CARIBIC-1 (Phase #1, abbreviated hereafter “C1”) was operational from November 1998 until April 2002 using a Boeing 767-300 ER operated by LTU International Airlines (Brenninkmeijer et al., 1999). Using a whole air sample (WAS) collection system, 12 air samples were collected per flight (of 8–10 h duration at cruise altitudes of 10–12 km) in stainless steel tanks for subsequent laboratory analysis of the mixing ratios (i.e. mole fractions) of various trace gases, including 14CO. Large air samples were required in view of the ultra-low number density of this mainly cosmogenic tracer (10–100 molecules cm-3 standard temperature and pressure (STP), about 0.4–4 amol mol-1). Hereinafter STP denotes dry air at 273.15 K, 101 325 Pa. Each C1 WAS sample (holding 350 L of air STP) was collected over 15–20 min intervals representing the number density-weighted average of the compositions encountered along flight segments of about 250 km. The overall uncertainty of the measured WAS CO is less than ±1 % for the mixing ratio and ±0.1 ‰ / ±0.2 ‰ for δ13C(CO)/δ18O(CO), respectively (Brenninkmeijer, 1993; Brenninkmeijer et al., 2001). Isotope compositions are reported throughout this manuscript using the so-called delta value δi=(iR/iRst-1) relating the ratios R of rare (13C, 18O or 17O) over abundant isotopes of interest to the standard ratios Rst. The latter are Vienna Standard Mean Ocean Water (VSMOW) for 18O / 16O (Gonfiantini, 1978; Coplen, 1994) and 17O / 16O (Assonov and Brenninkmeijer, 2003), and Vienna Pee Dee Belemnite (VPDB) for 17O / 16O (Craig, 1957), respectively. As we mention above, the oxygen isotope composition of the CO present in these WAS samples was corrupted, in particular when O3 levels were as high as 100–600 nmol mol-1.

CARIBIC-2 (Phase #2, referred to as “C2”) started operation in December 2004 with a Lufthansa Airbus A340-600 fitted with a new inlet system and air sampling lines, including perfluoroalkoxy alkane (PFA) lined tubing for trace gas intake (Brenninkmeijer et al., 2007). No flask CO mixing/isotope ratio measurements are performed in C2.

In addition to the WAS collection systems, both C1 and C2 measurement setups include different instrumentation for on-line detection of [CO] and [O3] (hereinafter the squared brackets [] denote the mixing ratio of the respective species). In situ CO analysis in C1 is done using a gas chromatography (GC)-reducing gas analyser which provides measurements every 130 s with an uncertainty of ±3 nmol mol-1 (Zahn et al., 2000). In C2, a vacuum ultraviolet fluorescence (VUV) instrument with lower measurement uncertainty and higher temporal resolution of ±2 nmol mol-1 in 2 s (Scharffe et al., 2012) is employed. Furthermore, the detection frequency for O3 mixing ratios has also increased, viz. from 0.06 Hz in C1 to 5 Hz in C2 (Zahn et al., 2002, 2012).

