Application of WRF/Chem over North America under the AQMEII Phase 2 – Part 2: Evaluation of 2010 application and responses of air quality and meteorology–chemistry interactions to changes in emissions and meteorology from 2006 to 2010

The Weather Research and Forecasting model with Chemistry (WRF/Chem) simulation with the 2005 Carbon Bond (CB05) gas-phase mechanism coupled to the Modal for Aerosol Dynamics for Europe (MADE) and the volatility basis set approach for secondary organic aerosol (SOA) are conducted over a domain in North America for 2006 and 2010 as part of the Air Quality Model Evaluation International Initiative (AQMEII) Phase 2 project. Following the Part 1 paper that focuses on the evaluation of the 2006 simulations, this Part 2 paper focuses on a comparison of model performance in 2006 and 2010 as well as analysis of the responses of air quality and meteorology– chemistry interactions to changes in emissions and meteorology from 2006 to 2010. In general, emissions for gaseous and aerosol species decrease from 2006 to 2010, leading to a reduction in gaseous and aerosol concentrations and associated changes in radiation and cloud variables due to various feedback mechanisms. WRF/Chem is able to reproduce most observations and the observed variation trends from 2006 to 2010, despite its slightly worse performance than WRF that is likely due to inaccurate chemistry feedbacks resulting from less accurate emissions and chemical boundary conditions (BCONs) in 2010. Compared to 2006, the performance for most meteorological variables in 2010 gives lower normalized mean biases but higher normalized mean errors and lower correlation coefficients. The model also shows poorer performance for most chemical variables in 2010. This could be attributed to underestimations in emissions of some species, such as primary organic aerosol in some areas of the US in 2010, and inaccurate chemical BCONs and meteorological predictions. The inclusion of chemical feedbacks in WRF/Chem reduces biases in meteorological predictions in 2010; however, it increases errors and weakens correlations comparing to WRF simulations. Sensitivity simulations show that the net changes in meteorological variables from 2006 to 2010 are mostly influenced by changes in meteorology and those of ozone and fine particulate matter are influenced to a large extent by emissions and/or chemical BCONs and to a lesser extent by changes in meteorology. Using a different set of emissions and/or chemical BCONs helps improve the performance of individual variables, although it does not improve the degree of agreement with observed interannual trends. These results indicate a need to further improve the accuracy and consistency of emissions and chemical BCONs, the representations of SOA and chemistry–meteorology feedbacks in the online-coupled models.


Introduction
Changes in meteorology, climate, and emissions affect air quality (e.g., Hogrefe et al., 2004;Leung and Gustafson, 2005;Zhang et al., 2008;Dawson et al., 2009;Gao et al., 2013;Penrod et al., 2014).As federal, state, and local environmental protection agencies enforce the anthropogenic emission control programs, ambient air quality is expected to be continuously improved.However, such an improvement may be compensated by adverse changes in climatic or meteorological conditions (e.g., increases in near-surface temperature, solar radiation, and atmospheric stability, or reductions in precipitation) that are directly conducive to the formation and accumulation of air pollutants and that may result in higher biogenic emissions.It is therefore important to examine changes in meteorology/climate and emissions as well as their combined impacts on air quality.The Air Quality Model Evaluation International Initiative (AQMEII) Phase 2 was launched in 2011 to intercompare online-coupled air quality models (AQMs) in their capabilities in reproducing atmospheric observations and simulating air quality and climate interactions in North America (NA) and Europe (EU) (Alapaty et al., 2012).The simulations over NA and EU with multimodels by a number of participants have been performed for 2 years (2006 and 2010) that have distinct meteorological conditions.Compared with 2006, 2010 is characterized by warmer summer conditions in the eastern US and less precipitation over NA (Stoeckenius et al., 2015;Pouliot et al., 2014).In addition, the emissions of key pollutants are reduced in 2010 relative to 2006, e.g., emissions of oxides of nitrogen (NO x ) and sulfur dioxide (SO 2 ) are reduced by 10-30 and 40-80 % for many regions in NA (Pouliot et al., 2014).Comparison of 2010 and 2006 simulations will thus provide an opportunity to examine the success of the emission control programs and the impacts of meteorological/climatic variables on air quality.Compared to model intercomparison during AQMEII Phase 1 (Rao et al., 2012) in which offline-coupled models were used, the use of onlinecoupled AQMs during AQMEII Phase 2 allows for study of the interactions between meteorology and chemistry through various direct and indirect feedbacks among aerosols, radiation, clouds, and chemistry (Zhang, 2008;Baklanov et al., 2014).The 2-year simulations further enable an examination of the responses of air quality and meteorology-chemistry interactions to changes in emissions and meteorology from 2006 to 2010 that was not possible with offline-coupled models.
Similar to offline AQMs, large uncertainties exist in online-coupled AQMs, which will affect the model predictions and implications.Such uncertainties lie in the meteorological and chemical inputs such as emissions, initial and boundary conditions (ICONs and BCONs), model representations of atmospheric processes, and model configurations for applications such as horizontal/vertical grid resolutions and nesting techniques.Several studies examined the uncertainties in emissions (e.g., Reid et al., 2005;Zhang et al., 2014) and BCONs (e.g., Hogrefe et al., 2004;Schere et al., 2012).There are also uncertainties in various chemical mechanisms and physical parameterizations used in AQMs such as gas-phase mechanisms (Zhang et al., 2012), aerosol chemistry and microphysical treatments (Zhang et al., 2010), microphysical parameterizations (van Lier-Walqui et al., 2014), convective parameterizations (Yang et al., 2013), boundary layer schemes (Edwards et al., 2006), and land surface models (Jin et al., 2010).Due to the complex relationships in online-coupled AQMs among the emissions, ICONs and BCONs, and model processes that may be subject to inherent limitations, it is difficult to isolate the contributions of model inputs or the representations of atmospheric processes to the model biases.In mechanistic evaluation (also referred to as dynamic evaluation), sensitivity simulations are performed by changing one or a few model inputs or process treatments, while holding others constant.This approach can help diagnose the likely sources of biases in the model predictions.
The Weather Research and Forecasting model with Chemistry (WRF/Chem) version 3.4.1 with the 2005 Carbon Bond (CB05) gas-phase mechanism coupled with the Modal for Aerosol Dynamics for Europe (MADE) and the volatility basis set (VBS) approach for secondary organic aerosol (SOA) (hereafter WRF/Chem-CB05-MADE/VBS) has been recently developed by Wang et al. (2014).The WRF/Chem-CB05-MADE/VBS has been coupled to the aqueous-phase chemistry scheme (AQChem) based on the AQChem version in CMAQ v5.0 of Sarwar et al. (2011) for both largescale and convective clouds (Wang et al., 2014).WRF/Chem-CB05-MADE/VBS also contains heterogeneous chemistry involving sulfur dioxide on the surface of aerosols based on Jacob (2000) and treats both aerosol direct and indirect effects.The applications of WRF/Chem-CB05-MADE/VBS to 2006 and 2010 in this work use the same model physical and chemical parameterizations as those in the Part 1 paper of Yahya et al. (2014) but with different emissions, meteorological ICONs and BCONs, and chemical ICONs and BCONs.The mechanistic evaluation by comparing WRF/Chem-CB05-MADE/VBS predictions for the 2 years would help in understanding the sensitivity of the model predictions and performance to different model inputs, and that by comparing WRF/Chem-CB05-MADE/VBS and WRFonly predictions would quantify the impacts of chemistrymeteorology feedbacks on the meteorological predictions.A comprehensive evaluation of the 2006 simulation has been presented in the Part 1 paper of Yahya et al. (2014).In this Part 2 paper, the differences in emissions, meteorological and chemical ICONs/BCONs, and meteorology between 2010 and 2006 are first examined briefly.The model performance in 2010 is then evaluated and compared with that in 2006.Finally, the responses of air quality and meteorologychemistry interactions to changes in emissions, chemical ICONs/BCONs, and meteorology individually and collectively from 2006 to 2010 are analyzed.The main objectives of this Part 2 paper are to examine whether the model has the ability to consistently reproduce observations for two separate years, as well as to examine whether the trends in air quality and meteorology-chemistry interactions are consistent for both years.Stoeckenius et al. (2015) carried out an extensive analysis of the trends in emissions and observations of meteorological variables, O 3 , SO 2 , and PM 2.5 concentrations between 2006 and 2010.This Part 2 paper complements the work of Stoeckenius et al. (2015) by examin-ing the changes in WRF/Chem predictions and chemistrymeteorology feedbacks in 2010 relative to 2006.Similar evaluations of 2010 and 2006 are performed for the coupled Weather Research and Forecasting -Community Multiscale Air Quality (WRF-CMAQ) system (Hogrefe et al., 2014).Unlike the coupled WRF-CMAQ system used in AQMEII Phase 2 that only simulates aerosol direct effects, WRF/Chem used in this work simulates both aerosol direct and indirect effects.In addition, the work by Hogrefe et al. (2014) involves nudging of temperature, wind speed, water vapor mixing ratio, soil temperature and soil moisture, while the model used for this study did not include any nudging.
2 Differences in emissions and ICONs/BCONs between 2006 and 2010

