In an initial test phase, important characteristics of the analyser performance were determined in accordance with Harris et al. (2020) to define a strategy for real-time measurements at WWTPs. Optimal averaging times to reach adequate precision levels (i.e. 0.2 ‰) and drift effects were determined using the Allan variance technique (Allan, 1966; Werle et al., 1993). For this, a gas mixture with known N₂O isotopic composition (Cal1, see Table 1) was dynamically diluted with N₂O-free synthetic air (SA) to a mole fraction of 12 ppm N₂O and a total flow of 300 mL min⁻¹ and measurements were recorded for more than 3.5 d. The Allan precision for δ-values for 5 and 10 min integration time were 0.08 ‰–0.09 ‰ and 0.06 ‰–0.07 ‰, respectively (Fig. 1). Instrumental drifts over 24 h periods were estimated on basis of the Allan analysis and accounted for 0.26 ‰, 0.36 ‰, 0.05 ‰, 0.40 ‰, and 0.85 ‰ for δ15N^α, δ15N^β, δ15N^bulk, δ18O, and SP, respectively.

From every 8 to 10 min measurement interval the first three to 5 min were discarded to assure complete exchange of the analyte gas and only the last 5 min were averaged before further data processing. To reduce drift effects and enable offset correction, every analyte gas measurement was bracketed by a calibration gas (Cal1, Table 1) measurement. A second calibration gas (Cal2, Table 1) was measured before and after each experiment or every 8 h during continuous measurements for two-point δ-calibration.

We acknowledge that δ-values of Cal1 and Cal2 do not cover most of the measurement range relevant for WWTP. The accuracy of measurement results, however, was assessed using target gas measurements (Table 1), at an isotopic composition typical for the WWTP application performed in four experiments distributed over the complete measurement period (24 September 2024, 30 and 31 January 2025, 11 September 2025). Each measurement block consisted of four 10 min analyses of the target gas and was performed in an identical manner as any experimental measurements, i.e. with two-point calibration before and after each target gas measurement block and bracketed with Cal1 measurements. Average results were -25.51 ± 0.3 ‰, -23.35 ± 0.9 ‰, -24.43 ± 0.5 ‰, 31.22 ± 0.2 ‰ and -2.15 ± 0.9 ‰ for δ15N^α, δ15N^β, δ15N^bulk, δ18O and SP, respectively. Consequently, OA-ICOS results were systematically (0.4 ‰ to 1.2 ‰) lower than δ-values reported by IRMS (Table 1). This discrepancy is higher than differences observed in past inter-laboratory comparisons between Empa and Science Tokyo (Mohn et al., 2014; Ostrom et al, 2018) was not further addressed as it is within the uncertainty targets of our study. It might be rationalized by the absence of Ar in the target gas tank, which was shown to lead to lower apparent δ-values for OA-ICOS (Harris et al., 2020) or a potential uncertainty introduced by extrapolating calibration scales.

The analyser's dependence of reported δ-values on variations in N₂O mole fraction was assessed by dynamic dilution of a reference gas (Cal1) with N₂O free dilution air to a sequence of N₂O mole fractions ranging from 0.5 to 90 ppm. Gas mixtures were prepared at a flow rate of 300 mL min⁻¹ and measurements recoded for 10 min per mole fraction step. To account for analyser drift, in-between every mole fraction step the Cal1 gas was diluted to 12 ppm and measured as a reference point. This experiment was repeated three times on different days prior to the measurement period at the pilot reactors and once after the measurement period of 12 months. This enabled us to assess the analyser-specific variation in N₂O mole fraction dependence in the short and long term.

Figure 2 displays apparent δ-values for Cal1 measured between 0.5 and 90 ppm N₂O. In accordance with earlier work (Harris et al., 2020) the mole fraction dependence (e.g. Δδ15N^bulk per Δ[N₂O]) showed largest nominal values towards low, i.e. ambient, mole fractions and less pronounced above 6 ppm N₂O. Only in a mole fraction range from 8 to 20 ppm N₂O could reported δ15N^α and δ15N^β values be described by a quadratic function. For δ18O the mole fraction range for which a quadratic correction function applies was even smaller, 8 to 16 ppm N₂O. Like (Plouviez et al., 2026), we observed significant temporal changes in the N₂O non-linearity over time, necessitating a dynamic dilution as well as corrections for the non-linear mole fraction dependence (see Sect. 2.4.1).

