Since there is currently no water vapor product derived from OMPS LP measurements, we instead use the MLS version 5 water vapor product as our ground truth. To prioritize times where SNPP and Aura have closely aligned orbits, we select dates that have

at least one orbit with 60 consecutive co-locations that are within 30 min and within 100 km, and

We apply the recommended MLS data screening criteria except for the period of 15 January–12 February 2022, as it has been shown that the recommended screening criteria filter out many valid profiles affected by Hunga . We limit the MLS water vapor profiles to ≤ 261 hPa. We log-linearly interpolate the water vapor profiles from the MLS pressure grid to OMPS LP's geometric height (11.5–40.5 km in 1 km steps) using the NASA Global Earth Observing System Forward Processing for Instrument Teams GEOS FP-IT; pressures interpolated in time and space to the LP measurements. We limit the altitude range to 11.5–40.5 km because 10.5 km can exceed 261 hPa and the H₂O sensitivity of OMPS LP becomes ∼ 0 above 40.5 km.

We also consider a similar methodology for ACE-FTS and SAGE III/ISS data with recommended screening criteria applied to both datasets. We utilize co-location criteria of within 1 d, within 2° latitude, and within 1113 km longitude (equal to 10° longitude at the equator), consistent with the criteria used in . These data sets are used to investigate whether it is viable to train exclusively on ACE-FTS or SAGE III/ISS data and whether training on a combination of MLS, ACE-FTS, and/or SAGE III/ISS data offers benefits over only training on MLS data.

For each co-located measurement, we construct an input–output pair to be used during NN training. The inputs are comprised of

LP radiances at 554, 596, 654, 720, 728, 824, 917, 929, 943, 956, 970, and 983 nm,

To evaluate the NN's typical accuracy and how well it generalizes to unseen data, we calculate the root mean square error (RMSE) and coefficient of determination (R2) for the validation and test sets. We validate our LP water vapor product by comparing with satellite measurements, balloon-borne measurements, and an assimilation/reanalysis product.

For satellite measurements, we consider the MLS version 5 , SAGE III/ISS version 6 , and ACE-FTS version 5.3 products. For MLS, we consider co-locations within 6 h, while for SAGE III/ISS and ACE-FTS we consider co-locations within 24 h. For all three, we only consider the co-location if it is separated by less than 1000 km and within 2° latitude. When multiple co-locations satisfy these criteria, we use the co-location with the shortest distance. For each of these instruments, we compute both a global median percent difference as well as zonal median percent differences in 5° latitude bins. We found anomalous values in the SAGE III/ISS and ACE-FTS data sets that differ by a factor of up to 10 000 compared with the layers above and below it, even after applying each product's recommended screening criteria. To screen out the most extreme of these unrealistic values, we apply a very conservative 20σ median rejection routine to the data set of percent differences.

We investigate whether our product shows similar properties as the MLS product by performing a multiple linear regression (MLR) on monthly means for a 5° × 5° grid, with proxies for a linear trend, seasonal cycle, quasi-biennial oscillation (QBO), and El Niño Southern Oscillation (ENSO), as these terms explain the majority of stratospheric H₂O variability. For the seasonal cycle term, we also include a phase offset term for lag in months, given that it takes time for the effects to propagate upwards through the stratosphere. For the ENSO term, we regress using the sea surface temperature anomaly with a fitted lag in months, as previous work showed this is necessary to maximize correlation e.g.,. For the QBO term, we use coefficients for the two leading empirical orthogonal functions for the QBO wind time series between January 1956 and February 2025.

To compare with balloon-borne measurements, we consider NOAA Frost Point Hygrometer and Cryogenic Frost point Hygrometer soundings from Boulder, USA; Hilo, USA; Lauder, New Zealand; San José, Costa Rica; Lindenburg, Germany; and Biak, Indonesia. For each sounding, we co-locate satellite measurements within 24 h, 2° latitude, and 1113 km longitude (equal to 10° at the equator). If multiple co-locations meet these criteria, we select the profile that minimizes the distance from the sounding. We calculate the median absolute difference and percent difference over the data set of co-locations between each satellite instrument and the frost point measurements.

