An extensive overview of the ORACLES project is presented by Redemann et al. (2021). Aspects of the project and ORACLES measurements relevant to this study are outlined here. In situ sampling aboard the NASA P-3 Orion aircraft covered the southeast Atlantic MBL and lower FT during periods of high aerosol loading. The high BBA concentrations in the Southern Hemisphere spring are due to BBA-loaded air in the African planetary boundary layer being carried out over the southeast Atlantic by lower-FT easterly flow (Garstang et al., 1996; Adebiyi and Zuidema, 2016). This air settles over the southeast Atlantic cloud deck due to large-scale subsidence, and that air may then entrain into the MBL. The large-scale subsidence also plays a role in the strong-inversion-topped MBLs of the region, which transition to cumulus-coupled MBLs toward the Equator as sea surface temperatures (SSTs) increase (Wood, 2012, and references therein).

Sampling flight tracks are shown in Fig. 2. Sampling took place over three intensive observation periods (IOPs): 31 August–25 September 2016 (14 flights), 12 August–2 September 2017 (12 flights), and 27 September–23 October 2018 (13 flights). Flights were typically every 2–3 d, 7–9 h in duration, and within 07:00 to 17:30 LT. Vertical profiles spanned the altitude range 100 m to 7 km, covering the MBL and the region of the FT where biomass burning plumes were present.

The total water mixing ratio (q) and heavy isotope ratio RD=qHDO/qH2O were measured with the Water Isotope System for Precipitation and Entrainment Research (WISPER) instrument package, which centered on two Picarro isotopic gas analyzers (models L-2120fxi and L-2120i). The isotope ratios are reported in delta notation and the deviation of RD from that of Vienna Standard Mean Ocean Water: δD≡(RD/RD,SMOW-1). A detailed description of those data and calibration methods is given by Henze et al. (2022). Potential temperature (θ) was computed from measurements of temperature (T, Rosemount 102-type non-deiced probe) and static and dynamic (ram) pressures. Refractory black carbon (rBC) and carbon monoxide (CO) are used as indicators of BBA loading in this study. rBC concentration from 53 to 524 nm mass equivalent diameter and adjusted to standard temperature and pressure was measured using a Droplet Measurement Technologies single-particle soot photometer (Stephens et al., 2003; Schwarz et al., 2006). CO concentrations were measured by an ABB–Los Gatos Research CO/CO2/H2O analyzer (Liu et al., 2017). Cloud condensation nuclei (CCN) concentrations at 0.3 % supersaturation were measured by a Droplet Measurement Technologies CCN-100 continuous-flow streamwise thermal-gradient CCN chamber (Roberts and Nenes, 2005). Sub-cloud CCN measurements were used to estimate the total quantity of hygroscopic aerosols present in sampled MBLs. Measurements for q, δD, rBC, and CCN were sampled from a forward-facing solid diffuser inlet outside the P-3 cabin in front of the wing (McNaughton et al., 2007), operated by the Hawaii Group for Environmental Aerosol Research and maintained at near-isokinetic flow. CO measurements were made from a rear-facing Rosemount inlet probe.

For each flight, time intervals of vertical profiles were flagged manually. To maximize the number of data used in this study, profiles were liberally identified as any flight segment with a progressively upward or downward trajectory from near the surface (100–200 m) to at least 2500 m (e.g., Fig. 2). For each profile, vertical profiles of T, θ, and relative humidity (RH) averaged into 50 m vertical bins were used to identify the well-mixed layer top and MBL capping inversion bottom as described below. Data below the well-mixed layer top are flagged as SCL, data between the mixed layer top and inversion bottom are flagged as CL, and data above the inversion are flagged as FT. For the FT analysis, it was desired to highlight hydrologic histories of distinct air masses from the African continent (e.g., different BBA plumes), rather than any subsequent mixing between them, as they advect westward to the southeast Atlantic. Therefore, once FT segments of each profile were identified, the portions of those segments determined to be distinct air masses were isolated (rather than intermediary mixtures between them) and used for the analysis in Sect. 3.2. This was done using a pseudo-conservative variable approach described in Appendix B.

