Assessment of the radiative effects of aerosols in an on-line coupled model over the Iberian Peninsula

The effects of atmospheric aerosol particles over the Earth’s climate mainly depend on their optical, microphysical and chemical properties, which modify the Earth radiative budget. The aerosol radiative effects can be divided into direct and semi-direct effects, produced by the aerosol-radiation interactions (ARI); and indirect effects, produced by aerosol-cloud interactions (ACI). In this sense the objective of this work is to assess whether the inclusion of aerosol radiative feedbacks in 15 the on-line coupled WRF-Chem model improves the modelling outputs over the Iberian Peninsula (IP). For that purpose, the methodology bases on the evaluation of modelled aerosol optical properties under different simulation scenarios. The evaluated data come from two WRF-Chem simulations for the IP differing in the inclusion/no-inclusion of ARI and ACI (NRF/RF simulations). The case studies cover two episodes with different aerosol types over the IP in 2010, namely a Saharan desert dust outbreak and a forest fire episode. The evaluation uses observational data from AERONET stations and 20 MODIS sensor, including aerosol optical depth (AOD) and Angström exponent (AE). Experimental data of aerosol vertical distribution from the EARLINET Granada station are used for checking the models. The results indicate that for the spatial distribution the best-represented variable is AOD and the largest improvements of including the radiative feedbacks are found for the vertical distribution. In the case of the dust outbreak, a slight improvement(worsening) is produced over the areas with medium(high/low) levels of AOD (-9%/+12% of improvement) when including the radiative feedbacks. For the 25 wildfires episode, improvements of AOD representation (up to 11%) over areas further away from emission sources are estimated, which compensates the computational effort of including aerosol feedbacks in the simulations. No evident improvement is observed for the AE representation, whose variability is largely underpredicted by both simulations. Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-473, 2016 Manuscript under review for journal Atmos. Chem. Phys. Published: 6 July 2016 c © Author(s) 2016. CC-BY 3.0 License.


