A total of seven MLF and seven Tedlar bags were prepared, with dry air (<0.003% H2O) from a cylinder; the mole fractions of CO2, CH4 and CO are listed in Table . Since the mole fractions of methane are significantly lower in the stratosphere than a typical value of around 2000 ppb in the free troposphere e.g., we prepared two samples (nos. 6 and 7) with low mole fractions (∼120 ppm CO2, ∼600ppb CH4 and ∼75ppb CO) by diluting air from a cylinder with N2. The CO mole fractions in the stratosphere are also low, but we did not make any storage test for samples with a mole fraction lower than ∼75ppb.

Directly after sample preparation, the air sample is analysed for CO2, CH4, CO and H2O mole fractions on a cavity ring-down spectrometer (CRDS, Picarro Inc., model G2401-m). During an actual balloon flight, it usually takes 3–5 h from sampling until the samples are retrieved and brought back to the laboratory for analysis. Therefore, we have chosen a period of 4 h as the storage time to represent this time delay; i.e. the bags are stored under laboratory conditions (∼ 20 ∘C, ∼1000hPa, ambient mole fractions of CO2, CH4, CO and H2O) for 4 h before they are analysed again. The 4 h drift during storage is defined as the difference between the measurement after 4 h of storage and the initial measurement: [X]4hours-[X]direct, where [X] is the measured dry mole fraction.

Previous studies show that the material of Tedlar bags is prone to water vapour diffusion , which leads to humidified air samples after 4 h of storage. H2O measurements are used to obtain dry mole fraction of CO2, CH4 and CO using the water vapour corrections described in and , before assessing drift of these species.

The difference between the measured mole fractions after 4 h and those measured immediately after filling are shown in Fig. , which captures the drift over 4 h of storage. The drift in CO2 after 4 h is comparable for both types of sampling bags for sample nos. 1–5, within a range of -0.2–0.2 ppm (Fig. a). Low mole fractions of CO2, i.e. samples nos. 6 and 7, are less stable in both types of sampling bags; however, these low mole fractions are not observed in the stratosphere, and hence the drift observed for samples nos. 1–5 is more representative than that observed for samples nos. 6 and 7 for the storage of stratospheric air samples. The CH4 mole fractions are preserved within the range of ±2ppb for all cases for both types of sampling bags (Fig. b). Although the Tedlar bags perform slightly better than the MLF bags, both are satisfactory for CH4 measurements when considering its large gradient in the stratosphere (500–2000 ppb, e.g. ).

The CO mole fractions appear to be stable in the MLF bags, with no clear indication of drift, independent of the mole fractions (Fig. c). The variability of CO differences may be in large part due to the repeatability of the CRDS analyser (1σ 7 ppb). In contrast, the CO mole fractions decrease in the Tedlar bags, coupled with a significant increase of water vapour mole fractions of up to ∼1% (Fig. d), which is due to the high permeability of the bag material to water vapour and has been observed in previous studies . The increase of water vapour mole fractions in the MLF bags is only up to 0.01 %. The observed decrease of CO mole fractions in the Tedlar bags, even when its mole fractions are lower than the ambient, cannot be explained by the permeability of the bag material, as diffusion would increase the CO mole fractions. Although mole fractions are corrected for water vapour to obtain dry mole fractions, we cannot exclude that there is still some remaining bias from the water vapour correction function, and this correction function was not tested before with low mole fractions. This would not affect the depicted results for the MLF bags since the water vapour content remains low. Further investigation is needed before Tedlar bags are used to collect samples for analysis of high-precision CO mole fractions at the ambient level.

We found out that it is necessary to precondition the MLF bags before use, because we observed a positive offset of ∼12ppm CO2, ∼8ppb CH4 and ∼30ppb CO between the immediately analysed results of un-preconditioned MLF bag samples after filling and the assigned cylinder values. This contamination issue could be overcome by preconditioning the bags with N2. The bags were filled with N2 from a cylinder and subsequently evacuated with a vacuum pump, prior to filling with test sample. In principle ambient air can be used to flush the bags, as long as it is dry.

Based on the storage test results, we choose to use the MLF bags for our sampler. The stability of the CO2 and CH4 mole fractions in both MLF and Tedlar bags is comparable; however, the observed CO mole fractions in Tedlar bags is less stable than those in MLF bags. In addition, the permeability of water vapour through the Tedlar bags causes a significant increase of water vapour, which may affect the isotopic composition, e.g. δ18O in CO2. Moreover, the aluminium layers of the MLF bag protect the air samples against radiation that could affect the stability of CO mole fractions. We emphasize the importance of preconditioning the MLF bags before use.

