We use the following notation for isotopic abundances, reported as deviations from the water isotope standard, Vienna Mean Standard Ocean Water (VSMOW): δ 18 O = R s a m p l e i R V S M O W i , δ D = R s a m p l e 2 R V S M O W 2 (1) where i = {17,18}, ⁱ R = n( ⁱ O)/n ¹⁶(O), ² R = n(²H)/n(¹H), and n denotes isotope abundance.

Deuterium excess, in its conventional linear (Merlivat and Jouzel, 1979) and empirical logarithmic (Uemura et al., 2012) forms, are given respectively by: d = δ D − 8 ( δ 18 O ) (2) and d l n = l n ( δ D + 1 ) − 8.47 ( l n ( δ 18 O + 1 ) ) − 28.5 ( l n ( δ 18 O + 1 ) ) 2 (3) The ¹⁷O excess (Δ¹⁷O) is defined as (Barkan and Luz, 2005): Δ 17 O = l n ( δ 17 O + 1 ) − 0.528 ( l n ( δ 18 O + 1 ) ) (4) All reported isotope values are normalized to the composition of SLAP (Standard Light Antarctic Precipitation), and are therefore on the VSMOW-SLAP scale. The isotopic values for SLAP, relative to VSMOW, are given by δ ¹⁸O = −55.5‰, δD = −428‰ and δ ¹⁷O = −29.6986‰, following convention (Gonfiantini, 1978; Coplen, 1988; Schoenemann et al., 2013). Note that the δ ¹⁷O of SLAP is defined by Δ¹⁷O ≡ 0 (Schoenemann et al., 2013). For all water δ and Δ values, VSMOW ≡ 0. Although Schoenemann et al. (2013) recommends using explicit terminology (e.g., δ ¹⁷O_VSMOW−SLAP) to make it clear that the VSMOW-SLAP normalization is used, for simplicity we use more concise terminology (e.g. δ ¹⁷O for uncalibrated data, δ ¹⁷O_VSMOW for calibrated data).

We obtained an ice core at 89.99^°S, 98.16^°W, approximately 2 km from the geographic South Pole, during the 2014–2015 and 2015–2016 austral summers (Johnson et al., 2020; Souney et al., 2020). The core, “SPC14”, was drilled to a depth of 1751 m, equivalent to an age of approximately 54 ka. The modern accumulation rate at South Pole is 8 cm w. e. yr⁻¹ (Casey et al., 2014), and the mean annual temperature is −51^°C (Severinghaus et al., 2001). A depth-age relationship for the core, tied directly to the timescale for the WAIS Divide ice core (WAIS Divide Project Members, 2013; WAIS Divide Project Members, 2015), was developed by Winski et al. (2019) and is termed the “SP19” chronology. A timescale for the gas within the core is also available (Epifanio et al., 2020).

We sampled SPC14 in 1 m long sections, with the resulting ice sticks measuring 1.3 × 1.3 cm in square cross section. The small volume of ice used allowed us to sample some sections of the core in duplicate, so that additional samples could be used for replicate measurements. For continuous-flow measurements, each 1 m ice stick was stored in a 1.9 × 1.9 cm square acrylic tube that was loaded vertically onto a melting device, described in the next section. For a limited number of replicate measurements, ice was melted in 1/2 m sections in clean HDPE bottles and decanted into 1 ml vials for later analysis.

We analyzed the entire South Pole ice core (depth 0–1751 m) for δ ¹⁸O and δD. For δ ¹⁷O, we analyzed the core from 556 to 1751 m only; the necessary instrument was not available at the time of the first analysis campaign.

