According to IASI satellite observations of the altitude of Bárðarbunga SO2 near Iceland (Fig. ), the injection height is lower than 4–5 km. Consequently, the Level-2 product of the ultraviolet-visible OMI/Aura satellite sensor for the SO2 total column is mostly preferred to hyperspectral infrared IASI/Metop observations whose sensitivity decreases below 5 km. In addition, the centre of mass of the SO2 cloud is assumed to be within the PBL. North-south gaps in snapshots of the SO2 cloud result from the so-called “row anomaly” of OMI detector (http://www.knmi.nl/omi/research/product/rowanomaly-background) which alters radiance data at all wavelengths for particular viewing directions. In a complementary manner, IASI captures on 21 September the front of the SO2 cloud which is largely missed by OMI due to the “row anomaly”. The major advantage of IASI is that it can track the altitude of SO2 , even from moderate eruptions .

A continuously operating ground-based platform, with various remote-sensing instruments, is installed on the roof of the Laboratoire d'Optique Atmosphérique in Lille-Villeneuve d'Ascq (northern France) and allows for tracking aerosols. It includes a micro-pulse CIMEL lidar measuring the radiation elastically backscattered by atmospheric particles and molecules at 532 nm. The BASIC algorithm allows for determining the vertical distribution of atmospheric particles over Lille as a function of time and distinguishing meteorological clouds from aerosols which are the focus of our study. A high load of low-tropospheric aerosols, lying at an altitude below 1.2 km, is highlighted and suspected to be partly or mostly of volcanic origin (Fig. ).

Lidar observations are also used here to follow the PBL dynamics. The PBL is detected by applying a wavelet covariance transform to lidar backscatter profiles averaged over 20 min . The PBL top is defined as the location of the maximum in the covariance profiles. As low-level meteorological clouds may disturb PBL height retrieval, a filter is applied so as to provide only cloud-free heights.

Complementary to lidar observations, the retrieval of ground-based sunphotometric observations, which are performed at two sites 80 km apart (Lille and Dunkerque/Dunkirk), allows for identifying and isolating the signature of Bárðarbunga aerosols from other atmospheric particles transported over the north of France, such as cirrus particles here. Due to frequent cloudy conditions, time variations of vertically integrated aerosol properties derived from level 1.0 (not cloud-screened) and 2.0 (cloud-screened and quality assured) sunphotometric data from the AERONET network are exploited using different inversion algorithms and a two-site approach (Lille and Dunkerque) (Fig. ). Fine (sub-micron) and coarse (super-micron) aerosol optical depths (AOD) at 500 nm are retrieved using spectral deconvolution algorithm (SDA) applied on AOD within the range 340 to 1640 nm (Fig. ). A sharp and significant increase in the fine mode AOD is highlighted in the early afternoon of 21 September (Fig. ), which will be shown later (Sect. ) to correspond to the arrival of the Bárðarbunga cloud over France. A persistent fine mode is then observed in the following days. Volume size distribution (VSD) of volcanic aerosols are determined using two different inversions: AERONET (version 2) standard algorithm which requires cloud-free almucantar observations and recently developed GRASP (Generalized Retrieval of Aerosol and Surface Properties) code . Over the period of study, there is only one almucantar in Lille fulfilling AERONET level-2.0 requirements (on 23 September). Therefore, for the other 2 days, VSD is retrieved using GRASP: on 21 September, GRASP inverts a (manually inspected) cloud-free principal plane as in AERONET almucantar standard inversion ; on 22 September, direct sun (DS) measurements (available without information of sky radiances) are inverted. For this latter inversion of DS observations, we assume VSD to be a bi-modal lognormal function and optical properties (i.e. refractive index and sphericity parameter) identical to those retrieved from the almucantar on 23 September. The consistency of these algorithms and strategies is shown in Fig. . Using a multi-site approach (ie., including AERONET VSD determined in neighbouring site of Dunkerque), the influence on the fine mode of cirrus co-existing with Bárðarbunga aerosols on 22 September in Lille is evidenced (bottom of Fig. ). SO2 modelling in Fig.  shows that the volcanic cloud passes over Dunkerque (close to Calais) a few hours before Lille. Therefore, the similarity of fine-mode components retrieved at Lille and Dunkerque indicates that cirrus in Lille weakly influence the fine-mode, which is in turn mainly associated with volcanic sulfate aerosols in this specific case.

