Two-way coupled meteorology and air quality models in Asia: a systematic review and meta-analysis of impacts of aerosol feedbacks on meteorology and air quality

Abstract. Atmospheric aerosols can exert influence on meteorology and air quality through aerosol-radiation interactions (ARI) and aerosol-cloud interactions (ACI) and this two-way feedback has been studied by applying two-way coupled meteorology and air quality models. As one of regions with high aerosol loading in the world, Asia has attracted many researchers to investigate the aerosol effects with several two-way coupled models (WRF-Chem, WRF-CMAQ, GRAPES-CUACE and WRF-NAQPMS) over the last decade. This paper attempts to offer bibliographic analysis regarding the current status of applications of two-way coupled models in Asia, related research focuses, model performances and the effects of ARI or/and ACI on meteorology and air quality. There are total 157 peer-reviewed articles published between 2010 and 2019 in Asia meeting the inclusion criteria, with more than 81 % of papers involving the WRF-Chem model. The number of relevant publications has an upward trend annually and East Asia, India, China, as well as the North China Plain are the most studied areas. The effects of ARI and both ARI and ACI induced by natural aerosols (particularly mineral dust) and anthropogenic aerosols (bulk aerosols, different chemical compositions and aerosols from different sources) are widely investigated in Asia. Through the meta-analysis of surface meteorological and air quality variables simulated by two-way coupled models, the model performance affected by aerosol feedbacks depends on different variables, simulation time lengths, selection of two-way coupled models, and study areas. Future research perspectives with respect to the development, improvement, application, and evaluation of two-way coupled meteorology and air quality models are proposed.