Emission trends
The emission variation trends are examined for major precursors for ozone (O 3 ) and secondary particulate matter (PM) (i.e., sulfur dioxide (SO 2 ), oxides of nitrogen (NO x ), ammonia (NH 3 ), volatile organic compounds (VOCs) including both anthropogenic and biogenic VOCs) and primary PM species (elemental carbon (EC) and primary organic aerosol or carbon (POA or POC)).As shown in Table S1 in the Supplement, emissions of most species decrease from 2006 to 2010 with domainwide averages of −10 to −24 %.Comparing to emissions in 2006, the annual emissions of SO 2 and NO x decrease significantly in 2010, especially at the point sources (Fig. S1 in the Supplement), with similar variation patterns in all seasons (figure not shown).The annual emissions of NH 3 decrease over most areas but increase in some areas in California (CA) and the midwest.Unlike the changes in the emissions of SO 2 and NO x , NH 3 and VOC emissions exhibit strong seasonal variations in the emission trends, as shown in Fig. S2.Although anthropogenic VOC emissions decrease over the continental US (CONUS) for all seasons (figure not shown), the VOC emissions increase in the southeast, which is dominated by enhanced biogenic emissions from vegetation as a response to temperature increases (Stoeckenius et al., 2015).The total annual emissions of EC and POA also decrease but to a smaller extent over most areas of the continental US.The changes in annual and seasonal emissions of those species between 2010 and 2006 will affect simulated air quality and meteorologychemistry interactions.In addition, there exist uncertainties in the NEI (National Emissions Inventory) emissions.The major sources of uncertainties or errors in the NEI emissions include (1) the emissions calculated using a bottom-up approach based on information provided by individual state, local, and tribal air agencies; and (2) improvements in emission estimation methodology over the years which may result in inconsistencies between years of NEI data (Xing et al., 2013).These will affect the accuracy of the model simulations.

