Measurements of VOC mixing ratios and fluxes were made during two intensive measurement campaigns (winter: 12 November–10 December 2016; summer: 15 May–24 June 2017) from a meteorological mast located on the campus of the Institute of Atmospheric Physics, Chinese Academy of Sciences (IAP) in Beijing (39∘58′33 N, 116∘22′41 E, 49 m a.s.l.). This site was used as the main sampling site for all APHH-Beijing programme activities (see Shi et al., 2019). The campus is situated north of central Beijing between the 3rd and 4th ring roads, with parkland to the east and west and a mix of dense residential and commercial (restaurants and shops) buildings to the north and south. The population density in the Chaoyang district is 7500 people km-2 (Amir Siddique et al., 2020). Busy roads are situated 120 m north and 300 m east of the mast. Jing et al. (2016) classified the major roads present in the footprint as “artery roads”. The average speed of roads in this classification peaked at ∼ 43 km h-1 at 04:00 LT (local time). Daytime average speeds peak at ∼ 33 km h-1 (12:00 LT), with a minimum average speed of ∼ 28 km h-1 (18:00 LT). Average traffic volume on artery roads peaked at ∼ 3100 vehicles h-1 (07:00 LT) with a minimum average flow of ∼ 250 vehicles h-1 (03:00 LT). In 2002, the dominant tree species in Beijing were Sophora japonica, Populus tomentosa and Juniperus chinensis (Yang et al., 2005). Large-scale tree planting occurred between 2012 and 2015, increasing forest cover by 10 %, with the main species planted being Pinus tabuliformis, Sophora japonica, Fraxinus chinensis, Salix spp. and Populus spp. (Yao et al., 2019). For comparative purposes, the sampling location can be described as an “urban background” site with respect to ambient concentrations for central Beijing with the main sampling inlet being significantly elevated above ground level.

Flux measurements of CO2 have been made previously at this site (Song and Wang, 2012; Liu et al., 2012a; Song et al., 2013). The fractional land cover within 2 km of the tower was assessed by Song and Wang (2012) and found to be buildings (0.65), vegetation (0.23) and roads (0.12). Using the local climate zone (LCZ) classification system described by Stewart and Oke (2012) this area can be classified as LCZ2 “compact mid-rise”, becoming LCZ1 “compact high-rise” to the southwest. The aerodynamic roughness lengths were calculated by Li et al. (2003) as cited by Liu et al. (2012a) and found to be 2.5, 3.0, 5.3 and 2.8 m from the northeast, southeast, southwest and northwest, respectively, with the zero-plane displacement heights of 12.3, 15.0, 26.4 and 13.2 m, respectively. The roughness length and displacement height at this site has remained consistent over the last ∼ 20 years (Cheng et al., 2018b). The flux inlet was located at 102 m; this is over twice the mean height (45 m) of the tall residential building to the southwest (Liu et al., 2012a) and is below the mixing layer height (Squires et al., 2020), positioning the inlet in the constant flux layer.

A PTR-ToF-MS 2000 (Ionicon Analytik GmBH, Innsbruck) was housed in an air-conditioned container ∼ 10 m from the base of the mast. The core operating principle of the instrument is described in detail by Jordan et al. (2009). The PTR-ToF-MS subsampled from two inlet systems, a common flux inlet line and a gradient switching system. The common flux inlet had an outer diameter of 0.5 in. (inner diameter: 3/8 in.) PFA and sampled air from ∼ 50 cm below a sonic anemometer (Gill R3, Gill Instruments, Lymington, UK) which was mounted on a 3 m horizontal boom positioned on the mast at a height of 102 m a.g.l. Air was drawn through the inlet line using a rotary vane pump (Model VT4.8; Becker, Hull, UK). The flow rate in the inlet line was recorded using a mass flow meter (TSI Mass Flowmeter 4043, TSI, Shoreview, USA) and remained in the range 90–103 L min-1 during both campaigns. At 90 L min-1 the Reynolds number in the inlet line is 12 900, indicating turbulent flow. Particulates were removed from the air flow by a 90 mm Teflon filter installed near the inlet. This filter was replaced each day.

