WRF-Chem version 3.8.1 is designed to predict mesoscale weather and atmospheric chemistry (Grell et al., 2005; Fast et al., 2006; Skamarock et al., 2008; Peckham et al., 2011). The aerosol size and mass distributions are calculated with the Modal Aerosol Dynamics Model for Europe (MADE; Ackermann et al., 1998) that includes three lognormal distributions for Aitken-, accumulation-, and coarse-mode particles. MADE considers the new particle formation process of homogeneous nucleation in the H2SO4 and H2O system (Wexler et al., 1994; Kulmala et al., 1998). The model also treats inorganic chemistry systems as the default option and organic chemistry systems as coupling options. Inorganic chemistry systems include the chemical reactions of three inorganic ionic species: SO4-2, NO3-, and NH3+ (Ackermann et al., 1998). The Secondary Organic Aerosol Model (SORGAM), an optional model to calculate secondary organic aerosol (SOA) chemistry processes (Schell et al., 2001), is coupled to MADE (MADE/SORGAM). MADE/SORGAM treats atmospheric aerosols as an internal mixture of sulfate, nitrate, ammonium, organic carbon (OC), elemental carbon (EC), sea salt, and dust aerosols. Additionally, gas-phase chemical processes are calculated in Regional Acid Deposition Mechanism version 2 (RADM2; Chang et al., 1989). RADM2 simulates the concentrations of air pollutants, including inorganic (14 stable, 4 reactive, and 3 abundant stable) and organic (26 stable and 16 peroxy radicals) chemical species.

For the microphysics calculation, we use the CCN activation parameterizations (Abdul-Razzak and Ghan, 2000, hereafter ARG) and Morrison double-moment microphysics scheme (Morrison et al., 2009). The CCN activation is determined by meteorological factors (e.g., updraft) and physicochemical properties of aerosols based on the assumption of internally well-mixed aerosols. Detailed model designs for the modeling studies of aerosol–cloud interactions in WRF-Chem can be found in Gustafson et al. (2007), Chapman et al. (2009), Grell et al. (2011), and Bar et al. (2015).

For the physics parameterization, we use the following configurations: the Rapid and accurate Radiative Transfer Model for GCMs (RRTMG) for the shortwave and longwave radiative transport processes (Iacono et al., 2008); the Yonsei University scheme (YSU scheme) for the atmospheric boundary layer processes (Hong et al., 2006); and the Unified NOAH (NCEP Oregon State University, Air Force, and the Hydrologic Research Laboratory) land surface model for land surface processes (Tewari et al., 2004).

The CCN activation parameterization is based on the Köhler theory, which is described using the water activity and the surface tension of the solution droplets. The water activity is estimated from detailed information on aerosols such as the van't Hoff factor, osmotic coefficient, molecular weight, mass, and density of aerosols. If aerosol chemical information is fully provided, CCN activation could almost be accurately calculated using the Köhler theory (Raymond and Pandis, 2003); however, it is very computationally expensive (Lewis, 2008). Petters and Kreidenweis (2007) proposed a single quantitative measure of aerosol hygroscopicity, known as the hygroscopicity parameter (κ). This method does not require detailed information on aerosol chemistry and, therefore, reduces the computational cost when calculating the water activity. For this reason, κ values are applied in many observational, experimental, and numerical studies (Zhao et al., 2015; Chang et al., 2017, Shiraiwa et al., 2017; Gasteiger et al., 2018). κ can be determined separately for the three lognormal modes (Aitken, accumulation, and coarse modes). That is, κi is the volume-weighted average of κj for mode i: 1κi≡∑j=1Jεijκj, where εij is the volume ratio of chemical j in mode i (=Vij/Vtot,i,Vtot,i=∑j=1JVij, and Vij is the volume of chemical j in mode i), and κj is the individual hygroscopicity parameter for chemical j. In Eq. (1), the temperature is assumed to be 298.15 K. The upper end of the κ value for hygroscopic species of atmospheric relevance is around 1.40 (Petter and Kreidenweis, 2007).

