The measurements used in this study constitute a quasi-continuous time series of COS atmospheric mixing ratios obtained at the monitoring site of Gif-sur-Yvette (GIF) located in the Paris region at 48.7109° N, 2.1476° E at 163 ma.s.l. with an inlet height of 7 ma.g.l. A commercial gas chromatograph (Varian 3800) coupled with a cryogenic preconcentrator (ENTECH P7100) for sample preparation and a mass spectrometer detector (Varian Saturn 2200) for COS detection was used to analyze this gas, as described by . Calibration and drift correction is done every 3 weeks using a calibration gas containing 1.013 ppm of COS in helium, with the occasional calibration using a standard of compressed air with 573 ppt of COS and another standard traceable to the NOAA ESRL standard of 448.6 ppt. Calibrations led to a repeatability of 1 % and a precision of 0.2 %.

The GIF time series of hourly data spanning August 2014 to December 2021 is available in . In this study, we only make use of the time series from August 2014 to December 2019 (Figs.  and  to ), which covers the period of availability of inputs required for simulations, in particular global concentration fields used to compute the background signal (see Sect.  and ).

The time series of flask measurements sampled by the National Oceanic and Atmospheric Administration (NOAA) network at Mace Head (MHD) in Ireland at 53.3° N, 9.9° W at 42 ma.s.l. described inwith one to five pairs of flasks per month, collected mostly between 08:00 and 17:00 UTC is also shown in Fig.  to illustrate the so-called “background”, i.e., the overall contribution of all the sources and sinks which are not in our area of interest. In our study, which focuses on western Europe and more particularly on the footprint of GIF, the mixing ratios at MHD are representative of the background when the air masses are advected from the west, which is a frequent meteorological configuration .

The contribution of the background to the simulated mixing ratios is computed using 3D fields at the global scale (see details in Sect. ). The COS mixing ratio fields used here are obtained from , at a horizontal resolution of 3.75° in longitude and 1.875° in latitude for 39 pressure levels at a 3-hourly time resolution (illustration in Fig. ). These global atmospheric inversions were designed to assimilate data from the background NOAA observation sites, such as MHD.

In the following, “emission” denotes fluxes which are sources as seen from the atmosphere and “uptake” denotes fluxes which are sinks as seen from the atmosphere; “flux” is used for (ensembles of) processes which can be either sources or sinks.

Our simulations, detailed in Sect. , take into account COS fluxes in the area of interest as provided by the data sets described here and listed in Table . The contributions from biomass burning emissions and the atmospheric sink are neglected in this study, as well as the emissions from anoxic soils. They play a significant role at the global scale for long-term studies e.g., but are negligible compared to ocean emissions (see Sect. ), biogenic fluxes (see Sect. ) and anthropogenic emissions (see Sect. ).

The modeled COS mixing ratios were obtained using the Lagrangian atmospheric transport model FLEXPART (FLEXible PARTicle) version 10.3 . The FLEXPART model is driven by meteorological data from the European Centre for Medium-Range Weather Forecast (ECMWF) ERA5 reanalysis with 3-hourly intervals and 60 vertical layers, retrieved using the FLEX-extract toolbox . The performance of ERA5-provided boundary layer height and wind direction is very good e.g.,. In our case, meteorological data are provided to FLEXPART at a 1° horizontal resolution. For the whole duration of the observation period, 2000 virtual particles are released every 6 h and followed backward in time for 10 d. The multiple FLEXPART simulations are driven by the GUI (graphical user interface) toolbox designed by . FLEXPART and ERA5 meteorological data are well-recognized tools in the atmospheric community, with the combination of the two being widely used in the atmospheric inversion community . The choices made for FLEXPART's configuration (number of particles released, frequency of releases, resolution) are consistent with the setups used for the atmospheric inversion fluxes e.g.,.

We assume a conversion of CS2 to COS with a maximum of 87 % molar conversion rate, as used by . The kinetics of the conversion is approximated by prescribing a half-life of 3 d to CS2 in the atmosphere. The amount of COS created from a CS2 source along the transport path of the air mass is then given as 0.87 × (1-e-tτln2) with τ = 3 d. Note that a lifetime of 1.5 d has also been tested (not shown), with almost no changes in the conclusions of this study.