When comparing the CO mixing ratios in relation to those of O3 for C1 and C2, differences are apparent in the LMS, where C2 [CO] values are systematically lower. This is illustrated in Fig. 1a which presents the LMS CO–O3 distribution of the C2 in situ measurements overlaid with the C1 in situ and WAS data. The entire C1 CO/O3 data set is presented in Fig. 2. For the in situ CO data sets we calculated the statistics (Fig. 1b) of the samples with respective O3 mixing ratios clustered in 20 nmol mol-1 bins, i.e. the median and spread of [CO] as a function of [O3] analysed. The interquartile range (IQR) is used in the current analysis as a robust measure of the data spread instead of the standard deviation. The LMS data exhibit large [CO] variations for [O3] between 300 and 400 nmol mol-1, which primarily reflect pronounced seasonal variations in the NH tropospheric CO mixing ratio. With increasing [O3], [CO] decreases to typical stratospheric values, and its spread reduces to mere 3.5 nmol mol-1 and less, as [O3] surpasses 500 nmol mol-1. Despite the comparable spread in C1 and C2 [CO], from 400 nmol mol-1 of [O3] onwards the C1 CO mixing ratios start to level off, with no samples below 35 nmol mol-1 having been detected, whereas the C2 levels continuously decline. By the 570–590 nmol mol-1 O3 bin, C1 [CO] of 39.7-1.3+0.7 nmol mol-1 contains some extra 14 nmol mol-1 compared to 25.6-1.1+1.2 nmol mol-1 typical for C2 values. Overall, at [O3] above 400 nmol mol-1 the conspicuously high [CO] is marked in about 200 in situ C1 samples, of which 158 and 69 emerge as statistically significant mild and extreme outliers, respectively, when compared against the number of C2 samples (n>3×105). The conventions here follow Natrella (2003), i.e. ±1.5 and ±3 IQR ranges define the inner and outer statistical fences (ranges outside which the data points are considered mild and extreme outliers) of the C2 [CO] distribution in every O3 bin, respectively. The statistics include the samples in bins with average [O3] of 420–620 nmol mol-1. None of C1 CO at [O3] above 560 nmol mol-1 agrees with the C2 observations. Because the CO–O3 distribution cannot have changed over the period in question, we find that an apparent relative excess CO of up to 55 % justifies and investigation into sampling artefacts and calibration issues.

Unnatural elevations in δ18O(CO) from WAS measurements are also evident, as shown in Figs. 3 and 4. The large δ18O(CO) elevations that reach beyond +16 ‰ are found to be proportional to the concomitant O3 mixing ratios (denoted with colour in Fig. 3) and are more prominent at lower [CO]. Lower δ18O(CO) values, however, are expected based on our knowledge of UT/LMS CO sources (plus their isotope signatures) and available in situ observations (Fig. 3, shown with triangles), as elucidated by Brenninkmeijer et al. (1996) (hereafter denoted as “B96”). That is, the greater the proportion of stratospheric CO, the greater its fraction stemming from methane oxidation with a characteristic δ18O of 0 ‰ or lower (Brenninkmeijer and Röckmann, 1997). This occurs because the sink of CO at ruling UT/LMS temperatures proceeds more readily than its production, as the reaction of hydroxyl radical (OH) with CO, being primarily pressure-dependent, is faster than the temperature-sensitive reaction of OH with CH4. Furthermore, as the lifetime of CO quickly decreases with altitude, transport-mixing effects take the lead in determining the vertical distributions of [CO] and δ18O(CO) above the tropopause, hence their mutual relationship. This is seen from the B96 data at [CO] below 50 nmol/mol that line-up in a near linear relationship towards the end members with lowest 18O / 16O ratios. These result from the largest share of the 18O-depleted photochemical component and extra depletion caused by the preferential removal of C18O in reaction with OH (fractionation about +11 ‰ at pressures below 300 hPa, Stevens et al., 1980; Röckmann et al., 1998b).

We are confident that the enhancements of C1 C18O originate from O3, whose large enrichment in 18O (above +60 ‰ in δ18O, Brenninkmeijer et al., 2003) is typical and found transferred to other atmospheric compounds (see Savarino and Morin (2012) for a review). In Fig. 3 it is also notable that not only the LMS compositions are affected but elevations of (3–10) ‰ from the bulk δ18O(CO) values are present in more tropospheric samples with [CO] of up to 100 nmol mol-1. These result from the dilution of the least affected CO-rich tropospheric air by CO-poor (however substantially contaminated) stratospheric air, sampled into the same WAS tank. Such sampling-induced mixing renders an unambiguous determination of the artefact source isotope signature rather difficult, because neither mixing nor isotope ratios of the admixed air portions are known sufficiently well (see below).