CCN activation is affected by κ values (e.g., Nenes et al., 2002; Kristjánsson 2002). H2SO4 has a κ value that is more than 2 times higher than (NH4)2SO4: 1.19 for κH2SO4 and 0.53 for κ(NH4)2SO4 (Clegg and Wexler, 1998; Petters and Kredenweis 2007; Good et al., 2010). Such large disparities in the κSO4 between different sulfate species could cause large variability in the estimation of ERFaci. However, many aerosol modules simplify the physical and chemical characteristics of aerosols, often neglecting some chemical species (Kukkonen et al., 2012; Im et al., 2015; Bessagnet et al., 2016). Sulfate aerosols are usually prescribed as a single species of either H2SO4 or (NH4)2SO4. Some models consider H2SO4 as the representative sulfate aerosol when the neutralization reaction between H2SO4 and ammonia is not considered or when only the binary sulfuric acid–water nucleation is considered (e.g., Wexler et al., 1994; Kulmala et al., 1998; Stier et al., 2006; Kazil and Lovejoy, 2007; Korhonen et al., 2008; Mann et al., 2010). Some other models consider (NH4)2SO4 as the representative sulfate aerosol when studying aerosol–CCN closure (e.g., VanReken et al., 2003), or when including the ternary sulfuric acid–ammonia–water nucleation process or the neutralization reaction between sulfate and ammonia (Kulmala et al., 2002; Napari et al., 2002; Grell et al., 2005; Elleman and Covert, 2009; Watanabe et al., 2010). To reduce the uncertainty of ERFaci, more speciated κSO4 parameters need to be utilized in the calculation of cloud droplet activation process – at least for the two main sulfate aerosols, H2SO4 and (NH4)2SO4. Here, we suggest a new method of representing κSO4 that considers both H2SO4 and (NH4)2SO4 using the molar ratio of NH4+ to SO42-. We also suggest an alternative method that utilizes the spatial distribution of κSO4, based on the distinct distribution patterns of NH4+ and SO42-over land and sea.

H2SO4 is completely neutralized as (NH4)2SO4 when ammonia is abundant (Seinfeld and Pandis, 2006). During the neutralization process of H2SO4, 1 mol of SO42- takes up 2 mol of NH4+ and forms 1 mol of (NH4)2SO4. Here, the assumption is that ammonia neutralizes SO42- ions prior to nitrate ions (Seinfeld and Pandis, 2006), and sulfate aerosols appear only in the form of H2SO4 and (NH4)2SO4. In the calculation of κSO4, the proportion of H2SO4 and (NH4)2SO4 is determined using the ammonium to sulfate molar ratio R=nNH4+/nSO42-, where nNH4+ is the molar concentration of NH4+ ions, and nSO42- is the molar concentration of SO42- ions. Generally, sulfate aerosols are completely neutralized as (NH4)2SO4 under high R conditions (R>2) and are partially neutralized under low R conditions (R<2) (Waggoner et al., 1967; Fisher et al., 2011). Using R and the Zdanovskii–Stokes–Robinson relationship (i.e., Vd=Vw+Vtot,Vtot=∑j=1JVj, where Vd is the droplet volume, Vw is the volume of water, and Vj is the volume of the chemical j), a representative κSO4 is defined as follows: 2κSO4=εH2SO4κH2SO4+ε(NH4)2SO4κ(NH4)2SO4, where εH2SO4 is the volume fraction of H2SO4 in the total volume of sulfate aerosols (defined as VH2SO4/VSO4, where VH2SO4 is the volume concentration of H2SO4, and VSO4 is the total volume concentration of sulfate aerosols), and ε(NH4)2SO4 is calculated in the same manner for (NH4)2SO4 (defined as V(NH4)2SO4/VSO4, where V(NH4)2SO4 is the volume concentration of (NH4)2SO4). In this study, we use 1.19 and 0.53 to represent κH2SO4 and κ(NH4)2SO4, respectively (Clegg and Wexler, 1998; Petters and Kredenweis 2007; Good et al., 2010). The volume fractions of H2SO4 and (NH4)2SO4 are calculated as follows: (i)ifR=0,thenεH2SO4=1andε(NH4)2SO4=0,(ii)if0<R<2,then3εH2SO4=1-R2×nSO42-×mH2SO4ρH2SO4VSO4andε(NH4)2SO4=R2×nSO42-×m(NH4)2SO4ρ(NH4)2SO4VSO4,(iii)ifR>2,thenεH2SO4=0andε(NH4)2SO4=1. Here, m and ρ indicate the molar mass and density of the specific chemical species, respectively. To be more realistic, ammonium bisulfate may also need to be considered: when the number of SO42- is smaller than NH4+, the sulfates appear as a mixture of ammonium bisulfates and sulfuric acids, and when the number of SO42- is greater than NH4+ but not twice as large as NH4+, the sulfates appear as a mixture of ammonium bisulfates and ammonium sulfates (Nenes et al., 1998; Moore et al., 2011, 2012). For simplicity, however, such partitioning is not considered in this study. As a result, sulfate aerosols are treated as (NH4)2SO4 when R is greater than two (R>2) and as H2SO4 when R is zero (R=0). This method is applicable to the models that consider both NH4+ and SO42- ions. If NH4+ data are not available in a model, we suggest an alternative method to represent κSO4 based on the typical geographical distribution pattern of sulfate aerosols available from observations, as discussed below.

