Sensitivity of spatial aerosol particle distributions to the boundary conditions in the PALM model system 6.0

. High-resolution modelling is needed to understand urban air quality and pollutant dispersion in detail. Recently, the PALM model system 6.0, which is based on the large-eddy simulation (LES), was extended with a detailed aerosol module SALSA2.0 to enable studying the complex interactions between the turbulent ﬂow ﬁeld and aerosol dynamic processes. This study represents an extensive evaluation of the modelling system against the horizontal and vertical distributions of aerosol particles 5 measured using a mobile laboratory and a drone in an urban neighbourhood in Helsinki, Finland. Speciﬁc emphasis is on the model sensitivity of aerosol particle concentrations, size distributions and chemical compositions to boundary conditions of meteorological variables and aerosol background concentrations. The meteorological boundary conditions are drawn from both a numerical weather prediction model and observations, which occasionally differ strongly. Yet, the model shows good agreement with measurements (fractional bias < 0 . 67 , normalised mean-square error < 6 , factor of two > 0 . 3 , normalised 10 mean bias factor < 0 . 25 and normalised mean absolute error factor < 0 . 35 ) in respect of both horizontal and vertical distribution of aerosol particles, their size distribution and chemical composition. The horizontal distribution is most sensitive to the wind speed and atmospheric stratiﬁcation and vertical distribution to the wind direction. The aerosol number size distribution is mainly governed by the ﬂow ﬁeld along the main street with high trafﬁc rates and in its surroundings by the background concentrations. The results emphasize the importance of correct meteorological and aerosol background boundary conditions, 15 in addition to accurate emission estimates and detailed model physics, in quantitative high-resolution air pollution modelling and future urban LES studies. LDSA tot , O PSD O MET O PSD better performance at the M MET M PSD underestimates LDSA, the supersite O MET O PSD overestimates LDSA.

High-resolution modelling is needed to understand urban air quality and pollutant dispersion in detail. Recently, the PALM model system 6.0, which is based on the large-eddy simulation (LES), was extended with a detailed aerosol module SALSA2.0 to enable studying the complex interactions between the turbulent flow field and aerosol dynamic processes. This study represents an extensive evaluation of the modelling system against the horizontal and vertical distributions of aerosol particles 5 measured using a mobile laboratory and a drone in an urban neighbourhood in Helsinki, Finland. Specific emphasis is on the model sensitivity of aerosol particle concentrations, size distributions and chemical compositions to boundary conditions of meteorological variables and aerosol background concentrations. The meteorological boundary conditions are drawn from both a numerical weather prediction model and observations, which occasionally differ strongly. Yet, the model shows good agreement with measurements (fractional bias < 0.67, normalised mean-square error < 6, factor of two > 0.3, normalised 10 mean bias factor < 0.25 and normalised mean absolute error factor < 0.35) in respect of both horizontal and vertical distribution of aerosol particles, their size distribution and chemical composition. The horizontal distribution is most sensitive to the wind speed and atmospheric stratification and vertical distribution to the wind direction. The aerosol number size distribution is mainly governed by the flow field along the main street with high traffic rates and in its surroundings by the background concentrations. The results emphasize the importance of correct meteorological and aerosol background boundary conditions, 15 in addition to accurate emission estimates and detailed model physics, in quantitative high-resolution air pollution modelling and future urban LES studies.

