The CMA new generation of weather forecasting model GRAPES is a fully compressible non-hydrostatic weather model system which uses a semi-implicit and semi-Lagrangian discretization scheme (Zhang and Shen, 2008; Xu et al., 2008; Chen et al., 2008). It uses an Arakawa-C staggered grid and the central finite-difference approach of second-order accuracy in horizontal coordination, a non-hydrostatic approximation method together with the staggered approach of Charney-Phillips to improve the accuracy of vertical pressure gradients and a semi-implicit and semi-Lagrangian scheme for temporal and advection discretion. A height-based terrain-following coordinate, which behaves as a natural height coordinate, is used to compromise the Lagrangian trajectory errors in spherical coordinates at high latitudes. The physical packages include cumulus convective, single-moment cloud microphysics, radiative, land surface, and boundary layer processes.

CUACE is a unified atmospheric chemistry environment with four major functional subsystems: emissions, gas-phase chemistry, aerosol microphysics, and data assimilation. It is designed to facilitate the establishment of a chemical weather forecasting system using near real time data in China (Zhou et al., 2012, 2008). Seven aerosol components, i.e., sea salts, sand/dust, elemental carbon, organic carbon, sulfates, nitrates, and ammonium salts are sectioned into 12 bins with detailed microphysics of hygroscopic growth, nucleation, coagulation, condensation, dry depositions and wet scavenging in the aerosol module. The gas chemistry module is based on the second generation of the Regional Acid Deposition Model (RADM II) mechanism with 63 gaseous species through 21 photochemical reactions and 121 gas-phase reactions applicable under a wide variety of environmental conditions especially for smog (Stockwell et al., 1990), and prepares the production rates of sulfate and secondary organic aerosol for the aerosol module. Emission inputs have been provided by the same emission subsystem (EMIS) of CUACE with the official basic emission sources data updated to the year 2010 (Zhou et al., 2012).

CUACE is fully coupled online to the regional version of GRAPES, establishing the comprehensive modeling system, GRAPES/CUACE, and has been used for radiative feedback research (Wang et al., 2010, 2015).

The sectional aerosol activated scheme developed by Hayder Abdul-Razzak and Ghan (HG scheme) provides a convenient platform to connect sectional aerosols into cloud physics (Abdul-Razzak and Ghan, 2002; Hayder and Ghan, 2000; Abdul-Razzak et al., 1998; Ghan et al., 1995, 1993). This scheme is derived from the basic theory of a Köhler curve to calculate how a particle can be activated under a certain supersaturation in an air parcel. The newly activated CCN would be parameterized in terms of environmental supersaturation and the upper and low limit of the critical supersaturation for each aerosol bin. This method is very useful for the precise determination of the concentration of droplets nucleated at the cloud base. However, droplet nucleation or aerosol activation can also take place above the cloud base as well, which is induced by increasing supersaturation above the cloud base and lateral entrainment of the surrounding air with dry aerosols (Khain et al., 2000). Therefore, it is not enough to calculate the CCN at the cloud base only in an air parcel, especially for stratus clouds where turbulence is more important than the vertical movement (Bodenschatz et al., 2010). For these reasons, the HG scheme has been online coupled with both the stratus scheme and the convective scheme in GRAPES. Consequently, aerosols from CUACE can be activated into CCN as the humid condition is satisfied, not just in the cloud base.

The WRF Double-Moment 6-class scheme (WDM6) is introduced into GRAPES since it can predict not only the mass but also the number of droplets and drops (Lim and Hong, 2010). It is developed from the WRF Single-Moment 6-class scheme (WSM6) and needs the CCN input for cloud droplets which provides a direct way for aerosol feedbacks into clouds and precipitation. The original activation scheme is expressed as (K2000) (Twomey, 1959; Khairoutdinov and Kogan, 2000) na=n+NcSwSmax⁡κ, where n is the aerosol number concentration, κ is the activation power exponent, Smax⁡ is the maximum supersaturation to activate all the aerosols, and Sw is the air supersaturation. The change rate of cloud droplet number concentration due to CCN activation is given by ∂Nc∂t=max⁡0,n+Ncmin⁡1,SwSmax⁡κ-NcΔt.