We estimate the measurement uncertainty based on the laboratory storage test results and the uncertainties associated with the sample analysis. The total uncertainty of CO2, CH4 and CO mole fraction measurements consists of three terms: sampling error, drift due to storage and analysis uncertainty. There are two contributions to the analysis uncertainty: (1) analyser precision (σi) and (2) calibration uncertainty (σc).

The sampling error encompasses any contamination introduced by the sampling system itself. This includes chemical production of the species of interest and residual air in any dead volumes of the manifold. The chemical production during sampling is likely to be very small for two reasons. First of all, the wetted surfaces, Kynar and EPDM diaphragm, are chemically inert. Secondly, the high flow rate minimizes exposure of the sample to materials used in the sampler and hence chemical interaction with the wetted surfaces is limited. In addition, the flushing procedure with high flow rates ensures multiple turnovers of the manifold, which reduces the surface effects on the sample. These effects are thus assumed to have no influence on the CO2, CH4 and CO mole fractions.

The dead volume in the tube connecting the bag to the manifold is a potential source of contamination bias. The dead volume is estimated to be 1.5 mL per sample and will be at local ambient pressure prior to sampling. The dead volume uncertainty, σv, is estimated using a dead volume of 1.5 mLstp, which prior to sampling is assumed to be at 50 hPa and 220 K. This volume might remain unflushed. Hence the air is of tropospheric origin, with concentrations of 400 ppm CO2, 1800 ppb CH4 and 150 ppb CO. The total sample volume is 200 mLstp and has mole fractions of 395 ppm CO2, 500 ppb CH4 and 30 ppb CO. The bias is then calculated as the resulting deviation after mixing the contamination with sample air.

The main factor likely to affect mole fraction measurements of the stratospheric air samples is the drift in the sampling bags, an effect that has been quantified in Sects.  and . In principle, one could correct for the drift as a systematic error. The drift, a consequence of diffusion through the bag's material, is governed by Fick's first and second law. A systematic correction for the drift would require the determination of the species-dependent diffusivity, which usually depends on pressure and temperature. Secondly, since diffusion depends on the concentration gradient, mole fractions of the sampler environment are needed. Hence, detailed information about storage conditions over the 3–5 h period between sampling and analysis is required to correct for the drift. Moreover, the information required is usually unavailable in the field. Therefore, we have not determined a correction function for the drift, but rather use the maximum observed drift as an estimate of the drift uncertainty, σd. The maximum observed drift in these tests was 0.11 ppm, 2 ppb and 2.7 ppb for CO2, CH4 and CO, respectively.

Assuming Gaussian error propagation, we compute a total uncertainty on the measurements: σs=σd2+σi2+σc2+σv2. The total uncertainty includes the analyser precision (1σ 0.04 ppm, 0.2 ppb and 7 ppb for CO2, CH4 and CO, respectively), the calibration uncertainty (0.07 ppm, 1 ppb, 2 ppb for CO2, CH4 and CO, respectively), and the aforementioned drift. The different uncertainties are summarized in Table . We compute the total uncertainty of the LISA sample measurements to be 0.14 ppm, 2.3 ppb and 7.8 ppb for CO2, CH4 and CO, respectively.

During sampling of each bag, the atmospheric pressure decreases as the payload ascends, and the volume flow rate drops due to a nonlinear increase of pressure in the bag. Therefore, not all atmospheric pressure levels contribute equally to the collected sample in size or mole fractions of trace gases. The integrated sample thus has an associated pressure-weighted mean altitude. The contribution of each pressure level to one sample is proportional to the number of moles of air sampled at that pressure level. In general, the first 19.7 s of sampling contributes the most and the end of sampling contributes the least to the collected sample. When pressure and temperature within the manifold are measured, the number of moles of air at each pressure level can be computed directly, and the weight of that pressure level will be wi=dnin, where dni is the number of moles of air sampled at the pressure level pi, n is the total number of moles of collected air samples. wi is then the weight of the air samples collected at the pressure level pi. The pressure-weighted mean pressure altitude p‾ can be calculated as follows: p‾=Σwipi. The first 19.7 s of sampling cannot be calculated directly, since dpi=0; i.e. no compression and the pressure inside the bag is the same as outside. We assign the first weight that can be calculated to the first 19.7 s of sampling.