We additionally compare with the Modern-Era Retrospective analysis for Research and Applications version 2 (MERRA-2) Stratospheric Composition Reanalysis of Aura MLS (M2-SCREAM) reanalysis product , which assimilates MLS products including H₂O, as this guarantees a co-location for all OMPS LP measurements. To assess whether our methodology blindly memorizes the days it sees during training (where co-locations with MLS are frequent) or if it learns to generalize to days it does not see during training (where co-locations with MLS are less frequent), we compute the mean differences and standard deviation of the differences between the LP product and M2-SCREAM for two subsets of 2021 data: one that contains the days where training data were drawn from, and another that contains the days where no data were used during training.

NOAA-21 OMPS LP has an insufficient period overlapping with MLS measurements (2023–present, with MLS only taking measurements for ∼ 6 d per month since May 2024), inhibiting the use of NOAA-21 OMPS LP data for our NN training methodology. Given the imminent termination of Aura MLS and that the SNPP satellite will presumably cease operations before the end of NOAA-21 or the subsequent JPSS satellites, we test the application of our SNPP-trained model to NOAA-21 OMPS LP measurements to determine whether our model can generalize to future iterations of the same instrument, as found this to be the case for OMPS LP aerosol retrievals.

Figure  shows the global mean and standard deviation of the R2 and RMSE metrics as well as the mean R2 and RMSE as a function of latitude for the NN ensemble when applied to the test set of data not seen during training. Globally, R2 is > 0.7 except between 15.5 and 18.5 km. Figure c shows this reduction in R2 is due to the upper troposphere in the tropics as well as localized reductions in the mid-latitudes, discussed in detail below. When considering only events outside the tropics, the global R2 for 15.5–16.5 km increases to ∼ 0.7, while considering only events within the tropics results in R2 reducing to ∼ 0.55 for these altitudes. In the stratosphere, the global RMSE is < 10 % of the H₂O VMR. Errors increase below 18.5 km as measurements increasingly occur in the troposphere where water vapor VMR increases substantially. The discontinuity in RMSE at 32.5 km is related to the discontinuity in MLS v5 a priori profiles .

Figure c exhibits local minima in R2 in the mid-latitudes. This reduction in R2 is a numerical effect associated with how the NN models are trained combined with the variability of MLS water vapor. The equation for R2 involves the ratio of the residual sum of squares (RSS) to the total sum of squares (TSS). The combination of the MSE loss function and data normalization procedure optimizes the NN models to make predictions such that the errors (RSS) are relatively evenly distributed (homoscedasticity). TSS, on the other hand, is proportional to the variance of the MLS water vapor values. When MLS water vapor variance is reduced, the R2 must also reduce if the agreement between LP and MLS remains constant. These mid-latitude regions are local minima in MLS water vapor variance, and thus they are also local minima in R2 given the expected homoscedasticity of the NN models. However, the absolute minimum in water vapor variance occurs in the SH mid-latitudes, while the absolute minimum in R2 occurs in the NH mid-latitudes. In the NH, there is an additional effect due to reduced water vapor sensitivity at certain combinations of altitudes, SZA, single scattering angle, aerosol, and H₂O. Based on RTM simulations, the spectral features of water vapor at 930–970 nm shift from reductions in radiances in the SH and tropics to increases in radiances in the NH. At the transition between these regimes, water vapor sensitivity is at its minimum, corresponding to the minimum in R2 in the NH mid-latitudes. For more details on this, see Appendix .

Table  summarizes the OMPS LP H₂O product characteristics. The along-track resolution is ∼ 1° for SNPP OMPS LP and ∼ 0.4° for NOAA-21 OMPS LP. Given that the MLS H₂O product has a vertical resolution of 2.8–3.8 km over the pressure range we consider and that the LP vertical field of view is 1.2–1.5 km, the LP H₂O product's vertical resolution may be between 1.2 and 3.8 km. However, more analysis is needed to better estimate the LP H₂O vertical resolution and how it changes as a function of altitude and under different conditions. At present, we conservatively report a vertical resolution ∼ 3.8 km.