In nitrifying zones of WWTPs, ambient air is injected to oxidize NH4+ to NO2- and NO3- as well as organic matter to CO₂. Consequently, O₂ mole fraction in the off-gas are at sub-ambient levels and the volumetric content of persistent gas components such as N₂ and Ar is enhanced since any produced CO₂ is removed prior to analysis (see Sect. 2.3.1 for details). We therefore tested the dependence of reported N₂O mole fraction and δ-values on O₂ mole fractions in the gas matrix. For this Cal1 was dynamically mixed with N₂O free SA, and high purity N₂ (99.999 %, Linde Gas AG, Switzerland) to incrementally vary the O₂ mole fraction in the analyte gas from 4 % to 20.95 %, while N₂O mole fractions were held constant at 12 ppm. As N₂ did not contain Ar, in parallel with a decrease in O₂ by e.g. 16 % Ar dropped by 0.7 %. Earlier studies (Harris et al., 2020) indicate that the gas matrix effect attributed to O₂ for an OA-ICOS analyser as tested here might therefore be overestimated by around 10 %, and our uncertainty estimates (Sect. 2.5.2) are rather conservative. Gas mixtures were prepared at a flow rate of 300 mL min⁻¹, and each sample was analysed for 10 min. To account for any analyser drift, in between every O₂ mole fraction step, the Cal1 gas was measured at ambient O₂ mole fractions (20.95 % O₂). The O₂ mole fraction dependence of apparent δ-values was linear over the entire range of O₂mole fractions (4 % to 20.95 % O₂; Fig. 3), with values of 1.45 ‰, 1.57 ‰, 1.51 ‰, 1.76 ‰, and -0.12 ‰ [% O₂]⁻¹ for δ15N^α, δ15N^β, δ15N^bulk, δ18O, and SP, respectively. Oxygen mole fractions in the reactors' off-gas ranged from 19.8 % to 20.95 % over the entire 1-year experimental period (O₂ mole fractions were measured using a PG-350E, Horiba, Japan). The average O₂ mole fraction in the actual analyte gas is somewhat higher at 20.6 % (± 0.2 %) as the off-gas is diluted with high-purity SA (see above) to set the N₂O mole fraction to 12 ppm. To account for a systematic reduction in O₂ mole fractions by 0.35 %, δ15N^α, δ15N^β, δ15N^bulk, δ18O, and SP values were corrected for 0.56 ‰, 0.53 ‰, 0.55 ‰, 0.83 ‰, and 0.04 ‰, respectively by using the linear models in Fig. 3. The variability in O₂ mole fractions was considered in the uncertainty budget.

As the apparent δ-values reported by the OA-ICOS analyser are subject to a strong, time-variant N₂O mole fraction dependence, as discussed in Sect. 2.2.3, we decided to implement a dynamic dilution system (Fig. 4) using N₂O-free SA (N₂O < 1 ppb) as dilutant to limit uncertainty contributions from N₂O non-linearity corrections. The target mole fraction of the dilution system was set to 12 ppm N₂O and data with N₂O mole fractions outside the range of 11 to 13 ppm were discarded (53 % of data). Data passing this criterion were corrected for non-linearity using average correction functions (Fig. 2). The target N₂O mole fraction (12 ppm) represents a compromise for analyser sensitivity and non-linearity, which both degrade data quality at low N₂O concentrations and temporal coverage, i.e. sufficiently low N₂O target concentration to cover relevant emissions.