Observational studies show the distinctly different distribution patterns of the two dominant sulfate aerosol species, i.e., (NH4)2SO4 over land and H2SO4 over sea (Fujita et al., 2000; Paulot et al., 2015; Kang et al., 2016; Liu et al., 2017). Such distribution patterns are related to the sources of sulfate and ammonium. In general, sulfate aerosols are emitted from land and sea, whereas ammonium is mostly produced from land. Sulfur dioxide is produced from fossil fuel combustion, volcanic eruptions, and dimethyl sulfide (DMS) via air–sea exchanges, and then forms sulfate aerosols (Aneja 1990; Jardin et al., 2015). Wind transportation of pollutants could also cause high concentrations of sulfate aerosols over the sea (Liu et al., 2008). In contrast, ammonium is emitted from livestock, fertilizer, and vehicles (Sutton et al., 2013; Paulot et al., 2014; Bishop et al., 2015; Liu et al., 2015; Stritzke et al., 2015); therefore, it is concentrated mostly on land. Ammonium is usually not abundant enough to fully neutralize H2SO4 in the marine boundary layer (Paulot et al., 2015; Ceburnis et al., 2016). Thus, when ammonium information is not available, the κSO4 can be alternatively estimated by considering the land and sea fractions as follows: 4κSO4=f×κSO4,land+1-f×κSO4,sea, where f represents the fraction of land at each grid point; unity means entire land, zero means entire sea, and the value in between represents the fraction of land at the grid points in coastal areas. κSO4,land and κSO4,sea represent κSO4 over land and sea, respectively (i.e., κSO4,land=κ(NH4)2SO4=0.53 and κSO4,sea=κH2SO4=1.19).

The simulated sulfate and ammonium distributions are compared with the observational data that were measured onboard the NASA DC-8 aircraft during the KORUS–AQ campaign (https://www-air.larc.nasa.gov/missions/korus-aq/, last access: 18 July 2019) in and around the Korean Peninsula in May and June of 2016. The measurements were taken within the boundary layer. The mass concentration of sulfate and ammonium were obtained using the method described in Dibb et al. (2003).

In Fig. 1, the mass concentration of sulfate and ammonium simulated by AS are compared with the KORUS-AQ aircraft observations (OBS) following the flight track. The simulated sulfate shows a positive bias but has a high temporal correlation with OBS (r=0.78). The simulated ammonium is less biased than sulfate but indicates a moderate temporal correlation with OBS (r=0.58). Overall, it seems reasonable to state that the WRF-Chem model can calculate the distribution of sulfate aerosols well enough.