To minimise the impact of the tower on flux measurements, the sonic anemometer was positioned so that air from the prevailing wind directions would not be disturbed by the tower. The sonic anemometer was installed on the north west side of the tower in the winter and the south east in the summer. Analysis of the turbulence characteristics (Squires et al., 2020) did not show an impact from the structure even when the air flow passed through the tower. This is likely due to the open lattice construction of the tower and the predominance of larger eddy scales at 102 m.

Gradient measurements were made by switching between five sampling heights (3, 15, 32, 64 and 102 m) and zero air. Air was drawn down five separate 0.25 in. outer diameter (inner diameter: 1/8 in.) PFA lines using a common second rotary vane pump (Model VT4.8; Becker, Hull, UK). For each height, the flow was sampled at 3.3 L min-1 of which ∼ 300 mL min-1 was drawn through a 10 L stainless steel container, and an additional bypass flow of 3 L min-1 was used to reduce the residence time in each line (Fig. 1). The five containers were heated to 30 ∘C to limit adsorption and desorption effects and had a turn-over time of approximately 30 min. For 2 min each hour the PTR-ToF-MS sequentially subsampled from the flow exiting each container via a manifold valve, providing an hourly 30 min average mixing ratio at each of the five heights. The PTR-ToF-MS was operated in an hourly cycle switching between the flux and gradient measurements. The PTR-ToF-MS sampled from the gradient system for the first 20 min of each hour and from the flux line for the final 40 min of the hour. A zero-air generator was built in house and made up of a platinum catalyst heated to 260 ∘C, and zero air was sampled for 5 min each hour at the end of the gradient cycle.

The PTR-ToF-MS was operated with an inlet flow rate of 30 sccm and an E/N ratio (where E is the electric field strength and N is the buffer gas density) of 130 Td. To achieve this, the drift tube was maintained at 60 ∘C with a pressure 1.9 mbar and a voltage of 490 V applied across it. Data were acquired by the PTR-ToF-MS at a 5 Hz time resolution, allowing VOC fluxes to be calculated using the eddy covariance method. In order to facilitate mass calibration, trichlorobenzene was introduced by diffusion through a needle valve into the inlet stream.

The PTR-ToF-MS was calibrated twice weekly using a 15-component VOC standard (National Physical Laboratory, Teddington, UK) during both the winter and summer campaigns. This standard contained methanol, acetonitrile, ethanol, 1,3-butadiene, acetone, isoprene, butenone, butan-2-one, benzene, toluene, m-xylene and 1,2,4-trimethylbenzene at 1 ± 0.1 ppmv each. The standard was dynamically diluted in zero air to provide a six-point calibration. The sensitivities to these compounds in the winter and summer campaigns are summarised in the Supplement.

During the winter campaign, the instrument was also calibrated three times using a second independent 15-component VOC standard (Ionicon Analytik, Innsbruck, Austria) containing methanol, acetonitrile, acetaldehyde, ethanol, acrolein, acetone, isoprene, crotonaldehyde, butan-2-one, benzene, toluene, o-xylene, chlorobenzene, α-pinene and 1,2-dichlorobenzene at 1 ± 0.05 ppmv each. At the end of the campaign the PTR-ToF-MS was also calibrated using an aldehyde standard (Ionicon Analytik, Innsbruck, Austria) made up of formaldehyde, acetaldehyde, acrolein, propanal, crotonaldehyde, butanal, pentanal, hexanal, heptanal, and octanal at 1 ± 0.05 ppmv each; nonanal at 600 ± 0.07 ppbv; and decanal at 500 ± 0.08 ppbv. Calibration using both Ionicon standards was performed following dynamic dilution in zero air to give a six-point calibration. In the summer campaign, four-point calibrations were performed using both Ionicon VOC standards once a week.

Over the course of the winter campaign 29 h (6 % of the PTR-ToF-MS operational period) was lost to instrument maintenance and calibration, and 133 h (16 % of the PTR-ToF-MS operational period) was lost to instrument maintenance and calibration in the summer.