Introduction
Exposure to outdoor air pollution is a major global threat resulting up to 0.8 million premature deaths in Europe (Lelieveld et al., 2019) and 3 million worldwide (Lelieveld et al., 2015;WHO, 2016) every year. Specifically aerosol particles can be extremely harmful, and based on a recent study by Burnett et al. (2018) outdoor fine particulate air pollution (PM 2.5 ) solely could have caused up to 8.9 million deaths worldwide in 2015. As over half of the global population lives in cities (55 % according to 5 UN, 2019), urban air quality is of major importance. In addition to high population densities, urban areas are characterized by major air pollutant sources, namely traffic exhaust and road dust, being at the same level where urban dwellers inhale outdoor air. The dispersion of these traffic-related pollutants is not, however, straightforward as buildings, trees and other obstacles modify the flow within the urban canopy and hence also pollutant dispersion (Tominaga and Stathopoulos, 2013) as well as the environment for aerosol dynamic processes and chemical reactions to occur. 10 As a consequence of the complex interactions between the urban morphology, meteorology, local emissions and air pollutant dynamics and chemistry, air quality is highly variable both in time and space, and strong concentration gradients are observed in urban areas. However, measurements from a single monitoring station nearest to the individual's residence, hospital, or primary health care clinic have commonly been applied in air pollution exposure studies (Andersen et al., 2012;Adam et al., 2015), which can lead to notable errors. Moreover, both the size and chemical composition of aerosol particles are of major 15 importance when it comes to their health impacts (Kampa and Castanas, 2008;Kelly and Fussell, 2012). For instance, particle deposition in lungs depends strongly on the inhaled particle size (Hussain et al., 2011), and thus the negative health effects of aerosol particles have been found to correlate more strongly with the surface area of particles than their number or mass (Brown et al., 2001;Oberdörster et al., 2005).
Computational fluid dynamics (CFD) models have been successfully applied in studying the air flow and dispersion of air 20 pollutants in urban areas. Mainly models based on either Reynolds-averaged Navier Stokes (RANS, e.g., Baik et al., 2009;Kwak et al., 2015;Santiago et al., 2020) or large-eddy simulation (LES, e.g., García-Sánchez et al., 2018;Letzel et al., 2012;Salim et al., 2011) have been utilised. While being computationally more expensive than RANS, LES has been shown to perform better in resolving instantaneous turbulence structures in a complex urban environment (García-Sánchez et al., 2018;Salim et al., 2011). Further, air pollutant concentrations can be significantly modified by their chemical and physical processes 25 (Kurppa et al., 2019;Nikolova et al., 2016;Zhong et al., 2020), especially as the residence time of air pollutants is increased in a complex urban environment (Gronemeier and Sühring, 2019;Ramponi et al., 2015). Therefore a detailed module describing the characteristics of air pollutants and their dynamics is needed to enable modelling aerosol particles of different size, chemical composition and harmfulness. To date, only a few LES models include a module for treating aerosol particles with a specific size distribution and chemical composition and their dynamic processes (Kurppa et al., 2019;Steffens et al., 2013;Zhong et al., 30 2020).
The sectional aerosol module SALSA (Kokkola et al., 2008) was recently implemented to the PALM model system (Kurppa et al., 2019) to consider the impact of aerosol dynamic processes on aerosol concentrations and size distributions, and to study the relative importance of pollutant dispersion and aerosol dynamic processes. A model evaluation in central Cambridge, UK, showed the model to be capable of reproducing the vertical distributions of aerosol size distribution in a simple street canyon.
However, due to the lack of observations, the capability of the model to reproduce the horizontal distributions of aerosol particles has not been studied yet. Also the meteorological conditions were limited to one single day and the examined street canyon had no vegetation.
Still, even if the air pollutant processes would be modelled accurately, correct boundary conditions for the meteorological 5 variables and air pollutant concentrations are vital for realistic air quality simulations. Boundary conditions can be drawn from observations, which however are typically point measurements that lack spatial representatives and also are prone to measurement errors. Another alternative is to use model data, which provide a good spatial coverage but not necessarily stable performance. Previously, CFD models have been successfully coupled with mesoscale models to study the impact of larger scale atmospheric features on microscale interactions (e.g., Baik et al., 2009;Heinze et al., 2017;Liu et al., 2012;10 Michioka et al., 2013;Wyszogrodzki et al., 2012) as well as to consider realistic air pollutant background concentrations (Kwak et al., 2015). Recently, Santiago et al. (2020) investigated the sensitivity of RANS-based urban PM 10 (particulate matter with aerodynamic diameter < 10 µm) simulations on the meteorological boundary conditions and showed the model performance to be improved when replacing the wind direction (WD) predicted by the WRF model with the observed WD. However, Santiago et al. (2020) only modelled passive PM 10 without taking into account chemical or physical transformation of aerosol particles. 15 Hence, it is still unclear how much uncertainties in aerosol particle concentrations and size distributions can model boundary conditions cause.
To further assess the performance of SALSA2.0 in the PALM model system 6.0 in simulating the spatial distribution of aerosol particle concentrations in an urban area and to examine the importance of meteorological and aerosol background boundary conditions, we will use observations made during an extensive measurement campaign in an urban neighbourhood 20 in Helsinki, Finland, in summer and winter 2017. The campaign focused on the spatial variability of aerosol particle number, surface area and mass both in horizontal and vertical as well as aerosol size distributions and chemical composition with a high temporal and spatial resolution measured using a mobile laboratory and a drone. The model evaluation is done at three observation periods with different prevailing meteorological conditions.