Figure 2 shows the 30 d averaged mass concentration of sulfate and ammonium and the molar ratio (R) of ammonium to sulfate over the model domain. During the KORUS-AQ campaign period, high-pressure systems often covered East China and the Yellow Sea, and this led to stagnating sulfate and ammonium concentrations. However, sulfate and ammonium are distributed differently due to different sources. Pollutants emitted from the Asian continent are often transported by westerly and southerly winds. Sulfate is highly concentrated in China and the northern part of the Yellow Sea, and DMS emission from the sea also contributes to the formation of sulfate aerosols over the sea. Ammonium is widely distributed throughout China due to the use of fertilizers over farmlands (Paulot et al., 2014; Van Damme et al., 2014; Warner et al., 2017). The concentration of ammonium is generally low over the sea, but it is high over the northern part of the Yellow Sea due to wind transport.

The distribution of R is associated with the distribution of sulfate and ammonium (Fig. 2). In general, R is high (R>2) over land on account of the high anthropogenic emissions of continental ammonium, and R is low (R<2) over remote seas because the ammonium concentration is small. However, high R is also shown over the Yellow Sea in Fig. 2. This is because the ammonium concentration increases when the westerlies carry continental pollutants over the Yellow Sea during the simulation period. Based on the distribution of R, sulfate aerosols are expected to be almost completely neutralized over land (e.g., (NH4)2SO4) and partially neutralized over sea ((NH4)2SO4+H2SO4).

Figure 3 shows the average κ of the accumulation-mode aerosols in AS and the difference between RA and AS and between LO and AS.

The accumulation mode is selected because sulfate aerosols are dominant in this mode. AS simulates κ values that are roughly consistent with the observed mean κ values in the literature (i.e., κ over land is about 0.3 and κ over sea is about 0.7; Andreas and Rosenfeld, 2008), but it varies significantly between land and sea. The κ over land is expected to be lower than the κ over sea because continental aerosols usually include more hydrophobic aerosol species such as black carbon and organic carbon from industry, whereas maritime aerosols consist mainly of hygroscopic substances, i.e., sea salt and non-sea salt sulfates originating from DMS. The variation in κ is also influenced by chemical reactions and meteorological factors, i.e., wind transportation of aerosols and scavenging of aerosols due to precipitation, as well as gravitational settling.

Compared with AS, RA and LO show a pronounced difference in κ over sea (Fig. 3b, c). That is, RA and LO produce significantly higher κ over the sea than AS does because the ammonium concentration is not sufficient to neutralize sulfate completely over the sea (i.e., R<2). RA predicts slightly higher continental κ following the coastal regions than AS because R occasionally becomes low due to the intrusion of maritime air masses that have very low concentrations of ammonium. Maritime κ of RA is lower than that of LO because the transportation of continental pollutants increases the portion of ammonium over the Yellow sea.

According to the Köhler theory, changes in κ directly influence CCN activation. In this study, the CCN activation rate (fCCN) is defined as the ratio of the CCN number concentration at 0.6 % supersaturation to the total aerosol number concentration. Simulated fCCN is compared with the aircraft measurements during the KORUS-AQ campaign (OBS). During this campaign, aerosol and CCN number concentrations were measured by a condensation particle counter (CPC; TSI, 3010) and a CCN counter (CCNC; DMT, CCN-100), respectively (Park et al., 2020). The CPC measures the number concentration of aerosols larger than 10 nm in diameter, and the CCNC measures the CCN number concentration at 0.6 % supersaturation.

The model simulations capture the temporal variation of fCCN well (r≈0.7 for the linear correlation with OBS; Fig. 4).