For continuous-flow analysis (CFA), we use the system developed and described in detail by Jones et al. (2017b). Briefly, ice core samples in acrylic tubes are placed vertically on an aluminum block melter that has two concentric catchments that channel meltwater into a small drain. The melter is maintained at 14.6±0.1 C. Liquid water is drawn away from the melter using a peristaltic pump and pushed through an 8 μm filter. The filtered water then enters a 2 ml open-top glass vial where bubbles can escape. The water enters a six-port selection valve, and is mixed with high-pressure (80 psi) dry air in a glass concentric nebulizer, where it is converted into a fine spray. The spray enters a 20.0 × 1.8 cm Pyrex vaporizing tube heated to 200 C, and additional dry air (<30 ppm H₂O) is added at a rate of 3.5 L/min to dilute the water vapor and achieve a final water vapor content of ∼25,000 ppm H₂O. For our measurements the target value of 25,000 was maintained within ± 3,500 ppm, with rare values below 10,000 ppm flagged and excluded. The water vapor enters two intake lines (3.175 mm OD × 2 mm ID × 10 cm stainless-steel tubing), each inserted approximately 5 cm into the Pyrex vaporizing tube, which acts as an open split such that excess water vapor is vented to the laboratory. Each intake line is plumbed into a different laser spectroscopy instrument, a Picarro L2130-i and a Picarro L2140-i.

Both the L2130-i and L2140-i instruments are based on the earlier design described by Crosson (2008). These instruments use cavity ring-down spectroscopy (CRDS), in which the abundance of a given target molecule in an optical cavity is determined by pulsing a laser source at 10²–10³ Hz though the cavity. At the end of each pulse, the intensity of infrared light decays exponentially; the e-folding length of the decay is the “ring-down” time, which is proportional to the abundance of the target molecule. In the L2130-i, the ¹H₂ ¹⁶O, ²H¹H¹⁶O, and ¹H₂ ¹⁸O isotopogues are detected from the ring-down time for a single laser operating at wavenumbers in the range of 7,200 cm⁻¹. In the L2140-i, a second laser, centered on 7,193 cm⁻¹ where there are stronger absorption lines for ¹H₂ ¹⁷O, is also used. Rapid switching between the two lasers allows the measurement of all four isotopologues. The L2140-i uses peak-area integration rather than peak height (as used by L2130-i) to determine isotopologue abundance ratios, and laser resonance in the optical cavity is achieved by a different method, as described in detail in Steig et al. (2014). We use both instruments in our measurements because the use of the L2130-i in continuous-flow analysis of ice cores is well established (Jones et al., 2017b), and serves as a reference point for the quality of the δD and δ ¹⁸O measurements in either instrument.

Measurements of the South Pole ice core using the CFA system follows essentially the same protocol described in Jones et al. (2017b), with the addition of the L2140-i. Water flow from the ice core can be switched off at a six-port Valco value, allowing flow from 30 ml glass vials that contain laboratory reference waters. At the end of each day, the laboratory reference waters whose values are known relative to the VSMOW standards are run for 40 min each (We also ran reference waters at the beginning of each day in the first year of analysis, but found that this provided no additional benefit). Values of the laboratory reference waters are given in Table 1. The reference waters “KAW”, “KPW” and “WW” were analyzed twice each day; “VW1f” was analyzed once. To characterize isotopic mixing, three small (20 cm × 1.3 × 1.3 cm) sections of laboratory-prepared ice are made from batches of isotopically distinct waters with δ ¹⁷O values of approximately −14‰ for the first and third sections and −5‰ for the second section. Additionally, each set of several cores is separated by a 20 cm section of ice with a δ ¹⁷O value of about −14‰, which we refer to as “push ice”. This provides an easily identified signal so that ice cores can be distinguished in the data stream. In addition to the daily standards and ice core samples, we also ran longer measurements of reference waters at the end of each week, both to ensure that water continued to flow through the system when it was not being used for ice-core analysis, and to obtain additional data for evaluation of the system for inter-laboratory comparison.

A number of routine performance metrics were used to monitor the quality of the CFA measurements and to produce “raw” data sets that are calibrated against the reference waters. Each meter of ice takes approximately 40 min to melt, and data are generated nominally at 1 Hz, resulting in approximately 2,400 data points per meter. Each data point is assigned to a specific depth using a depth-assignment algorithm; depths are provided by readout from a digital laser-distance meter. Raw data are flagged for exclusion where they do not meet metrics such as reproducibility and water concentration targets in the Picarro instrument, and other criteria such as the timing of valve switches (which can introduce spurious noise); details are given in Jones et al. (2017b) and Steig et al. (2020). A final data set is produced by averaging all non-flagged data points over 0.5 cm intervals, after calibration (see next section). Jones et al. (2017b) developed a methodology for correcting measurements for memory; this affects samples that are measured immediately after the “push ice”. We do not use memory correction but instead remove data obtained within 200 s of a large isotopic transition (e.g., after the switch from push ice to sample); the cut-off time is based on measurements of laboratory-created ice of known composition. Memory-correction data are available in the data archive (Steig et al., 2020).