Figure  shows a comparison between the NN, MLS, and LP uncertainties for a co-located event in the tropics in January 2019. The NN uncertainty is smaller than the MLS uncertainty by a factor of ∼ 2.5 in the stratosphere, while in the troposphere the NN uncertainty is larger than the MLS uncertainty by a factor of ∼ 3. Consequently, the combined LP uncertainty is dominated by the NN in the troposphere and by MLS in the stratosphere. We show two cases for this combined LP uncertainty, comprising the lower (uncorrelated) and upper (correlated) bounds for the combined NN-MLS uncertainty. They differ by ∼ 0.1 ppm, which is negligible relative to the volume mixing ratios of typical water vapor profiles. The uncertainties reported within the LP H₂O product are calculated assuming no correlation (see Eq. ), as the random and systematic errors of MLS are unlikely to be correlated with LP's systematic errors at the wavelengths considered.

The results of our experiments that omit certain years from the training set reveal that some years are more important than others for achieving a well-generalized model. When omitting 2015–2016, we find that the model is generally unaffected and still performs well during those years. However, when omitting 2024, we find that the model begins producing severely inaccurate predictions by March 2024. This behavior is likely explained by the difference between these considered periods: while 2015–2016 were ordinary years in terms of SWV, the continued evolution of elevated SWV from the Hunga eruption into 2024 is atypical and not represented by the data available during 2014–2023. By including a small fraction of 2024 data during training, we find that the model continues performing accurately up to the present time of writing this manuscript. As the stratosphere returns to pre-Hunga conditions, we expect that the NNs will continue producing accurate retrievals of H₂O, but continued comparisons with other instruments designed to measure water vapor, such as ACE-FTS, will be critical to ensuring that accuracy.

Since the input radiances are at wavelengths affected by aerosols, we analyze the model errors as a function of aerosol extinction reported in the OMPS LP aerosol product. We note that a weak anti-correlation exists between the predicted H₂O VMR and the aerosol extinction at 675 nm in the lower stratosphere, though this is a real phenomenon rather than an artifact of our model. Stratospheric aerosol extinction generally peaks immediately above the tropopause and decreases over the few kilometers above it, while stratospheric H₂O VMR typically is at a minimum immediately above the tropopause due to the cold trap and increases over the few kilometers above it. We find that the NNs' percentage error as a function of aerosol extinction at 675 nm is uncorrelated, indicating that the success of our approach is not dependent on aerosol conditions. See Appendix  for additional details.

In general, the differences with respect to MLS are independent of the presence of upper tropospheric clouds identified by the OMPS LP measurements . However, events affected by polar stratospheric clouds (PSCs) identified by OMPS LP show a median bias around -2 % at most altitudes. When considering the error as a function of distance below the PSC, the median bias can exceed -20 % at 17 km below the PSC, which only occurs for PSCs at ≥ 28.5 km. However, the standard deviation of these errors can be substantial, where it averages around 33 % between 15.5 and 22.5 km, with a maximum of 66 % at 22.5 km. Given this behavior, we recommend that events contaminated by PSCs should not be used for scientific studies; the data product includes quality flags for these events.

We also considered training on co-located ACE-FTS and/or SAGE III/ISS data as an alternative to the co-located MLS data set. Results for these experiments were not satisfactory.

When training exclusively on ACE-FTS or SAGE III/ISS data, we find that the resulting NN models do not properly generalize to unseen data, indicating that these data sets are not sufficient to solve this problem. Part of this may be explained by the typically significant measurement time differences between LP and the other instruments; LP measures in the early afternoon, while ACE-FTS and SAGE III/ISS measure at sunrise/sunset, and these times only coincide for select geolocations depending on the time of year. However, the number of co-locations seems to be a more limiting factor. When restricting the MLS data set to similar sizes as the ACE-FTS and SAGE III/ISS data sets, we find that the resulting performance is poor. Given the success when using the MLS-LP data set of ∼ 1 million co-locations but the failure when considering tens of thousands of co-locations, these results emphasize that our methodology relies on a large data set of co-located profiles.

Additionally, when including ACE-FTS and/or SAGE III/ISS data alongside the MLS data, we find significantly degraded performance, regardless of whether or not the ACE-FTS and SAGE III/ISS data were de-biased with respect to MLS. This is likely attributable to the variances between MLS, ACE-FTS, and SAGE III/ISS being on the order of the natural variability of water vapor, which inhibits NN learning.

Despite our negative results when training on ACE-FTS and/or SAGE III/ISS data, we cannot rule out that alternative ML approaches not considered here could utilize ACE-FTS and/or SAGE III/ISS data to derive water vapor profiles from LP radiances.

