A new chemistry option in WRF / Chem v . 3 . 4 for the simulation 1 of direct and indirect aerosol effects using VBS : evaluation 2 against IMPACT-EUCAARI data

Abstract. A parameterization for secondary organic aerosol (SOA) production based on the volatility basis set (VBS) approach has been coupled with microphysics and radiative schemes in the Weather Research and Forecasting model with Chemistry (WRF-Chem) model. The new chemistry option called "RACM-MADE-VBS-AQCHEM" was evaluated on a cloud resolving scale against ground-based and aircraft measurements collected during the IMPACT-EUCAARI (Intensive Cloud Aerosol Measurement Campaign – European Integrated project on Aerosol Cloud Climate and Air quality interaction) campaign, and complemented with satellite data from MODIS. The day-to-day variability and the diurnal cycle of ozone (O3) and nitrogen oxides (NOx) at the surface are captured by the model. Surface aerosol mass concentrations of sulfate (SO4), nitrate (NO3), ammonium (NH4), and organic matter (OM) are simulated with correlations larger than 0.55. WRF-Chem captures the vertical profile of the aerosol mass concentration in both the planetary boundary layer (PBL) and free troposphere (FT) as a function of the synoptic condition, but the model does not capture the full range of the measured concentrations. Predicted OM concentration is at the lower end of the observed mass concentrations. The bias may be attributable to the missing aqueous chemistry processes of organic compounds and to uncertainties in meteorological fields. A key role could be played by assumptions on the VBS approach such as the SOA formation pathways, oxidation rate, and dry deposition velocity of organic condensable vapours. Another source of error in simulating SOA is the uncertainties in the anthropogenic emissions of primary organic carbon. Aerosol particle number concentration (condensation nuclei, CN) is overestimated by a factor of 1.4 and 1.7 within the PBL and FT, respectively. Model bias is most likely attributable to the uncertainties of primary particle emissions (mostly in the PBL) and to the nucleation rate. Simulated cloud condensation nuclei (CCN) are also overestimated, but the bias is more contained with respect to that of CN. The CCN efficiency, which is a characterization of the ability of aerosol particles to nucleate cloud droplets, is underestimated by a factor of 1.5 and 3.8 in the PBL and FT, respectively. The comparison with MODIS data shows that the model overestimates the aerosol optical thickness (AOT). The domain averages (for 1 day) are 0.38 ± 0.12 and 0.42 ± 0.10 for MODIS and WRF-Chem data, respectively. The droplet effective radius (Re) in liquid-phase clouds is underestimated by a factor of 1.5; the cloud liquid water path (LWP) is overestimated by a factor of 1.1–1.6. The consequence is the overestimation of average liquid cloud optical thickness (COT) from a few percent up to 42 %. The predicted cloud water path (CWP) in all phases displays a bias in the range +41–80 %, whereas the bias of COT is about 15 %. In sensitivity tests where we excluded SOA, the skills of the model in reproducing the observed patterns and average values of the microphysical and optical properties of liquid and all phase clouds decreases. Moreover, the run without SOA (NOSOA) shows convective clouds with an enhanced content of liquid and frozen hydrometers, and stronger updrafts and downdrafts. Considering that the previous version of WRF-Chem coupled with a modal aerosol module predicted very low SOA content (secondary organic aerosol model (SORGAM) mechanism) the new proposed option may lead to a better characterization of aerosol–cloud feedbacks.

. RF50, RF55, RF56, RF57, RF58, RF61 6 and RF62 were conducted around Cabauw supersite, in order to study the origin and characteristic 7 of air masses collected at Cabauw. Other RFs were aimed at the study of aerosol properties along a 8 quasi-Lagrangian flight track, with west-east and north-south transects. ATR-42 was equipped with 9 instrumentation suitable for aerosol-cloud interaction measurements. We used the measurements 10 from a condensation particle counter (CPC), the CPC3010 with a cutoff size of 15 nm, a Cloud 11 Condensation Nuclei Counter (CCNC) for CCN number concentration measurements, and an AMS. 12 During the campaign a scanning mobility particle sizer (SMPS) was used to measure the number 13 size distribution of aerosol particles with diameter in the range of 0.02-0.5 µm, while the size 14 distribution of aerosol particles larger than 100 nm was sampled with a passive cavity aerosol 15 spectrometer probe (PCASP). SMPS and PCASP measurements were combined in order to 16 calculate the PM 2.5 concentration using an average aerosol density of 1.7 g/m 3 . A more exhaustive 17 and detailed description of the whole campaign and instrumentation is given by Crumeyrolle et al.