Differences between the WAS and in situ measured [CO] – a possible indication that the δ18O(CO) contamination pertains specifically to the WAS data – average at Δ‾(WAS – in situ)=5.3± 0.2 nmol mol-1 (±1 standard deviation of the mean, n=408). These differences also happen to be random with respect to any operational parameter or measured characteristic in C1, i.e. irrespective of CO or O3 abundances. The above-mentioned discrepancy remained after several calibrations between the two systems had been performed, and likely results from the differences in the detection methods, drifts of the calibration standards used (see details in Brenninkmeijer et al., 2001) and a short-term production of CO in the stainless steel tanks during sampling. The large spread of Δ(WAS – in situ) of ±3.5 nmol mol-1 (±1σ of the population) ensues from the fact that the in situ sampled air corresponds to (2–4) % of the concomitantly sampled WAS volume, as typically 6–7 in situ collections of 5 s were made throughout one tank collection of 17–21 min. The integrity of the WAS CO is further affirmed by the unsystematic distribution of the artefact compositions among tanks (in contrast to that for δ18O(CO2) in C1 discussed by Assonov et al., 2009). Overall, the WAS and in situ measured CO mixing ratios correlate extremely well (adj. R2=0.972, slope of 0.992 ± 0.008 (±1σ), n=408). However, both anomalies in [CO] and δ18O(CO) manifest clear but complex influences of the concomitant [O3]. That is, the C1 in situ and WAS [CO] and δ18O(CO) data very likely evidence artefacts pertaining to the same O3-driven effect. Below we discuss and quantify these influences.

We use the differential mixing model (MM, originally known as the “Keeling plot”) in combination with the parameterisation of the artefact CO component (Eq. 1) to derive the isotopic composition of the latter. This approach makes no assumptions on the isotope signatures of CO in the air portions mixed in a given WAS tank. The MM parameterises the admixing of the portion of artefact CO to the WAS sample with the “true” initial composition, as formulated below: [CO]=[COt]+[COc],δ(CO)[CO]=δ(COt)[COt]+δ(COc)[COc], where indices c and t distinguish the components pertaining to the estimated contamination and “true” composition sought (i.e. [COt] and δ(COt)), respectively. Here the contamination strength [COc] is derived by integrating Eq. (1) using the in situ C1 [O3] data for each WAS sample. By rewriting the above equation with respect to the isotope signature of the analysed CO, one obtains δ(CO)=δ(COc)+(δ(COt)-δ(COc))[COt]/[CO], which signifies that linear regression of δ(CO) as a function of the reciprocal of [CO] yields the estimated contamination signature δ(COc) at ([CO])-1 → 0 when invariable “true” compositions ([COt], δ(COt)) are taken (the Keeling plot detailing these calculations is shown in Fig. 5). We therefore apply the MM described by Eq. (4) to the subsets of samples picked according to the same reckoned [COt] (within a ±2 nmol mol-1 window, n>7). Such selection, however, may be insufficient: due to the strong sampling effects in the WAS samples (see previous Section), it is possible to encounter samples that integrate different air masses to the same [COt] but rather different average δ(COt). The solution in this case is to refer to the goodness of the MM regression fit, because the R2 intrinsically measures the linearity of the regressed data, i.e. closeness of the “true” values in a regarded subset of samples, irrespective of underlying reasons for that.

Higher R2 values thus imply higher consistency of the estimate, as demonstrated in Fig. 6 showing the calculated δ(COc) for [COt] below 80 nmol mol-1 as a function of the regression R2. The latter decreases with greater [COt] (i.e. larger sample subset size, since tropospheric air is more often encountered) and, correspondingly, larger variations in δ(COt). Ultimately, at lower R2 the inferred δ18O(COc) converge to values slightly above zero expected for uncorrelated data, i.e. C1 δ18O(CO) tropospheric average. A similar relationship is seen for the δ13C(COc) values (they converge around -28 ‰), however, there are no consistent estimates found (R2 is generally below 0.4). Since such is not the case for δ18O, the MM is not sufficiently sensitive to the changes caused by the contamination, which implies that the artefact CO δ13C should be within the range of the “true” δ13C(CO) values. Interestingly, the MM is rather responsive to the growing fraction of the CH4-derived component in CO with increasing [O3], as the δ13C(COc) value of -(47.2 ± 5.8) ‰ inferred at R2 above 0.4 is characteristic for the δ13C of methane in the UT/LMS. It is important to note that we have accounted for the biases in the analysed C1 WAS δ13C(CO) expected from the mass-independent isotope composition of O3 (see details in Appendix B).