The custom-built dilution system consisted of five mass flow controllers (MFC; Vögtlin Instruments GmbH, Switzerland), automated by a customized LabView programme (National Instruments, USA). Four MFCs were used to control flows of the analyte gas, Cal1, Cal2 and N₂O-free SA to dilute the analyte gas to the target N₂O mole fraction. One additional MFC set to 400 mL min⁻¹ was used to maintain a constant flow of analyte gas during Cal1 and Cal2 measurements to reduce lag times during subsequent analyte gas measurements (Fig. 4). The dilution ratio and MFC flow rates were calculated based on a 20 s running average of the actual N₂O mole fraction in the off-gas of the reactor provided by a NDIR analyser (X-STREAM X2XF, Rosemount Emerson, USA). The analyte gas was provided to the regulating MFC at 2.5 bar overpressure using a membrane pump (N86, KNF Holding AG, Switzerland). The intake of the pump was controlled with a needle valve, while the overpressure downstream of the membrane pump was monitored and manually adjusted with a pressure relief valve. To maintain stable operation of the MFC a critical orifice (bore diameter 150 µm) was installed in between membrane pump and MFC dampening pressure oscillations. The analyte gas was dehumidified by permeation drying (MD-070, Perma Pure, USA), CO₂ removed by absorption (Ascarite, mesh size 20–30; Sigma Aldrich, USA) and particles retained using an in-line sintered metal filter (pore size 2 µm; Swagelok, USA). Part of the analyte gas was supplied to the OA-ICOS spectrometer (about 200 mL min⁻¹), and the remainder (about 200 mL min⁻¹) exhausted to ambient air passing a NDIR CO₂ sensor (Sensair HPP, Sweden), to detect breakthrough of the CO₂ trap.

Based on the isotopic signatures, the fraction of residual, non-reduced N₂O (fN2O) for each measurement (δi) was estimated quantitatively by reformulating Eq. (1) to: 8fN2O=e(δR-δ0)/ϵ where δR is the isotopic signature (SP) of the residual N₂O fraction, δ0 represents the SP end-member value of the initial N₂O, i.e. N₂O not yet affected by reduction, and ϵ is the enrichment factor for SP (i.e. ϵ -5.9 ‰; Yu et al., 2020).

As relevant δ0 values for SP reported in literature span from -7.5 ‰ to 1.9 ‰ (hD: -7.5 ‰ to 3.7 ‰; nD: -13.6 ‰ to 1.9 ‰; Yu et al. 2020), we estimated our system specific δ0 value based on our dataset. First, it was ensured that the SP-δ18O regression slope fell within the expected range reported in literature (0.23 to 0.98; Yu et al. 2020), then the value of δ0 was estimated for SP by ranking observations according to the sum of their δ18O and SP values of which we retained the five lowest-ranked points. Further, the perpendicular distance of each of the lowest-ranked observations in the SP and δ18O space to the regression line was calculated and the observations with the smallest distance selected. The corresponding SP value was defined as δ0 (Fig. 5).

The fraction of reduced N₂O (fN2) was calculated as fN2=1-fN2O. Here, we used SP as input variables to calculate fN2, as SP is independent of the isotopic signature of the source. Signatures of δ15N^bulk however, are dependent on the isotopic signature of the substrate and may vary over the course of the aeration time and was therefore not used to calculate fN2.

The total uncertainty associated with the corrected δ-values (σi) was assessed by combining uncertainty contributions from (i) the N₂O mole fraction correction (σN2Oi), (ii) the correction for the gas matrix effect of O₂ (σO2i), and (iii) the repeatability of the instrument (σrepi). To do so, the law of error propagation was applied on the respective correction functions.

The uncertainty of the N₂O mole fraction correction was estimated from the uncertainty in the quadratic correction function. For this the 95th confidence interval of the fit was chosen at a mean N₂O mole fraction of 12 ppm. In contrast, the uncertainty related to the correction of the O₂ gas matrix effect was assumed to be dominated by the variability of O₂ mole fractions in the actual analyte gas (20.6 ± 0.2 %). Individual uncertainty contributions are shown in Table 2. To obtain the total uncertainty of the corrected isotope values the individual contributions were combined in quadrature: 9σi=σN2Oi2+σO2i2+σrepi2 The combined uncertainty (σi) represents the 1-sigma uncertainty of the final corrected δ-value. Combined uncertainties were 0.85 ‰, 1.08 ‰, 0.81 ‰, 0.48 ‰, and 1.09 ‰ for δ15N^α, δ15N^β, δ15N^bulk, δ18O, and SP respectively (see Table 2).