Daily ground-based MAX(Multi-AXis)-DOAS observations in the ultraviolet are performed by BIRA-IASB in Brussels-Uccle (Belgium) and provide time series of SO2 column amounts during daylight hours (time step of ∼ 12 min). The instrument is described in while retrieval method and settings can be found in .

Ground-level concentrations of SO2 and particles are routinely measured in France by a network of ground stations managed by accredited associations responsible for air quality monitoring. For this study, AIRPARIF provided observations at Neuilly-sur-Seine (near Paris) and Atmo Nord-Pas-de-Calais at Calais and Lille-Fives (northern France). Ground-level SO2 concentrations are monitored by ultraviolet fluorescence with a time step of 15 min. Mass concentration of particulate matter, with diameters less than 2.5 µm (PM2.5) and 10 µm (PM10), are measured by TEOM-FDMS (Tapered Element Oscillating Microbalances with Filter Dynamics Measurement Systems) (time step of 15 min) or by RST (regulated sampling tube) beta gauge automated air monitors (time step of 2 h) which account for both volatile and non-volatile PM fractions.

The atmospheric dispersion of volcanic SO2 is described using the CHIMERE Eulerian regional chemistry-transport model (CTM) . The model accounts for various physico-chemical processes affecting the SO2 released in the atmosphere, including transport, turbulent mixing, diffusion, dry deposition, wet scavenging and gas-/aqueous-phase chemistry. However, the conversion of SO2 to sulfate aerosols is not implemented in this study to avoid uncontrolled influence of uncertainties on the numerous factors governing this process in a volcanic cloud as they have not been specifically validated for application to a volcanic plume. Not accounting for this process leads to underestimating the actual volcanic emissions. CHIMERE CTM is driven by meteorological fields from the Weather Research and Forecasting (WRF) model , which is forced by NCEP (National Centers for Environmental Prediction) reanalysis data on a 6 h basis . The spin time up of WRF simulations is of 5 days. WRF meteorological fields have a 25 km × 25 km horizontal grid and 30 hybrid sigma-pressure vertical layers extending up to ∼ 19 km above sea level (a.s.l.). The dynamics of the PBL is described by the Yonsei University (YSU) parameterization scheme, which is the most widely used scheme implemented in WRF . It consists of a first-order, non-local scheme with an explicit entrainment layer and a parabolic K-profile in an unstable mixed layer. The calculated PBL height is then used as an input to CHIMERE. CHIMERE simulations are performed over the period 19–24 September 2014 on a large area extending from North of Greenland down to Spain. CHIMERE CTM has the same horizontal resolution as WRF but a finer vertical resolution with 29 hybrid sigma-pressure vertical layers extending up to 150 hPa (∼ 13 km a.s.l.).

SO2 emissions are poorly known. For simplicity, we model the source term as a step-function in time with an amplitude of 4700 t h-1, which roughly corresponds to peak values of the SO2 flux retrieved from ground-based UV-DOAS spectroscopy . SO2 is released along a Gaussian profile with a full width at half maximum of 100 m. Time of release and altitude of emissions are found by trial and error so as to reproduce by visual inspection first-order features of satellite and ground-level SO2 observations. As represented in Fig. , we find that two step-functions at (1) 1 km a.s.l. from 19 September 2014 12:00 UT until 24 September 2014 00:00 UT and (2) at 4 km a.s.l. from 20 September 2014 12:00 UT until 24 September 2014 00:00 UT, are sufficient to fit the two-wave behaviour of the Bárðarbunga cloud (Fig. ). This upper injection height is consistent with IASI level 2 products of SO2 altitude, which captured the Bárðarbunga SO2 cloud in the vicinity of the source on 19, 20, 22 and 23 September 2014 (Fig. ). Accordingly, this source term is not intended to reflect the full complexity of the actual emissions of Bárðarbunga but rather captures only the SO2 parcels travelling toward Western Europe.