The temperature of the air samples in the bag was not directly measured. For the calculation of the pressure-weighted mean sampling pressure of the samples, we assume constant temperature of the sampled air while sampling. In reality, the temperature of the air samples in the bag would be close to the ambient temperature as the bag is directly exposed to the ambient air. Since the observed variability of the ambient temperature during sampling is usually less than 1 K (1σ), the assumption of constant temperature during sampling causes insignificant uncertainty on the weighted mean sampling pressure.

Both the volume of the collected air samples (Fig. a) and vertical resolution decreases (Fig. b) with increasing altitude. The sample size achieved by the sampler is close to that estimated based on the empirical model shown in Sect. . The variability of the collected sample size can be mostly explained by the different settings for the sampling time and the maximum allowed pressure during different flights (see Table ). Furthermore, the cold temperatures in the stratosphere result in denser air, so the observed sample sizes are slightly higher, especially in the lower stratosphere.

The variability in the vertical resolution is the result of three factors: (1) varying sampling time; (2) varying ascending speed; (3) varying maximum allowed pressure. The ascending speed was typically around 7–9 m s-1 in the lower stratosphere and decreased to 4–5 m s-1 in the middle stratosphere. The varying ascending speed accounts for the observed deviations from the otherwise linear trend in Fig. b. In the lower stratosphere 10 to 15 km the maximum allowed pressure inside the bags was usually reached in a period shorter than the preset sampling time, leading to relatively high vertical resolution. In the region 10 to 15 km two samples deviate (4 September, 17 km and 26 April, 14 km), with lower resolution, which is due to a higher ascending speed. In the middle stratosphere, the sampling time was usually the limiting factor to vertical resolution. One sample in the middle stratosphere has a relatively good vertical resolution (5 September, 25 km), which is due to the relatively slow ascent speed.

While the sampler was still collecting the last sample during the flight on 6 September 2017, the balloon burst at a lower altitude (21.4 km) than previous flights. The vertical resolution of that particular sample was estimated to be 3.1 km (not shown), a number much larger than that of other samples due to the fast descending speed of 16.8 m s-1 after burst.

The vertical profiles of CO2, CH4 and CO mole fractions from both AirCore and LISA measurements are presented in Fig. . For the retrieval of the AirCore profiles we refer to Chen et al. (2018). AirCore and LISA measurements are compared based on the same pressure level. For a fair comparison, we averaged the AirCore profiles with the same weights that are used to calculate the weighted mean sampling pressure of the samples. The mean differences between LISA and AirCore measurements of CO2, CH4 and CO mole fractions are summarized in Table .

A relatively large difference in the CO2 mole fractions (>1ppm) between LISA and AirCore is clearly visible for the flight on 26 April 2017. The observed difference is much larger than the uncertainty caused by the drift of CO2 mole fractions due to storage in the MLF bag (shown in Fig. a) and cannot be explained by any known reasons. The differences of the CO2 mole fractions for other flights are significantly smaller. The summertime stratosphere is only affected by weak diabatic stirring , and can be considered relatively stable. Therefore, the flights on 4–6 September 2017 can be, to a large extent, considered duplicate measurements. This is supported by the excellent agreement between the AirCore profiles of CO2 and CH4 mole fractions measured on those dates. The AirCore data logger failure on 26 April 2017 may cause increased uncertainty in the altitude registration of the AirCore measurements, whereas the malfunction of the LISA data logger during the same flight may cause increased uncertainty in the weighted mean sampling pressure of the samples. Therefore, we also calculated the mean differences excluding the flight on 26 April 2017, which decreases the mean difference in CO2, but slightly increases the difference in CO.

CO mole fractions agree well during all flights, except that a small decrease with altitude was observed by LISA measurements in September 2017, but not captured by AirCore measurements. A good agreement between AirCore and LISA CO measurements is found for the flight on 26 April 2017. Besides this, an interesting CO plume at 13.5 km is observed by both AirCore and LISA during the flight on 5 September 2017.

The deviation between AirCore and sampler results is on average 0.84 ppm for CO2, a result that is comparable in magnitude to AirCore inter-comparisons e.g.. For methane, we find a mean deviation of 1.8 ppb, within the uncertainties of both AirCore and LISA (see Sect. ). CO also shows a good agreement within the measurement uncertainty of CO by the CRDS analyser. Several aspects are considered that could explain the observed differences. First of all, the altitude registration of AirCore measurements is associated with uncertainties, as outlined by , especially due to the manual selection of the start and the end of AirCore sample analysis or any potential loss of air samples in case of valve malfunction, which complicates the comparison between AirCore and the sampler. Secondly, there are uncertainties associated with the calculation of AirCore pressure-weighted mean and mole fractions. The AirCore profile needs to be weighted for a fair comparison, since air samples at different altitudes (or pressure levels) do not contribute equally to the sampler samples. The uncertainty in altitude of the AirCore profile adds a level of uncertainty to the AirCore pressure-weighted mean. Finally, the retrieved AirCore profiles are already smoothed due to molecular diffusion and Taylor dispersion, and smearing effects in sample renewal of the cavity of the CRDS. For more information on the uncertainties associated with AirCore profiles we refer to , , and .