Introduction
It is nowadays widely recognised that aerosol particles exert a substantial influence on Earth's climate changing the radiative budget (Charlson et al., 1992;Hansen et al., 1997;Ramanathan and Feng, 2009;Boucher et al., 2013, IPCC, 2013, among many others).The principal mechanisms by which aerosols cause these changes are: (1) scattering and absorption of solar radiation (aerosol-radiation interactions, ARI) (e.g.Ruckstuhl et al., 2008) and: (2) modification of clouds and precipitation, 5 thereby affecting both radiation and hydrology, or increasing the reflectivity of clouds (aerosol-cloud interactions, ACI) (e.g.Twomey, 1974;Albrecht, 1989;Twomey, 1991).In the first case, light scattering by aerosol particles such as sea salt or desert dust increases the solar radiation reflected by the planet, producing a cooling influence.Light-absorbing aerosols such as biomass burning exert a warming influence (e.g.Jacobson, 2001).These radiative influences are quantified as forcings (in W m -2 ), defined as the perturbation to the energy balance of the Atmosphere-Earth system.A warming influence is denoted a 10 positive forcing, and a cooling influence, negative (IPCC, 2013).Generally, modelling tools and observations indicate that anthropogenic aerosols have had a cooling influence on Earth since preindustrial time, with a total ARI+ACI mediumconfidence radiative forcing (excluding the effect of absorbing aerosol on snow and ice) of -0.9 (-1.9 to -0.1) W m -2 (Boucher et al., 2013).The large uncertainty quantifying these aerosol effects on the Earth radiative budget are much higher than for any other climate-forcing agent (IPCC, 2013).This happens because the physical, chemical and optical aerosol 15 properties are highly variable in space and time scales due to the aerosol particles short-lived and non-uniform emissions (Forster et al., 2007).
In order to reduce this uncertainty, the use of models is one of the most powerful tools to understand the different processes affecting the climate system.As aerosol may strongly drive the Earth's climate on global and regional scales, fully-coupled meteorology-climate and chemistry models allow for accounting the climate-chemistry-aerosol-cloud-radiation feedbacks 20 mechanisms between simulated aerosol concentrations and meteorological variables.It is also a promising way to go for future atmospheric simulation systems, leading to a new generation of models for improved meteorological, environmental and chemical weather forecasting (Baklanov et al., 2014).
Europe may be one of the most climatically sensitive world regions (Giorgi, 2006).Within the target domain, the role of aerosol particles may then be even more crucial over such regions as the Mediterranean basin, a crossroad that fuels the 25 mixing of particles from different sources (Papadimas et al., 2012).The Iberian Peninsula (IP), as a good example within the Mediterranean basin, can be affected by high aerosols concentration of different aerosol types.Due to its closeness to the Sahara Desert, the IP is frequently affected by dust outbreaks with large aerosol loads that modulate the aerosol climatology in different areas of this region, especially in Southern Spain (e.g.Toledano et al., 2007;Guerrero-Rascado et al., 2008, 2009;Córdoba-Jabonero et al., 2011;Antón et al., 2012;Pereira et al., 2014) and Portugal (e.g.Wagner et al., 2009;Preißler 30 et al., 2011).On the other hand, the Mediterranean climate, with high summer temperatures and dry soil-air conditions, encourage forest fires episodes over this region (Alados-Arboledas et al., 2010).Both types of emissions give major contributions to particle concentration in the atmosphere, particularly in the warmer season (Elias et al., 2006).Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-473, 2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.
There is a large number of studies assessing the aerosol feedbacks effects over the IP using different remote sensing measurement methods, using devices such as sun photometers (Lyamani et al., 2005(Lyamani et al., , 2006;;Toledano et al., 2007;Cachorro et al. 2008;Obregón el at., 2012), nephelometers (Pereira et al., 2008, 2011), lidars (Guerrero-Rascado et al., 2008) or a combination of these (Elias et al., 2006;Córdoba-Jabonero et al., 2011).Other studies using these instruments join satellite measurements to carry out this assessment (Cachorro et al., 2006;Guerrero-Rascado et al., 2009).Even, there are studies that 5 use these different measurements to estimate by a radiative transfer model the aerosol radiative forcing over some regions (Santos et al., 2008;Guerrero-Rascado et al., 2009;Valenzuela et al., 2012) or over the whole IP (Mateos et al. 2014).On the other hand, a number of studies (e.g.Myhre et al., 2007;Myhre et al., 2009) have tried to assess the aerosol feedbacks effects on a global scale, while other works (e.g.Péré et al., 2010;Meij et al., 2012;Curci et al., 2014, among others) have a more regional approach.However, no modelling studies of the aerosol radiative effects have ever been carried out for the IP. 10 According to Randall et al. (2007), the responses of the climate system to aerosols and their effects on the radiative budget of the Earth are the most uncertain climate feedbacks.
Therefore, the objectives of this work are (i) to assess whether the inclusion of aerosol radiative feedbacks in the on-line coupled WRF-Chem model improves the modelling outputs over the IP and (ii) to evaluate the representation of aerosol optical properties by this model over the target domain.15

Methodology
In this paper we evaluate the output of different simulations carried out by the WRF-Chem model (Grell et al., 2005) by using observational data provided by several instruments: two ground-based data networks (AERONET and EARLINET) and a sensor on-board a satellite (MODIS).Two different setups of the model have been considered, including/not including aerosol radiative feedbacks in the simulation.According to Boucher et al. (2013), the inclusion of these feedbacks involves a 20 change on the internal energy flows to the Earth system, affecting cloud cover or other components of the climate system such as aerosol particles, and, thereby, altering the global budget indirectly.
The evaluation has been performed by using classical statistics according to Willmott et al. (1985), Weil et al. (1992) and Willmott and Matsuura (2005).The individual model-prediction error or bias (e i ), the mean bias error (MBE), mean absolute error (MAE) and the correlation coefficient (r) have been calculated.All data needs have pre-processed and bilinearly 25 interpolated to a common working grid.This has a resolution of 0.1º and covers between 35º and 47º north and -15 and 5º east.The grid size is 6000 cells and the grid type is a regular lon-lat grid.After the interpolation, modelled data are evaluated against MODIS.The data to compare with AERONET and EARLINET are extracted from the model cell covering the corresponding station coordinates (Table 1) following a nearest neighbour approach.
First, in order to evaluate whether the inclusion of aerosol radiative feedbacks in the on-line coupled WRF-Chem model 30 produces significant changes on the studied variables (or changes are just mere signal noise), a surrogate variable, associated to the significance level of the changes modelled (S.L.), is defined (Eq.1).Therefore, high values of S.L. indicate whether Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-473, 2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.
the changes between simulations including (and not including) radiative feedbacks are noticeable with respect to the variability of the signal or not, and therefore, their significance: where S.L. is the significance level, x i is the value of the studied variable and S NRF 2 is the associated variance for the case not taking into account any radiative feedbacks (NRF).Moreover, NRF represents the base case and RF is the radiative 5 feedbacks simulation that includes the ARI+ACI.
Second, to evaluate whether the inclusion of the radiative feedbacks in the simulations leads to an improvement of the error of the model, the variable Improvement of the MAE is used (Eq.2): where |e i | is the absolute error of the simulations.10 Finally, to estimate whether the inclusion of the radiative feedbacks in the simulations produces an improvement of the vertical distribution of aerosols, the normalized improvement of MAE has been calculated (Eq.3):