Plain are the most studied areas. The effects of ARI and both ARI and ACI induced by natural 23 aerosols (particularly mineral dust) and anthropogenic aerosols (bulk aerosols, different chemical 24 compositions and aerosols from different sources) are widely investigated in Asia. Through the 25 meta-analysis of surface meteorological and air quality variables simulated by two-way coupled 26 models, the model performance affected by aerosol feedbacks depends on different variables, 27 simulation time lengths, selection of two-way coupled models, and study areas. Future research 28 perspectives with respect to the development, improvement, application, and evaluation of two-way 29 coupled meteorology and air quality models are proposed. The interactions between air pollutants and meteorology can be investigated by observational 49 analyses and/or air quality models. So far, many observational studies using measurement data from 50 a variety of sources have been conducted to analyze these interactions (Bellouin et Sekiguchi et al., 2003). 59 In the early stages, observational studies of ACI effects were based on several cloud parameters 60 mainly derived from surface-based microwave radiometer (Kim et al., 2003;Liu et al., 2003) and 61 cloud radar (Feingold et al., 2003;Penner et al., 2004). Later on, with the further development of 62 satellite observation technology and enhanced spatial resolution of satellite measurement comparing 63 against traditional ground observations, the satellite-retrieved cloud parameters (effective cloud 64 droplet radius, liquid water path (LWP) and cloud cover) were utilized to identify the ACI effects 65 studies on cloud scale. (Goren and Rosenfeld, 2014;Rosenfeld et al., 2014). Moreover, in order to 66 clarify whether aerosols affect precipitation positively or negatively, the effects of ACI on cloud 67 properties and precipitation were widely investigated but with various answers ( (droplet concentration and updraft velocity at cloud base, LWP at cloud cores, cloud geometrical 73 thickness and cloud fraction) to single out ACI under a certain meteorological condition, and found 74 that the cloudiness change caused by aerosol in marine low-level clouds was much greater than 75 previous analyses (Sato and Suzuki, 2019). Despite the fact that aforementioned studies had 76 significantly improved our understanding of aerosol effects, many limitations still exist, such as low 77 temporal resolution of satellite data, low spatial resolution of ground monitoring sites and lack of 78 vertical distribution information of aerosol and cloud  Sato and Suzuki, 2019; 79 Yu et al., 2006). 80 Numerical models can also be used to study the interactions between air pollutants and 81 meteorology. Air quality models simulate physical and chemical processes in the atmosphere (ATM) 82 and are classified as offline and online models (El-Harbawi, 2013). Offline models (also known as 83 traditional air quality models) require outputs from meteorological models to subsequently drive 84 chemical models (Byun and Schere, 2006;ENVIRON, 2008;Seaman, 2000). Comparing to online 85 models, offline models usually are computationally efficient but incapable of capturing two-way 86 feedbacks between chemistry and meteorology (North et al., 2014). Online models or coupled 87 models are designed and developed to consider the two-way feedbacks and attempted to accurately 88 simulate both meteorology and air quality (Briant et  Two-way coupled models can be generally categorized as integrated and access models based on 90 whether using a coupler to exchange variables between meteorological and chemical modules 91 (Baklanov et al., 2014). Currently, there are three representative two-way coupled meteorology and 92 air quality models, namely the Weather Research and Forecasting-Chemistry (WRF-Chem) (Grell 93 et al., 2005), WRF coupled with Community Multiscale Air Quality (CMAQ) (Wong et al., 2012) 94 and WRF coupled with a multi-scale chemistry-transport model for atmospheric composition 95 analysis and forecast (WRF-CHIMERE) . The WRF-Chem is an integrated 96 model that includes various chemical modules in the meteorological model (i.e., WRF) without 97 using a coupler. For the remaining two models, which belong to access model, the WRF-CMAQ 98 uses a subroutine called aqprep (Wong et al., 2012) as its coupler while the WRF-CHEMERE a 99 general coupling software named Ocean Atmosphere Sea Ice Soil-Model Coupling Toolkit (Craig 100 et al. 2017). With more growing interest in coupled models and their developments, applications 101 and evaluations, two review papers thoroughly summarized the related works published before 2008 102 (Zhang, 2008) and (Baklanov et al., 2014. Zhang (2008) overviewed the developments and 103 applications of five coupled models (WRF-Chem; Gas, Aerosol, Transport, Radiation, General 104 Circulation, Mesoscale, and Ocean Model; Community Atmosphere Model verison3; the Model for 105 Integrated Research on Atmospheric Global Exchanges; Caltech unified General Circulation Model) 106 in the United States (US) and the treatments of chemical and physical processes in these coupled 107 models with emphasis on the ACI related processes. Another paper presented a systematic review 108 on the similarities and differences of eighteen integrated or access models in Europe and discussed 109 the descriptions of interactions between meteorological and chemical processes in these models as 110 well as the model evaluation methodologies involved (Baklanov et al., 2014). Some of these coupled 111 models can not only be used to investigate the interactions between air quality and meteorology at 112 regional scales but also at global and hemispheric scales (Grell et al., 2011;Jacobson, 2001;Mailler 113 et al., 2017;Xing et al., 2015a), but large scale studies were not included in the two review papers 114 by Zhang (2008) and Baklanov et al. (2014). These reviews only focused on application and 115 https://doi.org/10.5194/acp-2021-855 Preprint. Discussion started: 28 October 2021 c Author(s) 2021. CC BY 4.0 License. evaluation of coupled models in US and Europe but there is still no systematic review targeting two-116 way coupled model applications in Asia. 117 Compared to US and Europe, Asia has been suffering more severe air pollution in the past three how ARI or ACI influenced climate/meteorology in Asia utilizing observations and climate models. 133 With regard to the impacts of aerosols on cloud, precipitation and climate in East Asia (EA), a 134 detailed review of observations and modeling simulations has also been presented by Li Z. et al. 135 (2019). Since the 2000s, substantial progresses have been made in the climate-air pollution 136 interactions in Asia based on regional climate models simulations, which have been summarized by 137 . Moreover, starting from year of 2010, with the development and availability of two-138 way coupled meteorology and air quality models, more and more modeling studies have been 139 conducted to explore the ARI or/and ACI effects in Asia (Nguyen et  as follows: (1) model-related keywords including "coupled model", "two-way", "WRF", "NU-175 WRF", "WRF-Chem", "CMAQ", "WRF-CMAQ", "CAMx", "CHIMERE" and "WRF-176 CHIMERE"; (2) effect-related keywords including "aerosol radiation interaction", "ARI", "aerosol 177 cloud interaction", "ACI", "aerosol effect" and "aerosol feedback"; (3) air pollution-related 178 keywords including "air quality", "aerosol", "PM2.5", "O3", "CO", "SO2", "NO2", "dust", "BC", 179 "black carbon", "blown carbon", "carbonaceous", "primary pollutants"; (4) meteorology-related 180 keywords including "meteorology", "radiation", "wind", "temperature", "specific humidity", 181 "relative humidity", "planetary boundary layer", "cloud" and "precipitation"; (5) region-related 182 keywords including "Asia", "East Asia", "Northeast Asia", "South Asia", "Southeast Asia", "Far 183 East", "China", "India", "Japan", "Korea", "Singapore", "Thailand", "Malaysia", "Nepal", "North 184 China Plain", "Yangtze River Delta", "Pearl River Delta", "middle reaches of the Yangtze River", 185 "Sichuan Basin", "Guanzhong Plain", "Northeast China", "Northwest China" "East China", "Tibet 186 Plateau", "Taiwan", "northern Indian", "southern Indian", "Gangetic Basin", "Kathmandu Valley". 187 After applying the search engines and the keywords combinations mentioned above, we found 188 943 relevant papers. In order to identify which paper should be included or excluded in this paper, 189 following criteria were applied: (1) duplicate literatures were deleted; (2) studies of using coupled 190 models in Asia with aerosol feedbacks turned on were included, and observational studies of aerosol 191 effects were excluded; (3) publications involving coupled climate model were excluded. According 192 to these criteria, not only regional studies, but also studies using the coupled models at global or 193 hemispheric scales involving Asia or its subregions were included. Then, we carefully examined all 194 the included papers and further checked the listed reference in each paper to make sure that no 195 related paper was neglected. A flowchart that illustrated the detailed procedures applied for article 196 identification is presented in Appendix A (Note: Although the deadline for literature searching is 197 2019, any literature published in 2020 is also included.). There was a total of 157 publications 198 included in our study. 199