Meteorology
6 Figure 2 shows the observed and modelled time series of hourly vertical profiles of temperature and 7 relative humidity at Cabauw supersite. WRF/Chem reproduces the day-to-day variation of 8 temperature, before, after, and during the wet period. As shown by statistical indices (Table 2) The biases of the temperature and relative humidity could be due to a misrepresentation of soil (and 1 sea) temperature and soil moisture or by misrepresentation of the clouds and rain. These two 2 problems are tightly coupled via land surface-atmosphere interaction. The errors in the simulation 3 of surface moisture and energy budget influences the fluxes of latent heat and moisture in the 4 atmosphere, affecting the local circulation, convective available potential energy (CAPE), cloud 5 formation and rain pattern (Pielke, 2001;Holt et al., 2006). Moreover, WRF/Chem tends to fail 6 simulating the thermodynamic variables near to coastline, because the uncertainties of land use data prediction is sensitive to used input data. They showed that varying the inputs used as initial and 10 boundary conditions, the mean daily model bias ranges from -2.71 to -0.65 K for the temperature 11 and from -0.81 to 0.50 g/kg for vapour water content.   in the afternoon and the evening, and is most likely due to the titration in these hours caused by 6 higher NO x concentration than observed. 7 WRF/Chem simulates the NO x , NO and NO 2 time series with a correlation of 0.70, 0.65, and 0.66 8 respectively. The timing of NO x daily cycle is reproduced. Indeed, the model captures the morning 9 and evening peaks as well as diurnal minimum of NO 2 . The mean bias of modelled NO 2 is +1.25 10 µg/m 3 (+20%) and occurs in the afternoon and evening hours. Moreover WRF/Chem reproduces the 11 morning peak and diurnal decrease of NO, but the daily cycle is affected by an average positive bias 12 of 0.28 µg/m 3 , with the average morning maximum overestimated of about 2 µg/m 3 (+33%). 13 Ammonia is reproduced with a correlation of 0.43. WRF/Chem underestimates the NH 3 during the 14 scavenging days and from 28 to 31 May. The model captures the daily cycle shape of NH 3 15 concentration average, but the modelled NH 3 concentrations are constantly underestimated. The 16 negative mean bias over the whole period is on average about -4.75 µg/m 3 (-28%). WRF/Chem 17 reproduces the observed HNO 3 with a poor correlation. The measured mean diurnal cycle is flat, 18 conversely the model predicts a nocturnal minimum and diurnal maximum. The origin of model 19 bias in simulating NH 3 and HNO 3 is discussed below, together with a discussion on inorganic 20 aerosols.

21
The nitrous acid concentrations are not well captured by the model and are underestimated by 95%.

22
This bias could be partly explained by the inefficiency of NO oxidation, the only important reaction  This is demonstrated, for example, by the larger NMGE for SO 2 than NO x (116% and 45%, 30 respectively). NO x is emitted near the surface by traffic and domestic heating. Therefore, NO x 31 emissions are subjected to a stronger temporal modulation than SO 2 point sources.

32
The different uncertainties found for the involved species may depend on the choice of the chemical 1 mechanism. Knote et al. (2014) compared several chemical mechanisms within a box model 2 constrained by the same meteorological conditions and emissions, and found that the prediction of 3 the O 3 diurnal cycle differs by less than 5% among the different mechanisms. Larger differences 4 were found for other species. For example, the key radicals exhibit differences up to 40% for OH, 5 25% for H 2 O 2 and 100% for NO 3 among the selected mechanisms. 6 Figure 5 shows the simulated and observed time series and diurnal cycle of aerosol sulphate, nitrate, 7 ammonium, and organic matter, at CESAR observatory. WRF/Chem simulates the measured SO 4 , 8 NO 3 , NH 4 with a correlation of 0.56, 0.68, and 0.66, respectively. 9 WRF/Chem captures the daily variations of SO 4 and its decrease during scavenging days. The shape 10 of diurnal cycle is also reproduced, with the nighttime minimum and diurnal maximum. The mass 11 concentration of SO 4 is overpredicted for the whole period with a mean bias of 1.04 µg/m 3 (+90%).