We derive the “best-guess” estimate of the admixed CO 18O signature at δ18O(COc) = +(92.0 ± 8.3) ‰, which agrees with the other MM results obtained at R2 above 0.75. Taking the same subsets of samples, the concomitant 13C signature matches δ13C(COc) = -(23.3 ± 8.6) ‰, indeed at the upper end of the expected LMS δ13C(CO) variations of -(25-31) ‰. Because of that, the MM is likely insensitive to the changes in δ13C(CO) caused by the contamination (the corresponding R2 values are below 0.1). Upon the correction using the inferred δ18O(COc) value, the C1 WAS δ18O(CO) data agree with B96 (shown with red symbols in Fig. 3). That is, variations in the observed C18O are driven by (i) the seasonal/regional changes in the composition of tropospheric air and by (ii) the degree of mixing or replacement of the latter with the stratospheric component that is less variable in 18O. This is seen as stretching of the scattered tropospheric values ([CO] above 60 nmol mol-1) towards δ18O(CO) of around -10 ‰ at [CO] of 25 nmol mol-1, respectively. The corrected C1 δ13C(CO) data (shown in Fig.7) are found to be in a ±1 ‰ agreement with the observations by B96, except for several deep stratospheric samples ([CO] below 40 nmol mol-1). The latter were encountered during “ozone hole” conditions and carried extremely low δ13C(CO) values, which was attributed to the reaction of methane with available free Cl radicals (Brenninkmeijer et al., 1996).

The contamination 18O signature inferred here (δ18O(COc) = +(92.0 ± 8.3) ‰) likely pertains to O3 and is comparable to δ18O(O3) values measured in the stratosphere at temperatures about 30 K lower than those encountered in the UT/LMS by C1 (see Table 1 for comparison). If no other factors are involved (see below), this discrepancy in δ18O(O3) should be attributed to the local conditions, i.e. the higher pressures (typically 240–270 hPa for C1 cruising altitudes) at which O3 was formed. Indeed, the molecular lifetime (the period through which the species' isotope reservoir becomes entirely renewed, as opposed to the “bulk” lifetime) of O3 encountered along the C1 flight routes is estimated on the order of minutes to hours at daylight (H. Riede, Max Planck Institute for Chemistry, 2010), thus the isotope composition of the photochemically regenerated O3 resets quickly according to the local conditions. Virtual absence of sinks, in turn, leads to “freezing” of the δ18O(O3) value during night in the UT/LMS. Verifying the current δ18O(O3) estimate against the kinetic data, in contrast to the stratospheric cases, is problematic. The laboratory studies on O3 formation to date have scrutinised the concomitant kinetic isotope effects (KIEs) as a function of temperature at only low pressures (67 mbar); the attenuation of the KIEs with increasing pressure was studied only at room temperatures (see Table 1, also Brenninkmeijer et al. (2003) for references). A rather crude attempt may be undertaken by assuming that the formation KIEs become attenuated at higher pressures in a similar (proportional) fashion to that measured at 320 K, however applied to the nominal low-pressure values reckoned at (220–230) K. A decrease in δ18O(O3) of about (6–8) ‰ is expected from such calculation (cf. last row in Table 1), yet accounting for a mere one-half of the (13–15) ‰ discrepancy between the stratospheric δ18O(O3) values and δ18O(COc).

