Response of atmospheric composition to COVID-19 lockdown in Paris (France)

The worldwide COVID-19 pandemic has led to lockdowns at national scales in the Spring of 2020. Large cuts in emissions occurred, but the quantitative assessment of their role from observations is hindered by weather variability. In order to circumvent this diculty, we developed here an innovative analog methodology and applied it to a comprehensive in-situ dataset of primary and secondary pollutants obtained at the SIRTA observatory, a suburban background site of the Paris megacity (France). We nd that concentrations of primary trac dropped by 69-76% during the lockdown period. Further, the decrease of NOx triggered a decrease of particulate nitrate (-41%), one of the main springtime aerosol components in North-Western Europe. We reveal a threshold effect highlighting the need of substantial NOx decrease to affect particulate nitrate. At the same time, the expected ozone increase (+21%) underlines the negative feedback of NO titration. Finally, an increase of residential wood burning sporadically compensated primary trac, and inuenced the oxidation state of secondary organic aerosols. Our results provide a quasi-comprehensive observation-based insight on mitigation policies regarding air quality in future low-carbon urban areas.

from source apportionment method) at the SIRTA-ACTRIS observatory 4 , located 20 km Southwest from Paris.

Compare apples with apples
The assessment of lockdown impact on air quality leads to completely different results depending on the chosen reference period (Fig. S1). On one hand, considering the weeks preceding lockdown (PreLP) as reference leads to the determination of signi cant concentration increases for various regulated pollutants during the lockdown period (e.g., + 83% in NO x , +439% in PM 1 ). This is mainly due to a drastic change of Western Europe meteorological conditions (from low-pressure to high-pressure system) concomitantly with the application of lockdown policy measures in France. On the other hand, considering the same weeks of previous years (e.g. 2017-2019, LP2017-2019) as reference period blindly encompasses interannual meteorological variability. As a matter of fact, April 2020 in France was the 3 rd warmest April months since 1900, as well as exceptionally dry (-50% of precipitation in the Paris region) and sunny (> +50% compared to normal). In the absence of climatological reference values, the representativeness of alternative references is a critical issue, and the major danger would be to apply methodologies unquestioningly, without verifying the strong inherent hypothesis that data are comparable. This may lead to erroneous, contradictory and counterintuitive results and can indeed be critical for semi-volatile and hygroscopic material, such as ammonium nitrate, which accounts for a signi cant fraction of PM during pollution episodes [5][6][7] .
That is why we developed a novel methodology, adapting for the rst time the "analog weather" approach 8 to in-situ measurements. Each day of lockdown is thus compared to a group of days with similar meteorology. The analogy is based on synoptic (atmospheric circulation), regional (backtrajectories), and local (ambient temperature and relative humidity) similarities (see Methods). As our methodology is based on a day-to-day comparison, it allows a valuable evaluation of its comparability along the lockdown period: for instance, Mean Bias of -0.17 m/s, -1.52°C, -1.3 hPa and 8.7% were respectively obtained for wind speed, ambient temperature, pressure and relative humidity, indicating a satisfactory analogy. Moreover, when comparing the air mass origins occurring during lockdown, our approach shows highest overall correlation (R=0.82) compared to PreLP (R=0.29) and LP2017-2019 (R=0.69). Therefore, our analog methodology enables an e cient estimation of business as usual conditions, and provides representative and robust insights of how the lockdown induced changes of the atmospheric composition and chemistry.
We then assess the lockdown impact on a variety of atmospheric variables. Our dataset (timeseries presented in Fig. S2) is detailed in Table 1. Changes of their concentrations due to lockdown relative to analog days are presented in Fig. 1 and discussed in the following sections. . This expected decrease is slightly higher than what has been found in an europeanscale study 2 , but in the same order of magnitude than urban areas in Spain 9 . Our results are consistent with tra c counting data (https://dataviz.cerema.fr/tra c-routier/), which reveals a drop of about 75% in France and the Paris region. Moreover, a clear week-end effect can be observed from these data with an additional decrease from -70% to -90%, since outdoor leisure activities were forbidden during lockdown. Interestingly, when applying a week day constraint on our analog list, we nd a similar behavior for NO x (average of -57% during the week, -81% during week end), BC ff (week:-58%, week-end: -78%) and HOA concentrations (week:-68%, week-end: -80%), reinforcing the consistency of our methodology.
