The ATom mission was designed to collect a multi-species, detailed chemical climatology that documents the spatial patterns of chemical heterogeneity throughout the remote troposphere. Figure S1 maps the 48 research flights, and the Supplement has tables summarizing each flight. We required a complete set of key species in each air parcel to initialize the models that calculate the CH4 and O3 reactivities. We choose the key reactive species (H2O, O3, CO, CH4, NOx, NOxPSS, HNO3, HNO4, PAN, CH2O, H2O2, CH3OOH, acetone, acetaldehyde, C2H6, C3H8,i-C4H10, n-C4H10, alkanes, C2H4, alkenes, C2H2, C5H8, benzene, toluene, xylene, CH3ONO2, C2H5ONO2, RONO2, and CH3OH) directly from the ATom measurements and then add corollary species or other observational data indicative of industrial or biomass burning pollution or atmospheric processing (HCN, CH3CN, SF6, relative humidity, aerosol surface area (four modes), and cloud indicator). We choose 10 s averages for our air parcels as a compromise and because the 10 s merged data are a standard product (Wofsy et al., 2018). A few instruments measure at 1 s intervals, but the variability at this scale is not that different from 10 s averages (Fig. S2). Most of the key species are reported as 10 s values, with some being averaged or sampled at 30 s or longer such as ∼90 s for some flask measurements.

Throughout ATom, gaps occur in individual species on a range of timescales due to calibration cycles, sampling rates, or instrument malfunction. The generation of the MDS uses a range of methods to fill these gaps and assigns a flag index to each species and data point to allow users to identify direct measurements and methods used for gap-filling. Where two instruments measure the same species, the MDS selects a primary measurement and identifies which instrument was used with a flag. The methodology and species-specific information on how the current MDS version 2 (MDS-2) is constructed, plus statistics on the 48 research flights and the 146 494 10 s air parcels in MDS-2, are given in the Supplement.

Over the course of this study, several MDS versions were developed and tested, including model-derived RDSs from these versions, some of which are used in this paper. In early ATom science team meetings, there was concern about the accuracy of NO2 direct measurements when at very low concentrations. A group prepared an estimate for NOx using the NO and O3 measurements to calculate a photostationary value for NO2 and thus NOx. This PSS-NOx became the primary NOx source in version 0 (i.e., MDS-0). With MDS-0, we chose to gap-fill using correlations with CO to estimate the variability of the missing measurement over the gap. The science team then rejected PSS-NOx as a proxy, and we reverted to the observed NO + NO2 resulting in NOx values that are 25 % larger on average than in MDS-0 (unweighted mean of 66 vs. 52 ppt). This change affected P-O3 most and L-CH4 least. We then estimated errors in the gap-filling and found that CO had little skill as a proxy for most other species. With MDS-2, we optimized and tested the treatments of gap-filling and lower limit of detection, along with other quality controls. With continued analysis of the unusually reactive eastern Pacific region, we determined that the method of long-gap filling for NOx resulted in propagation of high NOx levels from the over-land profiles into the over-water profiles in the tropics. We separated these two set of profiles used for long-gap NOx filling and created an updated version 2b. This experience points to the importance of having reliable, continuous NOx measurements. MDS-2b is fully documented in the Supplement.

The concept of using an MDS to initialize 3D global chemistry models and calculate an RDS was developed in the pre-ATom methodology papers (P2017; P2018). In this paper, we use the original six models for their August chemical statistics, and we use five of them plus a box model to calculate the reactivities; see Table 1. The RDS is really a protocol applied to the MDS. It is introduced in the Prologue, and the details can be found in P2018. A model grid cell chosen to be close to the measured parcel is initialized with all the core reactive species needed for a regular chemistry simulation. The model is then integrated over 24 h without transport or mixing, without scavenging, and without emissions. Each global model uses its own varying cloud fields for the period to calculate photolysis rates, but the F0AM box model simply takes the instant J values as measured on the flight and applies a diurnal scaling. We initialize with the core species and let the radicals (OH, HO2, and RO2) come quickly into photochemical balance. The 24 h integration is not overly sensitive to the start time of the integration, and thus models do not have to synchronize with the local time of observation (see P2018's Fig. S8 and Table S8).

The initial ATom-1 reactivities came from MDS-0 and six of the models in Table 1. Although these RDS-0 model results are now out of date because of the move to MDS-2b, they provide critical information on how models agree, or disagree, in calculating the RDS using the ATom protocol. Thus we include them here as a cross-model comparison. Given the excellent agreement at the parcel level using three models (GC, GMI, and UCI), and with a desire to avoid wasting the community's time, we continued the analysis of MDS-2b with just our local UCI CTM. This decision may need to be revisited.