According to OMI satellite observations, the model reproduces the correct timing of volcanic SO2 arrival in northern Scotland on 20 September 2014 descending down to the south-western coast of England on 21 September (Fig. ). SO2 observed to the north of 60∘ N and to the east of 5∘ W on 21 September is absent from simulations because these emissions, which were released at a high altitude (above 8 km a.s.l according to IASI images of the SO2 altitude in Fig. ), are not accounted for in the model source term as they are first transported toward Arctic then dispersed toward Fenno-Scandinavia, i.e. out of our domain of interest (i.e Western Europe). The model indicates a first SO2 wave (wave 1 in Fig.  and movie of the modelled dispersal of the Bárðarbunga SO2 cloud in the Supplement) reaching Belgium and northern France on 21 September, which cannot be confirmed by OMI observations, hampered above France due to north-south gaps resulting from detector “row-anomaly”. It is however captured by IASI (inset in top of Fig. ), which also indicates that this wave travels at a low altitude below 5 km a.s.l. (Fig. ). This first wave is then pushed and dispersed toward the Atlantic Ocean on 22 September. Note that some traces of SO2 detected by OMI over mainland Europe on 22 September are not associated with wave 1 but likely result from emissions released and then transported at high altitudes (above 8 km a.s.l.) toward Fenno-Scandinavia according to IASI images of the volcanic SO2 altitude (Fig. ). However, these emissions, as stated earlier, are not accounted for in the model. As a result, these traces of SO2 over mainland Europe cannot be reproduced by simulations. On 22 September, both model and observations depict a north-south elongated SO2 cloud reaching France for the second time (wave 2 in Fig.  and movie in the Supplement). Modelled SO2 column amounts are in agreement with OMI SO2 loading. However, the absence of SO2 above Fenno-Scandinavia in the OMI image contradicts the model. This inconsistency may result from inaccuracies of prescribed altitudes of SO2 injection or of meteorological forcing of the model. On 23 September, the model shows the arrival of SO2 above Norway/Sweden, after a long transport from Iceland up to northern Greenland. On the same day, model and observations both indicate some dispersed remnants (although of different intensity) of the second SO2 wave having hit western Europe above western France and southern UK.

The precise timing of arrival of the Bárðarbunga cloud in the French lower troposphere on 21 September 2014 is deduced from the synergetic analysis of volcanic SO2 modelling as well as observations from ground-based lidar and sunphotometers which remotely sense aerosols, on a continuous basis, above Lille-Villeneuve d'Ascq. Sunphotometry indicates the arrival of fine mode aerosols between 12:00 and 15:30 UT on 21 September (Fig. a1), presumably sulfate aerosols formed from volcanic SO2 in the atmosphere, according to their high single scattering albedo (∼ 0.98) derived from AERONET inversions indicative of non- or weakly absorbing aerosols. While fine mode AOD values remain below 0.1 at midday in Lille on 21 September, the arrival of volcanic sulfate aerosols marks an increase in AOD with values ranging between 0.3 and 0.45 in the afternoon (Figs.  and a1). Principal plane inversion also provides volume size distribution (Fig. a2), with an effective radius (reff) of these sulfate of 0.21 µm (mean volume radius rv of 0.26 µm). Simultaneously, lidar active observations, which characterize the temporal evolution of aerosol vertical distribution, indicate the presence above Lille of aerosols at 2 km a.s.l., with a decreasing altitude with time (Fig. b). This behaviour of aerosols coincides with the temporal decrease of the modelled altitude of the most concentrated layer of volcanic SO2 accompanying the first SO2 wave described in Sect.  (red line in Fig. b). This common evolution evidences the co-existence of SO2 and sulfate aerosols within the low-altitude volcanic cloud. Soon thereafter, ground-level sampling in Lille records the first significant increase of SO2 concentration up to ∼ 20 µg m-3 (compared to background values usually close to zero at this site except when contaminated by nearby urban heating plant) followed by a first rise in particulate matter abundance up to ∼ 35 µg m-3 (Fig. c). Hence, these four pieces of evidence (SO2 modelling, sunphotometry, lidar and ground-level air sampling) unambiguously confirm the arrival of the Bárðarbunga cloud in the French lower troposphere down to the ground in the early afternoon of 21 September.

After a period of quiescence, a second, more prolonged and intense episode of ground-level air pollution, in both SO2 and particles, is recorded from 22 to 23 September in Lille (Fig. c). During this second episode, the PM concentration exceeds the information and recommendation threshold prescribed by WHO of ∼ 50 µg m-3, defined as the hourly running 24 hour average value. Concomitantly, sunphotometry indicates a persistent fine-mode (Fig. a2) of weakly absorbing aerosols, which produce fine mode AOD values abnormally high for Lille and Dunkerque (up to ∼ 0.8, Fig. ). The size of Bárðarbunga sulfate aerosols (reff within 0.26–0.28 µm, rv within 0.21–0.24 µm) largely exceeds the radius characterizing typical urban aerosols in Lille (reff<0.2 µm; ). This size is also larger than values reported by sparse observations of volcanic tropospheric sulfate radius at distance from the volcanic source (rv within 0.12–0.16 µm in the Eyjafjallajökull cloud; ).