Vertical profiles of T, θ, and RH were chosen over liquid water content to flag the CL since the algorithm would also capture profiles through CLs with broken clouds, where by chance the P-3 avoided clouds during vertical ascent or descent. Mixed layer tops were roughly taken to be the height at which θ increased by 0.2 K from its 100–300 m mean. A potentially more rigorous identification of cloud base using lifting condensation levels was not used since data below 100 m were not always available. The MBLs in the study region were almost always characterized by a capping temperature inversion and a peak in RH at their top. Inversion bottoms were identified as the highest altitude below 2.5 km where a temperature minimum coincided with an RH maximum to within 150 m. The temperature minimum relative peak had to be at least 1 K and occur within a vertical region of 400 m or less. The RH absolute peak had to be greater than 80 % and occur within a vertical region of 300 m or less. Figure 2 shows an example flight using the above methods to identify the SCL, CL, and FT. Using this method, the majority of data are flagged as one of the three regions and are able to be used for analysis.

Most flights involved modules to directly target the CL. These included (1) level legs through the cloud layer at a single altitude (LLs), typically lasting 5–15 min, and (2) sawtooth patterns through cloud layer (SAWs), also typically 5–15 min, where individual saw teeth spanned the full vertical extent of the cloud layer (Fig. 2). Both LL and SAW modules were previously flagged (e.g., Diamond et al., 2018). Where available, these modules are preferred over the CL flagging in Sect. 2.3. For LLs, data were averaged into 3 min blocks. At an aircraft horizontal speed of 130 ms-1, each block represents almost 25 km of horizontal distance, which is half the horizontal grid spacing for higher-resolution general circulation model (GCM) simulations. SAW patterns included periodic portions above and below the CL. To ensure only the CL portions are used, times where RH drops below 99 % are discarded. The remaining data are averaged into 3 min blocks. Only LLs and SAWs below 2.5 km are used.

The δD of an air mass relative to its water vapor mixing ratio, q, will assume different relationships depending on the hydrologic processes the air mass experiences. Probability distribution functions (PDFs) in (q, δD) space can be compared to theoretical curves (e.g., Fig. 3) to constrain these processes. A survey discussion on the theory is presented in Noone (2012) (note that “q” in Noone, 2012, denotes water vapor, while here it denotes total water). Some key components are presented here to facilitate discussion below.

Figure 3 shows the (q, δD) value of asymptotic end-members for dry FT air and pure surface evaporation from land and ocean sources. Land surface evapotranspiration (ET) may have higher δD values than ocean evaporation since transpiration does not fractionate heavy isotopes. The mixing trajectories (black, gray) show the (q, δD) relationship for air where moisture from the respective evaporation source is diluted with FT air to varying degrees. In general, air that has not experienced condensation processes may be modeled as a mixture of two or all of these end-members. The blue circle in Fig. 3 shows an example MBL placement on the mixing line assuming 80 % relative humidity with respect to a sea surface temperature of 293 K (alternatively, the blue circle could represent planetary boundary layer air over the African continent). Further entrainment of FT air into the MBL (or conversely detrainment of MBL air into the FT) would continue along the mixing trajectory. On the other hand, precipitation preferentially removes heavy isotopes, resulting in a steeper trajectory. The dashed black curve in Fig. 3 shows the extreme case of 100 % condensate removal (Rayleigh process).

In the absence of rain droplet re-evaporation, the mixing and Rayleigh processes are limiting cases that bound the region of (q, δD) space in which we expect measurements to lie. If data lie directly on the mixing curve, we hypothesize that mixing was the primary process in the recent history of the sampled air (and likewise for Rayleigh). If data lie between the curves, then the air has a more complex history. In this study, we assume that the most frequent hydrologic history occurring for the BBA-laden air is that of a convective plume model with no entrainment and 100 % precipitation efficiency (e.g., Rayleigh), which detrains at various altitudes. A convecting air mass undergoing a Rayleigh process would first follow the dashed curve from the blue circle but then be diverted along the “convective detrainment” mixing curve (solid line) upon reaching a detrainment level, when the air mass begins to dilute with FT air. We use this model to estimate total precipitation amounts in sampled FT air based on their in situ (q, δD) values (as described further in Sect. 3.3 and Fig. 9).