So-called source-sensitivity (or footprint) maps are computed for every release date by counting the number of particles transported above a certain point below a threshold of 500 ma.g.l. The source contribution by process type is inferred by convolving source-sensitivity maps with flux maps from the ocean emissions, the biogenic fluxes and one of the two anthropogenic emission inventories (see Sect. –). One FLEXPART simulation is run for each process or sector and region of interest (see Fig. ) so that the simulations (Table ) can be combined to include or exclude particular sectors and/or particular regions when building the footprint maps.

The background mixing ratios (Fig. f) are calculated by combining the 3D source-sensitivity fields e.g., at the end of the backward trajectories with the available COS concentration field (see Sect. ). The background thus obtained represents the average of the mixing ratios in the grid cells where each particle trajectory terminated 10 d before the observation. For instance, as illustrated in Fig. f, for the given date, GIF observations are sensitive to background mostly in the northwestern Atlantic (higher sensitivity for dark-red contours). The computation of COS contributions from surface fluxes and the background was carried out using the Community Inversion Framework CIF;.

At GIF, the seasonal variations in COS mixing ratios are dominated by the contributions of the background and ocean, i.e., by large-scale fluxes; the variations at shorter timescales (weeks or days) are driven by the biogenic land contribution . Finally, depending on the wind speed and direction, anthropogenic emissions may dominate for short episodes of high concentrations see Sect. 3.3.2, “Selected winter episodes”, in. The contributions to COS mixing ratios at GIF due to the ocean, the biogenic land fluxes and the anthropogenic emissions are shown in Fig. .

In the following, we use the afternoon (12:00–18:00 UTC) daily averages of measured and simulated mixing ratios because the model is assumed to better represent the vertical mixing in the afternoon so that the comparison to measurements highlights the discrepancies in fluxes compared to the model's errors. More detailed results for the whole day or nighttime and per season are shown in Table .

By design, the background contribution (Sect. ) fits the main monthly variations as measured at a background station such as MHD (Fig. a and b). Its contribution at GIF ensures a good simulation of the variability (Table , Pearson's correlation of ≥ 0.75) and a baseline mean error of ≤ 35 ppt over the whole period of interest. As expected, adding the natural emissions from the ocean (Sect. ) and the biogenic land fluxes (Sect. ) reduces the bias (by almost 15 ppt as an absolute value) and the mean error is decreased by more than 0 %.

As mentioned in Sect. , we assess the impact of using temporally resolved biogenic fluxes on our simulations. Making use of vegetation uptake with a 3-hourly time resolution (only during the year 2016; see Sect. ) does not improve the statistical fit of the model to the observations: the three statistical indicators are almost unchanged (Table , period of 2016). The variability is not better reproduced when adding the natural emissions from the ocean and the biogenic land emissions from the soils and the vegetation to the background (correlations of ≤ 0.75 in all cases). This may be due to the monthly resolution of the ocean emissions (Sect. ) being too coarse compared to the variability in the transport and to an issue with the variability in the vegetation uptake or the soil exchanges, probably at the seasonal scale, which is assessed in Sect. .

The limited impact and even degradation of performance when using the diurnal cycle of biogenic flux suggest an issue in their representation in ORCHIDEE, potentially in the residual ecosystem (vegetation and soil) COS flux simulated at night.

The contributions of the anthropogenic emissions by Zumkehr degrade all three indicators of the fit to the measured concentrations at GIF: the bias and mean error are very high (both ≥ 100 ppt), and the variability is not reproduced anymore. The regional anthropogenic emissions located in the Paris and Rouen area explain the major part of the discrepancies between the simulation and the measurements (Table ). According to Zumkehr's inventory, COS emissions in Île-de-France, i.e., the Paris region itself, are mostly (≈ 62 %) due to the viscose industry, which is not consistent with the lack of any such factory declaring emissions in this region. The small (< 1 %) contribution by the coal-using industry is not consistent with the last coal power plant in the area being closed in 2012. Due to the same type of incomplete information or lack of a relevant proxy, the overall European COS emissions in Zumkehr's inventory may be overestimated, as suggested by the very high contributions (more than 100 ppt) due to anthropogenic emissions alone simulated at GIF (Fig. c).