Substantial increases in ground-level SO2 concentration are recorded by air quality monitoring networks not only in the north end of the country but also on a broad regional scale in France. Unseen for more than a decade, this makes this event of SO2 pollution exceptional (Fig. ). Interestingly, this pollution episode strictly follows a similar temporal pattern, except for a time lag, whichever the city of observation. As shown from the combined analysis of space-based SO2 observations and CTM simulations at a large scale (Sect.  and movie of the modelled dispersal of the Bárðarbunga SO2 cloud in the Supplement), this temporal behaviour results from the arrival of two successive waves of SO2 reaching France from 21 to 23 September. At ground-level, air quality measurements track the progressive transport of these two waves from the north to the centre of France (blue lines in Fig. ). SO2 concentrations up to 70 µg m-3 are associated with the second wave, which is recorded firstly in Calais, then successively 3 h later in Lille-Fives and 8 h later much further south near Paris in Neuilly-sur-Seine. While the modelled time series of SO2 column amounts reproduce this two-wave pattern (solid red line in Fig. ), simulations fail in correctly describing ground-level SO2 concentration as the second wave of pollution starting on 22 September is missed (dashed red line in Fig. ).

As illustrated by the broad agreement with OMI satellite data (Fig.  and Sect. ), chemistry-transport simulations, with a standard configuration here, are efficient at reproducing on a continental scale the dispersion of the Bárðarbunga SO2 cloud from Iceland toward France. The temporal dynamics of far-range events of air pollution, characterized for instance in France by the arrival of two distinct SO2 waves travelling from the north toward the capital city in ∼ 8 h, is hence well described (solid red line in Fig. ). A good agreement is also reached between model and observations of SO2 vertical column amounts (CA) by ground-based UV max-DOAS spectroscopy performed at long distance from the eruptive site in Uccle/Belgium, located less than 100 km from Lille/France (Fig. ). Model and observations find SO2 CA of the same magnitude on 19, 20, 22 and 23 September. Nevertheless, the model cannot capture the significant and abrupt SO2 CA increase (up to 14 DU) of very short duration (from 14:18 to 15:07 UT) observed on 21 September by DOAS, likely due to the insufficient (hourly) time resolution of the model.

However, the model has difficulty reproducing the correct intensity of the air pollution episode in remote areas. Similar issues have also arisen with independent modelling simulations using a Lagrangian approach forced with distinct meteorological reanalysis . Our model here completely misses the second wave of SO2 at ground-level in the north of France (dashed red line in Fig. ). This shortcoming results from an incorrect description of the vertical distribution of SO2 at long distance from the eruptive site. According to lidar observations capable to detect sulfate aerosols coexisting with SO2 (Sect. ), the model mimics correctly the drop in altitude above Lille of the first SO2 wave on 21 September (red line in Fig. b). This modelled wave hits the surface at about the same time as the first detection of air pollution at ground-level. But the second modelled wave, despite a similar pattern with a significant decrease in altitude with time from 6 km a.s.l., does not reach the ground and remains at an altitude ≥ 1.8 km above Lille on 22 September (red line in Fig. b).

Issues encountered for adequately modelling far-range air pollution episodes can arise from the difficulty of simulating both the long-range transport and/or dispersal of volcanic compounds and the meteorological dynamics at a local scale, as the latter controls the capture and mixing of the overlying volcanic cloud in the far-range planetary boundary layer (PBL). Simulations at higher spatial resolution of both CHIMERE CTM and WRF models may help to make progress along this path. For these reasons, we explore in the next sections the impact on far-range ground-level SO2 concentrations of both meteorological/chemistry-transport simulations at higher spatial resolution and of various PBL parameterization schemes in the meteorological model.

Meteorological and chemistry-transport simulations at higher spatial resolution require both high computation time and capacity, which challenges our current modelling capacities. We performed here WRF and CHIMERE simulations on two nested horizontal grids (Fig. ). The larger domain extends from north of Greenland down to Spain (as the low-resolution domain of simulations in Sects.  and ) with a horizontal resolution of 22 km × 22 km representing 209 × 229 grid cells. Note that this coarse resolution is nevertheless slightly higher than the low spatial resolution simulation performed on a 25 km × 25 km horizontal grid. The nested domain extends from Norway down to Central France, with a fine 7.3 km × 7.3 km horizontal resolution representing 217 × 232 grid cells in a Lambert projection. Except for one test run configured with 60 hybrid sigma-pressure vertical layers extending up to ∼ 19 km a.s.l. and ∼ 13 km for WRF and CHIMERE models respectively, most simulations are run assuming 30 vertical layers for both models, as in the simulations with a standard configuration in Sect. . The dynamics of the PBL is still described by the Yonsei University (YSU) scheme, as in Sect. .