Nitrogen oxides play a central role within atmospheric chemistry, enabling the formation of secondary pollutants, such as tropospheric ozone and secondary organic and inorganic aerosols (SOA and SIA). We nd ozone to increase by 22% (Fig. 1). This counterintuitive chemistry, due to the titration effect of NO, is well characterized 10 . Ozone has been also found elsewhere to rise to a similar order of magnitude 2,11,12 .
The variability of absolute changes is quite important (InterQuartile Range of 22 µg/m 3 ), but still inversely follows the change of NO x concentrations (Fig. 2). In addition to forecasted more frequent and intense heatwaves 13,14 , this lockdown is an observation-based reminder of this potential negative feedback 15  were neither stopped nor restrained. We can therefore reasonably assume business-as-usual ammonia concentrations, and since the formation regime of nitrate in Paris has previously been found to be NO xlimited 17 , a change of regime to NH 3 -limited is highly unlikely. On median, we nd that nitrate decreases by 41%, which is, as expected, linked to the decrease of NO x (Fig. 2). The behavior of nitrate contrasts with sulphate, showing little change (-8%), which indicates a similar in uence of long-range pollution advection 18,19 . Interestingly, we also nd a concomitant increase of the nitrogen-oxidation ratio (NO 3 /(NO 2 +NO 3 ), NOR, +53%), indicating a higher e ciency of HNO 3 formation, which may be related to higher ozone concentrations. Nevertheless, this enhanced oxidation capacity did not compensate for the decrease of NO x , suggesting the existence of a threshold effect, below which a decrease of NO x has statistically only little impact on NO 3 . From Fig. 2, we locate this threshold at around -15 µg/m 3 . This result indicates that in order to reduce the intensity of springtime PM pollution episodes, mitigation policies should impose substantial urban background NO x reduction.
Finally, considering secondary organic aerosols, we nd a decrease of 37%, which can be linked to numerous factors. The exceptional amount of sunshine during lockdown did positively in uence the availability of the hydroxyl radical for the initialization of SOA formation. But NO x steps in SOA formation when reacting with peroxy radicals (R-O 2 ), which can form less volatile products 20 . However, we don't nd here any signi cant change of the oxidation state of SOA (OSc SOA , -3.8%). The decrease of SOA seems to be primarily related to the decrease of NO x (Fig. 2), without any threshold effect.
The unexpected role of residential wood-burning Contrarily to primary tra c emissions, one striking feature of our results lies in the unexpected increase of wood burning tracers (BC wb and BBOA, +20% and +58% respectively), despite limited absolute change ( Fig.1). Converting BC wb to PM wb , and BC ff to PM ff (see Methods), we also highlight that during speci c days (12 out of 56), increased PM wb concentrations compensated or even exceeded the decrease of PM ff ( Fig. 3a). At the same time, we notice that the mean weekly variation of wood burning changed during lockdown (Fig. 3b), with increased concentrations during the week, compared to the relatively at variation in business-as-usual conditions (eg +67% on Fridays). Therefore, lockdown changed both intensity and temporality of the wood burning source. This behavior can be primarily related to the stayat-home order, enhancing emissions of residential heating 2 . Such increased in uence also mainly occurred at low wind speeds. This brings to light an unfavourable source-meteorology synergy which should be accounted for e cient mitigation policies, especially since the monotonic decreasing trend that was previously found at SIRTA (2011-2018) for tra c contrasts with the at trend of the wood burning source 21 .
Wood burning is also known to be associated to substantial SOA formation, due to the reactivity of their gaseous precursors 22 . Although no overall change of SOA oxidation properties is reported here, we nd that the temporal variability of the oxidation state of SOA (OSc SOA ) is inversely related to the fraction of BBOA in OA (Fig. 4). In other words, emissions from residential wood burning during lockdown unambiguously changed the chemical composition of SOA, towards less-oxidized material. Yet, SOA is a complex fraction whose understanding still remains limited because of the lack of speci c tracers' measurements. A SOA-tracer based source apportionment study during Spring at SIRTA previously pointed out less oxidized SOA from wood burning 23 . No clear change of behavior is found for the analog dataset, which stays within the observed variability (Fig. 4). This means that lockdown has changed the variability of the wood burning source during that period of the year, which induced a change in the chemical composition of SOA, but formation pathways seem to remain similar.