Analysis method 200
To summarize the current status of coupled models applied in Asia and quantitatively analyze 201 the effects of aerosol feedbacks on model performance as well as meteorology and air quality, we 202 carried out a series of analyses based on data extracted from the selected papers. We firstly compiled 203 the publication information of the included papers as well as the information regarding model name, 204 simulated time period, study region, simulation design, and aerosol effects. Secondly, we 205 summarized the important findings of two-way coupled model applications in Asia according to 206 different aerosol sources and components to clearly acquire what are the major research focuses in 207 past studies. Finally, we gathered all the simulated results of meteorological and air quality variables 208 with/out aerosol effects and their statistical indices (SI). For questionable results, the quality 209 assurance was conducted after personal communications with original authors to decide whether 210 they were deleted and/or corrected. All the extracted publication and statistical information were 211 exported into an Excel file, which was provided in Supplement Table S1. Moreover, we performed 212 quantitative analyses of the effects of aerosol feedbacks through following steps.
(1) We discussed 213 whether meteorological and air quality variables were overestimated or underestimated based on 214 their SI. Then, variations of the SI of these variables were further analyzed in detail with/out turning 215 on ARI or/and ACI in two-way coupled models.
(2) We investigated the SI of simulation results at 216 different simulation time lengths and spatial resolutions in coupled models.
(3) More detailed inter-217 model comparisons of model performance based on the compiled SI among different coupled 218 models are conducted. (4) Differences in simulation results with/out aerosol feedbacks were grouped 219 by study regions and time scales (yearly, seasonal, monthly, daily and hourly). Toward a better 220 understanding of the complicated interactions between air quality and meteorology in Asia, the 221 results sections in this paper are organized following above analysis methods (1) -(3) and 222 represented in Section 5, and the results following method (4) were represented in Section 6. In 223 addition, Excel and Python were used to conduct data processing and plotting in this study. 224 3 Statistics of published literature 225

Summary of applications of coupled models in Asia 226
A total of 157 articles were selected according to the inclusion criteria, and their basic 227 information was compiled in Table 1. In these studies, two commonly used two-way coupled models 228 were WRF-Chem and WRF-CMAQ, and two locally developed models global-regional assimilation 229 these schemes used in the selected studies were also summarized in Table 1. This table presents a  246 concise overview of coupled models' applications in Asia with the purpose of providing basic 247 information regarding models, study periods and areas, aerosol effects, scheme selections, and 248 reference.

249
It should be noted that in Table 1

275 3.2 Spatiotemporal distribution of publications 276
To gain an overall understanding of applications of coupled models in Asia, the spatial 277 distributions of study areas of the selected literatures and the temporal variations of the annual 278 publication numbers were extracted from Table 1 and summarized. Figure    ice. Su and Fu (2018b) described that the ACI effects of dust had less impacts on the radiative forcing 370 than its ARI effects and dust particles could promote (demote) ice (liquid) clouds in mid-upper (low-371 mid) troposphere over EA. With turning on both ARI and ACI effects of dust, less low-level clouds 372 and more mid-and high-level clouds were detected that contributed to cooling at the Earth's surface 373 and in the lower atmosphere and warming in the mid-upper troposphere (Su and Fung, 2018b).

374
Mineral dust particles transported by the westerly and southwesterly winds from the Middle East 375 (ME) affected the radiative forcing at TOA and the Earth's surface and in the ATM by the dust-376 induced ARI and ACI in the Arabian Sea and the India subcontinent, and subsequently changed the  is worth noting that how to partition dust particles into fine mode and coarse mode or initialize their 424 size distribution in coupled models can affect simulations in many ways and requires more detailed 425 measurements at the source areas and further modeling studies. absorbing aerosols, i.e., BC and brown carbon (BrC)) and aerosols originated from different sources. 453 The major findings in these research are outlined as follows, with respect to the effects of 454 anthropogenic aerosol feedbacks on meteorology and air quality. 455 Concerning the meteorological responses, most papers treated anthropogenic aerosols as a 456 whole to explore their effects on meteorological variables based on coupled model simulations with however until now no study had been performed to simulate the ACI effects of anthropogenic 482 aerosol serving as IN in Asia using two-way coupled models. Therefore, in Asia, further 483 investigations are needed that targeting cloud or/and ice processes involving anthropogenic aerosols 484 (including their size, composition, and mixing state) in two-way coupled models. Meanwhile, 485 several studies not only discussed aerosol feedbacks but also focused on the additional effects of obvious that ARI and ACI effects of different aerosol components are substantially distinctive, and 514 many other aerosol compositions (e.g., sulfate, nitrate and ammonium) besides BC and BrC should 515 be taken into considerations in future modeling studies in Asia.