Measurement campaign
The model evaluation and sensitivity study is conducted around an Helsinki Region Environmental Services Authority (HSY) air quality monitoring site, hereafter referred as the "supersite", in Helsinki, Finland (60 • 11'47"N, 24 • 57'07"E). The site is located 3 km north-northeast from the Helsinki city centre, and it is characterised as an urban street-canyon kerbside station with a traffic rate of around 28,000 on a workday, of which 12 % are heavy duty vehicles (City of Helsinki, 2018). The street 30 canyon is 42 m wide and the mean building height is around 19 m on the southwestern and 16 m on the northeastern side of the street (see Fig. 1 in Kuuluvainen et al., 2018) resulting in a height to width ratio of 0.42. The supersite consists of a container Table 1. Instrumentation of the mobile laboratory Sniffer. Abbreviations: PSD = aerosol particle number size distribution, Ntot = total aerosol particle number concentration and PM1 = mass of particulate matter with aerodynamic diameter < 1 µm. To get information about the spatial variability of air pollutants around the supersite, a two-week measurement campaign was conducted in summer (6-16 Jun) and winter (28 Nov-11 Dec) 2017. During both campaigns the horizontal distribution of air pollutants in the neighbourhood was monitored on non-rainy days using a mobile laboratory and additionally during two 5 intensive observation periods the vertical profiles of aerosol particles were measured using a drone.

Measured component
The mobile laboratory Sniffer (Pirjola et al., 2004) measured the horizontal distribution of trace gases and aerosol particle concentrations and size distribution. The measurements were done in one to two hour slots with a 1-s temporal resolution during the morning and afternoon rush hours, around noon and in the late evening. During each observation period, Sniffer was driving along a main street (Mäkelänkatu) and a side street as well as standing at the supersite, opposite the supersite and 10 on a field 185 m from the main street (hereafter "background"). The instrumentation of Sniffer is given in Table 1  During the intensive observation periods, a multi-rotor drone (X8, VideoDrone Finland Ltd) carried an electrical particle 15 sensor (Partector, Naneos GmbH) to measure the vertical distribution of the alveolar lung-deposited surface area (LDSA) of aerosol particles, which describes the total aerosol surface area penetrating to the deepest parts of lungs (see e.g., Kuula et al., 2020, and references within). The measurement were done on both sides of the street canyon when the Sniffer was simultaneously driving. The drone was flown ten times up-and-down between z = 2−50 m during one 30-minute measurement interval, after which measurements were repeated on the other side. Each intensive observation period started by measuring 20 LDSA at the supersite and ended on the other side. Measurements were started at 3 m from the building wall and the horizontal location was kept constant with a GPS sensor of the drone. Additionally, LDSA was measured at the supersite by a Pegasor AQ Urban sensor (Pegasor Ltd.) and on the other side by a DiSCmini (Testo Ltd.) or with another Partector at 1 m and in winter also at 14 m. For the details of the instrumentation, see Kuuluvainen et al. (2018).

Additional measurements
In addition to the Sniffer and drone measurements, we use stationary aerosol observations from the supersite and two urban background monitoring sites: Kallio site operated by HSY and SMEAR III (Station for Measuring Ecosystem Atmospheric relations, Järvi et al., 2009) around 1.0 km southwest and 0.8 km northeast from the supersite, respectively (Fig. 1). See Table S1 for the instrumentation. In addition to aerosol observations, meteorological data (wind speed, wind direction, air 5 temperature) from the SMEAR III measurement tower (z = 31 m) and Kivenlahti meteorological measurement mast 17.4 km west from the supersite (Wood et al., 2013) are used in the study.