Simulations at higher horizontal spatial resolution better resolve the long-distance transport of the Bárðarbunga SO2 cloud as well as its descent over France. Especially, an earlier and faster descent of the SO2 cloud over Lille is modelled on 22 September 2014, with the core of the plume reaching a significantly lower altitude than in low-resolution simulations (top panels of Fig. ). Subsequently, a clear improvement of modelled far-range ground-level SO2 concentrations is reached. Indeed, this earlier modelled descent of the SO2 cloud leads to the emergence of a second peak in ground-level concentrations on 22 September (bottom panel of Fig. , right), which was entirely missed by simulations at low spatial resolution as mentioned in Sect.  (bottom panel of Fig. , left).

The emergence of a second peak is also modelled at other air quality monitoring stations, i.e. Calais and Neuilly-Sur-Seine (Fig. ). A better agreement between model and observations is also noticed regarding the timing of the first peak concentration whichever the station (Fig. ). Note that only slight differences in ground-level concentrations were observed with simulations performed with a vertical resolution twice as high (i.e. 60 vertical layers in WRF and CHIMERE models) and are consequently not shown.

Nevertheless, although clear improvements in ground-level SO2 concentrations are achieved with simulations at higher spatial resolution, both timing and intensity of the second peak concentration are not perfectly reproduced by the model (Fig. ). The emergence of the second peak concentration is modelled late compared to measurements at Calais and Lille, and slightly too early at Neuilly-sur-Seine. The modelled intensity of this peak concentration is always under-estimated by a factor of 3 to 10 depending on the monitoring station.

These remaining discrepancies between model and observations may arise from various reasons. We explore in the next sections the impact on far-range ground-level concentrations of (i) a poor knowledge of the source emissions (assumed here to follow a simple pattern as described in Sect. ), and (ii) an incorrect modelling of the planetary boundary layer dynamics which may prevent one from correctly capturing the descending volcanic cloud down to the ground.

We do not aim to provide a detailed estimate of the source term but to show that a simple source term allows for representing the main features of this event of far-range air pollution triggered by the Bárðarbunga eruption (Sect.  and ). Instead, we investigate here whether variations in this simple source term could explain the current discrepancies between observed and modelled far-range SO2 concentrations at ground-level.

Whichever the monitoring station, we have deduced from SO2 modelling that the second peak in ground-level SO2 concentration results from the arrival in France of the second pulse of emissions, which are injected at the source at 4 km a.s.l. from 20 September 2014 12:00 UT until 24 September 2014 00:00 UT in low spatial resolution simulations (not shown). We vary the altitude of injection (between 3 and 7 km a.s.l.) as well as the start time of this second pulse of emissions (with a 5-hour earlier release), as both may impact the timing of the second peak concentration which is not correctly reproduced by the model (Sect.  and Fig. ). However, only slight modifications on far-range ground-level concentrations are obtained, unable to explain the discrepancy observed between model and observations.

Regarding the under-estimation by the model of ground-level SO2 concentrations by a factor of 3 to 10 (Sect. ), one could argue that it results from an under-estimation of the assumed SO2 emission flux by a similar factor given linear processes of large-scale transport/dispersion. Volcanic emissions may present rapid temporal fluctuations of large amplitude (e.g ). However, despite a poor time-resolved knowledge of the Bárðarbunga SO2 source relying on sparse ground-based measurements, assuming a SO2 flux five time stronger would better fit far-range ground-level concentrations but would also lead to far-range SO2 column amounts increased by the same amount (not shown). The latter would be in complete disagreement with SO2 column amounts retrieved from satellite observations (Fig. ) or ground-based MaxDOAS measurements performed in Belgium (Fig. ).

As a consequence, input model parameters characterizing the Bárðarbunga SO2 source (flux and altitude of injection) are shown to play a minor role on far-range ground-level concentrations over our relatively short period of study (19–24 September 2014). They do not allow us to solve the disagreement observed between model and observations. For this reason, we explore in the next section the impact of the PBL dynamics on air quality modelling.