To support our use of the above convective detraining plume, we also compare temperature and altitude measurements from the P-3 to detrainment altitudes and temperatures estimated from the in situ (q, δD) values and our model (Sect. 3.2). The algorithm used has the following features: (1) values of surface q, T, and elevation (e.g., starting values for air lying at the blue circle in Fig. 3) are estimated from Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2; Global Modeling and Assimilation Office, 2015a, b), reanalysis and used as initial conditions. (2) A temperature profile is computed from these surface values, assuming a dry adiabatic lapse rate up to a lifting condensation level and a moist lapse rate above. (3) For each q value along the Rayleigh trajectory, a dew point is calculated as the detrainment temperature and the height at which the dew point intersects the temperature profile is the detrainment height. Figure 3 shows a characteristic case with the dew point and height from this computation for the example MBL undergoing convection and detraining at q=6 gkg-1.

One objective of this study is to build a characterization of rainfall and evapotranspiration histories for FT air sampled during ORACLES using the humidity and isotopic data. For comparison, a characterization using air mass back trajectories and satellite products of monthly mean rainfall and evapotranspiration is also developed. The goal of this additional characterization is to provide a rough idea of the relative differences in precipitation and evapotranspiration which air experienced (on average) across sampling periods and across BBA concentrations. For this, 6 d back trajectories were computed with HYSPLIT (Draxler and Hess, 1997; Draxler, 1999; Warner, 2018) via the PySPLIT Python module (Warner, 2018) for FT air masses sampled by the P-3 (HYSPLIT runs used Global Data Assimilation System output on pressure levels at 1∘×1∘ resolution and 3 h frequency and outputs air parcel positions at 1 h frequency). Figure 4 shows the trajectories separated by sampling period and BBA-loading histories inferred from their in situ CO concentrations. Air masses where the in situ CO concentrations were higher vs. lower than the FT median for that sampling period are referred to hereafter as high BBA vs. low BBA, respectively. It is noteworthy that a large fraction of the “low-BBA” air in the FT had CO concentrations double that typically measured in the SCL during the campaign (not shown).

For 2016, high BBA originates from the highest latitudes out of the 3 years. For 2017, both high-BBA air and low-BBA air originate further north and their latitudes tend to overlap. For 2018, the low-BBA trajectories are on average displaced 3∘ north of the high-BBA trajectories, toward regions of higher rainfall. The high-BBA air originates from a region similar to that of 2016, but whereas the 2016 high-BBA air flows west off the continent, the 2018 high-BBA air tends to flow northwest. The low-BBA air in 2016 is distinct as it often originated from higher latitudes of the Atlantic Ocean.

Figure 5 provides a first coarse assessment of the possibility of precipitation formation underway across the air mass groups in Fig. 4. It shows PDFs of monthly mean rainfall values along the back trajectories for each group, and distinct differences are apparent. In 2016, high-BBA trajectories show the lowest monthly mean rainfall. The frequency along 2016 trajectories where satellite GPM monthly mean precipitation rates are greater than 0.25 mmh-1 is an order of magnitude less than that for the other 2 years. For 2017, there is no significant difference in the PDFs for high BBA vs. low BBA, while in 2018 the low-BBA trajectories are 2× to 3× more likely than high BBA to pass over regions with precipitation >0.25 mmh-1. As will be shown further below, this characterization aligns with conclusions drawn from the isotope analysis. However, two sources of uncertainty in this first characterization are acknowledged: first, it is not clear if precipitation did really form at the time of the air mass overpass over a given region, and second, it is not clear if the air mass being tracked is really the one involved in precipitation formation.

