The CISE-LOCEAN sea water isotopic database (1998-2021)

1 The characteristics of the CISE-LOCEAN sea water isotope data set (δ 18 O, δ 2 H, later designated 2 as δD) are presented. This data set covers the time period from 1998 to 2021 and currently 3 includes close to 8000 data entries, all with δ 18 O, three quarters of them also with δD, associated 4 with a time and space stamp and usually a salinity measurement. Until 2010, samples were 5 analysed by isotopic ratio mass spectrometry (IRMS), and since then mostly by cavity ring- 6 down spectroscopy (CRDS). Instrumental uncertainty on individual data in this dataset is 7 usually with a standard deviation as low as 0.03 and/ 0.15‰ for δ 18 O and δD, respectively. An 8 additional uncertainty is related to uncertain isotopic composition of the in-house standards that 9 are used to convert daily data into the Vienna Standard Mean Ocean Water ( VSMOW) scale. 10 Different comparisons suggest that since 2010 the latter have remained within at most 11 0.03/0.20‰ for δ 18 O and δD. Therefore, combining the two suggests a standard deviation 12 of at most (0.05 ,/ 0.25) ‰ for ( δ 18 O ,/ δD). 13 Finally, for some samples, we find that there has been evaporation during collection and 14 storage, requiring adjustment of the isotopic data produced by CRDS, based on d-excess (δD – 15 8 x )δ 18 O. This adds an uncertainty on the adjusted data of roughly 0.05 / and 0.10‰ on for 16 δ 18 O and δD, respectively. This issue of conservation of samples is certainly a strong source of 17 quality loss for parts of the database, and ‘small’ effects may have remained undetected. for subsets the dataset, when time series obtained (such as in the or North subpolar gyre). or temporally (over a year) averaged data on the order of or less than 0.03 / and 0.15‰ for 18 O 22


Abstract 1
The characteristics of the CISE-LOCEAN sea water isotope data set (δ 18 O, δ 2 H, later designated 2 as δD) are presented. This data set covers the time period from 1998 to 2021 and currently 3 includes close to 8000 data entries, all with δ 18 O, three quarters of them also with δD, associated 4 with a time and space stamp and usually a salinity measurement. Until 2010, samples were 5 analysed by isotopic ratio mass spectrometry (IRMS), and since then mostly by cavity ring-6 down spectroscopy (CRDS). Instrumental uncertainty on individual data in this dataset is 7 usually with a standard deviation as low as 0.03 and/ 0.15‰ for δ 18 O and δD, respectively. An 8 additional uncertainty is related to uncertain isotopic composition of the in-house standards that 9 are used to convert daily data into the Vienna Standard Mean Ocean Water (VSMOW) scale. 10 Different comparisons suggest that since 2010 the latter have remained within at most 11 0.03/0.20‰ for δ 18 O and δD. Therefore, combining the two suggests a standard deviation 12 of at most (0.05 ,/ 0.25)‰ for (δ 18 O ,/ δD). 13 Finally, for some samples, we find that there has been evaporation during collection and 14 storage, requiring adjustment of the isotopic data produced by CRDS, based on d-excess (δD -15 8 x )δ 18 O. This adds an uncertainty on the adjusted data of roughly 0.05 / and 0.10‰ on for 16 δ 18 O and δD, respectively. This issue of conservation of samples is certainly a strong source of 17 quality loss for parts of the database, and 'small' effects may have remained undetected. 18 The internal consistency of the database can be tested for subsets of the dataset, when time 19 series can be obtained (such as in the southern Indian Ocean or North Atlantic subpolar gyre). 20 These comparisons suggest that the overall uncertainty of the spatially (for a cruise) or 21 temporally (over a year) averaged data is on the order of or less than 0.03 / and 0.15‰ for δ 18 O 22 /and δD, respectively. On the other hand, 17 comparisons with duplicate sea water data analysed 23 in other laboratories or with other data sets in deep regions suggest a larger scatter. When 24 averaging the 17 comparisons done for δ 18 O, we find a difference close to the adjustments 25 applied at LOCEAN to convert salinety water data from the activity to the concentration 26 scaleproduced either by CRDS or IRMS. Such a difference is expected, but the scatter found 27 suggests that care is needed when merging datasets from different laboratories. Examples of 28 time series in the surface North Atlantic subpolar gyre illustrate the temporal changes in water 29 isotope composition that can be detected with a carefully validated dataset. 30 31 1. Introduction 33 Stable isotope analyses of ocean water (δ 18 O, δ 2 H later designed termed as δD) were first 34 discussed by Craig and Gordon (1965) as tracers of water masses, and of the different 35 components of the global hydrological cycle, in particular the signals gained through 36 evaporation, precipitation, the interaction with sea ice, and continental water inputs, for 37 example from the ice caps of Greenland and Antarctica, and ice shelves. Sea water stable 38 isotopes have been used to verify ocean model circulation and characterize processes 39 controlling their spatial variability (Xu et al., 2012). Sea water isotopes have also been used to 40 provide information on what controls the oxygen isotopic ratio of calcite plankton shells, in 41 order to reconstruct past ocean salinity and circulation. The GEOSECS program (Östlund et al., 42 1987) provided the first consistent global dataset of sea water isotopes, but with a limited data 43 coverage. The Global Seawater Oxygen-18 Database at GISS  has 44 assembled most water isotope data collected prior to 1998, with an effort to homogenize the 45 dataset, when possible, by estimating biases based on multiple measurements of deep-water 46 samples (Schmidt, 1999;Bigg and Rohling, 1999). A large part of the early analyses was done 47 by isotope ratio mass spectrometry (IRMS) and more recently using cavity ring-down 48 spectrometry (CRDS). Walker et al. (2016) illustrated that the two measurement techniques can 49 provide equivalent results with no obvious biases. 50 51 Since 1998, the CISE-LOCEAN isotopic platform facility at LOCEAN (later 'CISE-52 LOCEAN') has measured sea water isotopic composition of samples collected on a series of 53 oceanographic cruises or ships of opportunity, mostly in the North Atlantic, in the equatorial 54 Atlantic, in the southern Indian Ocean and the Southern Ocean. This LOCEAN data set of the 55 oxygen and hydrogen isotopes ( 18 O and D) of marine water covers the period 1998 to 2021,56 and is ongoing. Most data prior to 2010 (only  18 O) were produced using an Isoprime IRMS 57 coupled with a Multiprep system (dual inlet method), whereas most data since 2010 (and a few 58 earlier data) were obtained by CRDS, usually with a Picarro L2130-i, or less commonly on a 59 Picarro L2120-i. Occasionally, some data samples were also run on an Isoprime IRMS coupled 60 to a GasBench (dual inlet method) at the university of Iceland (Reykjavik We most commonly used a Picarro L2130-i CRDS, but at times, a Picarro L2120-i CRDS was 168 used, resulting in a larger standard deviation, in particular for D. On both CRDS analyzers, 169 when repeatability of the different injections of the sample was not sufficient or the daily run 170 presented a too largean unacceptably large drift, the samples were analyzed at least a second 171 time. In that case, either the best value or an average of the different values was taken/retained.

