In our experiment, an Isoprime mass spectrometer (GV Instruments, Manchester, United Kingdom) is used with an accelerating voltage of 4.1 kV to measure the isotopic abundance in O₂. Ion currents are collected continuously by Faraday cups for 32, 33, and 34 m/z with resistors being set at 5 × 10^–8, 5 × 10^–10, and 1 × 10^–8 Ω, respectively. With these settings, we found an overall ionization efficiency of our mass spectrometer of 0.5‰, corresponding to the detection of a single O₂ molecule out of 2,000 molecules entering the mass spectrometer. The number of detected O₂ molecules can only be increased at the expense of overall precision. Sample and reference gas were allowed to pass through a single open split. We used an open split closed at one end, i.e., a glass tube with a length of 7.3 cm having inner and outer diameters of 1 and 2 mm, respectively. Since fractionations are associated with open split settings (Elsig and Leuenberger, 2010), care must be taken with the correct referencing as well as with the protection against contamination with laboratory air. Five fused silica capillaries with varying diameter were used: (i) the reference capillary, attached to a movable piston, allows admission of a working standard (pure O₂); (ii) the background capillary is set to a fixed position, with a constant flow of a working standard of known isotopic composition in order to minimize adsorption effects; (iii) the He capillary, a high flow of helium enters the open split, acting as a dynamic seal to the laboratory air; (iv) the mass spectrometer capillary, connected to the mass spectrometer inlet valve; (v) the sample capillary, eluted O₂ signals from H₂O conversion are transferred through it. Only the reference capillary can be lifted and lowered within the open split. The dimensions and fluxes of all capillaries used in our experiment are summarized in Supplementary Table 1.

The setup used for this study is shown in Figure 1, manual or autosampler (H/Device) injections (2 μl H₂O) are possible. Prior to injection into a purified (carbagas, purity, 99.999%) helium flow through a high temperature septum (BGB analytic, part no SM11), the syringe (Hamilton, part no 80300-00) is flushed two times with 5 μL of the water sample to minimize the memory effect. The helium flow is maintained at a constant level by a mass flow controller (Analyt-MTC, model no 35827). Sample and helium lines are attached to the injection port (Swagelok, T-union). The reaction of H₂O with CoF₃ takes place in a nickel tube (0.250″ outer diameter (hereafter o.d.) × 0.056″ inner diameter (hereafter, i.d.) of 5.4″ in length). The upper and the lower end of the nickel tube are attached to the 1.7″ (0.250″ o.d × 0.056″ i.d.) and 2.2″ (0.250″ o.d × 0.056″ i.d.) length of a stainless steel tube through connectors (Swagelok. part no SS-400-6), respectively; 10 μm of steel frit (Grace, port no 703125) is attached to the lower end of the upper steel tube for the convenient filling of a new batch of CoF₃ to the inside of the nickel tube. Nickel wool (Säntis analytical) is placed at ∼2 cm above the bottom end of the nickel tube after filling CoF₃ (approximately 3 g) and Ni grains in such a way that the reagents are ideally placed to reach a constant temperature of 643 K (370 °C) in the reacting furnace. The HF produced as a byproduct of the reaction is removed by a NaF trap (Franz and Röckmann, 2004) by the following reaction:

The NaF trap consists of a 6 cm, 0.250″ o.d. 0.056″ i.d. nickel tube with a 10 μm steel frit at the front end and nickel wool at the rear end. It is completely filled with NaF (approximately 170 mg). A subsequent liquid nitrogen trap (coiled steel tube with ∼ 70 cm length 0.250″ o.d.) acts as a backup trap to remove any condensable left in the gas stream for safe venting. To minimize the chances of any condensable escaping from the liquid N₂ trap, a 7 μm filter (Swagelok, part no SS-F-7) is additionally attached with a 0.125″ o.d. steel tube to prevent fused silica capillaries. A manually operated 4-port valve (Swagelok, part no SS43Y6FS2) allows switching between waste and measuring mode. It has a sample inlet and outlet as well as waste and helium flushing port, the latter have a pressure gauge (Kistler) and a Pfeiffer vacuum pump attached (0.125″ o.d. steel tube) to read the pressure of the produced O₂ signal after the chemical reaction and to evacuate the sample line, respectively. A homemade gas coupling system (open split) is attached to the 4-port sample outlet by a 0.0265″ o.d. steel tube. The helium capillary leading to the open split is set to a flow rate of 6.25 ml/min while the background capillary is controlled through a mass flow controller at a flow rate of 0.2 ml/min. The reference capillary is set to 7.5 ml/min while the sample capillary is controlled through another mass flow controller at a flow rate of 15 ml/min for an improved Gaussian type sample peak. Only the reference capillary is attached to a movable piston to admit pulses of the working standard.