Differences in chemical and meteorological ICONs/BCONs
Large differences exist in the chemical and meteorological ICONs/BCONs used in the simulations.For example, Stoeckenius et al. (2015) reported that the mid-tropospheric seasonal mean O 3 mixing ratios are generally lower by several ppbs in 2010 as compared to 2006, especially during spring and summer.Less Asian mid-tropospheric fine dust was also transported over to the US in the spring of 2010 and less African dust reached the US in the summer of 2010 (Stoeckenius et al., 2015).As shown in Fig. S3, significant differences exist for January, February, and December (JFD)   (Chen, 2007).Pleim and Gilliam (2009)  For precipitation, the model performs consistently well against GPCC for both years with seasonal NMBs within −11 and −12 %, and annual NMBs of 0.3 and 1.3 %, respectively, for 2006 and 2010.The evaluation against NADP shows larger differences with NMBs of 22.2 and 2.5 % and Corr values of 0.43 and 0.1 for 2006 and 2010, respectively.As compared to other meteorological variables such as T2, SWDOWN, and WS10, the meteorological performance for precipitation does not follow a clear trend for all seasons or years against NADP and GPCC.For example, precipitation in JJA is underpredicted against NADP and GPCC for 2010 but this is not the case for 2006.In general, the reported biases in precipitation simulated by WRF from literature are significant.For example, Wang and Kotamarthi (2014) studied the precipitation behavior in WRF and showed that even with nudging, the precipitation biases remained up to a root mean square error (RMSE) of 62.5 % due to inherent weaknesses in the microphysics and cumulus parameterization schemes.Similarly, WRF/Chem gives large seasonal mean biases (up to 44 % in 2006 and up to −26 % in 2010) for simulated precipitation for most seasons in 2006 or 2010, although the annual mean biases are small to moderate (with NMBs of −2.2 to −1.3 % against GPCC and of 9.7-17.6% against NADP in both years).Yahya et al. (2014) compared and evaluated the full-year WRF and WRF/Chem 2006 simulations with the same physical configurations to analyze the effects of feedbacks from chemistry to meteorology.The results for 2006 show that for the evaluation of SWDOWN, T2, and WS10 against CASTNET and SEARCH, the Corr is almost identical for both WRF/Chem and WRF simulations.For evaluation of precipitation against NADP, WRF has a higher Corr compared to WRF/Chem.Unlike 2006, the 2010 WRF-only simulation has higher Corr values for all meteorological variables compared to the 2010 WRF/Chem simulation except for Precip against GPCC and CF against MODIS.This means that the emissions and chemistry-meteorological feedbacks play an important role in influencing model performance.Section 4.4 will explore this in further detail.Another obvious difference is that the NMBs for the meteorological variables for 2010 are smaller compared to 2006 for all the variables except for Precip against GPCC, while the NMEs are larger for 2010 compared to 2006 for all variables except for Precip against GPCC.A smaller overall averaged NMB but a larger NME may indicate compensation of over-and underpredictions leading to a small bias, but the magnitude of the differences are reflected in the NME values.
The same model physics and dynamics options are used for both years.In addition to different emissions, there are characteristic climate differences between the 2 years that lead to lower Corr values and larger NMEs for most meteorological fields in 2010 compared to 2006 for both WRF and WRF/Chem simulations.The year 2010 is reported to be the warmest year globally since 1895 according the National Climactic Data Center (NCDC) (http://www.ncdc.noaa.gov/cag/).Even though 2010 has high temperatures compared to previous years, a trend analysis of extreme heat events (EHE) from 1930 to 2010 showed that in 2010 there were more than 35 extreme minimum heat events (where temperatures are extremely low) over the southeastern US compared to about ∼ 10 events in 2006.In fact, the number of extreme minimum heat events is the highest overall for CONUS in 2010 compared to all the other years from 1930 onwards (Oswald and Rood, 2014).The Intergovernmental Panel for Climate Change (IPCC) reported that, since 1950, weather events have become more extreme likely due to climate change (IPCC, 2012).Grundstein and Dowd (2011) stated that, on average, by 2010 there would be 12 more days with extreme apparent temperatures than those in 1949.These studies imply that increased temperatures change the weather in unexpected ways with uncertainties in the state of science (Huber and Gulledge, 2011), including models.These high and low temperatures could contribute to the compensation of over-and underpredictions leading to smaller NMBs in general for 2010.To better simulate model extreme heat events, Meir et al. (2013) suggested using a higher spatial resolution with a grid size of 12 km or smaller, better sea surface temperature estimates, and enhanced urbanization parameterization.Gao et al. ( 2012) reported better results in reproducing extreme weather events with WRF over the eastern US at a 4 km × 4 km resolution.In this study, although the urban canopy model is used for both WRF and WRF/Chem simulations, a 36 km × 36 km grid resolution might not be sufficient to reproduce the extreme temperature events (highs and lows) in 2010.
As shown in Fig. S4, the spatial distribution of MB (mean bias) values for T2 for JFD 2010 by WRF/Chem show very large negative MBs over the southeastern US compared to JFD 2006.T2 is also generally underpredicted over the southeastern US in both years but with larger negative biases in 2010 than those in 2006.T2 biases also seem to be more extreme for JFD 2010 compared to JFD 2006, with dark red and dark blue colors for the MB markers, indicating large positive and large negative biases, respectively.This could explain the poorer correlation for T2 in 2010 compared to 2006 as shown in Table 1.On the other hand, the performances of T2 for JJA 2010 and 2006 are very similar, with MBs of ∼ −0.1 to 0.1 • C in the eastern US, large negative MBs at the sites in Montana and Colorado, and a large positive MB at the site in Wyoming.

Differences in chemical predictions for 2006 and 2010
The  According to Table 1 and Fig 2014), the inclusion of aerosol indirect effects also tends to reduce O 3 mixing ratios, comparing to the models that simulate aerosol direct effect only or do not simulate aerosol direct and indirect effects (i.e., offline-coupled models).
Figure 5 shows the spatial distribution of NMBs for PM  Figure 8 shows the time series plots for 24 h average concentrations of PM 2.5 , SO 2− 4 and NO − 3 against STN for 2006 and 2010.In 2006, the daily-average PM data were collected on a daily basis in 2006 but every 3 days in 2010.The model is able to predict most of the observed peaks and troughs for 2006 even though the observed and simulated magnitudes are significantly different for several days.For 2010, the model does not show large spikes and can reproduce the magnitudes well, although it does not predict the peaks and troughs as well as 2006 for some months (e.g., January-March and July-September for PM 2.5 ).This could be at- cuss in further detail the role of emissions, meteorology and chemical ICONs/BCONs on O 3 and PM 2.5 concentrations.