However, fCCN values are underestimated mainly due to the underestimation of CCN concentrations. The average aerosol (CN) number concentrations for the flight track in all simulations (AS, SA, RA, and LO) and the actual observed values during the flight are 5934 and 5794 cm-3, respectively. Thus, unlike Georgiou et al. (2018), who showed that WRF-Chem coupled with MADE/SORGAM tended to overestimate aerosol number concentrations, our simulations only slightly overestimated aerosol number concentrations. The average CCN number concentration at 0.6 % supersaturation for the AS, RA, and LO simulations are 982, 1027, and 1057 cm-3, respectively, but the observation was 2154 cm-3. Such underestimated CCN concentrations seem to be due to the systematic error in WRF-Chem. As discussed in Tuccella et al. (2015), the uncertainty of the updraft velocity parameterization and bulk hygroscopicity of aerosols lead to an underestimation of the CCN concentration and CCN efficiency (CCN/CN) by a factor of 1.5 and 3.8, respectively. Nevertheless, over land, AS, RA, and LO simulate similar values of fCCN because continental sulfate aerosols are generally expected to be a fully neutralized form of sulfate (i.e., (NH4)2SO4). This was not the case over sea. During KORUS-AQ, the aircraft passed over the Yellow Sea on 22 and 25 May 2016 (blue shading in Fig. 4). On this occasion, LO simulates the highest fCCN over the sea among all simulations because LO uses the prescribed κSO4 value of κH2SO4 over sea. RA simulates slightly lower fCCN over the sea because transportation of continental pollutants over the sea can be taken into account, as observed during the KORUS-AQ campaign. The transported air pollutants increase the ammonium concentration over the sea, neutralize H2SO4, reduce the hygroscopicity of sulfate aerosols, and consequently decrease fCCN. Simulated fCCN in RA has a high spatiotemporal correlation with the observation over the Yellow Sea (i.e., 0.83), whereas AS shows a rather lower correlation (i.e., 0.65). Such difference stems from the fact that R values vary significantly over the Yellow sea due to the transportation of anthropogenic chemicals by westerlies, and such variability is taken into account in RA. This improvement highlights the importance of appropriate chemical representation in atmospheric models. Compared with the RA and LO simulations, AS predicts the lowest fCCN because the lowest κSO4 (=κ(NH4)2SO4) is prescribed over sea as well as over land.

We conducted a reliability test that has been often used to evaluate the performance of air quality models. Kumar et al. (1993) proposed the following three criteria for judging model reliability: (1) the normalized mean squared error (NMSE) below 0.5; (2) the fractional bias (defined as 2×OBS‾-sim‾OBS‾+sim‾, where OBS indicates the observed values, sim indicates the simulated values, and the bar above the symbols indicates the average) between -0.5 and 0.5; and (3) the ratio of the model values to the observed values (defined as sim/OBS) between 0.5 and 2.0. These values for AS, RA, and LO are compared in Table 3. It indicates that RA and LO satisfy all three criteria, but AS does not satisfy two of the three criteria as it predicts a rather high normalized NMSE and fractional bias. Between RA and LO, LO seems somewhat closer to the observations than RA, but the difference is small for these calculations.

The variation in κSO4 almost directly influences the change in the column-integrated fCCN (Fig. 5).

RA predicts higher fCCN than AS over the coastal land regions because the occasionally very low ammonium concentration lowers R and affects the CCN activation. Meanwhile, SA prescribes κSO4 value 2 times as high as AS does and produces about 20 % higher fCCN values.

Different κSO4 parameterizations affect simulated cloud microphysical properties. Figure 6 shows the relative differences in the simulated column-integrated cloud droplet number concentration (CDNC) in RA, LO, and SA from AS. All three produce higher κSO4 values than AS and, therefore, simulate higher CDNCs. However, the differences in CDNC do not exactly correspond to the differences in fCCN (Fig. 5) because cloud droplet activation is also affected by in-cloud supersaturation and other meteorological factors. SA simulates higher CDNC than AS over both land and sea, but RA and LO simulate higher CDNC mostly only over sea. RA and LO produce similar CDNC distributions over the Yellow Sea (compare Fig. 6a and b) although RA produces smaller fCCN than LO (compare Fig. 5a and b). As in Moore et al. (2011), the reason for this may be that the sensitivity on fCCN decreases so much because supersaturation is so high that most aerosols can act as CCN regardless of their critical supersaturation. That is, the supersaturation over the Yellow Sea is high enough to activate most aerosols to cloud droplets. Over land, RA simulates higher CDNC (up to 12 %) than AS in southeast China and the Korean Peninsula, but LO simulates CDNC similar to AS. The results of RA seem to be related to the dilution of ammonium concentrations along the coastal land regions due to the intrusion of maritime air. However, such variation in ammonium cannot be taken into account in LO.