The contribution of the anthropogenic emissions from our homemade inventory increases the bias and mean error compared to the contributions of the background and natural fluxes only (Table ). The maximum simulated contributions are between 40 and 50 ppt for some months (Fig. c and d), i.e., 10 times smaller than with Zumkehr's inventory. Therefore, the bias and mean error computed over the whole period using our inventory remain more than 2 times smaller than with Zumkehr's inventory without the emissions of the Paris and Rouen area. This order of magnitude is more consistent with the expected anthropogenic contribution required to match the measured concentrations above the background at GIF. Overall statistics are very similar when using only viscose or coal or both. Further investigation is needed to narrow the range of anthropogenic fluxes in Europe. However, as anthropogenic contributions are observed as peaks, the magnitude of which is difficult to reproduce in transport models, a combination of several observation sites would be needed.

The information which can be retrieved from the time series of concentrations in one measurement site on the relevancy of our inventory compared to Zumkehr's is discussed in the following in terms of activity sectors and source regions.

The general performance of the model at GIF (Sect. ) shows that the total emissions of Zumkehr's inventory lead to a large overestimation of COS concentrations at GIF (Table ), almost half of which is due to the emissions located in the Paris and Rouen area. We show in Fig.  the distributions of observations and simulations from different sectors as violin plots. Above the overestimation due to the “natural” contributions, Zumkehr's inventory leads to a median overestimation of ≈ 52 ppt (Fig. , Z total vs. natural). The non-viscose-related emissions in the Paris and Rouen area explain almost half of this discrepancy (24 ppt, Z total vs. ZviW + ZnviW + ZviB + ZnviB + ZviU). The non-viscose-related emissions in “the rest of the world”, i.e., the whole world excluding the Paris and Rouen area and Benelux, explain a bit less than a quarter of the overestimation (12 ppt, ZviW + ZnviW vs. ZviW vs. natural). Of the remaining 12 ppt, 8 ppt is due to viscose industry emissions in the rest of the world and 1.5 ppt is due to viscose industry emissions in Benelux alone. The unrealistic COS emissions in the Paris and Rouen area lead not only to too high a median contribution but also to a large number of high concentration peaks, as shown by the upper elongation of the violin Z total (maximum actually > 1995 ppt). As expected, our homemade inventory is more compatible with the observations at GIF (Fig. ). In this case, the background and the oceanic and biogenic land fluxes contribute ≈ 40 % to the median overestimation (≈ 4 ppt, Fig. , GIF obs vs. natural) of the observations, compared to 60 % contributed by the homemade anthropogenic emissions (≈ 7 ppt, Fig. , HM total vs. natural). The distribution of extreme events, i.e., with very high or very low COS concentrations obtained with our homemade inventory, is closer to the observed one (Fig. , upper and lower elongations of GIF obs vs. Z total vs. HM total).

Additional continuous observation sites would be needed in different parts of Europe to clarify the relevancy of our inventory beyond the Paris region and vicinity. In addition, our inventory is based on coarse emission factors, both for the viscose industry and for coal power plants. The facility-level campaign as carried out around a viscose factory near Rouen in would significantly improve our understanding of COS anthropogenic fluxes and thus our inventory.

Disentangling the main contribution of mismatches between observations and simulations is very challenging with only one site. Differences at the synoptic scale, as represented in Fig. , can originate from an erroneous background, transport discrepancies or incorrect fluxes. Moreover, with only one site, absolute values of the biogenic fluxes cannot be systematically estimated. We therefore focus on the variations through time, particularly between day and night and seasonally. As represented in the time series in Appendix , fine-scale temporal variability, corresponding to the diurnal cycle, is well reproduced during some periods of the year, especially in spring. The diurnal cycle is dominated by local influences, with remote influence of the background and of distant fluxes transported to GIF filtered out. We therefore focus here on 12 h day or 12 h night differences in measured or simulated COS concentrations, defined as either the “daytime difference”, hereafter ΔdayCOS, between day D at 18:00 and day D at 06:00 or the “nighttime difference”, hereafter ΔnightCOS, between day D+1 at 06:00 and day D at 18:00. Thus, these differences are mostly influenced by a combination of small-scale meteorological conditions (mostly the diurnal cycle of the planetary boundary layer) and of regional fluxes (within the transport footprint around the station). In the present analysis, we assume that all discrepancies are attributable to fluxes, even though the diurnal cycle of transport in FLEXPART is imperfect. Discrepancies in the diurnal cycle of transport patterns would only impact the magnitude of the day–night differences but not the overall patterns discussed below and thus would not impact our qualitative conclusions.