The IonVantage software of the Isoprime mass spectrometer allows us to store the signals for beams 32, 33, and 34 m/z with a time step of 0.1 s. This raw data are passed to a template compiled by us that largely follows the steps described by Bock et al. (2010). Individual background correction and drift correction are considered in our template to improve the dataset.

The peak shape observed in the continuous flow system is not truly rectangular or Gaussian due to the rapidly moving capillary and the sticking nature of the H₂O molecule inside the tube (Franz and Röckmann, 2004), respectively. For the background correction of 32, 33, and 34 m/z, values are always taken before the beginning of each rectangular as well as the Gaussian peak. We observed that the signal depends on the determination of the peak start. Bock et al. (2010) believe that this observation relates to the fact that real peaks are not truly Gaussian and therefore the peak margins, in particular, are isotopically inhomogeneous even after a time shift correction (Elsig, 2009). To sort out these problems, each peak is individually evaluated in the template. To select the start and endpoint for the calculation, raw mass/charge (m/z) data are processed to get their time derivatives [d(m/z)/dt]. Two different procedures were performed for the reference calculation. The first one (Supplementary Figure 2A) is a calculation of the reference peak from its start to end. The second is based on the calculation of the reference peak when the time derivative gets stable which is the case on the plateau of the rectangular reference peak (Supplementary Figure 2B). The whole peak and plateau results of our internal St-91 standards are shown in Table 1. A better precision was obtained for plateau evaluation which may indicate that fractionations at the peak start lead to the poorer precision. However, the second calculation type can’t be used for the sample peaks due to its Gaussian shape peak character. Start and endpoint for the peak calculations were chosen according to the time derivatives of the mass/charge ratio (Figure 3) and are valid for each individual run [d(m/z)_x/dt, where x corresponds to 32, 33, and 34].

We also have used an improved CO₂-H₂O equilibration method that measures the triple oxygen isotopes of water sample. The detailed technical setup and experimental design of this method are available in Elsig and Leuenberger (2008) and are briefly summarized in the following:

A CO₂-H₂O equilibrium is necessary for precise isotope ratio measurements in particular for oxygen-17 and Δ¹⁷O. This can be achieved through the admission of an isotopically known CO₂ gas and proper shaking of the sample for time-efficient equilibration. A scaling factor r is used to determine the delta value correction which summarizes effects such as cross contamination or mass spectrometry sensitivity. Care must be taken during the optimization of the scaling factor since stretching or compressing of the r factor can increase or decrease its value from unity. For precise Δ¹⁷O measurements through the equilibrium method, 5 ml of the water sample was taken, equilibrated with about 7–8 mL STP of CO₂ (185 ± 10 mbar of CO₂), and admitted to a glass flask of 45 cm³ at 20°C temperature. The sample pH was optimized (<4.3 pH units) because it is also one of the most critical factors for the Δ¹⁷O value for this method. Applying the equilibrium method for samples with a Δ¹⁷O non-zero signal, the δ¹⁸O value has to be corrected. Assuming a false δ¹³C value of the admitted CO₂ leads to wrong Δ¹⁷O and δ¹⁸O values which also requires correction. Therefore, to obtain better precision through this method, all measurements were performed under a strict pH regulation, optimum sample amount, and a well-determined scaling factor to limit cross contamination.