In this activation scheme, n, κ, and max⁡Smax⁡ are usually preset just as in the Twomey formula, with the inherited shortcomings due to the insufficient information of real aerosols. To overcome these, the aerosol size and composition information from CUACE and the humid and thermal information from GRAPES are used in the HG scheme to calculate the CCN at each time step, which is then fed into WDM6 to replace the preset CCN by the scheme of K2000.

An aerosol particle can be quickly activated into CCN, usually in less than a second, as the atmosphere reaches the critical supersaturation of the aerosols (Kogan, 1991). In models with relatively coarse resolution like GRAPES, supersaturation cannot be easily satisfied because of the insufficient information of inhomogeneous turbulence and vertical movements even when the grid-mean relative humidity is over about 85 to 90 %. Local supersaturation around a particle can be satisfied, through which clouds can be formed. As the local supersaturation is not only decided by the mean fields, a parameterized scheme has been developed to add the effect of local turbulence and vertical movements to the supersaturation in HG scheme for WDM6 in GRAPES. The turbulence effect is assumed to be proportional to the mean relative humidity and inversely proportional to both the horizontal wind speed and the vertical height.

Stratiform and convective precipitation are the two precipitation regimes that cannot be mutually excluded. The former refers to precipitation with low vertical motions, and the latter to stronger precipitation (Houze, 1997). These two regimes can be fully explicit in mesoscale model, only under the condition that the model resolution is below several hundred meters (Molinari and Dudek, 1992), which is very difficult to realize for the online coupled modeling systems such as GRAPES/CUACE. A hybrid approach is often used in most mesoscale models, which can explicitly condense water for the stratus precipitation in grid scale, to parameterize convective-scale precipitation in sub-grid scale. In order to fully account for the ACI in the model, the HG scheme has also been coupled to the Kain–Fritch convective parameterization scheme (KF) in GRAPES/CUACE (Kain, 2003; Kain and Fritsch, 1990; Fritsch and Chappell, 1980). The sectional aerosols have been introduced and moved upward or downward as the air mass does in KF. They become activated as soon as the supersaturation satisfies the critical supersaturation by the HG scheme.

Since there is no size information for cloud and rainwater in the bulk convective scheme, a generalized gamma distribution has been introduced to describe the cloud-droplet and raindrop spectra (Cohard and Pinty, 2000; Walko et al., 1995; Clark, 1974). NX(DX)=NXαXΓνXλXαXνXDXαXνX-1exp⁡-λXDXαX Here λXis the slope parameter, and αX and νX are two shape parameters. λX=π6ρwΓ(νX+3/αX)Γ(νX)NXρarx, NX=N0+Nccn, Nccn is the newly activated aerosol number concentration, and N0 is the cloud number concentration which is set to be 30 cm-3 in terms of observations in fall in east China (Y. Zhang et al., 2012; Zhang et al., 2011). Normally αX can be specified a priori to be 1 for number concentration size distribution and 3 for mass concentration size distribution. Only νX is left to be tuned through measurements. A parameterization for νX has been proposed through the total droplet number concentration and liquid water content (Geoffroy et al., 2010) for size spectra of number concentration:αX=1,νX=14.5qc+6.7;for size spectra of mass concentration:αX=3,νX=1.58qc+0.72. Here qc is the liquid water content in the unit of g m-3. With Eqs. (4) and (5), droplet size distribution in Eq. (3) can be bulk-represented by the total liquid water content and total cloud droplet number concentration which can then include the newly activated aerosols from HG.