Figure  summarizes the global median percent differences between LP stratospheric H₂O profiles and co-located MLS, ACE-FTS, and SAGE III/ISS profiles. For this comparison, we filter all LP tropospheric measurements by using the nearest co-located tropopause altitude reported in the GEOS FP-IT product. The error bars show the standard error of the median, which is generally negligible except at low altitudes for ACE-FTS and at high altitudes for SAGE III/ISS. Where LP detects a cloud, we exclude any measurements at or below the cloud top, though we find that this criterion does not significantly alter the results. Note that for the MLS comparisons we include all co-located data, including those used during training; this choice does not bias the results, as discussed later in Sect. .

Differences with respect to MLS are less than 2 % at all altitudes ≥ 14.5 km, with a maximum difference of 4.1 % at 11.5 km. When considering only 2025 data, the most extreme difference is 7.7 % at 13.5 km, with a typical difference of ∼ 5 % below 22 km and < 2 % above 25 km. Given that we trained on MLS data, this close agreement is expected and shows the model has learned a good approximation to retrieve MLS-like water vapor profiles.

When comparing with ACE-FTS, we find agreement within 10 % except between 11.5 and 13.5 km, where differences can reach up to 19.3 %. In general, the differences increase with altitude from -19.3 % at 11.5 km up to 8.8 % at 40.5 km.

LP's differences with respect to SAGE III/ISS are generally around 6 % or less. Between 11.5 and 13.5 km, differences can reach up to 16.9 %. Given that a similar increase in differences at these altitudes is seen when comparing with both ACE-FTS and SAGE III/ISS but not with MLS, this suggests that either ACE-FTS and SAGE III/ISS are biased high in this regime, or MLS is biased low in this regime and the LP product has inherited this bias.

Figures  and  show the median percent differences and standard error of the median, respectively, in 5° latitudinal bins for the comparisons with MLS, ACE-FTS, and SAGE III/ISS. The results are generally consistent with those shown in Fig. . The notable exceptions occur at the lowest altitudes. For the comparisons with MLS, differences can reach up to 8 % between 11.5 and 13.5 km just outside of the tropics. For ACE-FTS, the deviations at these altitudes can exceed 21 %, but notably the standard error of the median is typically ∼ 11 % in this region, suggesting that this low bias may not be as substantial as it appears. However, comparisons with SAGE III/ISS also show this low bias at these altitudes, where the standard error is negligible. Together, this suggests that the LP product has a slight systematic low bias at 11.5–13.5 km just outside the tropics, but at latitudes ≥ 45°, the agreement with MLS suggests the biases when comparing with ACE-FTS and SAGE III/ISS are related to statistical differences between those products and MLS.

Figure  shows the “tape recorder” of alternating positive and negative anomalies in H₂O VMR, primarily attributable to seasonal changes in H₂O. For this plot, we subtract the pre-Hunga mean profile from each daily zonal mean to produce the daily anomaly. Before 2025, the OMPS LP and MLS tape recorders show excellent agreement throughout the stratosphere, with OMPS LP correctly capturing the increase in H₂O due to the Hunga eruption. For ∼ 3 months in early 2025, OMPS LP shows a positive bias > 1.0 ppm above 30 km that is not seen in the corresponding MLS data (Fig. c). This bias is likely due to the weak H₂O sensitivity at these altitudes, which inhibits the NNs' ability to reliably infer the H₂O VMR at these altitudes. Since the NNs correctly infer the H₂O at these altitudes before 2025, it suggests that the pre-2025 data were successfully predicted based on the shape of profiles at the lower altitudes that have sensitivity. With an absence of 2025 data in the training set, the NN model relies on profiles from the training data, primarily those influenced by elevated H₂O associated with the Hunga eruption. In 2025, the NNs perform reasonably well below 30 km, suggesting that there is sufficient sensitivity for the determined approximation to remain accurate when applied to unseen data, though there is a slight overestimation (∼ 0.25 ppm) in mid-2025 between 25 and 30 km. We therefore advise that users exercise caution when using the OMPS LP H₂O product above 30 km in the tropics. Additionally, as the excess SWV from Hunga continues to evolve, it is possible that those conditions will be different enough from the training data that the NNs' predictions become inaccurate; we will continue to monitor their performance to ensure the LP NN-based algorithm continues to provide reasonable H₂O profiles.