SOA evaluation for 2006 and 2010
The VBS framework in WRF/Chem of Ahmadov et al. (2012) provides a more realistic treatment of SOA compared to previous SOA treatments such as the 2product model by Odum et al. (1996)  The RTP site is located in a semi-rural area.Pasadena, CA, is located about 11 mi.from downtown Los Angeles (LA), and Bakersfield, CA, is located about ∼ 100 mi.from downtown LA.Both sites are classified as urban/industrial sites.OC concentrations were measured using an automated, semicontinuous elemental carbon-organic carbon (EC-OC) instrument.The observed SOA masses were determined from organic tracers extracted from filter samples (Lewandowski et al., 2013).The simulated OC concentration is calculated by summing up SOA and POA, and dividing the TOA by 1.4 (Aitken et al., 2008).
As shown in Figs. 9 and S5, the model overpredicts SOA but underpredicts OC at RTP in 2006 because (1) the SOA formed from alkanes and alkenes is excluded in the observations from RTP but simulated in WRF/Chem, and (2) WRF/Chem may have overestimated the aging rate co-efficient for both anthropogenic and biogenic surrogate VOC precursors (Wang et al., 2014).The SOA overprediction due to those reasons compensates the underprediction in SOA due to omission of SOA from POA, leading to a net SOA overprediction at RTP in 2006.By contrast, the VBS underpredicts SOA in 2010 with NMBs of −55.3 and −75.3 % at Bakersfield and Pasadena, respectively, which is mainly due to the omission of SOA formation from POA in the current VBS-SOA module in this version of WRF/Chem.As shown in Fig. S6, SOA to OC ratios at RTP in 2006 are in the range of 50-80 %, whereas they are < 20 % at Bakersfield, CA, and < 40 % Pasadena, CA, in 2010.This indicates that neglecting SOA formation from POA would have a much larger impact on SOA predictions at the two CA sites in 2010 than at RTP in 2006, due to the dominancy of POA in TOA at the two CA sites.As shown in Fig. 9, the model underpredicts OC at RTP in 2006 and significantly underpredicts OC at the two sites in CA in 2010.The differences in OC performance in both years are caused by different locations (i.e., RTP in 2006 and the two CA sites in 2010) that have different ratios of POC to OC as mentioned previously.OC performance thus largely depends on SOA performance at RTP but on POA performance at the two sites in CA.This is why the OC performance remains poor despite a relatively good performance in SOA at the two sites in CA.A poorer OC performance over the two CA sites in 2010 may also indicate potentially large underestimation of POA emissions over the western US.

Differences in aerosol-cloud predictions for 2006 and 2010
Figure  basis.The meteorological and chemical variables shown earlier are evaluated based on site-specific, and hourly, daily, or weekly data.

Differences in observed and simulated trends between 2010 and 2006
Table 2   decrease in O 3 mixing ratios from the ICONs and BCONs (Stoeckenius et al., 2015).The IMPROVE-observed EC concentrations decreased by ∼ 22 % from 2006 to 2010; however, WRF/Chem shows a small increase (by ∼ 2 %).For PM 2.5 concentrations, the simulated decrease from 2006 to 2010 by WRF/Chem is larger than the observed decrease for both STN and IMPROVE.Similar steeper decreases by WRF/Chem also occur for SO 2− 4 against STN, NO − 3 against IMPROVE, TC against STN, and OC against IMPROVE likely due to the influence of ICONs/BCONs and emissions.

Responses of 2010 predictions to changes in emissions and meteorology
The changes in emissions, boundary conditions, and meteorology between 2010 and 2006 lead to changes in simulated air quality and the chemistry-meteorology feedbacks, which in turn change meteorological and air quality predictions during the next time step.

Air quality predictions
Simulated air quality responds nonlinearly to the changes in emissions.Figures 11, show the seasonal changes between 2010 and 2006 in ambient mixing ratios of gases (SO 2 , NO 2 , NH 3 , O 3 , and hydroxyl -OH) and concentrations of PM species (SO 2− 4 , NO − 3 , NH + 4 , organic matter or OM, EC, POA, anthropogenic SOA or ASOA, biogenic SOA or BSOA, and PM 2.5 ).SO 2 and NO 2 concentrations tend to decrease for all seasons at most locations over CONUS due to the decrease in their emissions.The increases in NO 2 concentrations over urban areas in the eastern US in March, April, May (MAM) in 2010 relative to 2006 could be due to a few reasons including decreased photolytic conversion from NO 2 to NO due to a decrease in SWDOWN and less NO 2 conversion to nitric acid (HNO 3 ) due to decreased OH concentrations.The NO 2 hot spots also correlate to the decrease in hourly O 3 concentrations in urban areas.This could indicate an increased titration of nighttime O 3 by NO.This is an important result for policy implications, as reducing NO x emissions may reduce NO 2 concentrations overall for CONUS, but may not reduce NO 2 concentrations in several areas, especially in urban areas due to a combination of titration and complex interplay with local meteorology.NH 3 mixing ratios generally decrease in the US, except over the eastern US in MAM and September, October, and November (SON), where there are increases.NH 3 emissions decrease, however, over the eastern US in all seasons.The increase in NH 3 concentrations in MAM and SON could be attributed to a number of reasons including less NH 3 conversion to NH + 4 to neutralize SO 2− 4 and NO − 3 and less dispersion of NH 3 concentrations due to decreased wind speeds over the eastern and southeastern US in MAM andSON, respectively, in 2010 compared to 2006. In JJA andSON, high OM concentrations in Canada are attributed to the enhanced impacts of BCONs by increasingly convergent flow in this region.OM is made up of both POA and SOA. An increase in VOC emissions in the eastern US in MAM and SON leads to increases in OM concentrations.Decreases in VOC emissions in the western U.S. for all seasons lead to decreases in OM concentrations.The OM concentrations in some areas, however, do not follow a linear relationship with VOC emissions, such as for the southeastern US in JJA, where VOC emissions increase from 2006 to 2010 but OM concentrations decrease.A decrease in POA concentrations must dominate the overall decrease in OM concentrations, even under increased temperatures and biogenic VOC emissions in this area.PM 2.5 concentrations decrease for all seasons and most regions of the CONUS, which is attributed mainly to decreases in precursor gases, especially the inorganic precursors SO 2 and NO x in the eastern US.Increased PM 2.5 concentrations in JFD and MAM in the Midwest are due to surface temperature decreases, which are dominating in this region (Stoeckenius et al., 2015).This in turn leads to increased particle nitrate concentrations (Campbell et al., 2014).12, the decrease in SWDOWN from 2006 to 2010 is larger over the north-central and northwestern US and the increase in SWDOWN is smaller over the northeastern and southwestern US for MAM by WRF/Chem compared to MAM by WRF.For SON, the increase in SWDOWN from 2006 to 2010 simulated by WRF/Chem is larger over the eastern US than that by WRF.The differences between WRF and WRF/Chem are the largest for SWDOWN over the northeastern US in JFD with an increase in SWDOWN simulated by WRF but a decrease simulated by WRF/Chem from 2006 to 2010.The differences in SWDOWN are likely due to the differences in CF between the two sets of simulation pairs, as the spatial distribution for CF is consistent with that of SWDOWN.As expected, there are slight differences between T2 and PBLH between WRF and WRF/Chem  due to changes in radiation.There are also small differences between precipitation between WRF and WRF/Chem.The aerosol-cloudradiation feedbacks due to the differences between WRF and WRF/Chem for 2010 will be discussed in Sect.4.3.