For measurements of discretely sampled ice, we used a Picarro L2140-i with an autosampler, following the protocols given in Schauer et al. (2016). Briefly, 1.8 μl of water are injected using a 10 μl syringe into a vaporization module with a liquid autosampler. The vaporizer-provided mixture of water vapor and dry air is pulled into the laser cavity with a diaphragm pump. A “run” is defined as a set of vials automatically sampled sequentially from start to finish. Each run containing unknown samples is organized by having a set of five reference waters, n samples (typically n∼30 vials), and then another set of five reference waters. Reference waters used for these measurements are given in Table 1; the different and overlapping standards provides an additional quality control check on the calibration of the CFA data.

We calibrate the raw CFA data following established linear methods with some modification. An example of raw measurements for a typical analysis from the CFA system is shown in Figure 1, with ice core data, interspersed laboratory ice, and reference waters identified. The measured value of a reference water is given by the average across all data points, defined by the valve switch to that water, with the first 200 s removed to ensure the elimination of memory effects. Because each reference water is measured for a minimum of 40 min, this provides a minimum measurement time of 2,200 s, about twice the time to obtain reliable calibration for Δ¹⁷O, as established in previous work (Steig et al., 2014; Schauer et al., 2016). A simple calibration is defined by the linear relationship between these measured isotope values and their known values on the VSMOW-SLAP scale, given in Table 1. The resulting slope and intercept are then used to convert raw values for the measured ice core data to calibrated values on the VSMOW-SLAP scale.

We use two of the four reference waters (e.g., KAW and KPW, as shown in Figure 1) for calibration, reserving the other two (in this case, WW and VW1f) as unknowns or “traps”. The precision and accuracy of the calibration are then estimated, respectively, by the standard deviations of each trap and the mean difference with the known value of each trap. We repeat these calculations to obtain four independent estimates of precision and accuracy (one for each reference water).

We also use an adjusted calibration, based on the difference between the known and estimated values of reference waters used as traps, as follows. For each δ type (i.e., δ ¹⁷O, δ ¹⁸O, or δD), let X _ref =(x _1,ref ,x _2,ref …x _n,ref ) be the set of measured “raw” values of reference waters used for calibration, and Y _ref the corresponding known (calibrated against VSMOW-SLAP) values from Table 1. Let X _trap =(x _1,trap ,x _2,trap …x _n,trap ) be the set of measured “raw” values of reference waters used as traps, and Y _trap the corresponding known calibration values. Obtain the slope, m and intercept, b using m = ∑ ​ i = 1 n ( x i , r e f − X ¯ r e f ) ( y i , r e f − Y ¯ r e f ) ( x i , r e f − X ¯ r e f ) (5) b = Y ¯ r e f − m X ¯ r e f (6) where X ¯ , Y ¯ denote means. The calibration equation for a given δ type is, δ c a l i b r a t e d = m ( δ s a m p l e ,   m e a s u r e d ) + b (7) Equations 5–7 define the simple linear approach. The slope and intercept values are used to obtain estimated values for those reference waters that have been reserved as unknown “traps”: Y ¯ t r a p ,   e s t i m a t e d = m X ¯ t r a p + b (8) The adjustment to the intercept, b, is given by the difference between the estimated trap values Y _{trap,estimated} and their known values from Table 1: b a d j u s t e d = b − Y ¯ t r a p , k n o w n − Y ¯ t r a p , e s t i m a t e d (9) The adjusted calibration equation for a given δ type is then given by δ c a l i b r a t e d ,   a d j u s t e d = m ( δ s a m p l e ,   m e a s u r e d ) + b a d j u s t e d (10) The simple linear approach (Eq. 7) yields highly reproducible results for δ ¹⁸O, δD, and deuterium excess, as previously shown by Jones et al. (2017b), as well as for δ ¹⁷O. For Δ¹⁷O, results are significantly improved if the adjusted calibration approach (Eq. 10) is used. The a priori justification for this approach is that the slope of the calibration varies by less than 1%, while the intercept varies by as much as 50%, so that errors in the intercept have a much greater impact on the results. We find that using any combination of reference waters and traps yields essentially indistinguishable results. The final data set is based on the use of the two reference waters KAW and KPW as our primary reference waters and VW1f and WW as the traps. We note that the lowest isotope values in the South Pole core are more negative than the most depleted reference water (VW1f) or standard (SLAP).