PTR-ToF-MS mass calibration and peak fitting was performed on the 5 Hz data using PTRMS Viewer 3 (Ionicon Analytik, Innsbruck). VOC mixing ratios were then calculated using the method based on that previously applied by Acton et al. (2016), Tani et al. (2004) and Taipale et al. (2010). Counts per second (cps) of each protonated VOC species (RH+) were normalised against the primary ions (H3O+ and the H2O H3O+ cluster ion), and background counts were subtracted to give background-corrected normalised count rates I(RH+)norm (ncps): 1IRH+norm=IRH+InormIH3O++IH3O+H2O-1n∑i=1nIRH+zero,iInormIH3O+zero,i+IH3O+H2Ozero,i, where I(H3O+) and I(H3O+H2O) represent the measured count rate for H3O+ and the H3O+H2O cluster, respectively. Zero-air measurements were labelled using the subscript zero and the number of zero-air measurements was represented by n. The total reagent ion count rate (H3O+ and H3O+H2O) was normalised to a count rate of 106 cps (Inorm).

Following calculation of the background corrected normalised count rate, the volume mixing ratio (VMR) was calculated, as a wet mass fraction, as follows: 2VMRVOC=I(RH+)normSnorm, where Snorm is the normalised sensitivity (ncps/ppbv) for each mass calculated using a transmission curve as described by Taipale et al. (2008). Mixing ratios of specific compounds were then determined by summing the parent ion and fragment ion mixing ratios. For direct comparison with the molar flux, the volume mixing ratio was converted to a molar concentration (χVOC) using Eq. (3). 3χVOC=P×VMRVOCR×T, where P is the atmospheric pressure, T is atmospheric temperature and R is the molar gas constant (8.314 J K-1 mol-1). Note that in the following presentation of results and discussion, the abundance of VOCs in air are described in terms of volume mixing ratio (parts per billion; ppbv) although strictly speaking they should be described as the wet mass fraction (with the same units).

The continuous fast measurement of PTR-ToF-MS for all compounds allows the estimation of fluxes using direct eddy covariance rather than the virtual disjunct eddy covariance methods that had to be applied when using older PTR-quad-MS instruments (e.g. Karl et al., 2002; Langford et al., 2009, 2010a; Velasco et al., 2005, 2009). Eddy covariance fluxes of VOCs using PTR-ToF-MS were first made by Müller et al. (2010). A detailed description of flux measurement using eddy covariance and its limitations is provided by Aubinet et al. (2012). Here fluxes were calculated using the eddy4R routines (Metzger et al., 2017; R Core Team, 2019) with further analysis performed using the open air and ggmap packages (Carslaw and Ropkins, 2012; Kahle and Wickham, 2013). The flux (Fx) of each compound was determined by calculating the covariance function between the vertical wind velocity (w) and the VOC concentration (χx; nmol m-3): 4FxΔt=1N∑i=1Nw′(i-Δt/Δtw)χx′(i), where w′ and χ′ represent momentary deviations from the mean concentration or vertical wind speed (i.e. w′=w-w¯). N is the number of PTR-ToF-MS measurements in our half-hour averaging window (9000 for 5 Hz measurements), Δtw is the sampling interval between wind measurements (0.2 s), and Δt is the lag time between the vertical wind measurements recorded by the sonic anemometer at 102 m and the concentrations recorded using the PTR-ToF-MS instrument at ground level.

The lag time between the vertical wind velocity measurement and VOC concentration measurement is primarily controlled by the inlet line length and the flow rate in the line. However, small variations in temperature, pressure, humidity, and pump performance and also horizontal displacement between the anemometer and inlet, coupled with changes in wind speed and direction, can all cause deviation in the lag time. The lag time can be determined by assessing the covariance between w and χ as a function of time. The lag time can then be identified by selecting the maximum of this covariance function (Taipale et al., 2010). For many of the compounds recorded in this study, only a weak flux was observed. This results in a low signal-to-noise ratio, introducing a large uncertainty into the identification of the cross-covariance maximum. Adsorption and desorption rates to the inlet line are compound-specific and dependent on polarity. Therefore, the lag times for more polar compounds such as oxygenated VOCs will differ slightly from pure hydrocarbons, but weak covariance peaks made the identification of the lag-times difficult for many masses. Lag times were therefore determined by calculating the lag for compounds where a strong covariance peak was observed (isoprene and C2-benzene (C8H10) in the summer and winter campaigns, respectively) and applying their modal lag times as a fixed value to all masses. The calculation of lag time for eddy covariance data with low signal-to-noise ratio is described in detail by Langford et al. (2015).