The NNs are largely trained on normal conditions. While Hunga is an anomaly, its impact has lasted years and can be seen globally, and thus the NNs are exposed to various Hunga-affected conditions. Since our data selection process does not consider H₂O concentration or profile states, rare short-lived events are severely undersampled relative to nominal conditions. Despite this, we find that the NN models can still retrieve reasonably accurate profiles for rare cases not seen during training. Figure  shows the H₂O profiles retrieved by LP and MLS for an event over the Pacific Ocean on 8 January 2020. This event exhibits a water vapor enhancement in the lower stratosphere due to the intense wildfires in Australia during December 2019–January 2020. Pyrocumulonimbus clouds carried moist, tropospheric air into the lower stratosphere, where it was then transported eastward by winds. Despite that these H₂O concentrations were seen in < 0.045 % of the training data set, the NNs are able to identify this enhancement. The LP and MLS VMRs of this enhancement differ by up to a factor of two, but it is unclear whether this is due to LP underestimating the VMR, MLS overestimating the VMR, a combination of both, or differences in measured air masses.

In general, the results of our MLR analysis show similar behavior between the LP and MLS products, but the fitted coefficients for the LP data tend to be less than the corresponding MLS coefficients. The main exception to this is the seasonal phase offset, where both products closely agree. Figure  shows an example of the fitted coefficients for the seasonal amplitude at 14.5 km. The South Asian monsoon stands out clearly in both panels, though the LP product's fitted amplitudes for this region are around 1 ppm less than the corresponding fits for MLS.

Regarding water vapor trends, the LP product generally shows weaker trends in the stratosphere when compared with MLS. In some locations (particularly south and southeast Asia, central Africa, and central America), the LP product shows greater trends in the upper troposphere. Where both products show a trend of increasing H₂O, LP tends to show a lesser trend than MLS (Fig. ).

Figure  compares the OMPS LP and M2-SCREAM H₂O products for the year 2021. Note that M2-SCREAM assimilates MLS v4.2, which is biased high for water vapor, while LP is trained on MLS v5, resulting in a persistent ∼ 0.5 ppmv bias between the products. Days in which data were (Fig. a) or were not (Fig. b) included during training are shown separately but look almost identical, highlighting that the NNs' predictions are equally accurate whether or not they saw data from that day during training. Additionally, the standard deviation of the differences between OMPS LP and M2-SCREAM are consistently less than the standard deviation among OMPS LP or M2-SCREAM profiles, indicating that the OMPS LP product is more accurate than the natural variability of H₂O. Overall, our results suggest that the NN predictions are in good agreement with M2-SCREAM for data not seen during training.

Figure  shows the median differences between satellite instruments (MLS, SAGE III/ISS, ACE-FTS, and OMPS LP) and the frost point hygrometer soundings from the six stations considered (see Sect. ). The LP product agrees with the frost point measurements within 0.3 ppmv and within 10 % between 16.5 and 27.5 km; this is in close agreement with the MLS results, which is expected given that we trained on MLS data. The only notable difference in agreement between MLS and LP is that the LP product shows a slightly reduced bias between 16.5 and 21.5 km. Like in , we find that the satellite instruments show a dry bias in the upper troposphere compared to the frost point measurements, which may be due to spatiotemporal variability between the co-located measurements and/or reduced data quality in this regime.

Paralleling the SNPP comparisons in Sect. , Figs.  and  show respectively the global median percent differences and the 5° zonal median percent differences between NOAA-21 OMPS LP and MLS, ACE-FTS, and SAGE III/ISS. For global comparisons, NOAA-21 OMPS LP shows a persistent ∼ 5 % offset with respect to the corresponding SNPP comparisons at all altitudes. A similar low bias is seen in the NOAA-21 OMPS LP aerosol data, suggesting that this bias is attributable to differences in radiances between OMPS LP on SNPP and NOAA-21. Given our methodology, it is unclear whether the NNs have learned to implicitly account for a bias in the SNPP radiances or if the problem is related to the calibration of NOAA-21 radiances. However, this bias is not strictly a -5 % shift for all conditions; the zonal comparisons show that the tropics exhibit a positive bias not seen in the corresponding SNPP comparisons. Further investigation is necessary to understand the cause of these biases. If the origin of these biases is not able to be determined, they could be addressed via a soft calibration approach.