In the first flight on 26 April, an averaged difference of ∼1ppm in CO2 is observed, which cannot be explained by the associated uncertainties or by smoothing of the AirCore profile due to diffusion. The samples were taken at a distance no more than 1.5 km apart (determined from Vaisala GPS-data) in the stratosphere, with less than 1.5 h in between LISA and AirCore sampling. Such large horizontal mole fraction gradients are not expected in the stratosphere, although stratospheric dynamics in winter show a higher degree of variability in measured trace gases. The AirCore valve did not close during the 26 April 2017 flight. This complicates the altitude registration of the AirCore, especially on the lowest part of the profile. The influence on the stratospheric part of the profile is limited, which can be seen from the large degree of agreement in CH4 profiles between AirCore and LISA.

Another potential source for the bias is out-gassing from the packaging material and the balloon. As the ambient pressure decreases during ascent of the balloon flight, the desorption of trace gases from the surface of the packaging material and the balloon occurs, which potentially influences the mole fractions of the air samples. This would, however, not explain the good agreement during the September flights. Furthermore, the inlet is located at the top of the payload and any out-gassing from the packaging material would be flushed away from the inlet during ascent.

The seasonality in tropospheric CO2 that causes the difference between sampled air and its storage environment could contribute to the observed difference. The Northern Hemisphere winter CO2 mole fractions are typically 10 ppm higher those in summer. During the storage test with low mole fractions, e.g. sample nos. 6 and 7 in Fig. 2, a drift of up to 0.8 ppm was observed. Therefore, a typical seasonal difference of 10 ppm could only explain a difference of 0.03 ppm in the observed CO2 bias.

As seen from Table , the mean deviation for CO2 reduces from 0.84 to 0.55 ppm, when the 26 April 2017 flight is excluded. Still, the sampler shows consistently higher CO2 mole fractions than AirCore, which suggests that a small unexplained bias might exist in the LISA CO2 mole fraction measurements compared to AirCore. CH4 and CO on the other hand show an excellent agreement within measurement uncertainties, which suggests that no significant bias exists within the measurement uncertainties for CH4 and CO.

The vertical resolution of the collected stratospheric air samples ranges from 500 to 1500 m, and the sample size ranges from 800 to 180 mLstp. The vertical resolution and sample size thus outperform the AirCore. The vertical resolution and sample size of the LISA sampler is compared to the performance of AirCore sub-sampled air in Table . Its shows that sampler outperforms the sub-sample method described in and .

As mentioned previously, the vertical resolution depends on the ascending speed and the effective sampling time, and the sample size also depends on the effective sampling time. To this end, the effort in collecting more air samples by increasing the effective sampling time will compromise the vertical resolution. The vertical resolution can be improved by lowering the ascending speed and decreasing the sampling time. The pump works most effectively when the pressure difference across it is minimal. From the results shown in Fig. d, we see that after 200 s of sampling, the gain in sample amount decreases quickly. Therefore, the gain in sample amount, for example adding 10 extra seconds of sampling time, is small; however, the decrease in vertical resolution is significant.

During the experiments described in Sect. , the pump was at room temperature. The pump performance could be affected by the cold environment. First, the batteries could lose capacity and cause the power supplied to the pump to decrease. The temperature inside the thermally insulated package, where the pump is located, during flight ranged between 30 and -15 ∘C. Secondly, the diaphragm is exposed to the cold air passing through the pump. The elasticity of a rubber is temperature dependent, which could reduce the performance of the pump. On the other hand, heat is released during operation of the pump, which increases the temperature. Finally, the effect of air temperature on sample size follows the ideal gas law, and the sample size increases at low temperatures. As no experimental data are available to determine the performance of the pump at stratospheric temperature, we assume that the pump performs the same during flight as at room temperature for the calculation of the sample size.