12
The modelled SO 4 overestimation is directly attributable to SO 2 concentration overprediction. 13 Another potential source of the surplus of simulated SO 4 is related to an excessive production 14 within the clouds. Indeed, during scavenging days, the particulate sulphate is overestimated while 15 the predicted SO 2 does not show a bias in respect to the measurements. The overestimation of SO 4 , 16 moreover, explains in part the negative bias of predicted NH 3 . The excess of particle sulphate   Organic matter is reproduced with a correlation coefficient of 0.75. WRF/Chem reproduces the 1 right concentration during dry period, the decrease in the wet days, and following recovery. The 2 mean bias is negative by about 0.4 µg/m 3 and it is attributable to days from 23 to 26 May. The 3 discussion about the origins of OM bias is given in section 4.3, where we will discuss the model 4 evaluation with aircraft data.

5
The reader should consider that aerosol composition measurements performed with the AMS are 6 representative of particles with diameter between roughly 100-700 nm, whereas the model is 7 evaluated with aerosol concentration representative of PM 2.5 . Therefore, a bias could be present in 8 the comparison. This means that the bias found for inorganic aerosols could be smaller than that 9 reported above, conversely the OM bias could be larger of that found.  The results obtained here are statistically consistent with other modelling studies over Europe (e.g.,    planetary boundary layer (PBL) and free troposphere (FT) (see Figure 6). The height of the PBL 6 was lower than 1600 m during the whole campaign (Crumeyrolle et al., 2013). Therefore, we 7 considered for PBL and FT concentrations the data below and above 1600 m up to 3000-4000 m, 8 respectively. This rough approximation of PBL height could affect the comparison of the model to 9 data. 10 Figure 6 displays the observed and modelled box plots of the mass concentration of SO 4 , NO 3 , NH 4 , 11 and OM for PBL and FT. Their mean value, standard deviation, relative mass fraction, and 12 correlation coefficients, averaged over the whole period of interest, are reported in Table 3. 13 The average concentrations of inorganic aerosols show little absolute error (2-8%) with respect to 14 the observations in the PBL, while the NO 3 and NH 4 mean concentration presents a bias of +14% 15 and +20% (+0.3 and +0.2 µg/m 3 ), respectively, in the FT. The mean OM mass is biased low by a  Table 3. The predicted standard deviations for each species are lower than observed. In the PBL, the 26 observed and modelled standard deviations differ by 4-10% for inorganic ions and 55% for OM. In 27 the FT, the differences are higher than in the PBL. The model predicts standard deviations lower 28 than 10-40% for inorganic particles and lower than 65% for organic matter with respect to the 29 measurements.

30
For the purpose of this analysis, it is also interesting to explore how the model reproduces the 31 relative fraction of aerosol mass species with altitude (see Table 3). WRF/Chem overpredicts the relative fraction of the SO 4 and NO 3 by few percent in the PBL and about 10% in FT, while the 1 relative mass of NH 4 is overestimated by 3% along the whole profile. The relative amount of OM is 2 underpredicted by about 20% in both PBL and FT. The decrease of relative amount of NO 3 and 3 increase of SO 4 with altitude is captured by the model. The modelled relative mass of NH 4 and OM 4 is near constant with altitude as well as in the observations.

5
Looking at the individual flights, it is possible to note how the model captures the aerosol mass 6 trend as a function of the synoptic frame in both the PBL and FT, during the dry period, scavenging 7 days, and dust period characterized by southerly wind and passage of several fronts. The FT is a 8 layer mainly affected by long range transport and cloud contamination. Therefore, the relative small 9 bias in simulating aerosol inorganic mass in FT means that the model resolves quite correctly the 10 large scale transport and processes related to clouds. 11 Nevertheless, it should be noted that SO 4 is overestimated for 8 out of 14 RFs, while NO 3 and NH 4 12 are underpredicted for 7-8 out of 14 RF. This SO 4 overprediction is attributable to the SO 2 excess 13 and to a potential overproduction within the cloud chemistry scheme. The negative bias of NO 3 and 14 NH 4 could be explained by a low NH 3 regime, that limits the formation of the ammonium-nitrate.   Although the box plot and statistical summary (Table 3)