Modelling data: WRF-Chem
The evaluated data comes from regional air quality-climate simulations performed using the WRF-Chem online-coupled 15 meteorology and chemistry model (Grell et al., 2005), version 3.4.1,under the umbrella of the EuMetChem Cost Action ES1004.A detailed description of the simulations can be found in Forkel et al. (2015).However, a brief description of the modelling methodology taken from the aforementioned work is described below.
The following physics options were applied for both simulations, including (or not) radiative feedbacks: Rapid Radiative Transfer Method for Global (RRTMG) longwave and shortwave radiation scheme; the Yonsei University (YSU) PBL 20 scheme, the NOAH land-surface model and the updated version of the Grell-Devenyi scheme with radiative feedbacks.
Further description of the physics can be found in Grell et al. (2005).
For all simulations discussed in this paper the native modelling grid spacing is 23 km (270 by 225 grid cells, Lambert Conformal Conic projection with center at 50N and 12E).The modelling domain covers Europe and a portion of Northern Africa and as well as large areas affected by the Russian forest fires.However, because of the scope of the paper is the IP, 25 only data for a domain covering the IP has been used (Fig. 1).In the vertical direction, the atmosphere up 50 hPa is resolved into 33 layers with a higher resolution close to the surface.Initial and boundary conditions for the meteorological variables were obtained from 3-hourly data with 0.25° resolution (analysis at 00 and 12 UTC and respective forecasts 3/6/9 hours) from the ECMWF operational archive.3-hourly chemistry Research) from a recent update of the TNO MACC emissions inventory (http://www.gmes-atmosphere.eu/;Pouliot et al., 5 2012Pouliot et al., 5 , 2014;;Kuenen et al., 2014) were applied.
Biomass burning emission data have been calculated from global fire emission data that have been supplied from the integrated monitoring and modelling system for wild-land fires (IS4FIRES) project (Sofiev et al., 2009) with 0.1º x 0.1º spatial resolution.Day and night vertical injection profiles were also provided.WRF-Chem emission species have been calculated by speciation following Andreae and Merlet (2001) and Wiedinmyer et al. (2011).However, no heat release due 10 to the fires was taken into account.
Biogenic emissions are based on the Model of Emissions of Gases and Aerosols from Nature (MEGAN) model (Guenther et al. 2006).MEGAN is on-line coupled with WRF-Chem and makes use of simulated temperature and solar radiation.
The most important feature to bear in mind for this work is the aerosol module.This aerosol module is based on the modal aerosol MADE (Modal Aerosol Dynamics Model) (Ackermann et al., 1998) which is a modification of the Regional Particulate Model (Binkowski and Shankar, 1995).Here aerosol particles are represented by two lognormal size distributions, corresponding to an Aitken mode and an accumulation mode.They both describe submicrometer-diameter particles and micrometer particles.SOA have been incorporated into MADE in the SORGAM (Secondary Organic Aerosol Model) module (Schell et al., 2001).
Although the modelling domain covers all Europe, for the purpose of this work data from the IP with a resolution of 0.2º has 20 been extracted for two important aerosol episodes in 2010.One of these episodes consists of a Saharan desert dust outbreak (from 28 June to 12 July) and a forest fires episode (from 25 July to 7 August).These episodes are selected because they represent two situations with a high load of atmospherics aerosol particles, when the radiative budget can be strongly affected.No volcanic emissions were considered in spite of the Eyjafjallajökull eruption in spring 2010.However, the volcanic plume reached all the IP only in May 2010 (Sicard et al., 2012;Navas-Guzmán et al., 2013), which is out of the 25 scope of these case studies.
The simulations are run for two different configurations differing in the inclusion/no-inclusion of aerosol radiative feedbacks (ARI+ACI).The base case or NRF simulation, does not take into account any aerosol feedbacks and the RF simulation adds the ARI and ACI to the previous modelling setup.Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-473, 2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.