The effect of storage below the measurement height was calculated for flux averaging period t as follows: 5StorageFxt=hχxt-1-χxt+13600, where χxt-1 and χxt+1 are the concentration (nmol m-3) of compound X in the averaging periods before and after flux averaging period, respectively; t and h represent the measurement height in metres (102 m). This method was chosen so as to be comparable with the other flux measurements made during this project (e.g. Squires et al., 2020). This storage term was then added to the calculated turbulent flux term.

At the 102 m measurement height, large eddies mean that some of the flux may not be captured in the 30 min flux averaging period used here. This loss of low-frequency flux was investigated by Squires et al. (2020), who found that at the 102 m measurement height the flux loss was ca. 7 %. Due to the similarity of the transport mechanisms underlying the low-frequency flux loss (e.g. Mauder et al., 2020), we assume this estimate to also apply to the simultaneously measured VOC fluxes presented here. We did not correct the fluxes presented in this paper for low-frequency flux loss. We further investigated the effect of high-frequency spectral loss on the VOC fluxes using a wavelet-based methodology (Nordbo and Katul, 2013). Loss of high-frequency flux due to measuring at 5 Hz was estimated to be less than 10 %, which we did not correct for in the presented fluxes. Fluctuations in air density caused by changes in temperature and humidity can impact fluxes (Webb et al., 1980). As a closed path system was used in this study, the impact of air density variations caused by changes in sensible heat flux are negligible, but variations in air density caused by changes in the latent heat flux may impact the VOC flux calculation. The impact of changes in latent heat flux on the NOx flux at this site was assessed by Squires et al. (2020) and found to be significantly less than 1 % throughout both campaigns. As the sensitivity of a trace gas flux to density fluctuations is controlled by the ratio of its constituent density over its flux (Pattey et al., 1992), this ratio was assessed for the measured VOC species. This ratio was found to be comparable to that observed for NOx by Squires et al. (2020), suggesting that the uncertainty in VOC flux as a result of changes in latent heat flux was less than 1 %. Therefore the Webb–Pearman–Leuning correction (WPL: Webb et al., 1980) for changes in the density of air to the calculation of fluxes was not applied. Squires et al. (2020) estimated that the average time taken for an air parcel to reach the inlet point at 102 m was ∼ 68 s. The OH concentration in the summer campaign reached 2.8 × 107 molec. cm-3 (Whalley et al., 2020), and the ozone concentration reached 2.6 × 1012 molec. cm-3, at these concentrations ∼ 10 % of monoterpenes (the most reactive species recorded) would have reacted before the measurement height.

Half-hour averages were quality-assessed using three tests after a two-dimensional coordinate rotation to correct for tilting of the sonic anemometer. The limit of detection was calculated for each mass by determining the cross-covariance between 150–180 s, a region outside the expected time lag range (Spirig et al., 2005). In order to calculate a reliable flux, the averaging period must meet stationarity requirements such that this period is shorter than the timescale at which changes in atmospheric conditions occur. The stationarity of the flux across the half-hour averaging period was assessed using the method described by Foken and Wichura (1996). The half-hour averaged fluxes were filtered where the stationarity criterion (the difference between average flux and mean average flux of its components) exceeded 60 % and were above the limit of detection; as for fluxes below the limit of detection, a robust statement on their stationarity could not be made. Lastly, files were also flagged if the mean frictional velocity (u∗) was less than 0.175 m s-1, this threshold was derived from assessment of the u∗ dependence of the sensible heat flux. Data falling below this threshold were substituted by the campaign-average value for that hour so as not to introduce a positive bias to the VOC flux. The random error of each half-hour flux average was assessed using the method described by Lenschow et al. (1994) and fluxes with a relative random error greater than 150 %, averaged across the campaign, were discarded.