173
The typical daily run at LOCEAN currently includes one or two reference water samples 174 followed by three freshwater standards at the beginning to establish a slope calibration, as well 175 as regularly interspersed reference water samples afterwards (usually, from KonaDeep mineral 176 water with a value close to 0.8 and/ 2.0 ‰ in  18 O and/ D, respectively). In addition to these 177 freshwater in-house reference materials, a series can contain up to 12 isotopically-178 uncharacterized water samples, using a little over 1 ml of the sample placed in a cap-closed 179 vial. Until 2015, when samples were distilled, series typically included 12 water samples. Since 180 2015, when salt water was directly placed in the vials, we have mostly run not more than 9 181 samples in a run, because the deposit of salt in the liner induces water retention or release, and 182 thus noise in the measurements after roughly 60 injections of salty samples, as well as drifts in 183 the reference water ( Fig. 1) and possibly slope calibration. Another source of drift is the 184 appearance of condensation on the top cap of the vials after a few hours, which will result in 185 enriching the residual vial water, although it is by no means the largestis very likely a small 186 source of drift.

188
Each sea water sample is injected 6 times, and usually 9 to 12 times for the internal standards 189 at the beginning and end of the run. Whenever possible, samples expected to be in the same 190 range of values are placed together in the run to minimize the memory effect on the CRDS 191 which is largest for D. We reject the first injection, as well as later injections if they are not 192 stable, retaining between two and eight injections that we average. Two methods were tested, approach.

200
If a significant drift in the reference water values is noticed through the run, it is corrected, 201 usually by adjusting it linearly between the successive values of the reference water ( Fig. 1). 202 We thus assume that the estimated drift is independent of the  18 O, D values. In addition, in 203 2017-2019, the response slope of the Picarro CRDS was adjusted by interpolating between the 204 three-point slope estimate (based on 3 internal standards) at the beginning and at the end of the 205 runs, when that was deemed possible. However, this adjustment was discontinued in 2020 206 because the last internal standard samples were often not as reliably measured, with values more 207 sensitive to the number of injections, probably as a result of salt deposits in the liner. Since 208 2020, we only check the instrument's response at the end of the run with one of the freshwater 209 internal standards.