Meteorological predictions
The increase in SWDOWN from 2006 to 2010 does not necessarily translate to an increase in T2.However, in gen- eral, increases in SWDOWN lead to increases in T2, as shown in SON in Fig. 12, where SWDOWN generally increases over most of the continental US, T2 also increases over most of CONUS.In general, the largest differences in T2 between 2006 and 2010 occur in SON (increase) and JFD (decrease).The decrease in T2 in JFD in the north-central US and parts of Canada is significant as it results in a decrease in WS10 and PBLH.For JJA, there is an obvious pattern between SWDOWN and Precip, with an increase in SWDOWN corresponding to a decrease in Precip and vice versa.According to the IPCC (2007), in the warm seasons over land, strong negative correlations dominate as increased sunshine results in less evaporative cooling.Figure S12 compares wind vectors superposed with T2 in 2006 and 2010 from WRF/Chem and shows the largest differences are in JJA.
As expected, the spatial pattern of SWDOWN changes is anti-correlated with CF changes for all seasons between 2006 and 2010; however, the changes in the spatial pattern of CF do not correlate with changes in CDNC.CF in each grid cell is set to either 0 (no clouds) or to 1 (cloudy) if the total cloud water + ice mixing ratio > 1 × 10 −6 kg kg −1 (Wu and Zhang, 2005).In this study, the monthly CF is then normalized over the total number of time steps and vertical layers, giving a value of CF between 0 and 1 in each grid cell.In contrast, the calculations of CDNC in the model depend on the supersaturation, aerosol concentrations, aerosol hygroscopicity and updraft velocity (Abdul-Razzak and Ghan, 2004).The changes in CF are controlled by large-scale state variables including temperature and relative humidity, while CDNC depends on more complex changes in microphysical variables.The dominant CDNC decrease in MAM, JJA, and SON, is due to lower PM 2.5 concentrations, which in turn lower the effective number of cloud condensation nuclei.However, an exception occurs in the southeastern US where PM 2.5 decreases but CDNC increases.This is because CDNC also depends on other variables including the amount of liquid water in the atmosphere.The cloud liquid water path over the southeastern US increases, which may explain the increase in CDNC.The spatial pattern for precipitation correlates to that of CF.The spatial pattern of CWP also corresponds to a certain extent with CF.PBLH increases when the ground warms up during the day and decreases when the ground cools, so PBLH might be intuitively related to SWDOWN and T2.However, this consistent trend is now obvious in the plots, because the simulated growth of the planetary boundary later (PBL) also depends on the surface sensible latent and heat fluxes and the entrainment of warmer air from the free troposphere (Chen, 2007).

Meteorology-chemistry feedback predictions
As shown in Table 1, similar to 2006, comparison of the performance of most meteorological variables between WRF/Chem and WRF for 2010 is improved in terms of NMBs when chemistry-meteorology feedbacks are included.This indicates the importance and benefits of inclusion of such feedbacks in online-coupled models.However, unlike 2006 for which both WRF-only and WRF/Chem sim- ulations show similar values of Corr and NMEs, the 2010 WRF simulations give higher Corr and lower NME values than the 2010 WRF/Chem simulations.This indicates the impact of poorer chemical predictions on chemistrymeteorology feedbacks that can in turn affect meteorological predictions.These results indicate the need of further improvement of the online-coupled models in their representations of chemistry-meteorology feedbacks.Yahya et al. (2014) analyzed differences in meteorological performance between WRF/Chem and WRF for 2006.Figure S13 shows absolute seasonal differences between the meteorological predictions from WRF/Chem and WRF for 2010.The differences between WRF/Chem and WRF are consistent for both 2006 and 2010.SWDOWN in general is higher for WRF/Chem compared to WRF for all seasons, with larger differences over the eastern portion of the domain compared to the western portion.Other obvious similarities between 2006 and 2010 include the increase in T2 over the northern portion of the domain for MAM, SON and JFD; increase in PBLH over the ocean in the eastern part of the domain for all seasons; and increases over the ocean for CF for all seasons.The reasons for the differences between WRF/Chem and WRF in terms of meteorological variables have been discussed in Yahya et al. (2014).