In addition to precipitation, there is variation in mean evapotranspiration (ET) in the regions from which the trajectories originate. Because ET contribution can alter the isotopic content of an air mass, its expected fraction of moisture contribution for each sampling period was estimated using a back-of-the-envelope calculation outlined in Appendix A and the MODIS Terra Version 6 Evapotranspiration/Latent Heat Flux product (MOD16A2), taking the mean and standard deviation over the boxed regions (regions bound by solid and dashed black squares) of Fig. 4. The results are shown in Table 1 and show expected ET contributions of roughly 5 %–15 % at higher latitudes and 15 %–30 % for central Africa.

The grouping scheme in Sect. 3.1 (by IOP and BBA-loading history) is used to categorize the in situ data to determine if similar distinctions in their precipitation and evapotranspiration histories are derived using a (q, δD) space analysis (Figs. 6–8). The asymptotic end-members and theoretical curves reviewed in Sect. 2.5 are estimated and plotted in Figs. 6–8 for each sampling period. The symbols and curves for each quantity and process match those used in Fig. 3. Table 2 summarizes their computation. Some of the end-members require estimations of near-surface values over either the ocean or the African continent, as detailed in Table 2. Near-surface temperatures and water mixing ratios were estimated using monthly mean 10 m values from MERRA-2. The δD of near-surface air over Africa is taken from output of the isotopic version of the Community Atmospheric Model simulations (isoCAM5; Nusbaumer et al., 2017), which were run for the ORACLES time period (2016–2018) and nudged with the MERRA-2 winds in a configuration very similar to that described by Fiorella et al. (2021). FT air masses are split into high-BBA (red) vs. low-BBA (gray) as in Sect. 3.1. Data for high-BBA vs. low-BBA air in Figs. 6–8 are plotted as joint probability distributions (PDFs) of (q, δD), derived using a Gaussian kernel density estimation. SCL data are plotted in blue after 60 s averaging. Horizontally, the SCL data on average fall between 65 %–80 % RH with respect to mean SST, which is reasonable. A further comparison of the SCL data with theory is provided in Appendix C. In short, the agreement between the measurements and simple theory is robust enough to provide an anchor point for the measurements and the analyses which utilize them.

The blue mixing curves in Figs. 6–8 use a combination of ORACLES in situ and MERRA-2 datasets along with the evapotranspiration estimations to provide more observationally constrained moist mixing end-members. For the 2016 IOP where ET is thought to be minimal, the mixing curve simply starts at the mean SCL (q, δD) values. For the 2017 IOP, ET contribution is estimated as 25 %. The open box in Fig. 7 represents a 75 %/25 % mixture of the mean ORACLES SCL and a (q, δD) point taken from mean 10 m values derived from MERRA-2 and isoCAM5 simulations 10 m over the boxed region in Fig. 4 for 2017. It is taken as the moist end-member for the blue curve. The dashed boxed in Fig. 8 for the 2018 IOP is an analogous 75 %/25 % mixture for the dashed box in Fig. 4. The solid box is provided mostly for reference and is the point along the blue curve where q reaches its MERRA-2 10 m value over the 2018 solid box in Fig. 4.

While the majority of 2016 high-BBA data can be described by dilution of high-humidity continental air with clean FT air (the blue mixing curve), the 2017 data fall well below this curve and require more analysis. The assumed model here is a detraining convective plume discussed in Sect. 2.5. Assuming an ascending parcel model following a Rayleigh process, q and δD would fall monotonically with height as water in the parcel condensed and precipitated out. Detrained air at a specific height would have a characteristic δD, and subsequent mixing with ambient dry air would alter δD relatively little (by less than 20 ‰ until the air reached about 2 gkg-1 humidity). If Rayleigh distillation followed by detrainment into dry air is in fact the primary process for the 2017 air parcels, one expectation is that δD and altitude would have good correlation. For the 2017 IOP, a correlation of -0.67 is found between δD and altitude, with δD decreasing at a mean rate of -35 ‰ km-1. In comparison, 0.05 correlation is found for the 2016 IOP despite similar sampling altitudes. Further, using the method outlined in Sect. 2.5, detrainment heights and temperatures from the Rayleigh trajectory are estimated. The best-fit lines (r of 0.68 for each) have slopes of 0.54 (Fig. 7).