Samples were measured at our division Climate and Environmental Physics (CEP), the University of Bern. Measurements were performed using state-of-art Picarro L2140-i wavelength-scanned cavity ring-down spectroscopy (WS-CRDS) instrument (Picarro Inc., United States). Before an isotopic analysis, samples were transferred into 2 mL glass vials and sealed with rubber/aluminum caps to mount on a PAL COMBI-xt autosampler (CTC Analytics AG, Zwingen, Switzerland) that is connected to the Picarro L2140-i for δ¹⁷O, δ¹⁸O and δD analysis (Affolter et al., 2015). Six times injections were performed, and the first three injections were discarded to reduce the memory effect and obtained the reproducibility of typically 0.1‰ for δ¹⁸O and δ¹⁷O. The obtained precision for Δ¹⁷O is under 7 per meg.

To characterize potential influences of the carrier gas on the eluted O₂ sample signal, we performed an experiment where a 2 μl H₂O sample was allowed to react with CoF₃ and pass through the cryogenic cool trap without any carrier gas (Supplementary Figure 3). A pressure gauge was attached to the sample line to record the amount of O₂ as a function of time; an increase of 53.3%, 17%, 14.8%, 7.7%, and 7.3% in the eluted O₂ signal (pressure percentage) was observed at the time interval of 0–10, 10–20, 20–40, 40–60, and 60–90 min, respectively. Two-third of the sample signal were obtained within 30 min from the sample injection, while a small percentage of the sample was still obtained at the end of measurements (after 90 min) which indicates the sticky nature of water that can lead to fractionation effects within our system line and hence reduce precision. These results document the importance of using a carrier gas in an online system to obtain a Gaussian shape sample signal that significantly helps to minimize fractionation effects and reducing measurement time. The different behavior of the reactor composition on the 32 m/z ratio is shown in Figure 4. Multiple tests indicated that the carrier gas flow was dependent on the reactor composition. To maintain constant flow conditions, these experiments strongly suggest to use a mass flow controller in the carrier gas line (see Figure 1), since the reactor filled with CoF₃ and metallic frit at the bottom of the nickel tube generates high pressure inside the reactor that is most probably due to compaction of the reactor. The byproduct gas HF (HF is acidic and produced as a byproduct, see eq. 5) was able to back diffuse onto the injector and to react with the septum resulting in its damage. Hence, the lab air contamination probability increased which could easily be checked via N₂ detection in the IRMS. To solve this problem, we used nickel grains after CoF₃ chemical filling. Ni grains are non-reactive with CoF₃ and water. Blocking of the pores of the steel frit by CoF₃ particles resulted in a significant lowering of the carrier gas flow as observed in many of our experiments. To sort out this problem, Ni grains which have a bigger grain size compared to CoF₃ were placed at the bottom before CoF₃ fillings for each experiment.

In the case of Figure 4i, 2 micron steel frits and Ni grains were used at the top and the bottom, respectively, while a <2 g of CoF₃ amount was filled in between. However, a sample peak with a long tail was observed. This might be due to the sticking nature of water on the larger surface area of the nickel tube exposed to the H₂O molecule because of the lower CoF₃ filling (<3 g of CoF₃) resulting in a tailing of the eluted sample signal (Franz and Röckmann, 2004); 38 ml/min of the flow rate was measured at the end of the sample line. The signal obtained with this setup is higher in amplitude compared to other fillings but the long tailing of the eluted O₂ signal is the limiting factor for obtaining better precision.

In Figure 4ii, 2 micron steel frits were used at the top and the bottom and >3 g CoF₃ were filled together with Ni grains at the top, the middle, and the bottom of CoF₃. The reason for filling in more chemicals was to reduce the tailing effect of the eluted sample signal by reducing the surface area exposed to the H₂O molecule in the nickel tube. Nickel grains were also filled into the middle of the chemical to reduce compactness. However, no O₂ was eluted because of the high restriction (<1 ml/min of the flow rate) of the reactor most probably due to the increased chemical amount. Because of the restricted flow of helium through the reactor, reference signals (rectangular shape) with higher amplitude were observed due to a lower dilution effect by the very low carrier helium flow.