The variations in ΔdayCOS and ΔnightCOS over the available 5.5 years of observation are represented in Fig. , alongside COS fluxes from regional vegetation and soils, using both the ORCHIDEE and SiB4 models. The observed seasonal cycle of ΔnightCOS and ΔdayCOS is similar for all years. The observed ΔnightCOS is slightly negative in winter (JFM; -7 ppt) and then strongly negative in spring (AMJ; -10 ppt), with yearly peaks of ≈-20 ppt; then, summer positive peaks of 10–15 ppt in July are followed by another negative low in fall (-8 ppt). The positiveΔnightCOS in summer is most probably due to emissions from crops, such as rapeseed . Observed ΔdayCOS is opposite with slightly different magnitudes.

In winter (JFM), the vegetation is mostly inactive and soils are at their lowest level of sink activity with the net soil uptake becoming larger than the vegetation uptake. Thus the contribution of the uptake by the vegetation and soils in our simulations in both ORCHIDEE and SiB4 for both time resolutions (monthly or 3-hourly) is very small. Both the simulated ΔnightCOS and ΔdayCOS are therefore close to zero (on average, -2 ppt for nighttime and +1 ppt for daytime), contrary to observed ΔnightCOS and ΔdayCOS (-7 ppt for nighttime and +7 ppt for daytime). This suggests underestimated COS uptakes at night in winter in ORCHIDEE and SiB4 vegetation and/or soil data sets.

During spring (AMJ), vegetation COS uptake is activated in ORCHIDEE and SiB4. When using constant monthly fluxes with no diurnal cycle, their contribution to the simulated nighttime uptake leads to lower concentrations at nighttime compared to daytime, consistent with observations. In contrast, daytime concentrations are the result of compensating vegetation uptake and mixing in the boundary layer. However, the use of 3-hourly vegetation uptake by ORCHIDEE daily cycle results in simulated ΔnightCOS and ΔdayCOS close to zero, not consistent with the observations, most likely related to absent uptake at night in ORCHIDEE. On the contrary, in SiB4, persistent uptake occurs at night, leading to more realistic simulation than with ORCHIDEE. This discrepancy suggests a persistent nighttime uptake not adequately represented by ORCHIDEE's diurnal cycle. The ORCHIDEE minimal stomatal conductance at night (see Sect. ) has to be revised. Refining the ORCHIDEE model to improve the representation of nighttime processes may enhance its ability to reproduce observed dynamics, compared to SiB4. Summer (JAS) shows opposite behavior compared to spring, with positive nighttime enhancements. These enhancements are not properly reproduced by both ORCHIDEE and SiB4 when using constant monthly fluxes. With 3-hourly fluxes, both models perform well in summer. Using a realistic diurnal cycle is key to reproducing the positive ΔnightCOS and negative ΔdayCOS. The magnitude of fluxes in ORCHIDEE seems to be in better agreement with observations than with SiB4.

During fall (OND), the performance of our simulations exhibits variability across different years, regardless of the inclusion of a vegetation diurnal cycle. Notably, the simulations for the years 2016 and 2017 show significant discrepancies, whereas the results for the remaining 4 years align well with observational data. This disparity underscores the challenges associated with accurately reproducing the timing of vegetation senescence, particularly given the influence of intricate synoptic meteorological conditions on both vegetation dynamics and the performance of the transport model. The complexity of these interactions poses a significant hurdle in achieving consistent simulation outcomes across all years.