Model description
This study applies the PALM model system, version 6.0 (revision 4416) (Maronga et al., 2015(Maronga et al., , 2020, which features an LES core for atmospheric and oceanic boundary layer flows. PALM solves the non-hydrostatic, filtered, incompressible Navier-Stokes equations of wind (u, v, and w) and scalar variables (sub-grid-scale turbulent kinetic energy e, potential temperature θ, 5 and specific humidity q) in Boussinesq-approximated form. PALM is especially suitable for complex urban areas, owing to its features such as a Cartesian topography scheme and a plant canopy module, which are applied here to include the aerodynamic impact of both solid buildings and permeable vegetation on the flow. Furthermore, so called PALM-4U (short for PALM for urban applications) components have recently been implemented to PALM (Maronga et al., 2020), including the aerosol module SALSA, the online chemistry module, and the self-and offline nesting features, which are all applied in this study. pre-processing tool (Jöckel et al., 2010). The implementation is flexible, allowing the user to choose the chemical mechanism and components being considered. In this study, a simplified mechanism describing photochemical smog is applied (see Supplement, Section S3.2). Photolysis is parametrised based on Saunders et al. (2003). 20 To capture the dominant turbulent eddies of the atmospheric boundary layer (ABL) in LES, the horizontal extent of the modelling domain should span over several ABL heights, see e.g. (Fishpool et al., 2009;Chung and McKeon, 2010;Auvinen et al., 2020). At the same, to resolve most of the kinetic energy within street canyons, a high enough grid resolution is needed (Xie and Castro, 2006). Furthermore, uncertainty arising from the lateral boundary conditions usually decreases with increasing horizontal dimensions. To fulfill these contradicting requirements, a self-nesting feature has been included in PALM (Hellsten 25 et al., 2017;Maronga et al., 2020). In self-nesting, one or several child domains are nested within a parent domain and the child obtains its boundary conditions from its parent. Furthermore, PALM incorporates an automated mesoscale offline nesting with a mesoscale operational weather prediction model, which allows realistic, non-cyclic and non-stationary boundary conditions for the flow. To reduce the time and distance for the mesoscale flow field to adjust to the LES modelling domain, a synthetic turbulence generator within PALM can be applied.

Model domain and morphological data
The model simulations are conducted over a root domain of 6.9 km × 6.9 km, within which two smaller domains, parent and child, are nested progressively (Fig. 1). The dimensions (L x , L y , L z ), number of grid points (N x , N y , N z ) and grid resolutions Table 2. Dimensions (L), number of grid points (N ) and grid resolutions (∆) of the model domains in x-, y-and z-directions. (∆ x , ∆ y , ∆ z ) of each domain are given in Table 2. In this study, the focus is on the child domain which matches with the area of the spatial aerosol measurements around the supersite.
Information on the building and vegetation height and land surface elevation are drawn from high-resolution raster maps for Helsinki (Auvinen and Aarnio, 2019). The manipulation of the domain input files is done using the Python library P4UL (Auvinen and Karttunen, 2019). Only vegetation higher than z v,min = 4.0 m are included in the simulations. Due to the lack 5 of observational data on the leaf area density (LAD) of vegetation, a constant LAD value is applied for all tree crowns above z v,min . In summer, LAD = 1.2 m 2 m −3 for broad-leaf trees (Abhijith et al., 2017), while in winter LAD is decreased to 20 % of the summertime value.

Meteorological boundary conditions
We apply both modelled and observed data as meteorological boundary conditions, which are set dynamic, i.e., they change The initial conditions and dynamic meteorological boundary data are provided to PALM in a so-called dynamic driver. Of the MEPS data, the dynamic driver was created by the following procedure. First, the sigma-coordinates were translated to pressure coordinates and further to height coordinates applying the hypsometric equation. Then u, v, w, θ and water vapour mixing ratio q v were interpolated from the MEPS grid to the PALM grid: first in horizontal over a two-dimensional grid using the cubic spline method and then in vertical using the linear interpolation. The Kivenlahti mast observations, instead, were linearly interpolated in vertical until the highest observation level, after which a constant value was used. The dynamic driver created from the observational data does not include any horizontal variation.
A mesoscale interface, INIFOR, has been developed to transform mesoscale modelling data into PALM-readable boundary 5 data. However, it is currently only available for COSMO-DE/D2 datasets, which do not cover Finland.

