The observations of TTL water vapor are from the Earth Observing System (EOS) Aura Microwave Limb Sounder (MLS) . The MLS instrument has obtained continuous high-quality global observations of water vapor in the upper troposphere and stratosphere since August 2004. The data are available from https://mls.jpl.nasa.gov/.

Here, we use MLS version 4.2 level-2 water vapor retrievals from August 2004 to December 2016. The daily water vapor mixing ratio measurements are binned and averaged to produce monthly data on a 4∘× 8∘ latitude and longitude grid with the quality control following the instruction in . We focus on the interannual anomalies of water vapor from 30∘ N to 30∘ S at 100 hPa. Throughout this paper, the interannual anomalies at each grid point are calculated by subtracting the average annual cycle at that grid point.

We also use simulations of TTL water vapor from the Goddard Earth Observing System Chemistry Climate Model (GEOSCCM) in this study. The state-of-the-art GEOSCCM includes the GEOS-5 atmospheric general circulation model with a single-moment cloud microphysics scheme and the StratChem stratospheric chemical mechanism . The GEOSCCM simulation provides long-term simulations of temperature, water vapor, horizontal winds, diabatic heating rates, and convective ice content with a resolution of 2∘×2.5∘ in latitude and longitude on 72 vertical model levels, up to 0.01 hPa.

In this study, we investigate water vapor simulated by the GEOSCCM in the TTL during model years corresponding to the MLS period. As these simulations are from a free-running model, climate variability in the model is not synchronous with that in the observations, so the comparisons with MLS observations are done statistically – using regression models (discussed below).

We also produce simulations of TTL water vapor using a domain-filling forward trajectory model, which has been used in previous work to reproduce water vapor, ozone, and carbon monoxide anomalies in the TTL and lower stratosphere .

This model uses Bowman's trajectory code . The parcels are driven by 6-hourly horizontal winds and total diabatic heating rates from either reanalysis datasets or from the GEOSCCM. When comparing to MLS data, we use trajectory runs driven by two reanalysis datasets: the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-Interim reanalysis (ERAi) and NASA's Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) . When comparing to GEOSCCM output, we drive the trajectory model with meteorological fields from the GEOSCCM.

In all simulations, 1350 parcels are released every day from January 2000 to December 2016 on an equal area grid from 60∘ N to 60∘ S. The parcels are released on the 370 K isentropic level, which is above the zero net diabatic heating level over the tropics (∼ 355–360 K) but below the cold point (∼ 375–380 K). Each parcel travels forward following the horizontal winds and diabatic heating rate. Once a parcel has a pressure larger than 250 hPa, it is regarded as having descended back into the troposphere and is removed from the model.

Each parcel is initialized with a water vapor mixing ratio of 200 parts per million by volume (ppmv). Along the trajectory, a parcel will immediately be dehydrated to saturation once its water vapor mixing ratio exceeds a predetermined saturation threshold, which is 100 % in this study. The saturated water vapor mixing ratio is obtained from the thermodynamic equation with respect to ice based on temperatures from reanalyses or the GEOSCCM. The production of water vapor from methane oxidation is also included in these trajectory model runs but it has very little effect on water vapor in the TTL .

The water vapor mixing ratio from the trajectory model is gridded into 4∘×8∘ bins, just as the MLS data were. In the vertical, the trajectory output is binned by averaging the parcels in a pressure range around each MLS or GEOSCCM level. When comparing to MLS, the gridded water vapor mixing ratio is then re-averaged using the MLS averaging kernels following the instruction from . When doing this kernel averaging, grid boxes with no trajectory parcels (mostly at low altitudes) are filled with monthly water vapor mixing ratios from the reanalyses (ERAi and MERRA-2). Sensitivity tests confirm that changing water vapor mixing ratio from the reanalyses has no impact on the spatial distribution of the anomalies of TTL water vapor that are the focus of this paper.

We also use estimates of convective cloud occurrence produced by combining geostationary infrared satellite imagery and microwave rainfall measurements . The data have a horizontal resolution of 0.25∘×0.25∘, a temporal resolution of 3 h, and cover the period from 2005 to 2016. In this paper, we use the cloud-top height and cloud-top potential temperature to estimate the convective cloud occurrence frequency in the TTL, which we take to be an indicator of convective influence on the TTL. These data are available from https://bocachica.arc.nasa.gov/~lpfister/cloudtop/.