For the 2018 IOP, the high-BBA air lobe falls along a mixing line with evidence of ET contribution (Fig. 8), but the PDF extends downward and is bound from below by a detrainment curve at 3600 m. The main low-BBA PDF lobe falls near a Rayleigh path, even extending below (the majority of the signal lying below the Rayleigh curve is due to air sampled below 3000 m). An estimation of altitude and temperature was performed using the Rayleigh model separately for the high- vs. low-BBA data. Correlations of 0.5 with observed values were found for each group. The low-BBA air has best-fit slopes of 0.48 and 0.51 for altitude and temperature. The high-BBA air has best-fit slopes of 0.71 and 0.86. δD decreases with altitude by 19 ‰ and 12 ‰ km-1 for the low- and high-BBA groups, respectively.

Black carbon can serve as CCN and be permanently removed via precipitation. Carbon monoxide, on the other hand, has a low solubility in water and therefore is not altered significantly by clouds. Consequently, scavenging of BC by precipitation should cause the BC/CO ratio to decrease (e.g., Liu et al., 2020). The FT (q, δD) diagrams for the 2017 and 2018 IOPs show that air masses sampled had a variety of integrated precipitation amounts, providing an approach to test for evidence of BC scavenging (2016 was not used because there is minimal occurrence of precipitation). Unlike for the previous section, where distinct air masses were isolated in order to highlight the primary hydrologic processes occurring, all data that likely originate from the African planetary boundary layer are used here. The motivation is to diagnose the influence of scavenging on BC concentrations for an arbitrary parcel of air. FT data were averaged into 60 s blocks and data with CO less than 100 ppbv were discarded, as they reflect air of marine origin.

Figure 9 shows clear gradients in BC/CO coloring for (q, δD) space. If one uses the simple detraining convective plume model (outlined at the end of Sect. 2.5) to interpret the joint (q, δD) data, then the BC/CO decreases with increased precipitation amount, which would be expected for scavenging. Because the detrainment mixing curve has little δD variability for q>2 gkg-1, the boxed regions depicted in Fig. 9 correspond roughly to relatively low, moderate, and high precipitation amounts. PDFs of BC/CO for these three regions (Fig. 10a and b) separate with statistical significance (at α=0.05 using a t test) for both the 2017 and the 2018 IOPs. For the 2017 IOP the PDF means are 9.1, 5.9, and 3.9 ngkg-1ppbv-1. For 2018, the means are 5.1, 3.4, and 1.5 ng-1kg-1ppbv. As a control, the data separated by relatively low, moderate, and high values of q (bounds of (2,5), (5,8), and (8,11) gkg-1, respectively) show no statistically significant separation.

The change in q due to precipitation, Δqprecip, is computed for each data point in Fig. 9 using an approximation to the convective plume model. For q>2 gk-1, we take the observed δD to be equal to that of the plume when it detrained, so Δqprecip can be approximated with the Rayleigh model as 1Δqprecip∼|q0-qRay-1(δDobs)|, where q0 is the estimated initial q of the plume before precipitation, qRay-1 the Rayleigh model inverted to predict q as a function of δD, and δDobs the observed value. This difference is intuitively similar to the “rainout fraction”, F=q/q0, that appears in Rayleigh calculation. The correlation between Δqprecip and BC/CO is 0.71 for 2017 and 0.56 for 2018 (Fig. 10c). In comparison, the correlation between q and BC/CO is 0.21 for 2017 and -0.14 for 2018 and distinctly inferior. This result suggests a quantifiable path for deriving BC scavenging coefficients using isotopic data that are unavailable from measures of water concentrations alone.