In general, we use the reference waters analyzed at the end of a measurement day to calibrate the ice core data from that same day. For measurements of Δ¹⁷O with the L2140-i, it is critical to monitor changes to the currents provided in five channels to the piezoelectric material (lead zirconate titanate, PZT), that is used to vary the length of the laser cavity (Steig et al., 2014). Periodic adjustments to the laser cavity size, which is subject to drift and hysteresis, are made as part of routine instrument operation to achieve laser resonance at the target wavelength. Changes to the PZT values influence the instrument calibration, which can significantly affect Δ¹⁷O if it is not accounted for during calibration of raw measurements to obtain final values on the VSMOW-SLAP scale. We use the PZT values to ensure that calibration reference waters and ice core samples have been measured during the same “calibration window”, as described in Schauer et al. (2016). For example, if there is a change to the ordering of the PZT offset values during a measurement day, we use the reference water values from the previous day.

Continuous-flow analyses of the isotopic composition of H₂O in ice cores have advanced significantly since the early, discontinous measurements of discrete samples pioneered by Johnsen et al. (1972), Dansgaard et al. (1982), and others. The development of laser spectrometers (Crosson, 2008) has reduced the use of mass spectrometry, and allowed the development of continuous-flow analysis of ice-core water isotopes (Gkinis et al., 2010). This method has been successful both with Picarro cavity ring-down spectroscopy (Gkinis et al., 2010; Gkinis et al., 2011; Maselli et al., 2013; Jones et al., 2017b; Hughes et al., 2020) as we use here, as well as off-axis integrated cavity output spectroscopy (Los Gatos Instruments) (Emanuelsson et al., 2015). With the inclusion of δ ¹⁷O and Δ¹⁷O, our results represents a further advance of this earlier work.

In general, the analytical error of δ ¹⁸O and δD measurements on ice cores is much smaller than the signal being measured. Our results show that the same is true for δ ¹⁷O. For example, RMS errors for replicates (our most conservative estimate of precision) are <0.1, <0.9, and <0.1‰, respectively, for 10 cm averages of δ ¹⁸O, δD, and δ ¹⁷O. For most of the length of the South Pole ice core, 10 cm corresponds to just a few years. Interannual, decadal and centennial variations of δ ¹⁸O are typically 1–2‰ (e.g., Steig et al., 2013); variability in δ ¹⁷O is about half that of δ ¹⁸O, and variability in δD is about eight times greater. Millennial and glacial-interglacial changes are up to an order of magnitude larger (e.g., Grootes et al., 1993; EPICA Community Members, 2006; WAIS Divide Project Members, 2015). A measure of the signal to noise ratio (amplitude of the natural variablity over the RMS error) is therefore between 10 and 100 for all three types of δ measurement. This suggests that δ ¹⁷O, like δ ¹⁸O and δD, can be considered a routine measurement. Indeed, the measurement of δ ¹⁷O by CFA poses no particular challenge that has not already been addressed in previous work for δ ¹⁸O and δD. We would expect similar results with other available instruments such as the Los Gatos isotopic water analyzer (Berman et al., 2013) or other commerial or non-commerical instruments of similar design to the Picarro L2140-i used here.