Sensitivity simulations
The aforementioned differences in WRF/Chem predictions between 2006 and 2010 are caused by changes in emissions, meteorology, and meteorological and chemical ICONs/BCONs.Additional sensitivity simulations for the months of January and July in 2010 are carried out to estimate the individual contributions of each of those changes to the total net changes in model predictions.) for January and July, respectively.Since the impact of ICONs is only important at the beginning of the simulations whereas the impact of BCONs persists throughout the simulations, the changes due to changes in chemical BCONs will dominate over those due to changes in chemical ICONs/BCONs.Both Figs. 13 and 14 show that the differences in the meteorology due to the impact of meteorological ICONs/BCONs generated by WRF/Chem contribute to the largest differences in T2 and SWDOWN for both months (columns 1 and 4).For comparison, the changes in emissions and chemical ICONs/BCONs lead to less significant differences in T2 and SWDOWN (columns 2 and 3).The overall decrease in emissions from 2006 to 2010 results in a slight increase in both T2 and SWDOWN in January (column 2 in Fig. 13), and a larger increase in SWDOWN in July (column 2 in Fig. 14) due to decreases in aerosol loading.There is a small decrease in T2 and SWDOWN in January (column 3 in Fig. 13) due to influences of chemical ICONs/BCONs used for both years, but a larger decrease occurs in SWDOWN in July (column 3 in Fig. 14).As shown in Figs. 13 and 14 (column 1), changes in O 3 are influenced by all factors and the overall change of the O 3 mixing ratio is a combination of changes in emissions, and meteorological and chemical ICONs/BCONs.The O 3 mixing ratios are greatly increased due to the use of 2010 emissions as compared to 2006 emissions (column 2 in Fig. 13), indicating that using a different set of emissions can produce an increase of up to a domain mean of 6 ppb.Conversely, O 3 mixing ratios are greatly decreased (with a reduction of a domain mean of 6 ppb) due to the use of the 2010 chemical ICONs/BCONs compared to the 2006 chemical ICONs/BCONs (column 3 in Fig. 13).The use of different meteorological ICONs/BCONs also results in varying degrees of changes of O 3 mixing ratios domainwide as O 3 mixing ratios are influenced by photolysis and other meteorological parameters including wind and PBLH (column 4 in Fig. 13).In addition, T2 and SWDOWN influence the amount of BVOC (biogenic volatile organic compound) emissions produced, which also in turn influences O 3 mixing ratios.In VOC-limited urban centers over the eastern US (Campbell et al., 2014), a small increase in radiation or T2 will increase BVOC emissions, increasing O 3 mixing ratios and vice versa.In July (Fig. 14), the decrease in O 3 mixing ratios between 2006 and 2010 (column 1) is largely influenced by chemical ICONs/BCONs (column 3) and to a smaller extent by meteorological ICONs/BCONs (column 4).In this case, the difference in emissions (column 2) does not seem to significantly impact the changes of O 3 mixing ratios between July 2006 and 2010 (column 1).For January (Fig. 13), PM 2.5 concentrations decrease due to decreasing emissions and chemical ICONs/BCONs (columns 2 and 3).However, the use of 2010 meteorological ICONs/BCONs results in an increase in PM 2.5 concentrations over most of the domain except for the northeastern US (with a domain mean increase of 0.4 µg m −3 ) (column 4).The overall differences (column 1 in Fig. 13) are mainly due to net effects of emissions (column 2) and changes in meteorology (column 4).For PM 2.5 in July (Fig. 14), the net changes from 2006 and 2010 (column 1) are dominated entirely by changes in emissions (column 2) that increase in the southeastern and central US but decrease in the remaining domain, even though meteorological ICONs/BCONs also play a significant role (column 4).Table S2 in the Supplement shows the statistics for the NMB, NME, and Corr for a number of variables for the sensitivity simulations for January and July.The statistics in bold highlight the sensitivity simulations with the best performance (i.e., with the lowest NMB and NME and the highest Corr).The WRF/Chem performance of T2 against CASTNET improves to a large extent in terms of NME and Corr for Runs 3 and 4 (especially for January when Run 2 performs poorly), which use 2006 emissions.This indicates that, at least for January (and to a smaller extent for July), the inaccuracy of emissions may have contributed to the worse performance of T2 against CASTNET.Run 3 also gives the best performance of T2 against CASTNET, which indicates that improvement in both emissions and chemical ICONs/BCONs can improve meteorological performances for both January and July.For SWDOWN, Runs 3 and 4 improve the performance against CASTNET for January (with lower NMB and NME and higher Corr).The cloud-aerosol variables are affected to a smaller extent by changes in emissions and chemical ICONs/BCONs compared to the meteorological variables.The performance for CF remains relatively the same for January and July.The performance for COT and AOD improves slightly for January with a lower  The model performs relatively well for PM 2.5 concentrations.However, OC concentrations are significantly underpredicted against field data for 2010 in Bakersfield and Pasadena, CA, due mainly to underestimations in emissions of POA that contribute to most OC and also in part to underestimations in emissions of gaseous precursors of SOA and inaccurate meteorological predictions in 2010.The model also has significant biases for a few aerosol-cloud-radiation variables except for CF and QVAPOR; however, the model is able to reproduce the trends in aerosol-cloud-radiation variables for WRF/Chem with the CB05-MADE/VBS option used in this work has been incorporated into the WRF/Chem version 3.6.1 to be released in version 3.7 of WRF-Chem (available for download at http://www.mmm.ucar.edu/wrf/users/).The results in this work indicate a need to further improve the accuracy of emissions and chemical BCONs, and the representations of organic aerosols and chemistry-meteorology feedbacks in WRF/Chem.In addition, the improvements in aerosol-cloud treatments, such as the aerosol activation parameterization, and in the microphysics and cumulus parameterizations that affect the formation of precipitation are needed to improve the model's capability in reproducing the state of the atmosphere and also interannual trends.While the long-term air quality simulations using WRF/Chem with aerosol-cloud-radiation feedbacks in this work can provide guidance on future model development and improvement, they do not provide the impact of those feedback mechanisms on the model performance.Quantifying such impacts requires another set of simulations using a version of WRF/Chem that does not treat aerosol direct and indirect effects, which is not yet available to public.The simulations with and without aerosol direct and indirect effects have indeed been performed by Makar et al. (2014a, b) using a different model that was specially designed to quantify such impacts.It would be useful to develop a version of WRF/Chem that does not treat aerosol direct and indirect effects for this impact assessment.In particular, a comparison of the episodic or long-term simulation results using WRF/Chem that includes and excludes feedback mechanisms against observations of aerosol and cloud variables can provide further insight into whether inclusion of those aerosol direct and indirect effects can improve the model's capability in reproducing observations.Those simulations should be considered when the version of WRF/Chem without aerosol direct and indirect effects and computer resources become available.
The Supplement related to this article is available online at doi:10.5194/gmd-8-2095-2015-supplement.
Acknowledgements.This study is funded by the National Science Foundation EaSM program (AGS-1049200) at NCSU.The following agencies have prepared the data sets used in this study: the U.S. EPA (North American emissions processing), Environment Canada, Mexican Secretariat of the Environment and Natural Resources (Secretaría del Medio Ambiente y Recursos Naturales, SE-MARNAT) and National Institute of Ecology (Instituto Nacional de Ecología -INE) (North American national emissions inventories), the European Centre for Medium Range Weather Forecasting Global and Regional Earth-system (Atmosphere) Monitoring using Satellite and in situ data (ECMWF/GEMS) project and Meteo France/Centre national de recherches météorologiques (CNRM-GAME) for the Monitoring Atmospheric Composition and Climate (MACC) ICONs/BCONs.Meteorological ICONs/BCONs are provided by the National Center for Environmental Protection.Ambient North American concentration measurements are provided by several US networks (AQS, CASTNET, IMPROVE, SEARCH, and STN).North American precipitation-chemistry measurements are provided by several US networks (CASTNET, and NADP).GPCC precipitation data are provided by the National Oceanic and Atmospheric Administration's Earth System Research Laboratory in the Physical Sciences Division (NOAA/OAR/ESRL PSD), Boulder, Colorado, USA, from their web site at http://www.esrl.noaa.gov/psd/.The 2006 and 2010 SOA/OC data at RTP, NC, Bakersfield and Pasadena, CA, were provided by John Offenberg, U.S. EPA.Cloud variables were provided by MODIS.We thank Georg Grell, NOAA, Christian Hogrefe, US EPA, Paul Makar, Environment Canada, Christoph Knote, NCAR, and Patrick Campbell, NCSU, for helpful discussions on inputs and outputs of AQMEII model intercomparison.We would also like to acknowledge high-performance computing support from Yellowstone (ark:/85065/d7wd3xhc) provided by NCAR's Computational and Information Systems Laboratory, sponsored by the National Science Foundation and Information Systems Laboratory, sponsored by the National Science Foundation and Stampede, provided as an Extreme Science and Engineering Discovery Environment (XSEDE) digital service by the Texas Advanced Computing Center (TACC) (http://www.tacc.utexas.edu),which is supported by National Science Foundation grant number ACI -1053575.
The US Environmental Protection Agency through its Office of Research and Development collaborated in the research described here.The manuscript has been subjected to external peer review and has been cleared for publication.Mention of trade names or commercial products does not constitute endorsement or recommendation for use.
Edited by: J. Williams