In Figure 4iii, nickel grains were mixed with CoF₃ and 2 micron steel frits were replaced with 5 micron steel frits to increase the flow rate of the carrier gas and hence to increase the amplitude of the eluted sample signal. The idea behind mixing CoF₃ with Ni grains was to allow the carrier gas along with the eluted O₂ signal to easily pass through the reactor. A flow rate >5 ml/min was measured in parallel to a lower amount of the eluted sample signal and higher amplitudes of reference signals which shows us that the compactness of the reactor remained high even though the size of the steel frits was increased. Obviously, the mixture restricts the flow contrary to expectation.

In Figure 4iv, 5 micron filters as well as nickel grains at the top and the bottom of the reactor were used, with smaller amounts of Ni grains at the bottom than at the top. Eluted Gaussian shape sample signals were recorded with 30 ml/min of the carrier gas flow rate. A wider sample signal is observed due to the insufficient flow of the carrier gas which is unable to push the sample through the reactor to obtain an optimal eluted signal. Although the eluted O₂ signal obtained from this setup is not bad, a further improvement of the peak form yielded better isotope results.

Therefore, in Figure 4v, 5 micron steel frits at the top and nickel wool instead of steel frit at the bottom were used in the nickel tube reactor to increase the flow of the carrier gas with an eluted signal and to avoid compactness of the reactor. A flow rate of 60 ml/min was obtained with an improved Gaussian shape eluted sample signal. We also tested with 10 micron filters (figure not shown) at the top and nickel wool at the bottom with the same amount of chemical filling; however, the eluted O₂ signal area and shape were similar to Figure 4v, setup. Therefore, we continued with the reactor setup of version (v). Based on all these experimental designs, 5 and 10 micron steel frits and nickel wool are a better option to obtain optimal eluted Gaussian shape sample signals and along with it an improved precision.

After optimizing the reactor setup for the eluted O₂ peak, the mass flow controller was connected to the sample injection line to carry a constant flow and multiple sample injections were performed afterward which is shown in Figure 2.

Measurements with the equilibration method have been performed on precipitated water sampled at the High Altitude Research Station Jungfraujoch. They are compared to the measurements on the same water samples with the laser-based system (Picarro L-2140i) as shown in Figure 5.

To see the consistency between CO₂-H₂O equilibration (EQUI) and LBS system, a comparison of the last three decades (1983–2011) was performed for precipitation samples originating from Jungfraujoch station (JFJ: altitude 3580 m.a.s.l.). Here, we are restricting us to discuss only a comparison between EQUI and LBS using JFJ precipitation records (Figure 5). The detailed descriptions about the JFJ precipitation records can be found in a companion manuscript (Leuenberger and Ranjan, 2021, under review in this issue). For the EQUI method, δ¹⁷O, δ¹⁸O and Δ¹⁷O exhibit precisions of 0.0470, 0.0194, and 0.0458‰, respectively, while LBS shows a precision of 0.0140, 0.0174, and 0.007‰ for δ¹⁷O, δ¹⁸O, and Δ¹⁷O, respectively. These indicate that the EQUI method documents limited precisions for δ¹⁷O and Δ¹⁷O compared to δ¹⁸O which are significantly deviating from precisions to be expected under ideal conditions as discussed in Elsig and Leuenberger, 2008. However, recent work by Affek and Barkan (2018) claimed to achieve the precision of EQUI method close to the offline fluorination method (Barkan and Luz, 2005) using hot platinum for an isotopic exchange between O₂ and CO₂. The excellent correlation between δ¹⁷O and δ¹⁸O (r² > 0.9919) is not sufficient for getting good Δ¹⁷O values as obtained with the laser-based system for which the same correlation is significantly stronger (r² > 0.999957). Furthermore, correlation among isotopic records from both systems indicates that δ¹⁷O (r² > 0.937, p < 0.05) and δ¹⁸O (r² > 0.945, p < 0.05) show a strong correlation while no correlation was found for Δ¹⁷O. The poor correlation exhibited by Δ¹⁷O indicates the inconsistency and poor precision from EQUI methods (see Figure 5) in particular for δ¹⁷O. It is interesting to note that the uncertainty of Δ¹⁷O correlates neither with δ¹⁷O nor with δ¹⁸O and its uncertainty but shows a positive correlation with the uncertainty in δ¹⁷O. Based on the sensitivities investigated by Elsig and Leuenberger, 2008, it might point to an influence of the temperature control as this is the only positive dependence they state; yet, in this case one also would expect a correlation with the uncertainty in δ¹⁸O. Additionally, the calculated offset of 36 per meg between LBS and EQUI methods cannot be taken as robust due to the high uncertainty of the EQUI method. Even increasing the number of injections for the EQUI method to match the precision of the EQUI method to LBS (Picarro L2140i) through statistics is not guaranteed due to the fact that we do not understand where the limitation is coming from.