The variability in isotopic signals in the MBL is small in comparison to that in those of the FT. It was found that for MBLs sampled during ORACLES, the differences between the SCL and CL are small enough to be comparable to or less than SCL variability over the observation period and region. This range makes it more challenging to attribute variations in CL (q, δD) to specific CL processes (e.g., entrainment vs. precipitation) on an absolute scale given the measurement uncertainties. Therefore, CL (q, δD) data are plotted as deviations (Δq, ΔδD) from the mean values of the nearby sampled SCL, (q, δD)SCL, which subtracts out the SCL variations (Fig. 11). CL data are taken primarily from LL and SAW modules (processed as in Sect. 2.4). For these modules, (q, δD)SCL values were taken from temporally neighboring data below 500 m, which usually occurred within 10 min before or after the module, but in some cases ±20 min windows were necessary. CL data were taken secondarily from vertical profiles (Sect. 2.1) which did not include LL or SAW modules. For these profiles, (Δq, ΔδD) values were computed from the vertically averaged values of the SCL and CL for the respective profile. Only CL data for which the magnitude of the mean relative wind between the SCL and CL was less than 3.5 ms-1 were used to avoid noise due to large relative horizontal advection. Each block of SCL data used for computing (q, δD)SCL was also used to compute mean CCN values. Due to instrument limitations, it is necessary to use SCL rather than CL CCN measurements, but Diamond et al. (2018) showed that for the 2016 IOP, SCL CCN correlates well with CL droplet number concentration, implying a connection between the SCL and CL aerosol concentrations.

Figure 11 shows (Δq,ΔδD) diagrams colored by SCL CCN tercile groups. The CCN terciles are different for each year and are listed in Table 3. Distributions of SCL CCN are shown in Fig. 12. The solid and dashed black curves are mixing and Rayleigh curves. The end-members for the mixing curves are (q, δD)SCL and the same dry, isotopically depleted end-member used in Figs. 6–8. From here on out, this mixing trajectory is referred to as mixFT. Data which fall below mixFT are candidates for precipitation, but there are two possibilities. The first is that the CL has a recent history of precipitation. The other is that the air entraining in at CL top tends to align the data along an alternate mixing trajectory. To test this, mixing curves between (q, δD)MBL and the mean (q, δD) of the 250–500 m region above CL top (q, δD)250-500 for all vertical profiles in Sect. 2.3 are included in Fig. 11 and from here on out are referred to as mix250-500. They are expressed as (Δq,ΔδD) like the scatter points. The dark portion of the curves represent mixtures that have at most 25 % (q, δD)250-500 contribution. CLs sampled during the 2016 IOP occurred at higher latitudes in comparison to the other two IOPs, and the CLs were on average 2× to 3× shallower (Fig. 12). The majority of CLs show small deviations (Δq<0.5 gkg-1, ΔδD<4 ‰) from (q, δD)SCL. The majority of mix250-500 curves (30 out of 35) are centered on mixFT and clearly depart from the Rayleigh trajectory. Five of the curves fall closer to the Rayleigh trajectory.

In comparison to the 2016 IOP, a higher proportion of 2017 IOP data deviate from (q, δD)SCL, and the standard deviation is larger. There are fewer data points shown for 2017 – since three of the flights were eastward toward Ascension Island, days where the MBL transitioned to shallow cumulus rather than a stratocumulus-topped MBL were omitted. Two-thirds of the data lie within precision of one of the mixing curves or the Rayleigh curve. The other third cannot be described by the simple models presented here. As in the 2016 IOP, the majority of mix250-500 curves are clustered about mixFT (19 of 24), while 5 pass near the Rayleigh trajectory.

For the 2018 IOP, the alternate end-member mixing curves cover a wide range of (q, δD) space rather than being clustered about mixFT, which likely explains the middle-tercile CL data (yellow). Like in the 2017 case, the data which lie near a Rayleigh trajectory with ΔδD>5 ‰ (blue cluster) could not be explained by mixing curves unless the CL contained a fraction of FT air greater than 25 % in comparison to the SCL.