516
ARB is a common practice in many Asian countries after harvesting and before planting and 517 can deteriorate air quality quickly as one of the most important sources of anthropogenic aerosols, 518 so that it has been attracting much attention among the public and scientists worldwide ( South Asia (SA) (Singh et al., 2020). In general, when ARB occurred, the WRF-Chem simulations 524 from all the studies showed that the changes in radiative forcing induced by ARB aerosols were 525 greater than by those from other anthropogenic sources, especially in the ATM. Also all the modeling 526 studies indicated that ARB aerosols reduced (increased) radiative forcing at the surface (in the ATM), 527 cooled (warmed) the surface (the atmosphere), and increased (decreased) atmospheric stability 528 (PBLH aerosols were smaller over the source region (i.e., SEA) than the downwind region (i.e., SC) with 543 cloudier conditions. Annual simulations regarding the ARI effects of ARB aerosols from SA 544 (especially Myanmar and Punjab) indicated that CAs released by ARB reduced the radiative forcing 545 at the TOA but did not change the precipitation processes much when only the ARI effects were 546 considered in WRF-Chem (Singh et al., 2020). 547 Besides ARB, to our best knowledge, there were only a few research work quantitatively India to aerosols from five emission sectors (power, industry, residential, BB, and transportation), 551 and found that the power (residential) sector was the dominate contributor to the negative (positive) 552 DRF at the TOA over both countries due to high emissions of sulfate and nitrate precursors (BC) 553 and the total sectoral contributions were in the order of power > residential > industry > BB > 554 transportation (power > residential > transportation > industry > BB) for China (India) during 2013. 555 To pinpoint the ARI and ACI effects, Archer-Nicholls et al. (2019) reported that during January 556 2014, the aerosols from the residential emission sector induced larger SWRF (+1.04 W·m -2 ) than 557 LWRF (+0.18 W·m -2 ) at the TOA and their DRF (+0.79 W·m -2 ) was the largest, followed by their 558 semidirect effects (+0.54 W·m -2 ) and indirect effects (-0.29 W·m -2 ) over EC. This study further 559 emphasized a realistic ratio of BC to total carbon from the residential emission was critical for 560 accurate simulations of the ARI and ACI effects with two-way coupled models. 561 In terms of anthropogenic aerosol effects on air quality, the responses of PM2.5 had been widely (2018) indicated that even though the ARI effects had bigger impacts on PM2.5 during wintertime 579 than the ACI effects, the ARI and ACI impacts on PM2.5 were similar during other seasons and the 580 increase of PM2.5 due to the ACI effects was more noticeable in wet season than dry season. Using 581 the process analysis method to distinguish the contributions of different physical and chemical 582 processes to PM2.5 over the NCP area, Chen et al. (2019a) applied WRF-Chem with ARI and ACI 583 and found that besides local emissions and regional transport processes, vertical mixing contributed 584 the most to the accumulation and dispersion of PM2.5, comparing to chemistry and advection, and 585 the ARI effects changed the vertical mixing contribution to daily PM2.5 variation from negative to 586 positive. Regarding surface O3 concentrations, all the two-way coupled models with ARI, ACI, and 587 both ARI and ACI predicted reduced photolysis rate and O3 concentrations under heavy pollution 588 conditions, through the radiation attenuation induced by aerosols and clouds. to simulations without BC, the BC and PBL interaction slowed the O3 growth from late morning to 618 early afternoon somewhat before reaching its maximum value in afternoon due to less vertical 619 mixing in PBL, even though more O3 precursors were trapped in PBL that promoted photochemical 620 reaction of O3. also made the premature death numbers increased by 2 % in the NCP area, comparing to simulations 652 without the PM2.5 data assimilation. In India, WRF-Chem simulations with aerosol feedbacks and 653 updated population data revealed that the premature (COPD related) deaths caused by PM2. Even though there are a certain number of research papers using two-way coupled models to 681 quantify the effects of aerosol feedbacks on regional meteorology and air quality in Asia, model 682 performances impacted by considering aerosol effects varied to some extend. This section provides 683 a summary of model performance comparisons by using the SI (meteorology and air quality as 684 shown in data file (Table C2.xlsx)) collected from the published papers that supplying these indices 685 and being defined as "papers with SI (PSI) (listed in Appendix Tables B2-B3). As aforementioned 686 in Section 3, investigations of ACI effects were very limited and no former studies simultaneously 687 exploring aerosol feedbacks with and without both ARI and ACI turned on. Here, we only compared 688 the SI for simulations with and without ARI in the same study, as summarized in Appendix Tables  689 B4-B5. It should be pointed out that all the reported evaluation results either from individual model 690 or inter-model comparison studies were extracted and put into the Table C2.xlsx file.  Table  702 B2. In these two tables, we also listed the NS of positive (red upward arrow) and negative (blue 703 downward arrow) biases for the meteorological and air quality variables in parentheses in the MB 704 column. Note that NS in Fig. 3e-h and Appendix Table B4 counted the samples of SI provided by 705 the simulations simultaneously with and without ARI. Also the 5th, 25th, 75th and 95th percentiles 706 of SI are illustrated in box-and-whisker plots, and the dashed line in the box is the mean value (not 707 median) and the circles are outliers. 708 The evaluations for T2 (Fig. 3a) from PSI revealed that in Asia coupled models performed 709 rather well for temperature (mean R = 0.90) with RMSE ranging from 0.64 to 5.90 ℃, but 60 % of (limited papers with 12 samples) were improved somewhat and the mean correlation coefficient 720 increased from 0.93 to 0.95 and RMSE decreased slightly (Fig. 3e), but average MB of temperature 721 was decreased from -0.98 to -1.24 ℃. In short, temperatures from PSI or simulations with/without 722 ARI turned on agreed well with observations but were mostly underestimated, and the negative bias 723 of T2 simulated by models with ARI turned on got worse and reasons behind it will be explained in 724 Section 6. Compared with the correlation coefficients of T2, RH2 and SH2, mean R (0.59) of WS10 was 740 smallest with a large fluctuation ranging from 0.14 to 0.98 (Fig. 3d). The meta-analysis also physics. The 5 PSI with ARI effects suggested that the correlation of wind speed was slightly 747 improved (mean R from 0.56 to 0.57) and the average RMSE and positive MB decreased by 0.003 748 m·s -1 and 0.051 m·s -1 , respectively (Fig. 3h). The collected SI indicated relatively poor performance 749 of modeled WS10 (most wind speeds were overestimated) compared to T2 and humidity, but turning 750 on ARI in coupled models could improve WS10 simulations somewhat. 751 752 Figure 3. The quantile distributions of R, MB and RMSE for different meteorological variables from coupled 753 models performance data (a-d) and comparisons of the statistical indices with/out ARI (e-h). 754