Moderate Resolution Image Spectrometer (MODIS)
The satellite data chosen to evaluate the WRF-Chem simulations comes from MODIS (Levy et al., 2005) Level-2 Atmospheric Aerosol Product (MXD04_L2), collection or version 6 (C6) (Levy et al., 2013).The MODIS Aerosol Products monitor the ambient aerosol optical thickness over the oceans globally and over a portion of the continents.Daily Level 2 5 data have a spatial resolution of a 10x10 km.Two MODIS Aerosol data product files have been selected: MOD04_L2, containing data collected from the Terra platform; and MYD04_L2, containing data collected from the Aqua platform.In this case, the MXD04_L2 provides full global coverage of aerosol properties from the Dark Target (DT) aerosol retrieval algorithm, which is applied over ocean and dark land (e.g., vegetation) (Levy et al., 2013).
The variables used from MODIS are Aerosol Optical Depth (AOD) and Angström Exponent (AE). 10 AOD corresponds with AOD at a wavelength of 550 nm (AOD 550 ) for both ocean (best) and land (corrected) with best quality data (Quality Assurance Confidence = 3).The valid range of data is -0.05 to 5.0; that means a permission of small negative AOD values in order to avoid an arbitrary negative bias at the low AOD 550 end in long-term statistics.This is because MODIS does not have sensitivity over land to retrieve aerosol to better than ±0.05 + 15 % under very clean conditions.Negative values of AOD 550 have been considered as zero in this study.Over ocean the estimated error is -0.02 -15 10%, +0.04 + 10% (Levy et al., 2013).AE stands for AE for wavelengths between 550 and 860 nm (AE 550/860 ) over the ocean.The valid range for this variable is -1.0 to 5.0.In Collection 6, the preliminary estimated error for AE 550/860 is 0.45; pixels with an AOD 550 > 0.2 are expected to have a more accurate AE 550/860 representation (Levy et al., 2013).

Aerosol Robotic Network (AERONET) 20
The Aerosol Robotic Network (AERONET) collaboration (Holben et al., 1998) provides globally distributed observations of spectral AOD, inversion products, and precipitable water in diverse aerosol regimes.The highest quality data can be found in Version 2, Level 2.0 (cloud-screened and quality-assured) data products.
The data used from AERONET in this work comes from level 2.0 of AOD at different wavelengths (AOD 440 , AOD 675 , AOD 870 and AOD 1020 ) and AE (AE 440/870 ) from stations covering the IP available for the episodes studied (Table 1).25 Typically the total uncertainty for AOD data under cloud-free conditions is <±0.01 for λ>440 nm and <±0.02 for shorter wavelengths (Holben et al., 1998).

European Aerosol Research lidar Network (EARLINET)
EARLINET (Pappalardo et al., 2014) is the first aerosol lidar network, established in 2000, with the main goal to provide a comprehensive, quantitative, and statistically significant database for the aerosol distribution on a continental scale.30 used include backscatter profiles (BSCAT) at 355 and 532 nm (for the dates and times selected, no information is available at 1064 nm).The only station with available data for the studies cases in the IP during the year 2010 is Granada, and therefore is the only station included in this study.