In addition to the reference Yonsei University (YSU) scheme used in the low spatial resolution simulation, the impact on far-range ground-level concentrations of two additional PBL parameterization schemes, recently added to the WRF model, are tested: the Asymmetric Convective Model (ACM2) scheme as well as the improved Mellor-Yamada-Nakanishi-Niino level 3 model (MYNN 3) scheme . The ACM2 scheme is a first-order, non-local closure scheme which features non-local upward mixing and local downward mixing. The MYNN 3 scheme is a second order, local closure scheme tuned to a database of large-eddy simulations.

At first glance, time series of the PBL height above Lille do not seem to vary widely with the different PBL parameterization schemes (top of Fig. ). We note nevertheless a less marked diurnal cycle with the MYNN3 scheme. However, these slight differences are shown to be sufficient to produce up to a ten-fold variation of the ground-level SO2 concentrations (bottom of Fig. ).

As illustrated by Fig. , both timing and altitude of the encounter of the boundary layer top and the overlying volcanic SO2 cloud, which may vary with the PBL parameterization scheme, strongly impact the subsequent increase in SO2 concentration at ground-level some time later.

For our specific case-study, the top of the PBL encounters the overlying Bárðarbunga SO2 cloud above Lille at approximately the same time on 21 and 22 September, whichever the PBL scheme (Fig. ). After this capture of the volcanic cloud into the PBL, SO2 is mixed and diffused down to the ground triggering a noticeable increase of the ground-level SO2 concentration. The time delay between the capture of the volcanic SO2 at the top of the PBL and its record at the ground-level is estimated to be just a few hours (Fig. ).

However, only the ACM2 scheme allows the top of the PBL to encounter the core (i.e. the most concentrated part) of the Bárðarbunga SO2 cloud on 22 September 2014, with a PBL height at the time of encounter higher by just a few hundred metres compared to other schemes (top right of Fig. ). This substantial capture of volcanic SO2 by the boundary layer explains why the intensity of the second SO2 peak concentration is the highest with this scheme and the closest to observations (bottom right of Fig. ). Note that the best agreement between observations and model for the first SO2 peak concentration is also reached with the ACM2 scheme. These results demonstrate the crucial importance to correctly model both the PBL height and the vertical distribution of the overlying volcanic SO2 cloud with time. Note that this latter depends on a rigorous modelling of both long-distance transport and dispersion processes and of local PBL dynamics. Indeed, the PBL scheme influences the concentration of the overlying volcanic SO2 (top of Fig. ). We may even suspect a kind of “sucking” of the core of the volcanic cloud which seems to follow the PBL top, especially remarkable on 22 September.

Nevertheless, even if the ACM2 scheme provides the best fit to observations, none of the PBL schemes allows for precisely modelling the second SO2 peak concentration with correct timing and intensity (Fig. ). This difficulty likely results from the inaccuracy of the modelled PBL height which presents marked differences with observations, whichever the PBL scheme. The altitude of the PBL above Lille is retrieved from lidar observations and compared with model output in the bottom of Fig. . Simulations generally underestimate the PBL height over Lille (up to 1.5 km), especially in the mornings and evenings. Such underestimation is a relatively common feature of WRF PBL schemes , used here to force the CHIMERE chemistry-transport model.

In a context of urban air pollution where pollutants are injected at ground-level into the atmosphere, an underestimation of the PBL height favours an over-evaluation of the intensity of ground-level pollution. In our volcanic case-study, this may explain the overestimation by certain schemes of the SO2 ground concentration increase resulting from the first SO2 wave reaching the ground of Lille on 21 September in the evening (Fig. ). However, the PBL height underestimation by the model can also prevent from correctly capturing in the PBL the second SO2 wave which travels at a higher altitude than the first wave (Fig. ). In this context, the intensity of air pollution at ground-level is under-evaluated (bottom of Fig. ).

In our specific case study, our concern is that, whichever the scheme, the modelled PBL height increases too late and too weakly compared to lidar observations in Lille, which is especially problematic in the morning of 22 September (bottom of Fig. ). Indeed, this discrepancy explains both the delayed modelled timing of the second peak concentration and a substantial under-estimation of its intensity (by a factor of 2–3), as the modelled boundary layer captures too late a smaller fraction of the overlying volcanic SO2 than it should in reality. In other words, an earlier and higher modelled PBL height in the morning of 22 September, as expected according to lidar observations, would lead to an earlier and stronger capture of the overlying Bárðarbunga SO2 cloud at the top of the boundary layer. This would produce an earlier and stronger peak concentration at ground-level in better agreement with air quality monitoring observations.

Therefore, this case-study demonstrates the key role played by the PBL dynamics to rigorously estimate the magnitude of far-range volcanogenic air pollution.