Statistics for the three reactivities for six models using MDS-0 are given in Tables 2 and S8 for three domains: global (all points), Pacific (oceanic data from 53∘ S to 60∘ N), and Atlantic (same constraints as Pacific). The statistics try to achieve equal latitude-by-pressure sampling by weighting each ATom parcel inversely according to the number of parcels in each 10∘ latitude by 100 hPa bin, and each point is also cosine (latitude)-weighted. We calculate the means and medians plus the percent of total reactivity in the top 10 % of the weighted parcels (Table 2) and also the mean reactivity of the top 10 %, percent of total reactivity in the top 50 %, 10 %, and 3 %, plus the mean J values (Table S8).

These six-model version 0 statistics are shown alongside the version 2b results using the current UCIZ model but with a new protocol designated RDS*. While investigating sensitivities in the RDS, we found an inconsistency between the reported concentrations of both pernitric acid (HNO4) and peroxyacetyl nitrate (PAN) with respect to the chemical kinetics used in the models. High concentrations (100 ppt, attributed to instrument noise) were reported under conditions where the thermal decomposition frequency was >0.4 per hour in the lower troposphere (>253 K for HNO4 and >291 K for PAN). Thus, these species instantly become NOx. While these measurements are clearly spurious, there is no easy fix. We developed a new protocol, RDS*, that allows both species to decay for 24 h using their local thermal decomposition rate before being used in the model. This protocol avoids much of the fast thermal release of NOx in the lower atmosphere during the first 24 h of the RDS calculation but does not affect the release of NOx from photolysis or OH reactions in the upper troposphere where thermal decomposition is inconsequential. It is possible that some of the high concentrations of HNO4 and PAN in the lower troposphere are real and that we are missing this large source of NOx with the RDS* protocol, but we find no obvious sources of these species in the remote oceanic regions that would produce enough to match the thermal loss. Both this problem and its solution do not affect the initial NOx values.

We present the RDS-2b reactivities calculated under the RDS* protocol with the UCI CTM developed by Xin Zhu for P2017 and P2018 (designated UCIZ*) as our best results in the final column of Tables 2 and S8. We added diagnostics that give us confidence in our O3 reactivities: the approximate P-O3 and L-O3 based on the limited Reactions (rates 2a, b, and d and 3a, b, and c above) actually predict the calculated 24 h O3 tendency; see Fig. S6. Considering the ocean basin observations only, P - L (production minus loss) ranges from -12 to +15 ppb d-1. The mean error in P–L is about -0.01 ppb d-1, and the root-mean-squared error is about 0.04 ppb d-1, convincing us that we have correctly diagnosed the P-O3 and L-O3 terms. Following the practice of the GMI model, we also record the initial and 24 h abundances of all the ATom species to check that nothing unusual altered the species abundance in each cell over the 24 h.

Variations in reactivities due to clouds are an irreducible source of uncertainty: predicting the cloud-driven photolysis rates that a shearing air parcel will experience over 24 h is not possible here. The protocol uses 5 separated 24 h days to average over synoptically varying cloud conditions. The standard deviation (σ) of the 5 d, as a percentage of the 5 d mean, is averaged over all parcels and shown in Table S9 for the five global models. Three central models (GC, GMI, and UCI) show 9 %–10 % σ (Js) values and similar σ (Rs) values as expected if the variation in J values is driving the reactivities. Two models (GISS and NCAR) have 12 %–17 % σ (Js), which might be explained by more opaque clouds, but the amplified σ(R) values (14 %–32 %) are inexplicable. This discrepancy needs to be resolved before using these two models for ATom RDS analysis.

Inter-model differences are shown in the parcel-by-parcel root-mean-square (rms) differences for RDS-0 in Table 3. Even when models adopt standard kinetic rates and cross sections (i.e., Burkholder et al., 2015), the number of species and chemical mechanisms included, as well as the treatment of families of similar species or intermediate short-lived reaction products, varies across models. For example, UCI considers about 32 reactive gases, whereas GC and GMI have over 100, and F0AM has more than 600. The other major difference across models is photolysis, with models having different cloud data and different methods for calculating photolysis rates in cloudy atmospheres (H2018). The three central models (GC, GMI, and UCI) in terms of their 5 d variability (Table S9) are also most closely alike in these statistics, with rms = 20 %–30 % for L-CH4 up to 26 %–35 % for P-O3. These rms values appear to be about as close as any two models can get. The intra-model rms for different years (UCI 2016 versus 1997) is 10 %–13 % and shows that we are seeing basic differences in the chemical models across GC, GMI, and UCI. F0AM is the next closest to these central models, but it will inherently have a larger rms because it is a 1 d calculation and not a 5 d average. NCAR's rms is consistently higher and likely related to what is seen in the 5 d σ values in Table S9. GISS is clearly different from all the others (L-CH4 rms >100 %, while L-O3 rms <66 %).