The uncertainty estimate for the reduced N₂O fraction (σfN2) was provided, propagating the uncertainty of SP (σi), accounting for the number of samples (n) following first-order Taylor expansion: 10σfN2=fN2⋅σiϵ⋅n(10)

To determine the predominant microbial N₂O production pathway, we analysed the N₂O isotopic composition over several months (September to December 2024) under standard WWTP operation, i.e. at a dissolved O₂ concentration setpoint of 2 mg L⁻¹ (± 0.5 mg L⁻¹) and without N-substrate additions. Figure 7 displays the N₂O isotopic composition during standard operation of the two wastewater treatment reactors presented as a dual isotopic plot of SP versus δ18O (N₂O) values, after data processing (Sect. 2.3). Isotopic signatures of relevant production processes (Hy, nD, and hD) derived from laboratory studies are indicated as rectangular shaded areas for comparison (Yu et al., 2020). Concurrent changes in SP and Δδ18O (N₂O, H₂O) along a straight line with a slope of 0.37 indicate successive reduction of N₂O to N₂ as the slope falls within a range defining the reduction line (spanning from 0.23 to 0.98, see Yu et al., 2020) and therefore validating the activity of N₂O reduction. From Fig. 7 we can infer nD or hD as the dominant N₂O production process, as measurements fall within the range of isotopic signatures expected for denitrification processes with a minimal effect of N₂O reduction (Yu et al., 2020) and are in accordance with other studies on wastewater treatment (Gruber et al., 2022; Wunderlin et al., 2013). Furthermore, the absence of elevated SP values (> 30 ‰) suggests that Hy was not a source term during our measurements (Frame and Casciotti, 2010; Sutka et al., 2006; Wunderlin et al., 2013; Yu et al., 2020). This interpretation may strike the reader as odd, as the data points so obviously lay out-side of the shaded source signature areas that are traditionally visualised as boxes in dual isotope plots. These source signature areas (see Figs. 7 and 8a) are based on a limited number of pure culture studies and laboratory incubations (Yu et al., 2020) and therefore cannot be perceived as hard limits but require refinement to fully capture the dynamics of complex mixed microbial communities as are found in wastewater treat-ment reactors. Thus, the isotopic source signatures and their overlaps will be subject to change with increasing understanding of the complexities of managed and natural N₂O source systems. To determine system-specific source signature areas and to robustly differentiate between the contributions of nD and hD to N₂O formation requires further dedicated experiments that aim at stimulating both processes independently. A series of such experiments applying the analytical toolkit developed in this study is described in Strubbe et al. (2026). Controlling the concentration of ammonium, nitrite, and DO nitrifier denitrification was successfully isolated from heterotrophic denitrification and became more active at lower DO concentrations. Lower DO, higher organic carbon availability, and lower pH increased N₂O production by heterotrophic denitrification during aeration. These new insights provide a systematic framework for understanding N₂O dynamics and support the development of mitigation strategies at full-scale.

During heterotrophic N₂O reduction to N₂, the N₂O molecule is preferentially split between the N–O bond of lighter isotopes, leading to an enhancement in δ15N^bulk, SP, and δ18O in the remaining N₂O pool (Lewicka-Szczebak et al., 2017; Ostrom et al., 2007). The enrichment of SP, and δ18O caused by N₂O reduction in experiment 2 is characterised by a slope of 0.72 (Fig. 8a), which falls within the data range defined by Yu et al. (2020) for such a process, supporting the validity of the here reported N₂O reduction fractions. As the two experiments (1 and 2) are separated in time, the difference in the reduction line slopes is likely related to a shift in the active microbial community over time. This variability is within the range seen in scientific literature under similar settings (Gruber et al., 2022; Yu et al., 2020; Strubbe et al., 2026) but still represents an open research question worthwhile exploring in more detail. Dissolved O₂ concentration was identified as an important driver of N₂O to N₂ reduction by hD (Fig. 8b). In general, higher N₂O reduction was seen in the reactor set to a lower DO concentration. This can be seen in Fig. 8a, as the data from the reactor with a lower DO have progressed further along the SP / δ18O line compared to those from the other reactor. Intuitively, this can be explained by the enzymes responsible for N₂O reduction preferentially cleaving the bonds between lighter isotopes, i.e. between ¹⁴N–¹⁶O rather than between ¹⁵N–¹⁶O or ¹⁴N–¹⁸O, which leads to an accumulation of molecules with ¹⁸O and ¹⁵N in the central molecular position in the residual, unreduced N₂O fraction. When applying Eq. (8) to calculate the fN2 for both scenarios, we can quantify the difference in fN2. The low DO concentration setpoint of 0.5 mg L⁻¹ led to an increase in fN2 by 40 % (± 1.7 %) compared to the DO setpoint of 2 mg L⁻¹. Therefore, on-line isotopic analysis can highlight periods during which the DO set-points can be optimised for N₂O emission reduction.