Figure 2 .Figure 3 .
Figure 2. Spatial distribution of NMB plots for JFD and JJA 2006 and 2010 for maximum 8 h O 3 concentrations, based on evaluation against CASTNET, AQS and SEARCH.
chemical performance between 2006 and 2010 is more variable compared to the meteorological performance of surface variables.The lower Corr for 2010 compared to 2006 for meteorological variables has a large influence on the model performance for 2010.As shown in Table 1, all the chemical variables for all networks have lower a Corr in 2010 compared to 2006.As shown in Figs. 2 and 3, the maximum 8 h O 3 concentrations are underpredicted to a larger extent in 2010 compared to 2006, dominating the O 3 annual performance in 2010.These results are consistent with the results of Hogrefe et al. (2014).The large underpredictions of maximum 8 h O 3 in JFD 2010 over the southeastern US are attributed to larger cold biases in T2 shown in Fig. S4 and reduced NO x and VOC emissions in 2010 relative to their levels in 2006.While reduced NO x levels can result in an increase in nighttime O 3 concentrations due to reduced NO x titration of O 3, the impact of reduced NO x titration on the maximum 8 h O 3 is small.As shown in Fig. S4, the temperature biases for both years are relatively similar.Over the northeastern US, the T2 bias is generally less than −0.1 • C for JJA in both years.However, as shown in Fig. 2, O 3 concentrations over the northeastern US in JJA 2010 have negative biases whereas those over the northeastern US in JJA 2006 have positive biases.In this case, emissions might play a significant role in the underprediction of O 3 concentrations over the northeastern US in JJA 2010.Hourly average surface NO x emissions decrease significantly over the northeastern US in JJA from 2006 to 2010.As shown in Fig. 3, 2006 model performance for O 3 is generally good for all seasons and all networks.
. 1, WRF/Chem predicts SWDOWN to a lower extent in 2010 compared to 2006 against CASTNET.Khiem et al. (2010) reported that during the summer, a large percentage of the variations in peak O 3 concentrations during the summer can be attributed to changes in seasonally averaged daily maximum temperature and seasonally averaged WS10.Simulated WS10 is lower for 2010 compared to 2006 in general; therefore, WS10 does not seem to contribute to reduced O 3 concentrations (through dispersion, increased dry deposition) in 2010.Figure 4 shows diurnal variations of observed and simulated WRF/Chem T2 and O 3 concentrations from CASTNET in JJA 2006 and 2010.The diurnal averaging provides insight into whether the underpredictions of O 3 mixing ratios are a systematic bias during the daytime or nighttime or both.The diurnally averaged observed temperatures show a similar trend in JJA 2006 to 2010 against T2 measurements from CASTNET.This shows that the model is able to reproduce T2 for different years.The temperature trends also correlate strongly with the O 3 trends.At night, where the model has a cold bias, O 3 concentrations are underpredicted to a larger extent.The O 3 concentrations show a larger underprediction for JJA 2010 compared to JJA 2006.The underpredictions in O 3 in both 2006 and 2010 can be explained by several reasons.For example, Im et al. (2014) showed that the MACC (Monitoring Atmospheric Composition and Climate) model underpredicts O 3 mixing ratios, particularly in winter and spring during both day and night and in summer and fall during nighttime.As indicated by Wang et al. (2014) and Makar et al. ( 2.5 concentrations for JFD and JJA 2006 and 2010 against IM-PROVE, STN, and SEARCH.Overall, JJA 2006 and JJA 2010 have similar spatial distribution patterns of NMBs for all sites over CONUS except for several sites in the northwestern US where PM 2.5 concentrations are underpredicted for JJA 2010 but overpredicted for JJA 2006.However, many sites have positive NMBs over the eastern and central US for JFD 2006, whereas more sites have negative NMBs over the eastern and central US for JFD 2010.Statistics from Yahya et al. (2014) and Table 1 show that, in general, the simulated concentrations of PM 2.5 and all PM 2.5 species decrease from 2006 to 2010; however, the Corr values for PM 2.5 and PM 2.5 species become worse in 2010 compared to 2006.As shown in Fig. 6, PM 2.5 concentrations for 2006 can be overpredicted or underpredicted, depending on seasons and networks, with an equal distribution of positive and negative NMBs.However for 2010, PM 2.5 concentrations tend to be underpredicted for all seasons and for all networks except for JFD against SEARCH.As shown in Fig. 7, NMBs for PM 2.5 species for 2006 at individual monitoring sites range from −40 to 60 %, while those for 2010 range from −80 to 80 %.The markers are more spread out covering a wider range of NMBs and NMEs for 2010 with more extremes as compared to the markers for 2006 clustered around the zero NMB line.

Figure 5 .
Figure 5. Spatial Distribution of NMB plots for JFD and JJA 2006 and 2010 for average 24 h PM 2.5 concentrations based on evaluation against the IMPROVE, STN and SEARCH sites.