211
Accuracy is best when samples are distilled, and for D it is better on the Picarro CRDS L2130-212 i than on the Picarro CRDS L2120-i. Usually, the reproducibility of the δ 18 O measurements 213 between the different selected injections is within ± 0.05 ‰ and of the δD measurements within 214 ± 0.15 ‰, which should be considered an upper estimate of the random error on a measurement 215 with the Picarro L2130-i CRDS. Samples with a SD larger than 0.06 ‰ in δ 18 O were considered 216 too uncertain and were rerun, as well as often (after 2015) the first and last samples of each run.

218
In addition to the instrumental error of each sample δ 18 O and δD described above, other 219 uncertainties arise from the data processing and conversion of measured δ 18 O and δD into the The second source of uncertainty (for Picarro CRDS) is due to the way we process the data of 226 a daily run with salty water samples. As commentmentioned above, we first adjust the values 227 to compensate for the drift in reference water. Usually, this drift during the run is relatively 228 small, not exceeding 0.1 and /0.6 ‰ in δ 18 O and/ δD, respectively, during the run, but in about 229 10% of the runs, it exceeded 0.2 ‰ in δ 18 O over the whole run, or 0.10 ‰ in δ 18 O over 230 successive reference water samples (23 out of 214 daily runs over which statistics were 231 established from 06/2020 to 04/2021). When these large changes are encountered, the run is 232 estimated noisy and is usually rerun. However, even for the other runs, a drift is usually 233 observed with salty samples, and it often is a positive drift, in particular between the reference 234 water samples before and after the three initial internal standards (Fig. 1). The average (SD) 235 drift in reference water during a run was +0.081 (0.106) ‰ in δ 18 O, and +0.62 (0.53) ‰ in 236 δD in the 191 (out of 214) daily runs retained. The drift is also found in the internal standard 237 water analysed at the end of the run compared with the one analysed just after the initial 238 reference waters with an average (SD) drift of +0.069 (0.073) ‰ in δ 18 O, and +0.43 (0.34) ‰ 239 in δD for the same 191 daily runs subset. These values slightly differ from the drifts for the 240 reference water, not significantly at 99% for δ 18 O, but significantly at 99% for δD. This may 241 be indicative of errors resulting from linearly adjusting the drift, in particular for the initial 242 standard water samples. This suspicion of a slight non-linearity in the initial drift is reinforced 243 by 7 runs in 2020-2021 when the three standards were also measured at the end of the run. 244 However, as this is too uncertain, a correction has not been attempted for that, but in addition 245 to being a source of random error (at least 0.02 and /0. correction of a systematic bias has only been applied on the MIX2 value for analyses since 296 August 2020. For some internal standards, we witnessed larger differences for measurements 297 done in June 2020 after the L2130-i just returned from a cruise and long shipping and storage 298 for more than 9 months. We assume that this anomaly is instrumental, and did not last for a 299 long time, as the anomaly was not reproduced during later tests in August 2020, nor in 300 November 2021.

302
The two storage methods used successively for internal standard waters were designed to 303 minimize water vapor exchange. It is however possible that small isotopic drifts of the internal 304 standards have taken place with time, due to evaporation or possible oxidation of the tanks (rust 305 was found in one nearly empty tank). As mentioned, based on different comparisons over time, 306 sometimes over remnants of the tank waters, we could verify that these drifts have remained 307 smaller than 0.02 and /0.1 ‰ in δ 18 O and/ δD, respectively. Finally, standards for the daily runs 308 are temporarily stored, for up to a month, in glass bottles stored at 4°C, which are briefly opened 309 every day to extract water. Through its storage life, this water will slightly breath, by exchange 310 with the outside air that penetrates when the bottle is briefly opened. Back of the envelope 311 estimates suggest that the effect should be less than 0.01 and /0.05 ‰ in δ 18 O and/ δD, 312 respectively, even after a month. 313 314 2.4 Conversion to the Cconcentration scale 315 Both oxygen and hydrogen isotope compositions are reported in parts per thousand (‰) on the 316 VSMOW scale. One issue is that we analyse saline samples on the activity scale, while the 317 internal standards are fresh water standards, and the method of analysis has changed over time.

318
There is still a large uncertainty on the correction to be applied to account for the effect of salt 319 on IRMS and CRDS seawater analyses. Here we have applied the corrections provided by 320 Benetti samples, most of which (11.3%) correspond to unadjusted data with anomalously low d-excess 362 and thus suspected evaporation. There is of course also the possibility that for some samples, 363 too low or too high (for 1% of the cases) d-excess might just result from an occasional large 364 uncertainty in the analysis.