Significance level of simulated changes 5
First, to assess the effect of the inclusion of aerosol radiative feedbacks in the on-line coupled WRF-Chem model on the studied variables, the significance study described in Section 2 has been carried out.During the dust episode (Fig. 1), the inclusion of radiative feedbacks produces differences with a significance level (defined as the ratio for the NRF-RF differences and the associated variance for the case not taking into account any radiative feedbacks) for AOD 550 higher than 60% over the south-western IP.The rest of the domain presents S.L. ratios > 100 % in spite of the high AOD 550 variance 10 values (above 0.05).In the case of AE 440/870 , the entire domain shows significance levels higher than 100%.
The inclusion of radiative feedbacks during the simulated fire episode (Fig. 2) produces differences with a S.L. > 100%) for AOD 550 over most of the domain.Over the area of fire particles emissions, S.L. ranges between 50 and 100% due to the higher absolute changes (> 0.2) than variance values (> 0.05).Similarly, for the dust episode the AE 440/870 over the entire domain shows significance levels > 100%.15 Hence, over most of the domain the changes or differences due to the inclusion of radiative feedbacks have a high S.L., usually higher than 100%, and therefore the changes modelled are significant with respect to the variability of the studied variables.We can then state then that the changes discussed below are caused by the inclusion of the aerosol radiative feedbacks in our simulations, and not to the mere signal noise.

Model output vs. Terra-MODIS data 20
The results of the comparison between model outputs with MODIS data from Terra platform are shown in Figs.3-6.The results from Aqua Platform are similar to Terra, and are therefore not shown here (but included in the Supplementary Material).Fig. 3 (a) shows the mean values of AOD 550 from MODIS for the dust outbreak.In this episode, high levels (above 0.4) over the south and the east of the domain are found, due to the shape of the dust outbreak.On the other hand, for the fires episode (Fig. 4 (a)), the highest levels of MODIS AOD 550 (> 0.25) are shown over the north of Portugal due to the 25 presence of black carbon coming from wildfires, and over the south of the domain where a dust intrusion occurred at the end of this episode (> 0.3).
For AOD 550 over the entire domain, the model outputs present low values of the MBE (represented by Figs. 3 and 4, (c lower during the fires episode, a peak of positive bias (0.47 for both simulations) is evaluated over the Portugal fire area, thus the model overestimates AOD 550 for biomass burning particles for both model configurations, including or not aerosol radiative feedbacks.However, we should bear in mind that this fact may be conditioned by the MODIS underestimation of AOD 550 levels for high loads of this type of particles, which has been reported by Chu et al., 2002;Levy et al., 2005 andRemer et al., 2005, among others.The estimation of the MAE (Table 2) shows a slight increase for the RF simulation in both 5 episodes for maximum and minimum values of this statistical figure.
With respect to the correlation coefficients (Figs. 3 (e) and 3 (f) for the dust episode and 4 (e) and 4 (f) for the fires episode), both simulations show high levels (around 0.9) of this statistical figure during the dust episode, except for those areas with high levels of AOD 550 , where the correlations are lower (even with negative correlations values close to -0.5).Conversely, for the fire episode, correlation values are close to 1 both for both cases (NRF and RF) over the entire domain, especially 10 over the areas with high values of AOD 550 .
When considering the improvement (or not) of the AOD 550 when including radiative feedbacks in the simulations, the difference in the MAE of the simulations between NRF and RF is estimated as defined in Eq. ( 2).For the dust episode (Fig 3 is simulated close to the source of biomass burning aerosols.However, an improvement (up to +0.11) over areas further away from this source is estimated, which compensates the importance of including aerosol feedbacks in the simulations when assessing the improvement of worsening of simulations.
In the case of the AE 550/860 from MODIS, the results are analogous for both episodes.Low values (< 0.45, shown in Fig. 5 (a)) of this variable over the southeast of the domain are found.This, together with the high levels of AOD 550 (Fig. 3 (a)), is a 20 clear indication that natural dust aerosols coming from the Saharan desert govern the AOD 550 levels here.On the other hand, for the fire episode (Fig. 6 (a)), the highest levels (around 1.6) are found over the north of Portugal, coincident with the fires areas, representing thus the emissions of biomass burning particles.Generally, for both simulations in both episodes, the model underestimates the high values of AE 550/860 and overestimates the low values.
For the dust episode, the MBE (Fig 5 (c) and (d)) minimum values are found of -0.65 and -0.62 for NRF and RF simulations 25 (underestimation) and the maximum MBE takes values of 0.77 and 0.78, respectively (overestimation).Concerning the correlation coefficient (Fig. 5 (e) and (f)), also for both simulations the value of this statistic is lower than for AOD 550 .Over most of the domain negative values are found (around -0.7) and positive values found are low (< 0.3).
On the other hand, during the fires episode (Fig. 6.) MBE minimum values (underestimation) are found around -0.61 and -0.65 for NRF and RF simulation, respectively, and maximum MBE values around 0.68 and 0.66 for NRF and RF 30 simulations.With respect to the correlation coefficient, just for the dust episode, positive correlations (> 0.5) are located over the most of the domain, while negative correlations are estimated over the emission areas of biomass burning particles (with values around -0.8).However, for both episodes, a slight decrease for maximum and minimum MAE values (Table 2) are observed when the radiative feedbacks are taken into account.