The reactivities for three large domains (global, Pacific, and Atlantic) from the six-model RDS-0 are summarized in Tables 2 and S8. Sorted PDs for the three Rs and Pacific and Atlantic Ocean basins are plotted in Fig. 1 and show the importance of the most reactive “hot” parcels with deeply convex curves and the sharp upturn in R values above 0.9 cumulative weight (top 10 %). Both basins show a similar emphasis on the most reactive hot parcels: 80 %–90 % of total R is in the top 50 % of the parcels, 25 %–35 % is in the top 10 %, and about 10 %–14 % is in the top 3 %. The corollary is that the bottom 50 % parcels control only 10 %–20 % of the total reactivity, which is why the median is less than the mean (except for P-O3 in the Atlantic).

The enhancement factor for the top 50 % L-CH4 parcels is 2.0 (84 % of reactivity in 42 % of mass) given that our 53∘ S–60∘ N transects cover 83 % of the air mass below 200 hPa and assuming that L-CH4 is negligible poleward of these transects. This enhancement factor is a large-scale feature because the tropical lower troposphere, being warm and wet with high sun, dominates the budget. It is seen in previous model intercomparisons that calculate budgets in large tropospheric blocks like Voulgarakis et al. (2013) with 63 % of L-CH4 in 31 % of the air mass (500 hPa–surface, 30∘ S–30∘ N). The impact of the extremely hot parcels and the heterogeneity seen in the ATom 10 s parcels is evident in the steep slopes above the 90th percentile, yielding enhancement factors of 3 to 4.

Each R value and each ocean has a unique shape; for example L-O3 in the Atlantic is almost two straight lines breaking at the 50th percentile. In Fig. 1 the agreement across all models (except GISS) is clear, indicating that the conclusion in P2018 (i.e., that most global chemistry models agree on the O3 and CH4 budgets if given the chemical composition) also holds for the ATom-measured chemical composition. Comparing the dashed brown (UCI, RDS-0) and black (UCIZ, RDS*-2) lines, we find that the shift from MDS-0 to MDS-2b plus the new RDS* (HNO4 + PAN) protocol produces large reductions in P-O3 for all cumulative weights and small reductions in L-CH4 for the upper 5th percentile. We conclude that accurate modeling of chemical composition of the 80th and greater percentiles is important but that modest errors in the lowest 50th percentile are inconsequential; effectively, some parcels matter more than others (P2017).

How well does this ATom analysis work as a model intercomparison project? Overall, we find that most models give similar results when presented with the ATom-1 MDS. The broad agreement of the cumulative reactive PDs across a range of model formulations using differing levels of chemical complexity shows this approach is robust. The different protocols for calculating reactivities as well as the uncertainty in cloud fields appear to have a small impact on the shape of the cumulative PDs but are informative regarding the minimum structural uncertainty in estimating the 24 h reactivity of a well-measured air parcel.

A critical unknown for tropospheric chemistry modeling is what resolution is needed to correctly calculate the budgets of key gases. A similar question was addressed in Yu et al. (2016) for the isoprene oxidation pathways using a model with variable resolution (500, 250, and 30 km) compared to aircraft measurements; see also ship plume chemistry in Charlton-Perez et al. (2009). ATom's 10 s air parcels measure 2 km (horizontal) by 80 m (vertical) during most profiles. There are obviously some chemical structures below the 10 s air parcels. Only some ATom measurements are archived at 1 Hz, and we examine a test case using 1 s data for O3 and H2O for a mid-ocean descent between Anchorage and Kona in Fig. S2a. Some of the 1 s (200 m by 8 m) variability is clearly lost with 10 s averaging, but 10 s averaging preserves most of the variability. Lines in Fig. S2 demark 400 m in altitude, and most of the variability occurs on this larger, model-resolved scale. Figure S2b shows the 10 s reactivities during that descent and also indicates that much of the variability occurs at 400 m vertical scales. A more quantitative example using all the tropical ATom reactivities is shown in comparisons with probability densities below (Fig. 5).