The lockdown enforced during Spring 2020 in Paris corresponds to a real-life emission scenario, representing the extreme case of a quasi-total interruption of the vehicular tra c source. Up to now, no mitigation policy could have gone that far. We nd from our observations that NO x reduction can be an e cient mitigation policy regarding nitrate and SOA, which account for more than half of Reducing concentrations of secondary compounds is di cult, because mitigation policies can inherently only focus on the reduction of primary pollutants. Our understanding of the complex chemistry of the urban troposphere relies rstly on a better characterization of primary precursors. To that end, additional long-term, highly time-resolved measurements of Volatile Organic Compounds would be of high interest. Yet, the e ciency of mitigation policies is largely affected by on meteorology, which can limit the impact of reduction in source emissions. Regarding these issues, our analog methodology can be an e cient tool to monitor and quantify more precisely the impact of mitigation policies.

Methods
In The black carbon dataset consists of the harmonized gathering of 5-min AE31 (2011-2013) and 1-min AE33 (2014-2020) measurements at SIRTA. AE31 data were compensated using Weingartner algorithm 30,31 . AE33 has an online compensation approach, using the dual-spot technology 32 . In order to have comparable BC concentrations throughout both instruments, and since no C 0 value for AE33 has been recommended yet, compensated AE31 data were uncorrected from the C 0 of 2.14. Then, AE31 and AE33 data went through a common validation procedure 18 . Brie y, BC concentrations below the Limit of Detection (-100 ng/m 3 ) were set as invalid; for BC 950nm ≥ 200 ng/m 3 , the spectral dependance is calculated from the linear regression of ln(λ) versus ln(B atn ). Valid measurements are considered for an Aerosol Angström Exponent (AAE) between 0.8 and 3, and a r² higher than 0.9.
Daily Concentrations of nitrogen monoxide (NO) and nitrogen dioxide (NO 2 ) were retrieved from 1-min measurements performed with a T200UP Teledyne instrument, equipped with a blue light photolytic converter and a Na on dryer. The instrument has been regularly calibrated with a reference standard from National Physics Laboratory (Teddington, UK) and NO and NO 2 concentrations have been corrected from ozone interference.
Daily ozone concentrations from 2012 to 2020 from a peri-urban station in Les Ulis, located around 10 km away from SIRTA, and operated by the Ile-de-France air quality monitoring structure (Airparif, https://www.airparif.asso.fr).
Meteorological variables (wind speed and direction, temperature, relative humidity and pressure) were provided from the ReObs database 33 .
Source apportionment of carbonaceous aerosols. A source apportionment study of OA was carried out by The oxidation properties of secondary organic aerosols were characterized by removing the contribution of primary factors to the OA matrix, as follows: From there, O:C SOA , H:C SOA and OSc SOA were calculated [38][39][40] . These equations provide only qualitative information for ACSM data, but are su cient to characterize a change, since they are uniformily applied throughout the dataset.
Fossil-fuel and Biomass burning fractions were estimated 41 . Since the choice of ff and wb is critical and given the size of our dataset, their determination was based on the statistical hourly distribution of the AAE (Fig. S3). PM ff and PM wb were estimated from BC ff and BC wb concentrations, respectively, following the conversion factors of 2 and 10.3 found for SIRTA during the same season 43 .
Analog. Analogues of atmospheric circulation 8,44-47 have been widely used for different purposes. Here, circulation analogues are computed from daily sea-level pressure (SLP) data, to better characterize nearsurface atmospheric circulation and air masses origin, as our study covers near-surface pollutants. The SLP data is extracted from NCEP/NCAR reanalysis data 48 along the historical period that covers 2012 to 2019. The SLP elds considered here have a horizontal resolution of 2.5 x 2.5° and cover a spatial domain ranging -20°W to +15°E in longitude and +40°N to +60°N in Latitude. This region is chosen because it includes atmospheric pressure patterns that in uence near-surface wind 49 in our area of study.