During the aeration phases both reactors showed a considerable dynamic in fN2 over time leading to a difference in fN2 of 10 % to 15 % over the course of the experiment. In detail, the wastewater treatment reactor set to low DO showed an increasing trend in fN2 from 35 % to 50 % followed by a decline to 40 % over the course of the experiment. In contrast, the reactor set to high DO displayed a steady decrease in the fN2 from 10 % to 0 %. The substantial difference between the two treatments can be explained by the O₂-sensitive nature of the enzyme responsible for N₂O reduction (Pomowski et al., 2011). These findings align with previous studies demonstrating that low DO concentrations enhance N₂O reduction efficiency and vice-versa (Morley et al., 2008; Suenaga et al., 2018; Tang et al., 2022; Zhou et al., 2021). Different DO concentration thresholds were found to trigger N₂O reduction, Tang et al. (2022) found the N₂O reduction rate to increase exponentially with decreasing DO concentration below a threshold of 1.6 mg O₂ L⁻¹ in estuarine waters. In contrast, Rees et al. (2021) reported that in marine water samples, N₂O consumption was observed at DO concentrations as high as 8 mg O₂ L⁻¹ and Körner and Zumft (1989) reported a threshold of 5 mg O₂ L⁻¹ below which N₂O reductase expression was enhanced. This high variability of reported thresholds may be related to microbial species or community dependencies in the response of N₂O reduction to DO concentrations (Cavigelli and Robertson, 2001; Suenaga et al., 2018; Zhou et al., 2021). The type and availability of organic carbon represents another factor parameter affecting N₂O reduction, which likely controlled the decrease in N₂O reduction over time, particularly at the high DO setpoint (Azam et al., 2002; Liu et al., 2022; Morley and Baggs, 2010). This suggests that a higher availability of organic carbon enhances N₂O reduction. Isolating the N₂O-to-N₂ reduction step by on-line isotopic analysis for the first time provides insights into its key drivers, helping identify operational levers to reduce production and enhance consumption of N₂O, supporting the design of net-zero emission treatment systems. The detailed interpretation of positive or negative effects of operational factors on N₂O reduction beyond data shown here, i.e. pH, temperature, total air consumption, total suspended solids, will be published elsewhere.

Low-level ¹⁵N-labelling is a novel methodology that uses the addition of labelled substrate to slightly increase δ15N values to a level above natural abundance (100 ‰ to 200 ‰) to achieve clear traceability through biogeochemical reactions, but within a range that still allows the use of natural abundant fractionation factors as well as standard isotopic measurement methods (Deb et al., 2025). Here, the addition of ¹⁵N–NH₄ (δ15N–NH4+ = 100 ‰ in the total NH4+ pool) at the onset of the aeration phase 1 results in N₂O emissions with distinctly higher δ15N^bulk values compared to the bulk of our measurements outside of the low-level labelling experiments (Fig. 9a). An initial increase in δ15N^bulk to 6.5 ‰ as compared to around -40 ‰ under standard WWTP operation is consistent with progressive conversion of ¹⁵N-labelled NH4+ into NO2- and NO3-, which act as substrates for N₂O formation (Fig. 9b). After the NH4+ is depleted, δ15N^bulk of N₂O stagnates. In the second cycle, after a volume exchange of ca. 38 % and an anoxic feeding phase, denitrification likely converted a large portion of the labelled inorganic ¹⁵N nitrogen substrate to N₂O and N₂ of which most was removed from the system but the δ15N^bulk values of -27 ‰ are still slightly elevated compared to standard WWTP operation. These results demonstrate the suitability of the on-line isotopic measurement setup for low-level ¹⁵N-labelling studies as well as the benefit of labelled substrate addition to isolate contributions from specific microbial conversions to the N₂O that is produced. A more complete quantitative analysis of these results requires considering the δ15N of aqueous nitrogen species and will be published independently.