Figure  shows monthly and tropical averaged 100 hPa water vapor anomalies from MLS observations and from trajectory model runs driven by meteorology from ERAi (traj_ERAi) and MERRA-2 (traj_MERRA2). Similar to the results in at 82 hPa, there is good agreement between the observations and trajectory models at 100 hPa.

also showed that we can fit tropical average anomalies of 82 hPa water vapor with a simple linear model: H2O=a⋅BDC+b⋅QBO+c⋅ΔT+r, where BDC, QBO, and ΔT are indices representing the strength of the Brewer–Dobson circulation, the phase of the QBO, and the tropospheric temperature anomalies of the tropical climate system, respectively. verified that this approach is also valid in chemistry–climate models.

To gain additional physical insight into the regression result, in this paper, we perform a similar multivariable regression but at individual grid points in the TTL: H2O(xi,yj)=a(xi,yj)⋅BDC+b(xi,yj)⋅QBO+c(xi,yj)⋅ΔT+r(xi,yj).

Here, H2O(xi,yj) represents the H2O anomaly time series at 100 hPa in a grid box centered at longitude xi and latitude yj. The coefficients a, b, and c, as well as the residual term r, are also functions of latitude and longitude.

The regressors in Eq. () are the same tropical average time series used in : BDC is the Brewer–Dobson circulation index – here, we use the tropical averaged diabatic heating rate anomaly at 82 hPa, with units of Kday-1. QBO is a quasi-biennial oscillation index, and here we use the standardized monthly and zonally averaged equatorial zonal wind anomaly at 50 hPa, with units of ms-1. ΔT is the tropical averaged tropospheric temperature anomaly at 500 hPa, with units of Kelvin. Because these regressors are tropical average values, they do not vary with location. The QBO index is lagged by 2 months in the regression because the phase of the QBO takes time to impact the TTL temperature and then the water vapor at 100 hPa . There is no lag for the BDC and ΔT indices in this study.

We first analyze MLS water vapor observations. We run the regression on these observations twice: once using BDC and ΔT regressors from the ERAi reanalysis and again using regressors from the MERRA-2 reanalysis. The QBO index is the same in both regressions (we use observations downloaded from http://www.cpc.ncep.noaa.gov/data/indices/qbo.u50.index).

The BDC coefficients (Fig. a and d) are negative over the tropics, consistent with the idea that an enhanced Brewer–Dobson circulation cools the TTL and reduces water vapor . The QBO coefficients (Fig. a and d) are positive over almost all of the tropics, as the positive phase of QBO tends to decrease the upwelling in the TTL, thereby warming it .

We also run the regression on water vapor simulated by the trajectory model. We use BDC and ΔT regressors from the same reanalysis used to drive each trajectory model (i.e., we use ERAi fields to regress the ERAi-driven trajectory model); the QBO index is always from the NCEP observations.

The BDC coefficients from regression of the trajectory models (Fig. b and e) agree well with the coefficients from the regressions of the MLS observations (Fig. a and d). The average BDC coefficient in the MLS/ERAi regression is -2.6 ppmv(Kday-1)-1 (Fig. a), in good agreement with the average value from the accompanying trajectory regression, -2.4 ppmv(Kday-1)-1 (Fig. b). The grid-point-by-grid-point scatter plot (Fig. 2c) demonstrates this agreement in more detail.

The average BDC coefficient in the MLS/MERRA-2 regression is -2.3 ppmv(Kday-1)-1 (Fig. d). The average coefficient from the accompanying trajectory regression is -1.4 ppmv(Kday-1)-1 (Fig. e). This larger difference stems from what appears to be problems in the MERRA-2 heating rates. These heating rates disagree significantly with those from both ERAi as well as the original MERRA. Thus, we put more weight on the ERAi results for this coefficient and conclude that the BDC response is well simulated by the trajectory model.