366
We what we have previously measured in regions with repeated cruises, and outliers (6%) are 389 flagged as probably bad. The smaller (by half) proportion of flagged IRMS analyses than for 390 the CRDS analyses suggests either that this validation missed some evaporated IRMS samples, 391 or that these earlier data had evaporated less than the more recent ones (some were analyzed 392 sooner after collection), or that the IRMS runs had smaller uncertainties than the latter CRDS 393 runs. 394 395 3. Validation 396 As discussed in section 2, in addition to random errors or to issues related with evaporation of 397 samples, there is the possibility of shifts between subsets of the data, due to the different internal 398 standard waters, methods of processing, adjustment (for CRDS) or conversion from the activity 399 to the concentration scale (for IRMS We illustrate the dataset with time series of June (or July) data between 50° and 55°N in the 466 eastern North Atlantic subpolar gyre (NASPG) collected mostly during the OVIDE cruises 467 (Fig. 4). This scatter plot of cruise-averaged S and  18 O indicates a near alignment of the values. than for the OVIDE surveys (Fig. 4). However, there is some aliasing of the seasonal cycle in 483 the annual averages (see Reverdin et al., 2018b), which contributes to the scatter, as well as 484 noise on the data, and natural variability. On this plot the freshest year appears to be 2017, in 485 agreement with an analysis using a much more complete salinity dataset (Reverdin et al.,486 2018a). 2017 is also one of the lighter  18 O years. The corresponding d-excess versus S diagram 487 (Fig. 5b)  2012. There was however a short-term larger difference found for the most negative standard 502 (equal to 0.1‰ for  18 O ), most likely related to the readjustment of the instrument to laboratory 503 conditions in May 2021. When using the CRDS Picarro L2130-i, we also found periods with 504 quite uncertain analyses, in particular due to salt or particle deposit in the vaporizer or filters. 505 These samples could often be run again afterwards to reach lower resulting uncertainty.

507
Finally, there is the issue of possible evaporation during collection and storage. When the 508 analysis is done on a CRDS, we are usually able to detect possible biases larger than 0.05‰ in 509 δ 18 O, by comparing d-excess with the expected d-excess derived from regional d-excess-S 510 linear relationships. Attempts were made here to correct δ 18 O and and δD when the resulting 511 uncertainty does not exceed 0.05 and /0.1‰, respectively. In particular this was done for some 512 OISO cruise samples which were analysed many years after collection, or in the case of faulty 513 caps being used, or caps that were not properly closed and wrapped with no parafilm. This is 514 certainly a strong source of quality loss for part of the database, and 'small' effects may have 515 remained undetected. 516 517 Possible long-term drifts due to changes in internal standards, storage, instrumentation and 518 protocols are difficult to estimate. This is done here by checking the consistency of different 519 subsets of the database, for instance when time series can be obtained (such as in the southern 520 Indian Ocean or North Atlantic subpolar gyre), or by comparison with duplicate data analysed 521 in other laboratories, or with other datasets in deep regions commonly sampled. These 522 comparisons are encouraging. On one hand, they suggest that the internal consistency in the 523 database is usually within a (0.03, /0.15)‰ uncertainty for (δ 18 O, /δD). On the other hand, 524 although other datasets sometimes differ by much more with a large scatter between the 17 525 comparisons (with a standard deviation of 0.055‰ for δ 18 O), the average difference (+0.093‰) 526 found with them is close to the change that is applied to the LOCEAN data to report them on 527 the concentration scale (+0.09‰ for δ 18 O analyzed with a salt liner since 2015). Of course, 528 there is still the possibility of errors and biases in subsets that could not be compared in a similar 529 way, such as surface samples collected from ships of opportunity or sailing vessels in the 530 tropics, that could result from different handling of the samples during collection and more 531 uncertain storage conditions. There are also small errors originating from memory effects in the 532 Picarro CRDS runs that could be better corrected and taken into account (Vallet-Coulomb et 533 al., 2021). 534 535 We also illustrated the possibility of using this dataset to investigate ocean variability. Of 536 course, the interest of a data archive is to merge different institutes datasets such as this one, 537 while retaining a similar accuracy. This was attempted in the Global Seawater Oxygen-18 538 Database at GISS , although biases between subsets of this mostly δ 18 O 539 dataset remain at a level that makes the overall analysis of variability difficult to carry. The few 540 comparisons we could do suggest that differences with other datasets are at times large. The 541 effort to correctly adjust for these differences and produce a larger coherent archive is required 542 to get full use of the data collected.  being the same as on Fig. 4. 909