Air pollutant background concentrations
Similar to the meteorological boundary conditions (Section 3.3), both modelled and observed air quality data are used as background concentrations in the simulations. As in the previous model evaluation study (Kurppa et al., 2019), the modelled background aerosol particle number and trace gas concentrations are produced with the trajectory model for Aerosol Dynam-

Air pollutant emissions
In this study, air pollutant emissions only from traffic combustion are included, as traffic is the main pollutant source within the modelling domain (Helsinki Region Environmental Services Authority). Traffic-lane maps separating different road categories, 25 i.e., main streets, collector roads and residential streets, have been generated by combining lane and street type information from the Map Service (City of Helsinki). The lane width is 3.5 m. Emissions are introduced as dynamic surface fluxes.
Aerosol particle emission inventories are typically provided as total mass emissions EF PM2.5 . In SALSA, these would need to be translated to number emissions EF N , assuming some size distribution for the emitted aerosol particles. However, converting aerosol mass to number is highly sensitivity to the assumed size distribution. Therefore in this study we choose to 30 apply a number emission factor EF N = 4.22×10 15 kg −1 fuel based on fuel consumption and a number size distribution estimated by Hietikko et al. (2018) at the supersite in May 2017. Table 3. Unit emission factors for traffic combustion (s = solid and g = gaseous) on 9 Jun between 7:00-8:00 in units For gaseous compounds, mass composition of aerosol particles and fuel, unit emission factors EF [compound] (Table 3) are calculated using emission inventory by the European Environmental Agency for 2017 (Ntziachristos et al., 2016) and specifically the Tier 3 method, which applies information on the mileage per vehicle category and technology, and driving speed.
However, since no information on the cumulative mileage for different Euro classes was available, EF NH3(g) and EF N2O(g) are based on the Tier 1 method (see Ntziachristos et al., 2016, Eq. 28). Furthermore, the following estimates were applied:

Model set-up
The length of the morning simulations on 9 Jun and 12 Dec are two hours, and evening simulation on 9 Jun only one hour. 20 Simulation times correspond to the observation periods.  The aerosol and chemistry modules are run only within the child domain to limit computational costs. In all simulations, the aerosol processes of condensation and dissolutional growth, coagulation, dry deposition and sedimentation are included and calculated every 1.0 s. The aerosol particle size distribution is described by 10 size bins, of which three are within the first subrange between 2.5-15 nm and seven within the second subrange 15 nm-1 µm. Aerosol particles are assumed to be internally mixed and hygroscopic, and can contain H 2 SO 4 , OC, BC, HNO 3 , and/or NH 3 . The chemical reactions are calculated at every 10 time step of the PALM model.
The advection of both momentum variables and scalars is based on the fifth-order advection scheme by Wicker and Skamarock (2002) together with a third-order Runge-Kutta time-stepping scheme (Williamson, 1980). The pressure term in the prognostic equations for momentum is calculated using the iterative multigrid scheme (Hackbusch, 1985). The roughness height is z 0 = 0.05 m (Letzel et al., 2012) and the drag coefficient applied for the trees C D = 0.3. Kivenlahti mast, expect that T is roughly 2 • C higher at SMEAR III compared to the Kivenlahti mast and WD typically falls between the MEPS data and Kivenlahti observations. The difference in WD can be explained by flow distortion at SMEAR III due to the adjacent buildings to the north of the measurement site (Nordbo et al., 2012). The observed background aerosol particle number concentrations at SMEAR III are around 80 % lower and the modelled PSD shows a smaller peak diameter of D = 28 nm instead of D = 50 nm in the SMEAR III observations (Fig. S5). Furthermore, the observations show a secondary 20 peak at D = 14 nm, which is not captured by ADCHEM.
By the evening, the observed U on the Kivenlahti mast had increased to 2.0 − 2.5 m s −1 at z = 30 m (Fig. S6a) and the wind turned to south-west (Fig. S6b). The modelled and observed WD agree well (∆WD < 20 • ), whereas clear discrepancy is shown for U . The MEPS predicts a low-level jet with the maximum U at z = 100 m and shows up to 3 m s −1 higher values compared to the Kivenlahti observations at 8-9 pm. This low-level jet results in a strong wind shear and mechanical 25 turbulence production. Instead above, U is overestimated in the interpolated Kivenlahti data at 9-10 pm. The profiles of T D agree relatively well (Fig. S7), whereas MEPS predicts clearly lower T , with a difference up to −5 • C close to the ground (Fig. S6c). The SMEAR III observations agree with those from the Kivenlahti mast. The modelled and observed background PSD agree in shape, but the peak is observed at D = 79 nm compared to D = 100 nm in ADCHEM and observed total number concentration is around 35 % lower (Fig. S5).