The QBO coefficients from the regressions of the trajectory models are shown in Fig. b and e; grid-point-by-grid-point scatter plots are shown in Fig. c and f. For the MLS/ERAi comparison (Fig. c), the average QBO coefficients are 0.084 ppmv(ms-1)-1 for the MLS and 0.075 ppmv(ms-1)-1 for the trajectory model; for the MLS/MERRA2 comparison (Fig. f), the average coefficients are 0.14 ppmv(ms-1)-1 for the MLS and 0.15 ppmv(ms-1)-1 for the trajectory model. As with the BDC comparison, we conclude that the trajectory model does a good job reproducing the regressions of the MLS data.

Overall, the trajectory model accurately captures the impact of the BDC and QBO on TTL water vapor for both the tropical average and the spatial distribution. This supports the hypothesis that these processes mainly influence TTL water vapor by varying large-scale TTL temperatures and transport , which we expect the trajectory model to reproduce. For this reason, we will not focus any further on these coefficients.

Coefficients of ΔT from the MLS regressions are mostly positive, with large increases over the tropical warm pool region (TWP) and Indian Ocean (Fig. a and d), indicating that warming of the tropical troposphere increases TTL water vapor mixing ratio there. Over the central equatorial Pacific (CEP), however, a warming troposphere decreases TTL water vapor.

The decrease in TTL water vapor in the CEP is not entirely unexpected. TTL temperatures are usually coldest – and water vapor a minimum – above the convection maximum in the TWP. As ΔT increases in response to an El Niño event, this convective maximum, and its associated TTL cold pool, shifts eastward from the TWP to the CEP (corresponding to the shift from Fig. a to b) . Changes in TTL water vapor are expected to mirror this, with increases in water vapor in the TWP and decreases in the CEP as ΔT increases.

Both the MLS and trajectory model regressions show this dipole pattern. However, the MLS regressions yield ΔT coefficients that are systematically higher than those found in the trajectory model regressions throughout the tropics (Fig. c and f). For the MLS/ERAi comparison (Fig. c), the average coefficients are 0.43 ppmvK-1 for the MLS and 0.28 ppmvK-1 for the trajectory model; for the MLS/MERRA-2 comparison (Fig. f), the average coefficients are 0.20 ppmvK-1 for the MLS and 0.05 ppmvK-1 for the trajectory model. We have also done regressions using tropical average values (using Eq. 1, similar to what was done in ) and find that the ΔT coefficients from MLS and trajectory models are statistically different with probabilities of 85 and 70 % for ERAi and MERRA-2, respectively.

We hypothesize that the evaporation of convective ice accounts for the difference between the ΔT coefficients in the MLS and trajectory model regressions. As convection moves eastward during an El Niño event, there is an accompanying increase in convective ice in the TTL as seen in Fig. a and b; see also , where it evaporates and hydrates the TTL. The moistening from evaporation spreads throughout the tropics and increases the water vapor everywhere. The trajectory model, which does not include this process, simulates a smaller increase in water vapor, leading to smaller ΔT coefficients (Fig. c and f).

In support of this, in Fig. a, we show that the tropical average convective cloud occurrence frequency anomalies in the lower stratosphere increase with ΔT. This is consistent with the hypothesis that, as ΔT increases, we should also see an increase in evaporation of cloud ice.

To gain additional confidence in our hypothesis that evaporation of convective ice plays a role in the TTL water budget, we perform a parallel analysis with the GEOSCCM, a model where evaporation of convective ice is known to add water to the TTL . To do this, we run a regression on the GEOSCCM 100 hPa water vapor fields as well as on water vapor simulated by a trajectory model driven by the GEOSCCM meteorology.

Figure g and h show the spatial distributions of the BDC coefficients from the GEOSCCM and the corresponding trajectory model. This comparison is analogous to the comparison of the regressions on the MLS data and trajectory models driven by reanalyses. The coefficients are similar to each other and to the MLS regressions, suggesting that the GEOSCCM is accurately simulating the impact of BDC changes on TTL water vapor. The influence of the QBO in the version of the GEOSCCM analyzed here does not extend into the TTL, so we have not included a QBO term in the regression and there are consequently no GEOSCCM QBO coefficients in Fig. .