Comparisons of SI at different temporal scales for meteorology 755
To probe the model performance of simulated T2, RH2, SH2 and WS2 at different temporal 756 scales, the SI of these meteorological variables from PSI were grouped according to the simulation 757 time (yearly, seasonal, monthly and daily) and plotted in Fig. 4. Note that the seasonal results 758 contained SI values from simulations lasting more than one month and less than or equal to 3 months. 759 Here in Fig. 4, NP and NS were the number of PSI and samples with SI at different time scales, 760 respectively, and also their total values were the same as the ones listed in Appendix Table S2. The 761 correlation between simulated and observed T2 (Fig. 4a)  As shown in Fig. 4e, T2 underestimation mentioned above (Fig. 3a) appeared also in the seasonal, 765 monthly and yearly simulations (average MB = -0.87 ℃, -0.15 ℃ and -0.34 ℃, respectively), but 766 the daily T2 were overestimated (average MB = 0.07 ℃). It should be noted that T2 at the monthly 767 scale was underpredicted mainly during winter months (16 samples). Regarding the mean RMSE, 768 its value (Fig. 4i) at the daily scale was the largest (0.97 ℃) in comparison with that at the other 769 temporal scales. 770 Given that no SI was available for RH2 at the seasonal scale, results at other time scales were 771 discussed here. Figure 4b presented that simulated RH2 at the daily scale had the best correlation 772 coefficient (mean R= 0.74), followed by those at the monthly (0.73) and yearly (0.71) scales. Except 773 overestimation (average MB = 3.6 %) at the yearly scale (Fig. 4f), modeled RH2 were 774 underestimated at the monthly (average MB = -1.1 %) and daily (average MB = -0.2 %) scales, 775 respectively. Therefore, coupled models calculated RH2 reasonably well in short-term simulations. 776 However, at the daily scale, RMSE of modeled RH2 (Fig. 4j)  the yearly scale (Fig. 4k). Also, both over-and under-estimations of modeled SH2 (Fig. 4g) were 783 reported at different time scales with average MB values as 0.15 g·kg -1 , -0.02 g·kg -1 , and -0.14 g·kg -784 1 for yearly, seasonal and monthly simulations, respectively. Generally, the long-term simulations of 785 SH2 agreed better with observations than the short-term ones. 786 As seen in Fig. 4d, the modeled WS10 at the monthly scale (mean R = 0.68) correlated with 787 observations better than that at the daily, yearly and seasonal scales (mean R = 0.62, 0.48 and 0.46, 788 respectively). The simulations at all temporal scales tended to overestimate WS10 comparing 789 against observations (Fig. 4h) and their average MB were 0.80 m·s -1 (seasonal), 0.86 m·s -1 (monthly), 790 0.64 m·s -1 (yearly) and 0.62 m·s -1 (daily), respectively. The short-term simulations of WS10 better 791 matched with observations compared to the long-term ones. At the same time, the largest mean 792 RMSE (1.79 m·s -1 ) of simulated WS10 (Fig. 4l) appeared at the seasonal scale.  generally larger than that involving WRF-CMAQ, except for SH2. 804 As seen in Fig. 5a, the modeled T2 by both WRF-CMAQ and WRF-Chem was well correlated 805 with observations but WRF-CMAQ (mean R = 0.95) outperformed WRF-Chem (mean R = 0.90) to 806 some extent. On the other hand, WRF-CMAQ underestimated T2 (mean MB = -1.39 ℃) but WRF-807 Chem slightly overestimated it (mean MB = 0.09 ℃) (Fig. 5e). The RMSE of modeled T2 by both 808 models was at the similar level with mean RMSE values of 2.51 ℃ and 2.31 ℃ by WRF-CMAQ 809 and WRF-Chem simulations, respectively (Fig. 5i). 810 Both WRF-Chem and WRF-CMAQ performed better for SH2 (mean R = 0.96 and 0.97, 811 respectively) than RH2 (mean R = 0.75 and 0.73, respectively) (Figures 5b and 5c), which might be 812 due to the influence of temperature on RH2 (Bei et al., 2017). Also the modeled RH2 (SH2) by 813 WRF-Chem correlated better (worsen) with observations than those by WRF-CMAQ. The mean 814 RMSE of modeled RH2 (Fig. 5j) by WRF-Chem (11.1 %) was lower than that by WRF-CMAQ 815 (14.3%) but the mean RMSE of modeled SH2 (Fig. 5k) by WRF-Chem (2.25 g·kg -1 ) higher than 816 that by WRF-CMAQ (0.71 g·kg -1 ). It was seen in Figures 5f and 5d that WRF-CMAQ overestimated 817 RH2 and SH2 (average MB were 5.30 % and 0.07 g·kg -1 , respectively), and WRF-Chem 818 underpredicted RH2 (average MB = -0.32 %) and SH2 (average MB = -0.06 g·kg -1 ). Generally, the 819 modeled RH2 and SH2 were reproduced more reasonably by WRF-Chem than those by WRF-820 CMAQ.