Aloft aerosol particles 1
The comparison of WRF/Chem output with aircraft measurements of the number concentration of 2 condensation nuclei (CN) and of cloud condensation nuclei (CCN) at 0.2% of supersaturation is 3 done by using the boxplots as for aerosol mass. In this case the modelled and measured data are 4 smoothed by using a 10 seconds running mean. 5 Figure 9 reports the comparison of observed and modelled CN within PBL and FT. The measured 6 and predicted average, standard deviation, and correlation of CN number over the whole period of 7 our analysis are reported in Table 3. 8 The model resolves the decrease of a factor 5-6 of CN concentration between the PBL and the FT. The differences in simulated concentrations between land and sea (RF51 and 52) are also captured 10 by the modelling system. Nevertheless, WRF/Chem overestimates, on average, the observed CN by  Table 3. The bias of 23 simulated CCN 0.2 appears more contained with respect to CN prediction, especially in the free 24 troposphere. Figure S3 Figure 11 shows the comparison between the AOT at 550 nm measured by MODIS and the   Liquid water path (LWP) was calculated by vertically integrating liquid cloud mixing ratios (water 25 and rain water), while liquid cloud optical thickness (COT) was estimated from LWC and R e . Since 26 MODIS L2 data provide the total cloud water path (CWP), combined effective radius for all cloud 27 types and total (water and ice) cloud optical thickness (COT), the observed contribution to the 28 liquid water cloud was separated by using the retrieved top cloud phase, i.e. were discarded mixed 29 clouds.

30
The comparison between the predicted and observed R e , LWP and liquid COT was done in the 31 scavenging background and long-range transport periods in the days when MODIS cloudy pixel coverage was larger than 60%. Figure 12   As shown in Table 4, R e values averaged over the entire domain is underestimated by the model by underestimates the maximum and overestimates the higher and lower end of the distributions. Both 2 variables show a variability higher than the observations. The predicted standard deviations (Table   3 4) are about 2-3 and 1.5 times larger than those observed for LWP and liquid COT, respectively.

4
This probably stems from the large variability in simulated CCN.

5
Now it is interesting to analyse the model behaviour in reproducing the total CWP and COT given 6 by contribution of all cloud phases. Modelled CWP was calculated by vertically integrating all 7 cloud mixing ratio (water, rainwater, ice, snow and graupel). Predicted COT is given by the 8 contribution of the liquid water and ice. The contribution of the liquid water was calculated as 9 described above for liquid water cloud. The contribution of ice phase to COT was calculated 10 following Ebert and Curry (1992). Figure 14 displays the comparison between observed and predicted CWP and COT in P1, whereas 12 the same figures for P2 and P3 are reported in the Supplement (Figures S8 and S9). Although for all 13 three cases, the model reproduces with good approximation the shape and localizations of the cloud 14 systems, CWP and COT are systematically overestimated (except COT in P2). As shown in Table   15 5, the predicted domain average of CWP presents, indeed, a bias of 62%, 41%, and 80% for P1, P2 16 and P3, respectively, whereas the bias of COT is about 15% in P1 and P3.  and in situ observations. Indeed, analysing a system of thin cumulus clouds during EUCAARI 1 campaign, they also found that MODIS overestimates the droplet effective radius by a factor 2-3 2 and COT is 2-3 times lower than in situ measurements.  Table 4. NOSOA runs show a domain averaged R e larger than CTRL.    Nevertheless, SO 4 (NO 3 and NH 4 ) mass is overpredicted (underpredicted) in more than half of the 32 flights. SO 4 bias is attributable to the SO 2 excess and to a potential overproduction within the cloud chemistry scheme. The negative bias of NO 3 and NH 4 could be explained by a low concentration of 1 NH 3 that limits the formation of the ammonium-nitrate. The simulated OM concentration is at 2 lower end of the observed mass. The bias is attributable to the missing aqueous chemistry processes 3 of organic compounds, uncertainties in meteorological fields, to assumptions on the VBS approach 4 such as the SOA formation pathways, oxidation rate and dry deposition velocity of organic 14 For the future, there is still large space for improvements. For example, a more advanced treatment 15 of deposition of organic condensable vapours is desirable. Moreover, the missing production of 16 SOA in cloud is a gap that should also be filled. Finally, the extension of aerosol-cloud interaction 17 to the ice-phase would lead to a complete representation of the aerosol indirect effects.

Code availability
1 The code updated, described, and evaluated here will be incorporated in the next available release 2 of WRF/Chem. The users will be able to freely download the code from the WRF website 3 (http://www2.mmm.ucar.edu/wrf/users/download/get_source.html). A general WRF/Chem user's 4 guide is also available online (http://ruc.noaa.gov/wrf/WG11/).
where X is a generic vector, Z(X) is its standard score, and σ X is the standard deviation.     sulphur dioxide (SO 2 ), particle sulphate (SO 4 ), particle nitrate (NO 3 ), particle (ammonium), and 5 particle OM, collected at Cabauw tower.