The study period covers 92 days from March 1 st until May 31 st 2020, while the historical period covers the same months in 2012-2019. Fifty best circulation analogues are sought for each day of the study period, using the pattern correlation as a way to measure similarity between SLP elds. The calendar distance between the day in the study period and days in the historic period is maximum 30 days. Out of 50 potential days with analogue atmospheric circulation, we keep only those with an analogue score larger than C > 0.6.
An additional ltering of the daily analog list was performed from a backtrajectory-based approach (see backtrajectory analysis), preventing the air mass origin of the synoptic analog from being signi cantly different from the air mass origin of the observation day. A minimum Pearson correlation coe cient of 0.2 between trajectory densities was used here. An example of satisfactory and unsatisfactory analogs is shown in Fig. S4. When correlating the overall trajectory density during lockdown and all the analog days, Pearson correlation is 0.77 (without trajectory ltering). The additionnal lter leads to an increase to 0.82, which demonstrates the relevance of the approach.
Further, we implemented a speci c constrain on local ambient temperature and Relative Humidity (RH). Indeed, both variables are key drivers of the partitionning of semi-volatile material, such as ammonium nitrate. Therefore, a satisfactory representation of local meteorological conditions by the analogs is needed in order to robustly capture and characterize any change of concentration. Over the study period, the analog performance regarding temperature and RH respectively ranges from -15°C to +15°C, and from -36% to +64%. Concretely, analogs that are much colder and wetter than the observation day may be associated to enhanced condensation of semi-volatile compounds; which would lead to a overestimation Finally, we keep for comparison only days that have at least more than 5 atmospheric circulation analogues. Analog days are satisfactorily distributed over the 2012-2020 dataset (Fig.S5), avoiding one speci c year of driving business as usual conditions. Backtrajectory analysis. 120-h backtrajectories ending at SIRTA (49.15°E, 2.19°N) at 500m a.g.l. were calculated every 6h from 2012 to 2020 with the PC-based version of HYSPLIT 50 using 1°x1° Global Data Assimilation System (GDAS) les. For each day, trajectory density (log of the occurrence of trajectory endpoints) was calculated over a 0.5°x0.5° grid covering Western-Europe (see Fig.S3). Calculations were controlled by ZeFir 51 .
Sensitivity tests. The results presented in this paper primarily depend on the list of analog days that is calculated. The overall analog number is at rst determined by the strictness of the correlation coe cients of atmospheric circulation and air mass origin. Selection of best analogy leads to poor statistical representativeness (Fig. S6), with a low number of analog days. It is instead preferable to remove analog days associated with worst trajectory correlation (Fig. S4). It is noteworthy that little change in the correlation coe cients (Table S1, Scenario 1-4) has little impact on the results (for all variables) presented in this manuscript (Fig. S7). This can be mainly related to the reasonable change in the number of analog days.
Similarly, the impact of subsequent ltering with temperature and relative humidity were also investigated. To this end, two additionnal scenarios were considered (S5-6, Table S1). S5 and S6 have limited impact on tra c-related variables (Fig. S7), on the contrary to wood-burning tracers and secondary compounds. This especially highlights the essential role of meteorological representativess in order to characterize the changes of secondary pollution. Indeed, when no temperature and RH ltering is performed (Scenario 5), highest decrease of NO 3 is linked to analog days that are associated to higher RH ( Fig. S8a) and lower temperature (Fig. S8b), which favor the partitioning of nitrate in the particulate phase. Scenario 6 has a stricter ltering than the Base one, and exhibits very good performance regarding the reconstruction of meteorological conditions (Table S2). However, we show that both scenario have very similar daily NO 3 concentration change despite slight discrepancies in meteorological performance.
This underlines that the thresholds used in the Base scenario are su cient to provide robust results.   NOx dependence of O3, NO3 and OOA concentration changes. Concentration change (µg/m3) of O3, NO3 and OOA versus concentration changes of NOx (µg/m3). Line and markers and shaded areas correspond to mean ± standard deviation.