821
The modeled WS10 by both WRF-Chem and WRF-CMAQ (Fig. 5d) correlated with 822 observations on the same level with the mean R of 0.56. The mean RMSE of modeled WS10 by 823 WRF-Chem and WRF-CMAQ were 1.54 m·s -1 and 2.28 m·s -1 , respectively, as depicted in Fig. 5l. 824 Both models overpredicted WS10 to some extend with average MBs of 0.55 m·s -1 (WRF-CAMQ) 825 and 0.84 m·s -1 (WRF-Chem), respectively. These results demonstrated that overall WRF-CMAQ 826 and WRF-Chem had similar model performance of WS10. 827 In general, WRF-CMAQ performed better than WRF-Chem for T2 but worse for humidity 828 (RH2 and SH2), and both models' performance for WS10 was very similar. WRF-Chem 829 overestimated T2, RH2 and WS10 and underestimated SH2 slightly, while WRF-CMAQ 830 overpredicted humidity and WS10 but underpredicted T2. Compared to WRF-Chem and WRF-831 CMAQ, the very few SI samples indicated that for the meteorological variables excluding SH2, 832 WRF-NAQPMS simulations matched with observations better than GRAPES-CUACE simulations 833 but more applications and statistical analysis of these two models are needed to make this kind of 834 comparison conclusive.  The results of the overall statistical evaluation for the online air quality simulations are 841 presented in Figure 6, and all labels and colors indicating SI were the same as those for 842 meteorological variables. In Fig. 6a, the correlation between the simulated and observed PM2.5 843

Model performance for air quality variables
concentrations from PSI showed that in Asia coupled models performed relatively well for PM2.5 844 (mean R = 0.63), but RMSE was between -87.60 and 80.90 and more than half of samples of 845 simulated PM2.5 were underestimated (mean MB = -2.08 μg·m -3 ). With the ARI turned on in the 846 coupled models, modeled PM2.5 concentrations (limited papers with 15 samples) were improved 847 somewhat and the mean R slightly increased from 0.71 to 0.72 and mean absolute MB decreased 848 from 4.10 to 1.33 μg·m -3 (Fig. 6c), but RMSE of PM2.5 concentrations slightly increased from 35.40 849 to 36.20 μg·m -3 . In short, PM2.5 with/without ARI agreed well with observations but were mostly 850 underestimated, and PM2.5 bias simulated by models became overpredicted. 851 Compared with PM2.5, mean R (0.59) of O3 was relatively smaller (Fig. 6b). The statistical 852 analysis also showed the most modeled O3 concentrations tended to be overestimated (76 % of the 853 samples) with the average MB value of 8.05 μg·m -3 , and the mean RMSE value was 32.65 μg·m -3 . 854 The 14 PSI with ARI effects suggested that the correlation of O3 was slightly improved (mean R 855 from 0.58 to 0.64) and the average RMSE and MB were decreased by 15.93 μg·m -3 and 1.55 μg·m -856 3 , respectively (Fig. 6d). The collected studies indicated relatively poor performance of modeled O3 857 compared to PM2.5, but turning on ARI in coupled models improved O3 simulations somewhat.   Figure 7 depicted the SI of simulated PM2.5 and O3 at yearly, seasonal, monthly and daily scales. 863
O3 simulation was available at the daily scale, and the RMSE at the yearly scale (Fig. 7f) Figure 8 showed the SI for PM2.5 and O3 from different coupled models, and only WRF-Chem 880

Impacts of aerosol feedbacks in Asia 897
Aerosol feedbacks not only impact the performances of two-way coupled models but also the 898 simulated meteorological and air quality variables to a certain extent. In this section, we collected 899 and quantified the variations (Table C3.xlsx) of these variables induced by ARI or/and ACI from the 900 modeling studies in Asia. Due to limited sample sizes in the collected papers, the target variables 901 only include radiative forcing, surface meteorological parameters (T2, RH2, SH2 and WS10), PBLH, 902 cloud, precipitation, and PM2.5 and gaseous pollutants. 903 6.1 Impacts of aerosol feedbacks on meteorology 904

Radiative forcing 905
With regard to radiative forcing, most studies with two-way coupled models in Asia had 906 focused on the effects of dust aerosols (Dust), BC emitted from ARB (ARB_BC) and anthropogenic 907 sources (Anthro_BC), and total anthropogenic aerosols (Anthro). Figure 9 presents the variations of 908 simulated SWRF and LWRF at the bottom (BOT) and TOA and in the ATM due to aerosol feedbacks. 909 In concentrated in the range of -50.00 to -0.45 W·m -2 . The SWRF variations due to anthropogenic 916 aerosols in the ATM and at the TOA were -2.00 to +120.00 W·m -2 and -6.50 to 20.00 W·m -2 , 917 respectively. There were much less studies reported LWRF variations caused by anthropogenic 918 aerosols, which ranged from -10.00 to +5.78 W·m -2 , -1.91 to +3.94 W·m -2 , and -4.26 to +1.21 W·m -919 2 at the BOT and TOA, and in the ATM, respectively. 920 Considering BC from anthropogenic sources and ARB, they both led to positive SWRF at the induced by aerosol feedbacks in Asia.