Model output vs. AERONET data
This section shows the results of the comparison between model output and AERONET data.First, a linear regression is estimated (Figs. 7, 8 and 9) and the correlation coefficients are calculated for the daily averages (Table 3).For AOD at different wavelengths during the dust episode, the results indicate that the stations where the model show higher skills are Barcelona and Sagres (maximum correlation coefficient 0.72) and, in for the fire episode, Caceres and Evora (maximum 10 correlation coefficient 0.9 and 0.85, respectively).For AE 440/870 , during the dust episode, the best-represented stations are Caceres and Sagres (maximum correlation coefficient 0.62 and 0.57, respectively) and, for the fire episode, Autilla (0.75) and Evora (0.66).Results do not indicate a clear improvement or worsening for both variables in both episodes when including the radiative feedbacks in our modelling configuration.
High levels of AERONET AOD are found between 2-10 of July 2010 due to the dust outbreak in Barcelona and Sagres (Fig. 15 10 (a) & (c)).The time series of these stations have been selected as representative among all AERONET stations in the IP affected by the Saharan dust outbreak (see Supplementary Material for information on the rest of AERONET stations over the IP).Maximum values of AERONET AOD occur between 7 and 10 July 2010.For AOD 1020 , AOD 870 and AOD 675 , the model underestimates the highest levels of AOD, represented by the minimum bias values (Table 3).On the other hand, between 2 and 6 of July 2010 (medium levels of AOD) the model overestimates the values of this parameter (Table 3).When 20 an underestimation(overestimation) is produced, the bias is lower(higher) for lower wavelengths.Sagres stations lacks of AOD 675 and AOD 440 data.Finally, the behaviour of AOD 440 in Barcelona is different from the other wavelengths due to the location of the station, close to a main street of the city where fine particles are emitted because of the road traffic.
For the fire episode, the shown stations are Caceres and Evora (see Supplementary Material for the rest of stations).In both stations, AOD presents the highest levels from 28 to 30 of July due to the wildfires occurred in Portugal (Table 3).Except 25 for the first two days, the model tends to underestimate the AOD values.For all wavelengths the bias or error, in both stations, increases when the wavelength decreases.
Regarding the AE 440/870 , for the dust episode the AERONET values shows low values corresponding to large particles (generally between 0 and 1) in Sagres station, indicating the dust origin of the particles at this site.For the fires episode, values generally range between 1.5 and 2.5 at Evora station, revealing the small size of the biomass burning particles.30 Generally for all stations in both episodes, the model overestimates(underestimates) AE 440/870 values when there are low(high) values of this variable.Hence, the model strongly undepredicts the variability of this variable for the two configurations.