Natural variations in d in ice cores are typically smaller than those in δ ¹⁸O, while precision, about 0.5–1‰, is similar to that for δD. For example, millennial-scale variabilty in d in Antarctic ice cores is only about 1–2‰ (Markle et al., 2017); in contrast δD varies by 10–20‰ on the same timescale. Ice core d data is thus comparatively “noisy”, and relatively long averaging windows are typically required. Similar to d, the signal to noise ratio is low for Δ¹⁷O. For our measurements on the South Pole ice core, depth averages of at least 50 cm are needed to reduce the analytical error to 0.01‰ (10 per meg). The best mass-spectrometry measurements of Δ¹⁷O have reported precisions of as little as 4 per meg, though this is generally the standard error of multiple replicate measurements on discrete samples (e.g., Barkan and Luz, 2007; Schoenemann et al., 2013); more typical reproducibility is 6–8 per meg (Landais et al., 2008; Schoenemann et al., 2014; Landais et al., 2017). Previous work shows that millennial variations in Greenland Δ¹⁷O are of order 30 per meg (Guillevic et al., 2014), and our South Pole ice core record shows variations of the same magnitude. Considering these timescales, the signal to noise ratio of our data for 50 cm averages is therefore about ∼3, similar to that for d. In the South Pole core, 50 cm averages corresponds to a time resolution of only a few decades, so longer averages can be used to further reduce the noise when examining variability on longer timescales. For applications where the very highest resolution is sought (e.g., resolving the annual cycle, or obtaining measurements at high temporal resolution in ice that has been significantly thinned, such as near the bed of an ice sheet), continuous-flow techniques for Δ¹⁷O may be less valuable, and discrete measurements, while labor intensive, are still warranted.

In summary, our results show that measurements of ice-core δ ¹⁷O and Δ¹⁷O can be achieved by continous-flow methods that are competitive, in terms of the signal-to-noise, with traditional δ ¹⁸O and deuterium excess measurements, respectively. Our results for the South Pole ice core, while somewhat lower quality than the best-possible mass-spectrometery or discrete laser-spectroscopy measurements, provide useful information for the investigation of climate variability and other applications.

We now consider the complete water-isotope records from the South Pole ice core, and provide two examples of the use of these data. One important application of high-resolution isotope data is the calculation of the water-isotope diffusion length. As dry snow densifies to firn and ice, water vapor diffuses through interconnected pathways, smoothing the original isotope signal (Johnsen et al., 1972; Whillans and Grootes, 1985; Cuffey and Steig, 1998). The diffusion length, σ, is given by the solution to the differential equation (Johnsen, 1977) d σ 2 2 d t − ϵ ˙ ( t ) σ 2 = D i * ( t ) (13) where t is time, ϵ˙ is the vertical strain rate (due to compression of the firn, and dynamic thinning of the ice), and D _i ^*(t) is the effective diffusivity in the firn or ice of isotopic species i. Measurements of the diffusion length in ice core records can be used, in combination with models of the firn-densification process, as an independent measure of paleotemperature (Johnsen et al., 2000; Simonsen et al., 2011; Gkinis et al., 2014; van der Wel et al., 2015), and for correction of high-frequency paleoclimate signals (Jones et al., 2018). Because of the dependence on strain, diffusion-length measurements also provide information on ice thinning (Gkinis et al., 2014). In Figure 8 we show the diffusion length measurements, calculated using the spectral analysis-based method described in Kahle et al. (2018). Of particular interest is that the diffusion length for δ ¹⁷O is systematically greater than that for δ ¹⁸O, which is again greater than that for δD, as expected from theory. Johnsen et al. (2000) and van der Wel et al. (2015) have suggested that using the ratio (e.g. σ17/σD) or difference (e.g. σ17−σD) of diffusion lengths for different isotopologues can constrain firn temperature because temperature is the stronger control on those differences or ratios than on individual diffusion lengths. However, the small differences between diffusion lengths has made such calculations unsuccessful in practice (Holme et al., 2018). In this respect, the greater magnitude of the difference σ17−σD than σ18−σD could be advantageous, as Simonsen et al. (2011) have suggested. In principle these data could be also used to independently verify laboratory-based measurements of isotopic diffusivity, which remain uncertain (Hellmann and Harvey, 2020).

Figure 9 shows the δ ¹⁸O, deuterium excess and Δ¹⁷O, averaged to 100-years resolution, from the South Pole ice core, on the timescale from Winski et al. (2019). We include both the conventional linear deuterium excess (d, Eq. 2) and the logarithmic form (d _ln , Eq. 3). For the Holocene (most recent ∼12,000 years), 100 years resolution is equivalent to about 5 m averages; for older portions of the core, snow accumulation rates were lower and 100 years corresponds to about 2 m averages.