In the winter campaign, 25 % of the flux-averaging periods fell below the u∗ filter. Of the masses contributing more than 0.75 % to the total flux, the average percentage of flux files filtered for failing the stationarity test is 69 %. This ranged from 73 % for methyl vinyl ketone and methacrolein formaldehyde to 62 % for acetaldehyde. In the summer field campaign, 12 % of the flux-averaging periods fell below the u∗ filter. Of the masses contributing more than 0.75 % to the total flux, the average percentage of flux files filtered for failing the stationarity test is 25 %. This ranged from 52 % for formaldehyde to 13 % for methanol.

During the winter measurement campaign (7 November to 10 December), wind speeds were low, ranging between 0.3 and 9.7 m s-1, with a mean value of 2.4 m s-1, and with the highest wind speeds observed from the northwest (Fig. 2). The predominant wind directions were southerly and northwesterly. Temperatures, measured at 102 m on the IAP tower, ranged from -7 to 15 ∘C, with a mean value of 3.6 ∘C. Relative humidity ranged from 15 % to 92 % with a mean value of 45 % (Fig. 3). Precipitation was small during the campaign period with light rain on 20 November and snow on 21 November.

During the summer measurement campaign (20 May–22 June 2017), wind speed ranged between 0.5 and 7.9 m s-1, with a mean value of 2.3 m s-1. The predominant wind direction was southeasterly (Fig. 2). Temperatures at 102 m ranged from 15 to 37 ∘C, with a mean value of 25 ∘C, and the relative humidity ranged from 13 % to 93 %, with a mean value of 44 % (Fig. 3). Precipitation was low during the campaign period with light rain observed on 22 and 30 May and on 2, 6, 18, and 21–23 June.

The flux footprint was calculated for each half-hour flux averaging period of both field campaigns using the method described by Kljun et al. (2004) and Metzger et al. (2012). The calculation of the flux footprint for this campaign is described in detail by Squires et al. (2020). During both the summer and winter field campaigns 90 % of the measured flux originated from an area within 7 km of the IAP tower; however, 90 % of the contribution to the campaign-average flux footprint extended just 2 km from the tower. Mean flux footprint climatologies for the summer and winter campaigns are displayed in Fig. 4, with contour lines showing the distances from the tower where the surface contributions to the measured fluxes cumulate to 30 %, 60 % and 90 %, respectively. In the winter campaign the main contribution to the average flux was predominantly from the northwest and southeast, encompassing two large roads, the Jingzang Expressway and the Beitucheng West Road, and a mix of commercial and residential buildings and urban park land. In the summer campaign, the largest contribution to the flux came from regions approximately 1 km southwest of the tower and approximately 1 km northeast of the tower now encompassing different sections of the Jingzang Expressway, Beitucheng West Road, residential buildings, commercial buildings (shops and restaurants) and a larger contribution from urban park land compared to the winter campaign.

In the winter campaign the detected VOC species with the largest mixing ratios were the small oxygenated VOCs (OVOCs) methanol, acetaldehyde and acetone, with median mixing ratios of 19.29, 4.59 and 2.57 ppbv, respectively. Large mixing ratios of aromatic compounds were also observed, with median mixing ratios between 1.5 and 2.0 ppbv each for benzene, toluene and C2-benzenes. Mixing ratios of these compounds tracked the pollution events in a distinctive “sawtooth cycle” described by Jia et al. (2008) where mixing ratios increased over a period of 3–4 d when the wind speed was low and prevailing wind direction was southerly before dropping rapidly when the wind direction moved to the northwest. In the summer campaign, the dominant VOC species observed were again methanol, acetaldehyde and acetone (26.39, 3.50 and 3.67 ppbv, respectively), but the mixing ratios of aromatic compounds were lower, with median values of 0.30, 0.37 and 0.47 ppbv observed for benzene, toluene and C2-benzenes, respectively. Methanol and ethanol + formic acid together account for approximately one-third of the total oxygenated VOC mixing ratio. These compounds have been observed at high concentrations in many cities (Langford et al., 2009, Manchester, summer; Shao et al., 2009, Beijing, summer; Valach et al., 2015, London, August and September; Sahu and Saxena, 2015, Ahmedabad, winter), with ethanol reported to be the most abundant VOC in London with a mean mixing ratio of 5 ppbv in summer and winter (Dunmore et al., 2016).