Model output vs. EARLINET data
Finally, the results of the comparison between model output and EARLINET data are shown.In this section, only the dust episode is studied because the only station with available data for both the study episodes in the IP during the year 2010 is Granada.At this site, dust has an important contribution to aerosol loads.Several specific days (6 and 12 July 2010) are shown for the sake of brevity, but this discussion is valid for other days of this episode.It is important to notice the 5 differences between both discrete profile resolutions: the model profile with 33 levels from the ground to approx.20 km; and the profile measurement, which much higher vertical resolution (7.5 m).So the results below should be considered mainly from a qualitative perspective.However, in order to provide a more quantitative approach, the MAE of the model versus lidar observations is estimated.
As for the particle backscatter (BSCAT) at 532nm for 6 July 2010 (Fig. 11 (a)), the lidar detects a peak between 1.5 and 2 10 x10 -6 m -1 sr -1 around 3250 m above sea level caused by a dust layer.Although the model outputs overestimate the BSCAT values, simulations capture the profile of BSCAT.Although NRF and RF model configurations perform similarly, there is a slight improvement in the MAE of the vertical profile (estimated after Eq. 3) when the radiative feedbacks are taken into account (Fig. 11 (a)).Average MAE is 6.37 x10 -7 and 6.22 x10 -7 m -1 sr -1 for NRF and RF simulations, respectively.
Henceforth, the normalized MAE is improved by 2.4% when radiative feedbacks are included in WRF-Chem simulations.15 For the BSCAT for 12 July 2010 (Fig. 11 (b)), the model overestimates the BSCAT values of the vertical profile, as aforementioned.However, the shape of the vertical profile is correctly reproduced.Mean MAE is 3.14 x10 -7 and 3.12 x10 -7 m -1 sr -1 at 355nm for NRF and RF simulations, respectively; and 4.1 x10-7 m -1 sr -1 at 532nm for both cases.Here, the improvement when including radiative feedbacks is very limited, and estimated as 0.63% and 0.14% at 355 and 532nm, respectively.20

Conclusions
The use of on-line coupled models is one of the most powerful tools to understand the different processes influencing the climate system.In particular, for the study of atmospheric aerosol particles realistic simulation of the combined ARI and ACI are needed, irrespective of the aerosols source, where the interactions of aerosols, meteorology, radiation, and chemistry are coupled in a fully interactive manner.The use of modelling tools requires the observational study of physical, chemical 25 and optical properties of aerosol particles to establish its behaviour and to assess how good these properties are represented in on-line coupled models.
In this study, two configurations including/not including (NRF/RF simulations) the aerosol radiative feedbacks have been assessed against a number of remote sensing observations for two episodes characterized by dust and biomass burning aerosols, respectively.For the comparison between model output and MODIS data, the best-represented variable is AOD, 30 with low values of mean bias and high values of correlation coefficient both for NRF and RF simulations.The inclusion of the radiative feedbacks produces a slight improvement in the model representation for medium values of this variable and a Atmos.Chem. Phys. Discuss., doi:10.5194/acp-2016-473, 2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.
worsening for the lowest and highest values.At the same time, the model output of AE representation leads to underestimate the variability of this variable.This occurred for both episodes and may be related to the fact that the size distribution of the aerosol function within WRF-Chem considers a medium size of particles, smaller for dust and larger for fires particles.The inclusion of aerosol feedbacks does not produce a clear benefit, taking into account the expensive computational cost required for including the ARI and ACI in the model.5 As well as for MODIS, for the comparison between model output and AERONET data, the results indicate that the bestrepresented variable is AOD.Generally, for both episodes, the model underestimates the levels of AOD, but the highest levels of this variable for dust episode are underestimated.It is important to note that the bias is usually higher for low wavelengths.In both episodes, the AE is overestimated for low levels and underestimated for high levels, since the modelled variability is strongly underestimated.For both variables, there is not a clear improvement of the model outputs for the 10 radiative feedbacks simulation for any station in both episodes.
For the comparison between model output and EARLINET data, the results show a general slight improvement in the representation of vertical aerosol profiles when the radiative feedbacks are taken into account for all studied wavelength.
It is important to take into account these considerations to improve the time-efficiency when running the simulations, because the inclusion of radiative feedbacks in the simulations has a notable increase of the computational time.The 15 improvements observed, in particular related to the vertical distribution of aerosols, justify the inclusion of radiative feedbacks in WRF-Chem on-line coupled model and the much higher time devoted to running the simulations.
Atmos.Chem.Phys.Discuss., doi:10.5194/acp-2016-473,2016   Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.boundary conditions for the main trace gases and particulate matter concentrations were available from the ECMWF IFS-MOZART model run from the MACC-II project (Monitoring Atmospheric Composition and Climate-Interim Implementation, Inness et al., 2013) 1.125° spatial resolution.Anthropogenic emissions for the EU domain provided by the TNO (Netherlands Organization for Applied Scientific EARLINET data include particle backscatter and extinction coefficient profiles at 355, 532 and 1064 nm.EARLINET data Atmos.Chem.Phys.Discuss., doi:10.5194/acp-2016-473,2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.
) and (d)) for both NRF and RF simulations.During the dust episode the model underestimates MODIS AOD 550 (MBE minimum values for NRF and RF simulations, respectively, -0.31 and -0.36) over the locations with important dust loads (high 30 AOD 550 ) and overestimates (MBE maximum values 0.32 and 0.31) the low levels of AOD 550 .Although the bias is generally Atmos.Chem.Phys.Discuss., doi:10.5194/acp-2016-473,2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.
(b)), a slight improvement(worsening) is produced over the areas with medium(high/low) levels of AOD 550 , taking these changes values between -0.09 and +0.12.For the fire episode(Fig 4 (b)), a worsening of MAE (difference NRF-RF of -0.02) 15 Atmos.Chem.Phys.Discuss., doi:10.5194/acp-2016-473,2016 Manuscript under review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.At the same time, there is a slight improvement for RF simulations for the dust episode over the areas where the AE 550/860 is overestimated (reaching values of improvement of MAE of 0.13) and a slight worsening (values of improvement of MAE around -0.09) over the areas where this variable is underestimated (Fig 5 (b)).For the fire episode, a slight improvement (values of improvement of MAE of 0.16) is found over the south-eastern part of the domain and a slight worsening (around -0.18) over the rest of the IP (Fig 6 (b)).5