Figures 4 and 5 show that FT air from the 2016 sampling period experienced the least amount of precipitation while over Africa of the three sampling periods. The in situ (q, δD) data show that in fact there are almost no signs of precipitation in the sampled high-BBA air in 2016. The PDF is enveloped from above by the theoretical ET mixing curve, suggesting ET contributions for some of the measurements, although the dominant signal is simple mixing with air close to the observed SCL values. This is interpreted as evidence of minimal ET contribution to the African planetary boundary layer air, and therefore its moisture content closely resembles the original marine source which advected onto the continent. The low-BBA data are shown for completeness, but since back trajectories suggest they often originate from the southern Atlantic rather than Africa, they are not analyzed in detail.

Both high-BBA and low-BBA data for the 2017 IOP show signs of precipitation (Fig. 7). The data are bound from above by a mixing curve with a moist end-member including 25 % ET contribution and from the right by a family of Rayleigh curves. The data support a depiction of the air mass hydrologic history as convective detrainment followed by dilution; air masses follow the Rayleigh curve during moist convection (i.e., condensation with high precipitation efficiency), then begin to detrain remaining total water at their level of neutral buoyancy, and subsequently form a diluted mixture with surrounding dry FT air. The low-BBA air is associated with convection which reached higher altitudes. A correlation of -0.67 for δD vs. height is obtained, whereas the correlation is close to 0 for the 2016 IOP despite sampling at a similar altitude range. For the 2017 IOP, detrainment heights and temperatures predicted from the Rayleigh model both have correlations of 0.67 with the observations, which is remarkable considering the simplicity of the parcel-type model used to describe the entire study period. That is, a single choice of values for surface T, q, and height; no entrainment during convection; and all condensate converted to precipitation provide a reasonable first approximation. The best-fit lines have slopes of 0.54, however. Possible reasons for why this slope is less than 1 include variation in surface conditions during convection throughout the observation period and precipitation efficiency less than unity. With respect to the latter, entrainment into an updraft would decrease δD more than expected from the Rayleigh process alone. Therefore, to achieve the same δD with only a Rayleigh process, the parcel would have to be lifted higher to compensate. Additionally, δD along detrainment curves was taken as constant to simplify the calculations. However, δD decreases by about 20 ‰ once detrained updraft air has been diluted to around 2 gkg-1, resulting in the same bias for the Rayleigh model as caused by precipitation efficiency <1.

The October 2018 sampling period comprises elements of both the 2016 and the 2017 sampling periods. The high-BBA comes from higher latitudes that are similar to those of 2016, while the low-BBA air is further north with more rainfall. However, unlike in 2016, the high-BBA air flows northwest over Africa before flowing offshore. The high-BBA air (q, δD) signal reflects this – the PDF lobe falls along a mixing line with evidence of up to 25 % ET contribution – but the distribution extends below the mixing line much more so than in 2016, bounded beneath by detrainment at 3600 m. Meanwhile, a large fraction of the low-BBA data lie near or below a Rayleigh curve, suggesting convection with a high precipitation efficiency and little subsequent mixing and potentially re-evaporation of falling droplets. Data with this signal are almost exclusively at heights below 3000 m, while data above 3000 m fall either directly on a Rayleigh curve or in the region hypothesized to be detrainment. Perhaps rain droplets falling from these higher altitudes are partially evaporating into the air below 3000 m. Such refined hydrologic history is not readily extracted without more detailed modeling.