The sample size can be increased by using an alternative pump that can deliver a higher flow rate than the current 8 Lstp min-1 and using sampling bags with a larger size than the current 2.58 L. It will be mostly practical to increase the size of the sampling bag because this does not add significant weight or power consumption. An alternative more powerful pump could potentially increase the sample size, especially for the samples from high altitudes; however, it would also likely add more weight and consume more power that in turn increases the weight due to the need for more batteries.

Alternatively, to increase the amount of sample retrieved during one flight, additional bags can be considered. Currently the system is idle during several stages of the ascent as can be inferred from Fig. . This will however be more demanding on battery power. Furthermore, care has to be taken to avoid overlapping sampling schemes; i.e. sampling of a sample at altitude P1 is still ongoing while the set-point altitude for sample two, P2, is reached. This is complicated further with variable ascent speed, which is typical for these balloon flights.

The accuracy in sounding of the Vaisala RS92-SGP pressure sensors is 1 hPa respectively at 200 hPa, and 0.6 hPa in the range 3–100 hPa . The uncertainty of RS92-SGP pressure altitude is discussed in detail by . The uncertainty of the vertical position of the RS92-SGP radiosonde is 20 m and hence is also the uncertainty of the pressure-weighted mean altitude. Since the vertical resolution is calculated as the altitude where sampling stops minus the altitude at which sampling starts, the uncertainty in vertical resolution of the sampler is 29 m, calculated using Gaussian error propagation.

The uncertainty in the estimated sample amount is a result of the uncertainty in pressure and temperature measurements. The pressure sensor in the manifold has an uncertainty of 2.6 hPa whereas the radiosonde temperature measurements have a 1σ uncertainty of 0.25 K in sounding . This results in an error in estimated sample amount of 7.6 mLstp.

Since the manifold pressure was not logged during the flight on 26 April, the pressure-weighted mean altitude of the samples had to be estimated using Eqs. () and (). As mentioned earlier, Eq. () assumes a constant upstream pressure, which is not the case during flight where the pressure decreases. This results in errors in both the estimated sample amount and the estimated sampling altitude. The error in the fit parameters is included in the calculation. The error in pb can be calculated using the uncertainties of the fit (Eq. ) and the pressure and measurements can be calculated using standard error propagation. The error in sounding of the pressure sensors is 1 hPa at 200 hPa . The total uncertainty after 200 s of sampling is 9 mLstp, slightly higher than the effect found above.

The stratospheric air samples can be used for analysis of isotopic composition measurements of trace gases. Here we take CO2 and CH4 as an example to estimate the uncertainties of isotopic composition measurements due to the storage bias (see Table ) or the LISA–AirCore bias (see Table ), and the estimated isotopic signatures associated with the assumed contamination source.

The measured mole fraction (Cm) is the sum of the mole fraction of the original stratospheric air sample (Cs) and that of the contaminated air sample (Cc) in the bag, weighted with their respective contributions: Cm=fsCs+fcCc, with fc+fs=1 and solving for fc yields, fc=Cm-CsCc-Cs=biasCc-Cs, given that the Cm-Cs is the observed bias. The isotope composition after the mixing of the tropospheric contamination into the sample air, can be approximated with δm≈δsfs+δcfc, where δm is the final isotopic composition, and δs and δc represent the isotope composition of source and contamination and fs and fc are the fractional contributions to the total number of molecules after mixing. We further define the bias of the isotopic composition measurement as follows: Δδ=δm-δs. Combining Eqs. () and (), and using fc+fs=1 again, we derive Δδ=(δc-δs)fc. For the calculation, we regard the mean differences between AirCore and LISA measurements (Table , e.g. 0.84 ppm for CO2 and 1.8 ppb for CH4) as the upper limit of bias induced in the stratospheric samples. Another estimate is performed based on the storage test results that showed maximum drifts of 0.11 ppm for CO2 and 2 ppb for CH4, presented in Table .

The fraction fc can be calculated according to Eq. () with Cs being the typical stratospheric mole fraction, which is taken to be 395 ppm for CO2 and 500 ppb for CH4. For the contaminated air, we use typical tropospheric values of 405 ppm for the CO2 mole fraction and 1800 ppb for the CH4 mole fraction. The isotopic compositions δc and δs are taken from various references, and are presented in Table .

We can readily see that the estimated biases due to the storage are relatively small compared to the typical analytical precisions, also presented in Table . The estimated biases in stable isotope measurements based on the observed differences between AirCore and LISA may be significant in certain cases, but should be considered an upper estimate, since these are based on maximum differences between the troposphere–stratosphere mole fractions and isotopic composition. Hence the LISA sampler provides a viable sampling tool for useful measurements of stable isotopes in CO2 and CH4.