The condensed water of the KF scheme is allowed to be removed as precipitation based on an empirical relationship (Ogura and Cho, 1973): δrc=rco(1-e-c1δz/w). Here c1 is a constant; δz is the height of the layer. The precipitable water δrc in Eq. (4) is inversely proportional to the vertical velocity w, which means that a small vertical velocity would produce more δrc. As smaller vertical velocity means weaker convection and less rco (total condense), less precipitable water δrc would be produced. Therefore, rco and w are offset against each other, which may cause false large convective precipitation. In order to overcome this problem, the size spectrum information of cloud drops and droplets is introduced into the convective clouds, which resolves the total cloud water content and number concentrations, including the newly formed CCN. The new precipitable rainwater, indicated by rp, can be calculated by integrating Eq. (3) to replace rco in Eq. (6). The threshold for cutoff radius between the droplets and raindrops in the integration is 50 µm for deep convective precipitation (Berry and Reinhardt, 1974; Seifert and Beheng, 2001), and 25 µm for shallow convective precipitation (Khairoutdinov and Kogan, 2000). Terminal velocity wT for the raindrops has also been introduced. Finally, Eq. (6) is reformed as δrc=rp(1-e-wT/w). Now, the comprehensive interactions of aerosol–cloud–precipitation have been established in GRAPES/CUACE with activated aerosols from CUACE linked into both the stratus and convective clouds to participate in the cloud and precipitation processes.

A series of long-lasting heavy haze-fog events hit the east part of China in January 2013. Climatology analysis shows that the mean number of the hazy days in this period is much higher than the mean value from 1981 to 2010, especially in the three major pollution zones of the North China Plain, Yangtze River Delta, and Zhu River (R. H. Zhang et al., 2013). Meanwhile, the values of the stagnant polluted parameter PLAM (parameter linking air quality to meteorological conditions/haze), a threshold value to distinguish clear and polluted weather, are over 80 in most parts of east China, which indicates strong static weather conditions for pollutant accumulation (X. Y. Zhang et al., 2013; Wang et al., 2012). Surface daily mean PM2.5 concentrations are in the range of 100–150 µg m-3 and AOD (aerosol optical depth) is above 1.0 in many surface stations (Che et al., 2014; Wang et al., 2014).

Eight precipitation events also occurred in January 2013, with six of them sweeping over regions south of the Yangtze River as low-pressure systems moving out of the Sichuan Basin each time (Table 1). Two other cases are related to cold fronts and affect the whole east of China. This forms a very good period to study ACI under high-pollution conditions.

Three sets of experiential runs are designed: T1 with the single-moment microphysics scheme of WSM6 and original KF without aerosol activation; T2 with the WDM6 scheme and the activation scheme K2000; and T3 with the HG activation scheme connected online with WDM6, KF, and CUACE. Five target regions are selected for the evaluation in terms of typical heavy pollution regions over China (Fig. 1). R1 and R2 cover the whole of China and east China respectively. The areas of R2, R3, R4, and R5 are representative of three typical polluted zones: northeast China, the North China Plain, and south China.

The meteorological initial and boundary conditions, at the resolution of 0.5∘, are interpolated from the forecasting outputs of the CMA medium meteorological model T639 in a 6 h interval. The surface daily and hourly PM2.5 concentrations from CMA Atmosphere Watch NETwork (CAWNET) are used to evaluate the performance of aerosols (Wang et al., 2008; X. Y. Zhang et al., 2012). Precipitation data from the rain gauges in meteorological stations over China are used for the precipitation threat scoring (Wang et al., 2008). Temperature, geopotential height, humidity, and cloud water mixing ratio of NCEP Reanalysis at standard pressure levels from 1000 to 10 hPa are used to evaluate the outputs of GRAPES/CUACE (Kalnay et al., 1996).

Regional monthly mean vertical profiles of the total cloud water (the total mass of cloud water, rainwater, ice water, snow, and graupel) and temperature for T1, T2, and T3 together with the NCEP Reanalysis in R1, R2, R3, R4, and R5 are shown in Fig. 2a1–a5 and b1–b5. Compared to T1, T2, and NCEP Reanalysis in the layers below 600 hPa, the total cloud water increases obviously in all five regions for T3 with a clear peak at about 850 hPa. These results indicate that with more realistic aerosols interacting with clouds (T3), more water vapor condenses into cloud water due to the activation of aerosol particles.