Cloud layer data for the 2017 and 2018 IOPs deviate further from their respective (q, δD)SCL values in comparison to the 2016 IOP. This is in line with the deeper CLs (Fig. 12) transitioning to cumulus-coupled CLs, becoming less well-mixed with the SCL, and experiencing higher entrainment and precipitation rates. CL (q, δD) measurements for all IOPs are poorly constrained by the simple mixFT and Rayleigh models – around 50 % of data for the 2016 and 2017 IOPs lie within these models, compared to only 30 % for the 2018 IOP. For the 2018 IOP, the majority of data outside these bounds can be explained by mixing with air masses above the MBL which have varying (q, δD) values. For the 2016 and 2017 IOPs, even alternate end-member mixing cannot explain the data, and these exceptions merit further study since the fraction of the data they make up is non-negligible. One possibility is the choice of a single representative (q, δD) pair for the SCL when generating the mix250-500 curves. If either or both of q and δD for overlying air were higher than the SCL values, the resulting mixing curves could explain the data in question. Differential horizontal advection of the MBL and FT could also be a contributing factor.

Although the data show signals of local precipitation in the 2016 and 2017 IOPs, the null hypothesis that all data falling below the mixFT curve came from mix250-500 curves cannot be fully ruled out with this analysis alone. However, Fig. 11 (right column) highlights some cases where the data fall near a Rayleigh trajectory and cannot be accounted for by mixing models with overlying FT air sampled on those days. Previous work has highlighted the uncertainties involved with assuming an MBL is in significant contact with the FT air above it at the time of sampling and neglecting horizontal advection (Diamond et al., 2018). Therefore, it cannot be definitively concluded that the cloud layers sampled during the highlighted flights recently precipitated. However, these flights would be good candidates to examine with advanced modeling.

Assuming that at least some of the signal is due to precipitation, a question arises regarding the time period prior to the measurement during which the precipitation occurred. Simple bulk model studies (Schubert et al., 1979; Bretherton et al., 2010; Jones et al., 2014) estimate a thermodynamic re-equilibration timescale of a stratocumulus-topped MBL at 0.5–1 d. Therefore, at longer timescales it is assumed that any isotopic evidence due to precipitation has been “erased” by subsequent turbulent mixing, including influences from surface evaporation, and consequently any signals detected are from precipitation that occurred within the past day. Previous observational studies show stratocumulus-topped MBLs precipitate primarily at night or in the early morning hours (Burleyson et al., 2013; Leon et al., 2008; Comstock et al., 2004). It may be the case that the majority of precipitation signals detected with isotope measurements are from the morning on the same day as sampling.

By inspection, there is weak SCL CCN clustering in Fig. 11 for the 2016 IOP. The two higher-CCN-tercile groups cluster about the origin, while the lowest-CCN-tercile group extends to more negative values. If only a fraction of the signals come from mix25-500 processes, the data would indicate that aerosol loading has induced precipitation suppression. The converse could also be the case: aerosol scavenging in MBLs with previous precipitation results in lower SCL CCN. Microphysical modeling would be required to quantify the causality. Further, no clear clustering is present for the 2017 IOP, in part due to the smaller signals. Indeed, using either thermodynamic or isotopic measurements to assess cloud aerosol interactions in the MBL is not necessarily achievable with simple analytical models. As a case in point, one confounding factor could be locally enhanced entrainment due to precipitation (see, e.g., Fig. 9 in Wood, 2012), which could replenish BBAs which were scavenged.

For the 2018 IOP, a consistent explanation develops from Figs. 8 and 11 (which present the 2018 IOP FT and CL data, respectively). CL data in the groups of the middle and top SCL CCN terciles are largely clustered around mixFT or the region below the Rayleigh curve (Fig. 11, shaded regions). Of these data, the highest-tercile group is primarily along the former while the middle group is along the latter. This is consistent with the (q, δD) location of high- vs. low-BBA FT PDFs in Fig. 8 if they were to serve as the above-MBL mixing end-members. Therefore, the CL data in the middle- and top-tercile groups are likely both mixtures of SCL and FT air, with the differences in MBL CCN due to the differences in FT BBA levels. Data in the lowest-tercile cluster closer to a Rayleigh trajectory, possibly indicating local precipitation and aerosol scavenging. This self-consistency in the joint (q, δD) measurements builds a degree of confidence that CL isotopic signals can be connected to entrainment processes, and potentially precipitation, even if the connection is more detailed, as for the 2017 IOP.