Figure 6 .Figure 7 .Figure 8 .
Figure 6.Comparison of seasonal plots of NMB vs. NME for average 24 h PM 2.5 concentrations where the different shapes represent different seasons (diamond -MAM, circle -JJA, triangle -SON and square -JFD) and the different colors represent different observational data (purple -IMPROVE, black -STN and green -SEARCH).

Figure 9 .
Figure 9. Scatter plots of SOA (left column) and OC (right column) concentrations at various sites.

Figure 10 .
Figure 10.Comparison of soccer plots for JFD and JJA 2006 and 2010 evaluation of aerosol and cloud variables.Multiangle Imaging SpectroRadiometer (MISR) AOD, and Surface Radiation Budget (SRB) CF Obs data were not available for 2010.

Figure 11 .
Figure 11.Changes in hourly average surface concentrations of O 3 and PM species from 2010 to 2006 (2010-2006).

Figure
Figure S10 compares the seasonal changes between 2010 and 2006 in several meteorological variables that affect air pollution including SWDOWN, T2, WS10, PBLH, and Precip simulated by WRF-only simulations without considering chemistry feedbacks.Large changes occur in those variables between the 2 years, e.g., 10-50 W m −2 increases in SWDOWN in the western and midwest US in JJA, generally warmer in JJA and SON over most areas but cooler by 3-10 • C in the eastern US in JFD, and with reduced Precip in the eastern or southeastern US in JJA and SON but increased Precip in the northwestern US in MAM and JJA and in the western US in JFD.ICONs and BCONs for skin temperatures shown in Fig. S3 greatly influence T2 shown in Fig. S10 for JFD and JJA.Figures12 and S11show the seasonal changes between 2010 and 2006 in several meteorological and cloud variables (SWDOWN, T2, WS10, Precip, PBLH, AOD, COT, CF, CWP, and CDNC) for WRF/Chem, which accounts for meteorology-chemistry feedbacks.The relationships between various meteorological variables have been discussed inYahya et al. (2014).Comparing to the differences in predictions of SWDOWN, T2, WS10, Precip, and PBLH between 2010 and 2006 WRF-only simulations shown in Fig.S10and WRF/Chem simulations shown in Figs. 12 and S11, the differences in those meteorological variables do not vary significantly in terms of trends of average sea-

Figure 12 .
Figure 12.Changes in hourly average predictions of aerosol-cloud variables at the surface from WRF/Chem simulations from 2010 to 2006 (2010-2006).
2006 and 2010.The variation trends for most meteorological and chemical variables simulated by WRF and WRF/Chem are overall consistent with the observed trends from 2006 to 2010, but for 2010 WRF/Chem performs slightly worse than WRF.Similar to 2006, the inclusion of chemistrymeteorology feedbacks reduces NMBs for most meteorological variables in 2010, although WRF gives higher Corr and lower NME values than WRF/Chem.A number of sensitivity simulations are also conducted for January and July 2006 and 2010 to quantify the relative impact of emissions, chemical ICONs/BCONs, and meteorology on model performance of major meteorological and chemical species as well as on the variation trends between 2006 and 2010.Using more accurate emissions and chemical and meteorological ICONs/BCONs will help improve the performance of some individual chemical and meteorological surface variables.Although the 2006 emissions may not represent the true emissions for 2010, the 2010 sensitivity simulations using the 2006 emissions show improved model performance.However, using 2006 emissions for 2010 simulations does not improve the degree of agreement with observed interannual trends as the consistency between the 2006 and 2010 emissions are affected between the simulations.The baseline simulations for 2006 and 2010 reproduce the observed trends the best, as a consistent set of 2006 and 2010 emissions are used.The current 2006 and 2010 emissions were developed taking into account the interannual trends; the improvement of emissions need to be carried out consistently for all individual simulation years when simulating multiyear cases.

Model performance in 2010 and its comparison with 2006
and June, July, and August (JJA) 2010-2006 in averaged meteorological ICONs and BCONs of skin temperature and soil moisture fractions 100-200 cm below ground extracted from the National Center of Environmental Prediction (NCEP).
Yahya et al. (2014)n 2010 respond to changes in emissions, BCONs, and meteorology.The model performance for both meteorological and chemical predictions in 2010 is evaluated and compared with that in 2006.The surface observational networks used to evaluate 2010 results include the Clean Air Status and Trends Network -CASTNET (rural sites), the Southeastern Aerosol Research and Characterization -SEARCH (southeastern US only, rural and urban sites), the Speciated Trends Network -STN (urban sites), the Interagency Monitoring of Protected Visual Environments -IMPROVE (rural sites), the Air Quality System -AQS (rural and urban sites) and the National Atmospheric Deposition Program -NADP (rural and urban sites).(Sim)valuesaswellascorrelation coefficients (Corr) between the observed and simulated meteorological variables from the 2010 WRF/Chem and WRF simulations.Similar statistics from the 2006 WRF/Chem and WRF simulations can be found in Table1inYahya et al. (2014).Figure1

Table 1 .
Annual performance statistics for 2010 Predictions of WRF and WRF/Chem.

Table 2 .
Percentage changes in observed and simulated variables between 2010 and 2006.
1 The percentages are calculated according to this formula: [(2010 value − 2006 value) /2006 value] × 100 %.CASTNET -the Clean Air Status and Trends Network; AQS -the Aerometric Information Retrieval System/Air Quality System; SEARCH -the Southeastern Aerosol Research and Characterization; GPCC -the Global Precipitation Climatology Centre; MODIS -the Moderate Resolution Imaging Spectroradiometer; IMPROVE -the Interagency Monitoring of Protected Visual Environments; STN -the Speciated Trends Network.Note that IMPROVE did not contain NH4+ data for 2010."-" indicates that the results of those variables are not available from the WRF-only simulation.

Table 3 .
Summary of the setup of sensitivity simulations.

Table 4 .
Absolute and percentage differences between monthly mean of observed/satellite-retrieved data and sensitivity simulations.In general, the model performs well in terms of Corr and NMEs for almost all meteorological and chemical variables in 2006 but not as well in 2010 despite lower NMBs for most variables in 2010, due mainly to inaccuracies in emission estimates and chemical BCONs and ICONs used for 2010 simulations.The model is able to reproduce the observations to a large extent for most meteorological surface variables.