934
As shown in Fig. 9, SWRF variations at the BOT caused by total aerosols (sum of Anthro, 935 Anthro_BC, ARB_BC and Dust) had been widely assessed in Asia. Therefore, we further analyzed 936 their spatiotemporal distributions and inter-regional differences, which are displayed in Fig. 10.   Fig. 10  938 are listed in Appendix Table B1) at different time scales. In Asia, almost 41 % of the selected papers 939 investigated SWRF towards its monthly variations, 36 % towards its hourly and daily variations, 940 and 23 % towards its seasonal and yearly variations. Most studies reported aerosol-induced SWRF 941 variations were primarily conducted in NCP, EA, China, and India. At the hourly scale, the range of 942 SWRF decreases was from -350.00 to -5.90 W·m -2 (mean value of -106.92 W·m -2 ) during typical 943 pollution episodes, and significant variations occurred in EA. The daily and monthly mean SWRF 944 reductions varied from -73.71 to -5.58 W·m -2 and -82.20 to -0.45 W·m -2 , respectively, with relative 945 large perturbations in NCP. At the seasonal and yearly scales, the SWRF changes ranged from -946 22.54 to -3.30 W·m -2 and -30.00 to -2.90 W·m -2 with mean value of -11.28 and -11.82 W·m -2 , 947 respectively, with EA as the most researched area. 948 To identify the differences of aerosol-induced SWRF variations between high-(Asia) and low-949 polluted regions (Europe and North America), their inter-regional comparisons are depicted in Fig.  950 10b. This figure does not include information about temporal resolutions of data reporting and 951 durations of model simulations with ARI or/and ACI, but intends to delineate the range of SWRF 952 changes due to aerosol feedbacks. The SWRF variations fluctuated from -233.00 to -0.45 W·m -2 , -953 100.00 to -1.00 W·m -2 , and -600.00 to -1.00 W·m -2 in Asia, Europe, and North America, respectively. 954 It should be pointed out that the two extreme values were caused by dust (-233.00 W·m -2 ) in Asia 955 and wildfire (-600.00 W·m -2 ) in North America. Overall, the median value of SWRF reductions due 956 to ARI or/and ACI in Asia (-15.92 W·m -2 ) was larger than those in North America (-10.50 W·m -2 ) 957 and Europe (-7.00 W·m -2 ). 958 959 960 961 Figure 10. Responses of SWRF to aerosol feedbacks in different areas/periods in Asia (a) and the 962 inter-regional comparisons of SWRF variations among Asia, Europe and North America (b). In the included publications, only a few papers focusing on the effects of aerosol feedbacks on 992 cloud properties (cloud fraction, LWP, ice water path (IWP), CDNC and cloud effective radius) and 993 precipitation characteristics (amount, spatial distribution, peak occurrence and onset time) using 994 two-way coupled models in Asia, as shown in Among all the variables describing cloud properties and precipitation characteristics, the 1013 variations of precipitation amount were studied the most using two-way coupled models in Asia.

1014
How turning on ARI or/and ACI in coupled models can change precipitation amount is not 1015 unidirectional and depends on many factors, including different aerosol sources, areas, emission 1016 levels, atmospheric humidity, precipitation types, seasons, and time of a day. Under the high (low) 1017 emission levels as well as at slightly different humidity levels of RH > 85 % (RH < 80 %) with 1018 increasing emissions, the ACI effects of anthropogenic aerosols increased (decreased) precipitation 1019 in the MRYR area of China. In PRD (SK), wintertime (summertime) precipitation was enhanced 1020 (enhanced and inhibited) by the ACI effects of anthropogenic aerosols but inhibited (enhanced and 1021 inhibited) by ARI. In locations upwind (downwind) of Beijing, rainfall amount was raised (lowered) 1022 by the ARI effects of anthropogenic aerosols but lowered (raised) by ACI. Both ARI and ACI 1023 induced by anthropogenic aerosols had positive impacts on total, convective, and stratiform rain in 1024 India during the summer season and the increase of convective rain was larger than those of 1025 stratiform. Summertime precipitation amounts could be enhanced or inhibited at various subareas 1026 inside simulation domains over India, China, and Korea and during day-or night-time due to ARI 1027 and ACI of anthropogenic aerosols. Over China, dust-induced ACI decreased (increased) springtime 1028 precipitation in CC (western part of NC), and over India, dust aerosols from local sources and ME 1029 had positive impacts on total, convective, and stratiform rain through ARI and ACI. Simulations in 1030 India also revealed that precipitation could be increased in some subareas but decreased in another 1031 and absorptive (non-absorptive) dust enhanced (inhibited) summertime precipitation via ARI and 1032 ACI. The ARI (ACI) effects of BC from ARB caused precipitation reduction (increase) in SEC but 1033 CAs emitted from ARB (ARB_CAs) caused rainfall enhancement in Myanmmar. During pre-1034 monsoon (monsoon) season, ARI induced by anthropogenic BC could lead to +42 % (-5 to -8 %) 1035 variations of precipitation in NEI (SI

Impacts of aerosol feedbacks on air quality 1047
Aerosol effects not only gave rise to changes in meteorological variables but also air quality. 1048 Table 5 (the minimum, maximum and mean values were defined in the same way as in Table 3 Figure 11. A schematic diagram depicting aerosol-radiation-cloud interactions and quantitative 1090 effects of aerosol feedbacks on meteorological and air quality variables simulated by two-way 1091 coupled models in Asia.