30
In the winter morning on 7 Dec, easterly flow was observed and the wind was turning to south-east with both height and time ( Fig. S8a-b). Winds were stronger than in the summer morning, around 2 m s −1 at z = 30 m. The observed and modelled WD agree, but MEPS predicts up to 3 m s −1 lower U above the canopy. An inversion layer above ground is captured both in  MEPS and observations (Fig. S8c), yet it is stronger in the observations especially during the first hours. As a contrast to T , MEPS predicts down to −3 • C lower T D compared to the Kivenlahti observations at 9 am (Fig. S9). Similar to the summer morning, the observations on the Kivenlahti mast and SMEAR III are in agreement. Both the modelled and observed PSD peak at D = 31 nm, but the observed total number concentrations are around 60 % higher (Fig. S10).

5
The model is evaluated against observations at three different observations periods, and in both summer and winter morning the evaluation is done separately for both modelling hours. The following performance measures are applied in the evaluation: fractional bias (FB), normalised mean squared error (NMSE), factor of two (FAC2) (Chang and Hanna, 2004), normalised mean bias factor (NMBF) and normalised mean absolute error factor (NMAEF) (Yu et al., 2006). See Appendix A for the definitions and Table 5

Horizontal distribution of total aerosol particle number concentration
In order to compare the data, both the mobile Sniffer measurements containing its geographical coordinates and the PALM data output have been horizontally aggregated to a 5 m× 5 m grid, with a threshold of at least three measurement points per grid to calculate the median value. A comparison between the measured and modelled total aerosol particle number concentration (N tot ) values is illustrated in Fig. 3. In general, the model captures the large concentration gradient between the main street 10 (in the middle from northwest to southeast) and the side street on the northeast side of the main street. However, the model overestimates N tot at the north-western end of the Sniffer route at all simulation times. Along the side street, the modelled N tot is slightly higher than the observed in M MET M PSD during the first hour in the summer morning (Fig. 3f) and in O MET O PSD during the first hour in the winter morning (Fig. 3n).  Table S2 in the Supplement).   The grey indicates that the value exceeds the acceptance criteria given in Table 5. See Table 4 for the simulation names. Note that in the summer evening and winter morning, only two simulations have been conducted. within the acceptance criteria (see Table S3 in the Supplement). Also at the background, the Sniffer measurements agree with the model based on FB, NMSE and FAC2 even though the concentration of the smallest (the mean bin diameter D mid < 25 nm)

Aerosol size distribution
aerosol particles is underestimated. Instead along the side street, the modelled values are clearly lower than the observed and, for instance, based on NMBF the model underestimates the EEPS and ELPI observations by a factor of 3.45 and 5.48, respectively.
Comparing the two simulations with different boundary conditions, M MET M PSD performs better along the main street and 5 hence also at the supersite and opposite it during the first hour of the summer morning (Tables S3 and S4)

Aerosol chemical composition
The chemical composition of aerosols was measured at the supersite by an ACSM and the black carbon (BC) concentration by a MAAP (see the Supplement, Table S1). Additionally Sniffer measured BC within PM 1 . In general, the modelled and observed horizontal distribution of BC compare tolerably well based on the performance measures FB, NMSE and FAC2, while NMBF and NMAEF are not within the acceptance criteria (see Fig. S11 in the Supplement). Overall, the performance is

Background aerosol size distribution
Regarding all variables used in the evaluation, only minor differences due to using modelled or measured PSD as a bound-  (Fig. 11). Only a small decrease in model performance is observed opposite the supersite when applying the observed PSD as the boundary condition (e.g., FB is increased from 0.20 to 0.34 and NMSE from 0.07 to 0.14 during the first hour, Fig. 8). However, above z > 30 m, the difference gradually approaches 65-160 %, i.e., the relative difference in the background N tot between the modelled ADCHEM values