The reported δ values for ConFlow and LBS are shown in Table 2 and this corresponds to a single peak of single acquisitions with respect to water standard (ORSMOW) (Winkler, 2012). In the case of ConFlow, the overall precisions from each acquisition for δ¹⁷O, δ¹⁸O, and Δ¹⁷O are 0.077‰, 0.075‰, and 93 per meg, respectively. The reported precision for δ¹⁷O and δ¹⁸O are under 0.1‰ for five peaks and it further improves to 0.050‰ and 0.074‰, respectively, by skipping the first peak to minimize the memory effect (Figure 6). Thus in four peaks, overall precision for Δ¹⁷O improved to 49 per meg, which is two times better than the previous one. By skipping the second and third peak, surprisingly the Δ¹⁷O precision improves by a factor of four and seven (Table 2) while δ¹⁷O improves to two and four times, respectively. A drastic improvement in precision by skipping initial peaks indicates that the H₂O sticking nature inside the reactor is a limiting factor. In the case of LBS, the precision obtained by five subsequent injections in a single run (n = 5) for δ¹⁷O, δ¹⁸O, and Δ¹⁷O are 0.020‰, 0.035‰, and 2 per meg only. The obtained precision for δ¹⁷O, δ¹⁸O, and Δ¹⁷O are, respectively, >three times, >two times and >forty times, better than ConFlow system. The Δ¹⁷O precision for LBS is much better than the ConFlow system and the latter can only improve by skipping the first three injections to minimize the memory effect and its associated fractionation. Correlation among δ¹⁷O and δ¹⁸O shows a tight correlation (r² > 0.979, p < 0.05) for continuous flow and laser-based system (r² > 0.961, p < 0.05), respectively (Figure 7). Furthermore, an offset of only 0.136, 0.002, and 0.136‰ were observed between δ¹⁷O, δ¹⁸O, and, Δ¹⁷O of both systems, respectively. By removing the first two peaks from ConFlow, the offset between ConFlow and LBS becomes insignificant that further confirms the influence of memory effect and its associated fractionation role in δ¹⁷O, δ¹⁸O, and Δ¹⁷O precision.

Coherency between continuous flow and laser-based systems were tested using correlation analysis among δ¹⁷O, δ¹⁸O, and Δ¹⁷O of both systems. δ¹⁷O, δ¹⁸O, and Δ¹⁷O exhibit a significant correlation of 0.680, 0.494, and 0.333, respectively. The observed δ¹⁷O and δ¹⁸O correlation further improve to 0.690 and 0.822 by skipping the first peak while correlation among Δ¹⁷O of both systems further degraded. These results indicate that δ¹⁸O of both (continuous flow and laser-based systems) systems are comparable but δ¹⁷O and Δ¹⁷O need further improvement to match the precision level of laser-based system. Further improvement of the continuous flow system is certainly possible through the optimization of the reactor, tubing replacement from a nickel to steel, optimizing the tubing length between the reactor and the detector (IRMS). Therefore, more than 15 injections with the exclusion of the first three peaks should be applied to match the precision of a laser-based system.