The amount of the increasing cloud water for T3 has very unique regional characteristics. The increase in R1, i.e., nationwide, with plenty of remote dry areas such as Tibet and northwest China, is much less than in R2 which represents the most developed regions in China with a lot of emissions. Of the three typical polluted regions, R3 covers most parts of northeast China with below-freezing temperatures, which is not favorable for warm cloud formation; therefore no obvious increases are found compared to between T2 and T3. While the temperature in R4, the most polluted area in China, is near or just above freezing, the condition is favorable for the long-lasting haze formation but not a good condition for cloud and precipitation formation with the relatively little cloud water. In R5 where abundant cloud water exists and the temperature under 700 hPa is above freezing, this region has the most enhancement for aerosol activation and the development of warm clouds.

The amount of total cloud water by T1 and T2 is compatible with each other because there is no aerosol feeding into the single-moment scheme T1 and the aerosol activation is based on the prescribed aerosol numbers without spatial variation in T2. Profiles of cloud water from NCEP are close to that of T1 and T2 in all of the five regions. Even though basic variables such as wind, temperature, relative humidity, height, and pressure are most analyzed from measurements, the cloud water data from NCEP Reanalysis are not from direct observations but from cloud physics' diagnosis. Cloud water data from NCEP Reanalysis should be carefully used in the regions affected by high aerosol concentrations.

For all regions, the increases by T3 in total cloud water contents are driven by liquid water contents (Fig. 3a1–f1). The solid cloud water for T1, T2, and T3 is in the same order of 1.0 × 10-3 g kg-1 in all the three regions and follows almost the same vertical distribution shapes with tiny difference of height and thickness in each region, showing relatively small effects from aerosols (Fig. 3a2–f2).

The aerosol activation in T3 can also increase the cloud number concentrations which affect the mean diameters of droplets (MDD). The profiles of MDD in R4 and R5 are shown in Fig. 4. MDDs for T3 are all below 10 µm, about 1 order less than those of T2, and change little with height in the lower troposphere under 700 hPa (Fig. 4a), showing a clear decrease after aerosols activated into clouds in the North China Plain (R4) and south China (R5). Conventionally, smaller droplets (usually less than 20 µm) need more time to grow into raindrops than the larger drops, which should result in more rains produced in T2 than in T3 after a decrease in MDDs. However, due to the self-collection growth process, which depends on both the MDDs and number concentrations of cloud droplets, the simulations in T2 do not see more rain produced (as discussed in Sect. 4.2).

In order to investigate the combined impacts of both MDDs and number concentrations on rain production, the self-collection kernels for the two-moment WDM6 from Long's work (Long, 1974) is used. As shown in Fig. 4e, self-collection rates (SCRs) of R5 for T3 are the largest among all three tests among the three regions of R3, R4, and R5. They are over 100 m-3 s-1 under 700 hPa, about 1 order of magnitude higher than in T2. This indicates that in R5 even though the activated aerosols from CUACE decrease the MDD which may in some way decrease the self-collection process, the high concentration of cloud droplets from aerosol activation, 4 orders of magnitudes higher than that of T2, enhances the chance of collision and compromises the decreasing collision trend by MDD. Therefore, the SCRs can explain the phenomenon that the precipitation simulation ability increases obviously for T3 in R5 (in Sect. 4.2). While in the North China Plain (R4), SCRs for T3 and T2 are almost the same, they varied around the value of 2 m-3 s-1, meaning that a lot of aerosols activate as small cloud drops which grow very slowly into larger ones and stay in the atmosphere for a relatively long time, forming long-lasting haze there (Fig. 4d). In northeast China (R3), SCRs for T3 and T2 are so small that the SCRs could be ignored, which is also consistent with the cold cloud formation (Fig. 4c).