Figure 1 :
Figure 1: Dust episode (top: AOD 550 ; bottom: AE 440/870 ).From left to right: (a) modelled value of the variable, (b) value of the absolute differences between NRF-RF simulations, (c)variance value of NRF simulation and (d) S.L. values.

Figure 3 :
Figure 3: Comparison of AOD 550 model output vs. AOD 550 from MODIS data for the dust episode.(a) AOD MODIS values.(b) Improvement of MAE due to the inclusion of RF (MAE in RF-NRF simulations).(c) and (d) MBE for NRF and RF simulations, respectively.(e) and (f) correlation coefficient for NRF and RF simulations.

Figure 5 :
Figure 5: Comparison of AE 550/860 model output vs. AE at same wavelength from MODIS data for the dust episode.(a) AE MODIS values.(b) Improvement of MAE due to the inclusion of RF (MAE in RF-NRF simulations).(c) and (d) MBE for NRF and RF simulations, respectively.(e) and (f) Correlation coefficient for NRF and RF simulations.

Figure 7 :
Figure 7: Linear regression between AERONET (x) and simulations daily data (y; NRF in circles and RF in squares) for the dust episode: (a) AOD 1020 (b) AOD 870 (c) AOD 675 and (d) AOD 440 .

Figure 9 :
Figure 9: Linear regression between AERONET (x) and simulations daily data (y, NRF in circles and RF in squares) of AE 440/870 : (a) dust episode; (b) fire episode.

Figure 10 :
Figure 10: AERONET (dots), NRF (line) and RF simulations (dashed line).AOD at different AERONET wavelengths: (a) and (c) Barcelona and Sagres stations for the dust episode and (b) and (d) Caceres and Evora stations for the fire episode.(e) AE 440/870 in the Sagres station for the dust episode and for the (f) Caceres station for the fire episode.

Table 3 : Values of correlation coefficient for the linear regression between AERONET and simulation daily means. Values in italic 5 indicate the highest correlation coefficient among the different AOD/AE.
Atmos.Chem.Phys.Discuss., doi:10.5194/acp-2016-473,2016Manuscriptunder review for journal Atmos.Chem.Phys.Published: 6 July 2016 c Author(s) 2016.CC-BY 3.0 License.