Isoprene, which has both biogenic and vehicular sources (Borbon et al., 2001; Jaimes-Palomera et al., 2016), had median mixing ratios of 1.08 and 0.38 ppbv in the winter and summer campaigns, respectively. High winter isoprene concentrations have been observed previously in Beijing (Li et al., 2019), where an average value of 1.01 ppbv was observed, as well as during haze periods (Sheng et al., 2018). High winter isoprene has also been reported in other Asian countries, with isoprene mixing ratios of 1.6 and 1.1 ppbv recorded in Ahmedabad (Sahu and Saxena, 2015) and Kathmandu (Sarkar et al., 2016), respectively. These studies are all characterised by the use of a PTR-MS; it is possible that an isomer of isoprene contributes to the high winter time concentrations observed in Asian cities. In contrast, other studies have reported higher summertime isoprene concentrations in Beijing (e.g. Cheng et al., 2018a). In the winter, isoprene mixing ratios peaked during haze events and correlated well with anthropogenic and oxygenated VOCs. R2 values of 0.82, 0.77 and 0.74 were observed for the correlation of isoprene with propenal, benzene and toluene, respectively.

In contrast to the winter campaign, mixing ratios of most compounds during the summer campaign showed a clear diurnal cycle, with mixing ratios peaking at night and dropping during the day as the planetary boundary layer expanded during the daytime and contracted at night-time (Fig. 5). A summary of VOC mixing ratios for the principal VOC species observed during the two campaigns are displayed in Table 1. For a more detailed discussion of VOC mixing ratios recorded during this campaign, see Zhang et al. (2020).

The mixing ratio values represent the ambient concentrations observed at the sample inlet averaged over the sampling period. However, by combining these mixing ratios with the three-dimensional wind vector, the eddy covariance flux calculation method allowed for quantification of the net exchange (flux) of the observed compounds from a surface “footprint” to the point of sampling. The size of this footprint depends on the height of measurement and surface roughness, as well as a number of meteorological factors, including wind speed and atmospheric stability. Under the conditions in Beijing, 90 % of the campaign average measured flux originated from an area within 1–2 km of the tower (see Fig. 4). The fluxes observed at 102 m are therefore predominantly controlled by local emissions, unlike the mixing ratios observed at the same point which are influenced by both local and more distant emissions, chemical processing and meteorology. Fluxes of CO2 have been previously recorded at the IAP field site (Liu et al., 2012a; Song and Wang, 2012; Song et al., 2013). Liu et al. (2012a) reported four years of CO2 eddy covariance measurements showing a seasonal variation in emissions driven by winter heating and uptake of CO2 by vegetation in the summer. Spatial variation in emission was shown to be determined by the surface cover, and the diurnal profile of emissions was largely dependent on traffic volumes.

Greater surface solar heating in the summer led to stronger turbulence than was observed in the winter, with median u∗ values of 0.41 and 0.26 m s-1, respectively. This meant turbulent transport was more easily measured, allowing the fluxes of a larger number of compounds to be quantified in summer than in the winter. Of the 179 masses observed in the summer campaign, 64 masses gave a flux with a median relative random error of less than 150 %. These masses all showed a net positive flux, with no masses showing net deposition. Molecular formulae were assigned to 50 of these masses using their exact mass, while 15 masses with a small negative mass defect could not be allocated a formula, based on the combination of C, H, N, O, S and Cl within 0.1 mDa of the recorded mass. Mass spectral peaks observed at m/z 47.019 and 47.053, assigned to formic acid and ethanol, respectively, could not be sufficiently resolved, so these compounds were considered together.