The δ ¹⁸O time series in Figure 9 shows the typical pattern of variation seen in other ice cores from East Antarctica, with an early Holocene maximum (HM), the “Antarctic Cold Reversal” (ACR) between 14.7 ka and 13.0 ka (ka = thousands of years before present), and the Antarctic Isotope Maximum (AIM) events, the most prominent two of which are labeled. Such features provide a useful reference, when discussing the significance of the novel deuterium excess and Δ¹⁷O results.

Of particular interest is the relationship between Δ¹⁷ O and d _ln , and d. The traditional formulation for deuterium excess, d, is subject to artifacts resulting from the linear approximation to non-linear processes. The d _ln calculation is intended to be more faithful to the physics of the system (Uemura et al., 2012); this also motivates the non-linear definition of Δ¹⁷O (Landais et al., 2008). Consistent with this, 100-year average Δ¹⁷O is better correlated with d _ln (r = 0.3) than with d (r = −0.15). However, neither correlation is statistically significant at millennial frequencies, after accounting for autocorrelation.

Markle et al. (2017) showed that in the WAIS Divide ice core, d _ln is in phase with δ ¹⁸O for long timescales, but shows abrupt transitions that correspond to the abrupt changes recorded in Greenland ice cores–the Dansgaard-Oeschger (D-O) events. Markle et al. (2017) showed that d _ln variations can be explained as a combination of slowly varying sea surface temperatures (SST), in phase with Antarctic δ ¹⁸O, and faster-varying atmospheric shifts of moisture sources for Antarctic precipitation, in phase with D-O events. Later work verified the same relationships for additional Antarctic ice cores (Buizert et al., 2018; Svensson et al., 2020). A key element of this argument is that millennial scale changes in Antarctic d _ln are associated with changes in the evaporative conditions at the ocean surface–both SST and humidity–which affect the initial d _ln of water vapor over the ocean surface at the point of evaporation. Δ¹⁷O, like d _ln , is sensitive to humidity variations, but not directly to SST (Landais et al., 2008). We would therefore expect to see similarities between d _ln and Δ¹⁷O at millennial periods, if humidity variations dominate the d _ln signal, but less agreement if SST variations dominate. As already noted, the correlation between these variables is not significant on millennial timescales. Furthermore, the relative scaling between d _ln and Δ¹⁷O is inconsistent with an attribution to humidity variations. For example, the millennial-scale variations in d _ln we observe in the South Pole ice core have an amplitude of about 1‰, while the Δ¹⁷O amplitude is about 0.01‰ (10 per meg), where we estimate the amplitude from the standard deviation of highpass filtered data, with a 5,000-year cut-off period. A change of 10 per meg in Δ¹⁷O could be attributed to a ∼20% change in humidity at the ocean surface; all else being equal this would also cause a change of about 10‰ in d _ln – that is, variability about an order of magnitude larger than we observe — based on the data in Uemura et al. (2008; 2010). This suggests that the attribution of d _ln variations to the SST of the moisture sources for Antarctic precipitation (Markle et al., 2017) is more consistent with the evidence. However, this leaves the millennial-scale Δ¹⁷O variations unexplained. We note that these findings do not preclude a significant role for humidity in explaining Δ¹⁷O variations in Greenland ice cores (Guillevic et al., 2014), in which all these variables are more highly correlated with one another than they are in the South Pole record.

Finally, we note two interesting features of the Δ¹⁷O time series that warrant further study. There is a significant drop in Δ¹⁷O between 12.8 ka and 11.7 ka, corresponding to the Northern Hemisphere Younger Dryas, and also characterized by a rapid increase in d _ln , and a more gradual increase in δ ¹⁸O that defines the end of the Antarctic Cold Reversal. This is one of the few features in the record where all water-isotope variables show correlated changes. In contrast, another prominent feature in Δ¹⁷O, an abrupt drop in values at about 28 ka, has no obvious counterpart in d _ln or δ ¹⁸O (This event occurs at about 1200 m depth in the South Pole ice core, and is verified by triplicate measurements — see Figure 5). Interpretation of these relationships is not straightforward and would likely benefit from isotope-modeling studies that incorporate the complete set of water-isotope variables.