1092
Two-way coupled models are proved to be a valuable tool for investigating the aerosol-1093 radiation-cloud interactions. These models have been applied in US and Europe extensively and 1094 then in Asia due to frequent occurrences of severe air pollution events accompanied with rapid 1095 economic growth in the region. Until now, no comprehensive study is conducted to elucidate the 1096 recent advances in two-way coupled models' applications in Asia. This paper provides a critical 1097 overview of current status and research focuses of related modeling studies using two-way coupled 1098 models in Asia between 2010 and 2019, and summarizes the effects of aerosol feedbacks on 1099 meteorological and air quality variables from these studies. 1100 Through systematically searching peer-reviewed publications with several scientific-based 1101 search engines and a variety of key word combinations and applying certain selection criteria, 157 1102 relevant papers were identified. Our bibliometric analysis results (as schematically illustrated in Fig.  1103 11) showed that in Asia, the research activities with two-way coupled models had increased 1104 gradually in the past decade and the four mainstream two-way coupled models (WRF-Chem, WRF-1105 CMAQ, WRF-NAQPMS and GRAPES-CUACE) were extensively utilized to explore the effects in 1106 Asia with focusing on several high aerosol loading areas (e.g., EA, India, China and NCP) during 1107 wintertime or/and server pollution events, with less investigations looking into other areas and 1108 seasons with low pollution levels. Among the 157 papers, nearly 84 % of them focused on ARI (72 1109 papers) and both ARI and ACI effects (60 papers), but papers that only considering ACI effects were 1110 relatively limited. The ARI or/and ACI effects of natural mineral dust, BC and BrC from 1111 anthropogenic sources and BC from ARB were mostly investigated, while a few studies 1112 quantitatively assessed the health impacts induced by aerosol effects. Turning on aerosol feedbacks 1113 in two-way coupled models impacted the model performance differently in regard to models, 1114 simulation time periods and areas, meteorological and air quality variables, and ARI or/and ACI 1115 effects. Compared to US and Europe, the aerosol-induced decrease of the shortwave radiative 1116 forcing was larger due to higher air pollution levels in Asia. For other meteorological and air quality 1117 variables, the overall decrease (increase) of T2, WS10, PBLH and O3 (RH2, PM2.5 and other gaseous 1118 pollutant concentrations) caused by ARI or/and ACI effects were reported from the modeling studies 1119 using two-way coupled models in Asia. The ranges of aerosol-induced variations of T2, PBLH, 1120 PM2.5 and O3 concentrations were larger than other meteorological and air quality variables.

1121
Even though noticeable progresses toward the application of two-way coupled meteorology 1122 and air quality models have been made in Asia and the world during the last decade, there are still 1123 https://doi.org/10.5194/acp-2021-855 Preprint. Discussion started: 28 October 2021 c Author(s) 2021. CC BY 4.0 License. many pressing issues facing the modeling community. The latest advances in the measurements and 1124 research of cloud properties, precipitation characteristics, and physiochemical characteristics of 1125 aerosols (e.g., origin, morphology, size distribution, optical property, hygroscopicity, mixing state, 1126 and chemical composition) that play pivotal roles in CCN or IN activation mechanisms can guide 1127 the improvements and enhancements in two-way coupled models, especially to abate the 1128 uncertainties in simulated ACI effects. At the same time, computational costs should be considered 1129 with any new/enhanced parameterization schemes concerning ACI related processes, since running 1130 two-way coupled models is more expensive than running models without turning on feedbacks. 1131 Further inter-comparisons of multiple coupled models need to be conducted in Asia and other 1132 regions to comprehensively assess the model performances with/without aerosol feedbacks and how 1133 ARI or/and ACI affect meteorology and air quality. Besides the four two-way coupled models 1134 mentioned in this paper, more models capable of simulating aerosol feedbacks (such as WRF-1135 CHIMERE and WRF-GEOS-Chem) have become available and should be included in future inter-1136 comparisons. Future assessments of the ARI or/and ACI effects should pay extra attention to their 1137 impacts on dry and wet depositions simulated by two-way coupled models. So far, the majority of 1138 two-way coupled models' simulations and evaluations focuses on episodic air pollution events 1139 occurring in certain areas, therefore their long-term applications and evaluations are necessary and 1140 their real-time forecasting capabilities should be explored as well.   Note that the No.* is consistent with the No. in Table 1, and ↑ and ↓ mark over-and underestimations of variables, respectively, along with 1153 their number of samples.