Diurnal cycles of the six compounds with the largest fluxes are shown in Fig. 6 (solid lines), showing the large VOC flux observed in the summer campaign, in comparison to the much smaller fluxes observed in the winter campaign. This figure also displays fluxes without the storage term applied (dashed lines) and the gap-filled u∗ filtered flux (dotted lines). As shown, these corrections have only a small impact on the fluxes calculated. The fluxes of these six compounds are summarised in Table 2. As shown in Table 2, fluxes of all compounds were larger in the summer than in the winter. This was not the case for NOx and CO, which were larger in the winter than in the summer (Squires et al., 2020). There may be several reasons why the measured VOC fluxes were lower in winter than summer. In the winter, the high concentrations of VOCs at this site, mainly resulting from advection from outside the city, might lead to deposition of VOCs. This would have the effect of suppressing the net VOC emission rate measured at the site. In the winter campaign, OH concentrations peaked at 8 × 106 cm-3 (Slater et al., 2020), and the average O3 concentration was 16.4 µg m-3 (Shi et al., 2019). Using rate coefficients from Atkinson and Arey (2003), the expected atmospheric lifetimes of benzene and 1-butene with respect to OH would be 1.2 d and 1.1 h, respectively. The lifetimes of benzene and monoterpenes with respect to O3 would be over a year and 6 d, respectively. The lifetime of less reactive compounds such as aromatics is therefore long enough to allow for transportation from outside the flux footprint. The average temperature in the winter campaign was 3.6 ∘C, compared to 25 ∘C in the summer, and this would result in lower volatilisation of liquid VOCs (for example in gasoline), suppressing the emission rates. Sources such as cooking and solvents in cleaning products, cosmetics, and paints contribute to urban VOC emission (Karl et al., 2018) and these sources are likely to be reduced in the winter when cold weather causes ventilation of buildings to be reduced. Measured VOC emissions from indoor and outdoor sources may therefore be lower in winter. In addition, the lower atmospheric turbulence in winter and generally lower fluxes result in greater uncertainty in our winter fluxes compared to the summer fluxes. Our detailed analysis, therefore, focuses on the fluxes recorded during the summer campaign.

The Multi-resolution Emissions Inventory for China (MEIC; Li et al., 2017; Qi et al., 2017) is a comprehensive inventory of the emission of atmospheric pollutants across China and provides the emissions input to many modelling studies (e.g. Hu et al., 2016). Inventories such as MEIC use emissions estimates from individual sources scaled by activity to provide a “bottom-up” emissions estimate. The emission factors used by the inventory are drawn from a broad range of studies (Li et al., 2014; Zheng et al., 2018), with locally derived emission factors used where possible. These emission factors have been summarised by Li et al. (2017b). Whilst activity data is often relatively well constrained at a national level, its spatial disaggregation often relying on simplified proxies such as population density adds significant additional uncertainty for the estimate of the emission for a given location. The VOC flux measurements presented here provide a “top-down” measurement at the cityscape scale that can be used to evaluate the accuracy of the emissions inventory encompassed by the flux footprint. Zhao et al. (2017) assessed the uncertainties in emissions inventories for China and recommended that field studies were used to improve emission estimates.

The MEIC includes emissions from five sectors: power plants, transportation, industrial, agricultural and residential sources, with VOC emissions collated by functional group. As no agricultural land was present within the flux footprint, agricultural emissions are not considered here. Monthly emissions data are provided, with emissions assumed to be consistent across each month. A diurnal cycle based on local activity data is applied to each emission sector to give a diurnal profile. Within the MEIC inventory VOCs are grouped into classes. Nine of these groupings were considered, but those containing alkanes were not considered as alkanes have a proton affinity less than that of water and could not be detected using PTR-MS. The groupings, based on the speciation used by the RADM2 chemical mechanism, are summarised in Table 3. The emission ratio of different VOC species present within each grouping could represent a source of error in the inventory.