In summary, the aerosol effects on cloud formation are very different in the three typical polluted regions: R3, R4, and R5. As shown above, R3 is controlled by cold cloud formation processes; little impact of aerosol activation can be seen in this region. In R4, the relatively humid layer from the surface to 600 hPa indicates a favorable condition for the formation of long-lasting haze as proven by both the mean diameters of droplets and collection rates. In R5, liquid water for T3 is also 1 order higher than that in R4 and 3 orders of magnitude higher than that in R3. Together with the highest collection rate, R5 consists of good conditions for the formation of precipitation, which is consistent with the improvement of the precipitation simulation ability as discussed below.

Results of regional mean temperature profiles in the five target areas above 700 hPa are almost the same as from NCEP Reanalysis (Fig. 2b1–b5). The differences between simulations and NCEP Reanalysis are mostly under 700 hPa in all of the five regions and decrease with height. Temperature is from 5 to 10 ∘C higher for all the three tests (T1–T3) compared to that of NCEP near the surface in R3 and R4, indicating some problems for GRAPES in the North China Plain. In south China (R5) the regional mean temperature profiles are also almost the same as from NCEP Reanalysis, showing a good performance of GRAPES there. The regional mean difference of temperature for different tests in each region is not very obvious, with the highest value of about 0.2 ∘C near the surface located in south China between T3 and T1. Above the surface or in other regions, the mean temperature difference is not significant (< 0.05 ∘C). This means that the regional mean temperature changed by the aerosol is not obvious. This is also the case for the geopotential heights.

The time series of differences in the regional mean temperature biases between T3 & T1 and T2 & T1 in the three typical polluted areas (R3, R4, and R5) in east China at the three layers are shown in Fig. 7. T3 can clearly decrease the temperature biases below 700 hPa most of the time in all three regions. The difference magnitude decreases from 1000 to 700 hPa, showing the clear impact of aerosol below 700 hPa in these layers. The largest decrease is about 3 ∘C near the surface and about 1 ∘C even at 700 hPa in R5 during the precipitation event from 19 to 22 January. The difference is less than 0.5 ∘C in R3 and R4 most of the time. It also shows that T2 increases the biases with the largest increase in R5. T2 might even at some times decrease the biases, but with a lower magnitude compared to that of T3.

The above analysis also shows that WDM6 with the activation scheme K2000 can introduce errors to the atmospheric temperature fields because of the missing details of aerosol size and component information. The errors would increase with the intensity and frequency of the precipitation. Similarly to the conclusion from the threat scoring analysis, only the two-moment scheme with the real-time aerosol activation can reduce the temperature bias and increase the model's ability to simulate precipitation.

To realistically simulate the aerosol impacts on clouds, a good performance of PM2.5 simulation, including aerosol sizes and vertical distributions, is critical. Due to limited available observations during the simulation period, only daily averages of PM2.5 concentrations from GRAPES/CUACE have been interpolated and compared with the observations in 32 stations from CAWNET, which have been sorted into the five target regions from R1 to R5. Scatter plots in Fig. 8a–c for PM2.5 simulation for R1, R4, and R5, the nationwide region and the two regions more significantly affected by aerosols, show that the daily average particle mass concentrations between the observations and model outputs are within a factor of 2. This is compatible with the simulation level of particle matter in most models.

Total correlation coefficients for T1, T2, and T3 in R1 to R5 have been plotted in Fig. 8d. Correlation coefficients for R1, R2 and R5 are above 0.54 for all three tests. Correlation coefficients for R4 are about 0.48 for the three tests. This shows a relative stable and reasonable performance for PM2.5 over these four regions. Correlation coefficients for R3 are only about 0.2 for the three tests, showing a relatively poor simulation there. These results are a little higher than the result by Zhou 2012 for the same region (Zhou et al., 2012). Figure 8d also indicates that the differences of correlation coefficients of the three tests in each region are very small. ACI can increase the correlation coefficient by about 2 % in terms of monthly mean.