Before we discuss the ΔT coefficients, it is worth pointing out that the GEOSCCM has realistic El Niño–Southern Oscillation (ENSO) variability in TTL temperatures and convective ice. Figure c and d show that the monthly convective cloud ice water content (IWC) anomalies at 118 hPa and cold anomalies at 100 hPa in the GEOSCCM shift eastward as ΔT warms from a cold phase (Fig. c) to a warm phase (Fig. d), just as they did in observations (Fig. a and b).

The ΔT coefficient fields from the GEOSCCM and associated trajectory model regressions (Fig. g and h) show the same structural differences as do the ΔT coefficients from the MLS and accompanying trajectory model regressions – that the ΔT coefficient is larger in the GEOSCCM regression than in the trajectory model regression – as in the observations; the tropical average ΔT coefficients from the GEOSCCM and trajectory model are significantly different at the 85 % confidence level.

In the last section, we hypothesized that this difference in the coefficients was due to evaporation of convective ice in the MLS data, a process not included in the trajectory model. To directly test this hypothesis in the GEOSCCM, we run a second version of the trajectory model that includes the evaporation of convective ice from GEOSCCM, referred to hereafter as traj_CCM_ice. To do this, we use the 6-hourly three-dimensional convective cloud IWC field from GEOSCCM and linearly interpolate it to each parcel's position at every time step. We then assume instantaneous and complete evaporation of this ice into the parcel by adding the IWC to the parcel's water vapor, although we do not let parcels' water vapor exceed 100 % relative humidity with respect to ice. This is the same procedure used to simulate convective ice evaporation by .

We then run the regression on the traj_CCM_ice's water vapor field. The scatter plot of GEOSCCM vs. traj_CCM_ice BDC coefficients (Fig. k) shows larger scatter than the comparison without ice (Fig. i). The increase in scatter is likely the result of the crudeness of our microphysical assumptions, particularly the assumption that convective ice evaporates instantaneously. However, the comparison between the tropical average GEOSCCM BDC coefficient, -6.2 ppmv(Kday-1)-1, and those from the trajectory models, -5.8 and -6.9 ppmv(Kday-1)-1 without and with convective ice evaporation, respectively, is similar.

The scatter plot of GEOSCCM vs. traj_CCM_ice ΔT coefficients (Fig. k) similarly shows larger scatter than the comparison without ice (Fig. i). Adding ice does, however, increase the average ΔT coefficient (seen by comparing Fig. i to k), from 0.16 to 0.32 ppmvK-1, bringing the trajectory model into closer agreement with the GEOSCCM, which has a corresponding value of 0.31 ppmvK-1. There are also some interesting changes in the spatial pattern of the traj_CCM_ice (Fig. j). For example, negative ΔT coefficients appear in the TWP and Indonesia in the traj_CCM_ice regression; the cause of this is unknown but also is likely linked to the trajectory model's ice evaporation assumption.

We showed in the previous section that the convective cloud occurrence frequency in the TTL increased as ΔT increased (Fig. a) and we also see that the GEOSCCM simulates a similar correlation between convective cloud IWC and ΔT (Fig. b). While these are not exactly the same quantity, they show a consistency that provides confidence that the behavior of the model is realistic.

Finally, to quantify convective ice evaporation, we calculate the evaporation rate of convective ice at 100 hPa in the trajectory model. To do this, we save the amount of water added to each parcel by ice evaporation in every time step. We then bin and average the amount evaporated to come up with the distribution of the amount evaporated per day. Note that much of this water added will be lost in subsequent dehydration events, so this does not represent net water added to the stratosphere.

Figure c shows that the tropical average evaporation rate of convective ice also increases with ΔT, which provides further evidence that the difference in ΔT coefficients between the GEOSCCM and the trajectory model is due to evaporation of convective ice. Figure  shows the distribution of the monthly averaged evaporation rate during ENSO-like cold and warm phases in the GEOSCCM. We see that, as ΔT increases and we transition from a cold to warm phase of variability, the location of ice evaporation shifts from the TWP and Indian Ocean to the CEP. This is consistent with the analysis of .