VOC emissions estimates for the summer campaign period were calculated through combination of the flux footprint (100 m × 100 m) and high-resolution (3 km × 3 km) MEIC v 1.3 inventory for 2013. First, the footprint for each flux aggregation period and the inventory grid for the corresponding hour of day were aligned by transforming the flux footprint into the coordinate reference system of the inventory. Subsequently the inventory values were extracted at the centre of each footprint grid cell, creating a pseudo 100 m × 100 m inventory. This was multiplied by the footprint grid, weighting each cell by their contribution to the measured flux. There was little variation between VOC emission factors from the main four emissions inventory cells contributing to the flux observed at the IAP tower site.

Figure 11 shows the comparison between measured VOC emission (summer campaign) and the VOC emission predicted by the MEIC inventory for this period with the percentage contribution of each group to the total flux displayed in Fig. 12. Measured emissions of aromatic compounds were 3 % and 4 % of those predicted by the inventory for TOL and XYL (low mass and high mass aromatics). Squires et al. (2020) compared NOx and CO fluxes recorded during the APHH-Beijing campaigns with the MEIC inventory and also observed a similar overestimation of emissions by the inventory, with the inventory overestimating the measured flux 11 and 10 times, respectively. Industrial emissions make the largest contribution to aromatic VOC emission in the inventory. However, inspection of the flux footprint reveals few potential industrial sources, with the footprint encompassing roads and residential buildings, as well as shops and restaurants. At a coarser resolution of 9 × 9 km the TOL emission peaked at 25 nmol m-2 s-1, 9 times higher than the measured flux, indicating that the overestimate was in part caused by the proxies used to downscale the inventory to 3 km. Previous studies have shown that the inventory underestimates aromatic emission at the Chinese scale (Liu et al., 2012b; Cao et al., 2018), again indicating that the overestimation observed here is a result of the downscaling of the inventory. The allocation of industrial emissions to this residential area is likely a result of downscaling the emissions using proxies such as population (Zheng et al., 2017). Even discounting the industrial emissions, the predicted emissions of TOL and XYL are still 8 times larger than the measured flux at the 3 km × 3 km resolution. The inventory does capture the diurnal cycle in emissions, with a rapid increase in aromatic VOC emissions at 07:00 LT and emissions remaining high throughout the day before decreasing after 17:00 LT.

Isoprene emissions are significantly underrepresented in the inventory, with measured fluxes 20 times higher than inventory emissions. This is expected as the inventory only considers anthropogenic emission, and in the summer isoprene emissions are dominated by predominantly biogenic isoprene. Adding the isoprene emission predicted by MEGAN to the MEIC emission (Fig. 11h) shows that the inclusion of biogenic isoprene enables a more accurate estimate of total isoprene emission. The park at the base of the tower means that a larger biogenic source may be present at the IAP site than would be expected for Beijing on average.

The performance of the inventory in relation to oxygenated VOCs varied considerably between chemical classes. The inventory was more accurate when estimating emissions of ethanol, formaldehyde and ketones, with measured emissions being 71 %, 20 % and 81 % of inventory emissions, respectively. When industrial emissions are removed, these increase to 132 %, 30 % and 194 % of the inventory emissions, suggesting that, despite the PTR-ToF-MS only capturing two ketones, ketone emissions are significantly underestimated from residential and transport sources. Measured methanol, aldehydes and organic acid fluxes were also underestimated by the inventory with the measured flux being 13, 1.8 and 107 times higher than the inventory, respectively.

It is possible that some underestimation of methanol, aldehydes and organic acid fluxes can be explained by their formation on surfaces following the oxidation of primary-emitted predominantly biogenic (e.g. Lee et al., 2006; Acton et al., 2018b) or anthropogenic compounds (Atkinson, 2000). However, given the ∼ 68 s time taken for an air parcel to reach the inlet point, it is not likely that formation in the atmosphere contributes significantly to the flux of these compounds. Consumer goods are known to be a large source of oxygenated VOCs (Dinh et al., 2015; Nematollahi et al., 2019). The use of these products is likely to vary considerably between households, making bottom-up emission hard to estimate.