Background meteorological conditions
In Section 4.2, the simulation using the observed data as boundary conditions (O MET O PSD ) was shown to perform worse than M MET M PSD in the summer morning. As the observed wind speed at Kivenlahti and the one modelled by MEPS differ (see Section 4.1), we separately investigate the influence of the incoming wind direction on the model sensitivity.
In general, O MET O PSD and O WD,mast O PSD , for which the incoming wind direction is replaced with the one measured on the 5 Kivenlahti mast but the wind speed is the same, result in a similar pattern for the difference in the horizontal distribution of N tot compared to M MET M PSD (Fig. 10a,c). However, the differences are larger for O MET O PSD , for which the incoming upperlevel wind speed is up to 2 m s −1 slower during the first hour (Fig. 2a). This results in the aerosol particles being transported more to the southwest side of the main street. As the wind is more from the north in O MET O PSD than in O WD,mast O PSD , also the impact of wind direction on the street canyon vortex along the main street is observed by clearly lower (higher) All ∆LDSA profiles gradually approach ∼100 %, which results from using different PSD for M MET M PSD and for rest of the simulations.

5
Similarly for the aerosol size distribution, changing the modelled wind direction by MEPS to the one measured at Kivenlahti improves model performance at the background and slightly decreases elsewhere during the first hour, whereas during the last hour PSD is modelled better along the side street and at the supersite (see Tables S3 and S6 in the Supplement). Applying the wind direction from SMEAR III generally does not improve model performance (Table S7). Along the main street, the difference in PSD is mainly governed by the emission, which is shown by the peak at D mid ≈ 30 nm in Fig. S12 (in the Supplement), while at the background ∆N follows the difference in the boundary conditions for aerosol particles.

Discussion and conclusions
This study provides an extensive evaluation of the SALSA2.0 module in the  By the evening, MEPS shows only slightly more southerly but clearly stronger winds at z < 200 m than what is observed on the Kivenlahti mast. To be precise, MEPS predicts a low-level jet with the maximum wind speed at z = 100 m. Still, both simulations perform nearly equally. Slightly higher N tot and overestimation of near-surface LDSA in O MET O PSD can be explained by the difference in the wind speed. This is also reflected in the aerosol chemical composition at the supersite.
MEPS predicts a more stable stratification, which might justify why the difference in the spatial variability of aerosol particle Consequently, meteorological boundary conditions are particularly important for quantitative urban air quality modelling 15 using LES, and therefore the inlet meteorology should be evaluated prior to conducting CFD simulations (Santiago et al., 2020). However in our case, we are unable to evaluate the modelled meteorology. In Santiago et al. (2020), the meteorological observations were made within the simulation domain, while in our case the closest measurements at SMEAR III are 800 m away from the supersite. Observations are available also from the Kivenlahti mast, which has several measurement levels, but is located over 17 km away from the supersite and represents more semi-urban to rural area. Another problem with the Kivenlahti 20 data is the lack of wind observations above 217 m in summer, which presumably leads to, for instance, underestimation of the incoming wind speed during the first simulation hour in the summer morning and around 9 pm in the summer evening. Consequently, neither observations are optimal for evaluating the modelled meteorology nor providing meteorological boundary conditions for the simulations.
Of the aerosol metrics applied, LDSA directly estimates the health effect of aerosol exposure. The mean modelled LDSA However, LDSA is often overestimated near the ground in our simulations. One limitation of this study and in general in urban LES is omitting vehicle-induced turbulence (VIT), which would enhance vertical pollutant transport and mixing near the surface and very likely decrease concentrations near ground. The research to include VIT in LES without extensive computational costs is on-going and currently no freely available VIT-model exists for LES. Neglecting the thermal turbulence in the simulations is another important limitation of our study. We acknowledge that omitting the influence of anthropogenic heat and heating by incoming solar radiation leads to overestimation of the vertical stability near the ground, which can partly explain the overestimation of the modelled surface concentrations. However, the spatial variability has been shown 5 less dependent on a detailed heating distribution Nazarian et al. (2018) and therefore the horizontal distribution is mainly determined by the predominant inflow conditions. Lastly, condensation of the biogenic volatile organic compounds on aerosol particles and their consecutive growth are not considered in PALM.

Appendix A: Performance measures
Performance measures calculated using the modelled M i and observed O i values. N = number of samples.

FB = 1
FAC2 = fraction of data that satisfy