Overall, high CDNCs in RA, LO, and SA (Table 1) result in less precipitation but larger liquid water path (LWP), compared with AS (Table 2). Precipitation reduction is more pronounced over sea because of larger relative differences in CDNC. These results agree well with some previous studies – i.e., high CDNCs suppress local precipitation, prolong cloud lifetime, and consequently increase net LWP, which is known as the cloud lifetime effect (Albrecht, 1989). Obviously, SA, which assumes sulfate aerosols are all H2SO4 particles, produces the highest CDNC and also the largest differences in all other properties in Table 2. Less rainwater in SA than in any other simulations may also imply that precipitation scavenging of aerosols was less efficient and, therefore, that more aerosols (CCN) were retained to produce more cloud drops and a longer cloud lifetime. On average, SA has 103 cm-3 more aerosols over sea and 116 cm-3 more aerosols over land than LO. These surplus aerosols certainly have the potential to simulate a higher number of CCN in SA than in LO.

For the same LWP condition, high CDNC induces small effective radii (re). RA, LO, and SA simulate smaller re than AS, and the maximum difference in re amounts to 1.46, 1.38, and 1.48 µm, respectively. However, the domain-averaged differences in re are not as substantial as the differences in other cloud microphysical properties (Table 2). This may be related to somewhat larger LWPs in RA, LO, and SA than in AS as well as the sufficient water supply during droplet growth. All simulations in this study have high water vapor path (WVP) conditions (WVP>30kgm-2) throughout the whole domain. According to Qiu et al. (2017), cloud droplets have low competition for water vapor and a high chance of collision–coalescence under high WVP conditions (i.e., WVP>1.5cm or 15kgm-2). If LWP is similar, the re difference could be larger among the simulations than those that are shown herein.

Cloud microphysical properties determine cloud optical properties and, therefore, control the cloud radiative effects. For a fixed LWP, high CDNC is usually associated with low re but high cloud optical thickness. Then optically thick clouds reflect more sunlight and strengthen the cloud radiative cooling effect at the top of the atmosphere (TOA), which is known as the cloud albedo effect (Twomey, 1974). We calculate the cloud radiative effect at the TOA (CRE) by subtracting the clear-sky downward radiation from the net all-sky downward radiation (including clouds) (Hartmann, 2016).

RA, LO, and SA simulate optically thicker clouds that reflect more sunlight and exert stronger cooling effects at the TOA than AS (Fig. 7). For the domain average, the differences in CRE for RA, LO, and SA from AS amount to about -1.7, -1.5, and -2.1Wm-2, respectively. These differences are most pronounced over sea (Fig. 7b, c, d). Such pronounced difference over sea may be affected by the large cloud fraction around the East China Sea due to the East Asian summer monsoon (Pan et al., 2015). That is, a large cloud fraction exerts a large CRE cooling, so the impact of the new parameterization of κSO4 on CRE could be substantial under large cloud fraction conditions. Note that CRE is similar over land and over the sea in the latitude band from 25 to 28∘ N in AS (Fig. 7a), but the CRE differences between RA, LO and SA, and AS are much higher over the sea than over land (Fig. 7b, c, d). Such an enhanced cooling effect over the sea can be explained by increases in CDNC (Fig. 6) and, somewhat, by increases in LWP (Table 2). According to some previous studies, the contribution of CDNC and LWP to CRE could be larger than 56 % (Sengupta et al., 2003; Goren and Rosenfeld, 2014).