAQ and observations from MODIS and surface networks suggest that the model is capable of reproducing the severe dust storm event in terms of the spatial distribution, transport, and concentration of dust particles, with a possible underestimation of dust emission from the Taklamakan.

Despite the improvements of model performance demonstrated in the previous sections, it is necessary to note that there are some important remaining uncertainties within the modeling system. The first type of uncertainty is related to the anthropogenic emissions. The assessment of the dust prediction capability of the model was primarily performed by comparing the simulation with observations, yet the bias caused by anthropogenic emissions would affect the bias from the dust prediction. So it is difficult to distinguish the uncertainties that arise from dust treatment in the model. Figure 9 displays the dust emission rate (Tg day-1) from the Dust_Chem scenario (blue rectangles, with blue dashed line indicating the trend), the simulation bias of PM10 at the API stations (red circles, with red dashed line indicating the trend), and the simulation bias of PM10 at EANET stations (green diamonds, with green dashed line indicating the trend), with all variables averaged on a monthly scale. The prediction from the CMAQ suggests a slightly increasing trend of dust emission from 2006 to 2010, which is consistent with the decadal increase in dust reported by Kurosaki et al. (2011) due to changes in soil erodibility over Mongolia and northeastern China. The recent decadal trend of dust emission over East Asia remains an open-ended question that will rely on investigations with longer-term simulations. Simulation biases of PM10 agree fairly well with the dust emission at both the API and EANET stations, indicating that the overall underprediction of PM10 over East Asia has a smaller discrepancy for years with stronger dust events. Assuming there is a persistent underestimation of primary PM emissions in the anthropogenic inventory, more dust emissions will apparently help to reduce the modeling bias for total PM10. This is also consistent with previous studies (Wang et al., 2011; Dong and Fu, 2015a) which reported a systematic underestimation of anthropogenic emission of primary particles over China.

The second type of uncertainty lies within the friction velocity threshold u∗t, which is affected by the soil moisture fraction, which may be overestimated by the WRF. Although in this study the simulation performance is improved with the initial threshold friction velocity constants u∗c adjusted by avoiding the double counting of the soil moisture effect, there are still non-negligible biases as shown in Sect. 3. Both the 5-year average modeling bias shown in Fig. 3d and temporal variations shown in Fig. 8 suggest possible overestimated dust emission from the Gobi and underestimated dust from the Taklamakan. The averaged u∗t calculated by the CMAQ is 0.19 and 0.14 m s-1 over the Taklamakan and Gobi, respectively, with the soil moisture factor fsoilm being 1.21 and 1.13, respectively, indicating that the Taklamakan needs higher friction velocity in order to generate dust because of a more significant soil moisture impact. However, some recent field measurement studies suggest that the u∗t in the Taklamakan is lower than that over the Gobi. He et al. (2011) conducted measurements at three sites inside the Taklamakan and reported the value of u∗t as 0.25, 0.27, and 0.21 m s-1 at three different sites, and Yang et al. (2011) also reported the value of u∗t as 0.24  m s-1 at Tazhong (∼ 39.03∘ N, 83.65∘ E). For the Gobi, Li and Zhang (2011) reported the value of u∗t as 0.34–0.42  m s-1 based on measurements made in April 2006 and 2008. Field measurements defined u∗t as equal to the value of friction velocity u∗ when dust concentration is increased by 20 % for at least half an hour (Li and Zhang, 2011); thus, the reported values of u∗t from the measurement studies are higher than the calculations from the model. But the comparison between the Taklamakan and Gobi measurements suggests that the model may either underestimate u∗t in the Gobi or overestimate it in the Taklamakan. Since fsoilm is determined by the soil moisture fraction, we compare the soil moisture from FNL (NCEP final analysis data), which is used to drive WRF in this study, with another reanalysis dataset GLDAS (Global Land Data Assimilation System; Rodell et al., 2004). Figure 10 demonstrates the 5-year averages (for March and April) of the soil moisture fraction in the top 10 cm from (a) FNL and (b) GLDAS. Soil moisture is estimated to be 10–15 % by FNL in both deserts, while the values from GLDAS are less than 5 % in the Taklamakan and 5–10 % in the Gobi. Zender et al. (2003) reported that soil moisture from NCEP is too high over active dust emission areas and leads to a negative AOD bias of the model on a global scale. With the WRF-NMMB/BSC-Dust model, Haustein et al. (2012) conducted simulations with meteorology driven by FNL and GLDAS over north Africa and reported that the predictions with GLDAS had better agreement with the AERONET's AOD observations due to smaller friction velocity and slightly faster surface wind speed due to lower values of soil moisture. But no such sensitivity studies have been made over East Asia, and unfortunately there are no publicly available observation data for the period of 2006–2010 to examine the potential overestimation of soil moisture by FNL in our modeling domain. Although we previously reported a consistent negative bias of surface temperature at 2 m height for 2006–2010, which may be due to excessive soil moisture (Dong and Fu, 2015b), more research efforts are required to verify the uncertainties caused by using FNL soil data.

The last type of uncertainty lies within the mass contribution of fine aerosols within dust emission. Elevated dust particles are distributed into fine- and coarse-mode aerosols with mass ratios of 0.2 and 0.8, respectively, in the CMAQ dust scheme. In this study, however, the ratio of PM2.5/ TSP derived from observations at Duolun and Yulin is 0.42 and 0.39, respectively, indicating that fine particles should have a higher mass contribution within East Asian dust. The data from Huang et al. (2010) indicated that the ratio of PM2.5/ TSP at Tazhong was 0.45 in spring 2007, which suggested an even higher fine-particle mass contribution in the Taklamakan. Model evaluation results shown in Fig. 5 also demonstrate the systematic underestimations of both trace metals and total PM2.5 concentrations in both dust source regions and downwind areas, while the concentrations of PM10 are slightly overestimated near the source region as demonstrated in Fig. 3. Consequently, it is highly possible that the ratio of fine particles within dust emission should be higher. But since TSP also includes all large particles > 10 µm, observations of both PM2.5 and PM10 in active dust regions are urgently needed to help clearly characterize the ratio in the model.