Lower δ18O(COc) values could result from possible isotope fractionation accompanying the production of the artefact CO. Although not quantifiable here, oxygen KIEs in the O3 → CO conversion chain cannot be ruled out, recalling that the intermediate reaction steps are not identifiable and the artefact CO represents at most 4 % of all O3 molecules. Furthermore, the yield λO3 of CO from O3 may be lower than unity (see details in Appendix A). On the other hand, the inference that the contamination strength primarily depends on [O3] indicates that the kinetic fractionation may have a greater effect on the carbon isotope ratios of the artefact CO produced (the δ13C(COc) values) in contrast to the oxygen ones. That is because all reactive oxygen available from O3 becomes converted to CO, whilst the concomitant carbon atoms are drawn from a virtually unlimited pool whose apparent isotope composition is altered by the magnitude of the 13C KIEs.

Besides KIEs, selectivity in the transfer of O atoms from O3 to CO affects the resulting δ18O(COc) value. The terminal O atoms in O3 are enriched with respect to the molecular (bulk) O3 composition when the latter is above +70 ‰ in δ18O (Janssen, 2005; Bhattacharya et al., 2008), therefore an incorporation of only central O atoms into the artefact CO molecules should result in a reduced apparent δ18O(COc) value. Such exclusive selection is, however, less likely from the kinetic standpoint and was not observed in available laboratory studies (see Savarino et al. (2008) for a review). For instance, Röckmann et al. (1998a) established the evidence of direct O transfer from O3 to the CO produced in alkene ozonolysis. A reanalysis of their results (in light of findings of Bhattacharya et al. (2008)) suggests that usually the terminal atoms of the O3 molecule become transferred (their ratio over the central ones changes from the bulk 2:1 to 1:0 for various species). Considering the alternatives of the O transfer in our case (listed additionally in Table 1), the equiprobable incorporation of the terminal and central O3 atoms into CO should result in the δ18O(O3) value in agreement with the “crude” estimate based on laboratory data given above.

Furthermore, the conditions that supported the reaction of O3 (or its derivatives) followed by the production of CO are vague. A few hypotheses ought to be scrutinised here. First, a fast O3 → CO conversion must have occurred, owing to short (i.e. fraction of a second) exposure time of the probed air to the contamination. Accounting for the typical C1 air sampling conditions (these are as follows: sampled air pressure of 240–270 hPa and temperature of 220–235 K outboard to 275–300 K inboard, sampling rate of 12.85 × 10-3 mol s-1 corresponding to 350 L STP sampled in 1200 s, inlet/tubing volume gauged to yield exposure times of 0.01 to 0.1 s due to variable air intake rate, [O3] of 600 nmol mol-1), the overall reaction rate coefficient (kc in Eq. (A3) from Appendix A) must be on the order of (6 × 10-15/τc) molecules-1 cm3, where τc is the exposure time. Assuming the case of a gas-phase CO production from a recombining O3 derivative and an unknown carbonaceous compound X, the reaction rate coefficient for the latter (k in Eq. (A2) in Appendix A) must be unrealistically high, at least 6 × 10-10 molec-1 cm3 s-1 over τc = 1/100 s. This number decreases proportionally with growing τc and [X], if we take less strict exposure conditions. Nonetheless, in order to provide the amounts of artefact CO we detect, a minimum mixing ratio of 20 nmol mol-1 (or up to 4 µg of C per flight) of X is required, which is not available in the UT/LMS from the species readily undergoing ozonolysis, e.g. alkenes.

Second, a more complex heterogeneous chemistry on the inner surface of the inlet or supplying tubing may be involved. Such can be the tracers' surface adsorption, (catalytic) decomposition of O3 and its reaction with organics or with surface carbon that also may lead to the production of CO (Oyama, 2000). Evidence exists for the dissociative adsorption of O3 on the surfaces with subsequent production of the reactive atomic oxygen species (see, e.g. Li et al., 1998, also Oyama, 2000). It is probable that sufficient amounts of organics have remained on the walls of the sampling line, exposed to highly polluted tropospheric air, to be later broken down by the products of the heterogeneous decomposition of the ample stratospheric O3. Unfortunately, the scope for a detailed quantification of intricate surface effects in the C1 CO contamination problem is very limited.