How important is it for the models to represent the extremes of reactivity? While the sorted reactivity curves (Fig. 1, Tables 2 and S8) continue to steepen from the 90th to 97th percentile, the slope does not change that much. Thus we can estimate the 99th + percentile contributes <5 % of the total reactivity. Thus, if our model misses the top 1 % of reactive air parcels (e.g., due to the inability to simulate intensely reactive thin pollution layers) then we miss at most 5 % of the total reactivity. This finding is new and encouraging, and it needs to be verified with the ATom-2, 3, and 4 data.

The spatial structures and variability of reactivity as sampled by the ATom tropical transects (central Pacific, eastern Pacific, and Atlantic) are presented as nine panels in Fig. 2. Here, the UCIZ RDS*-2 reactivities are averaged and plotted in 1∘ latitude by 200 m thick cells, comparable to some global models (e.g., GMI, NCAR, and UCI). We separate the eastern Pacific (121∘ W, research flight (RF) 1) from the central Pacific (RFs 3, 4, and 5) because we are looking for contiguous latitude-by-pressure structures.

In the central Pacific (Fig. 2a, d, g), highly reactive (hot) P-O3 parcels (>6 ppb d-1) occur in larger, connected air masses at latitudes 20–22∘ N and pressure altitudes 2–3 km and in more scattered parcels (>3 ppb d-1) below 5 km down to 20∘ S. High L-O3 and L-CH4 coincide with this 20–22∘ N air mass and also with some high P-O3 at lower latitudes. This pattern of overlapping extremes in all three Rs is surprising because the models' mid-Pacific climatologies show a separation between regions of high L-O3 (lower–middle troposphere) and high P-O3 (upper troposphere, as seen in P2017's Fig. 3). The obvious explanation is that the models leave most of the lightning-produced NOx in the upper troposphere. The ATom profiling seems to catch reactive regions in adjacent profiles separate by a few hundred kilometers, scales easily resolvable with 3D models.

In the eastern Pacific (Fig. 2b, e, h), the overlap of outbound and return profiles enhances the spatial sampling over the 10 h flight. The region of very large L-O3 (>5 ppb d-1) is extensive, beginning at 5–6 km at 10∘ N and broadening to 2–8 km at 28∘ N. The region of L-CH4 is similar, but loss at the upper altitudes of this air mass is attenuated because of the temperature dependence of L-CH4 and possibly because of differing OH:HO2 ratios with altitude. Large P-O3 (>3 ppb d-1) occurs only in the center of this highly reactive L-O3/L-CH4 region, suggesting that NOx is not as evenly distributed as HOx is. Highly reactive (hot) P-O3 parcels (>4 ppb d-1) occur only in the upper troposphere (8–12 km) and only in the sub-tropics. ATom-1 RF1 (29 July 2016) occurred during the North American Monsoon when there was easterly flow off Mexico; thus the high reactivity of this large air mass indicates that continental deep convection with lightning NOx is a source of high reactivity for both O3 and CH4.

In the Atlantic (Fig. 2c, f, i), we also see similar air masses through successive profiles, particularly in the northern tropics. The Atlantic P-O3 shows high-altitude reactivity similar to the eastern Pacific. Likewise, the large values of L-O3 and L-CH4 match the eastern Pacific and not the central Pacific. Unlike either Pacific transect, the Atlantic L-O3 and L-CH4 show some high reactivity below 1 km altitude. Overall, the ATom-1 profiling clearly identifies extended air masses of high L-O3 and L-CH4 extending over 2–5 km in altitude and 10∘ of latitude. The high P-O3 regions tend to be much more heterogeneous with greatly reduced spatial extent, likely of recent convective origin as for the eastern Pacific.

Overall, the extensive ATom profiling identifies a heterogeneous mix of chemical composition in the tropical Atlantic and Pacific, with a large range of reactivities. What is important for those trying to model tropospheric chemistry is that the spatial scales of variability seen in Fig. 2 should be within the capability of modern global models.

The ATom data set provides a unique opportunity to test CTMs and CCMs in a climatological sense. In this section, we compare ATom-1 data and the six models' chemical statistics for mid-August used in P2017. The ATom profiles cannot be easily compared point by point with CCMs, and we use statistical measures of the three reactivities in the three tropical basins: mean profiles in Fig. 3 and PDs in Fig. 5.

This paper opens a door for what the community can do with the ATom measurements and the derived products. ATom's mix of key species allows us to calculate the reactivity of the air parcels and hopefully may become standard for tropospheric chemistry campaigns. We find that the reactivity of the troposphere with respect to O3 and CH4 is dominated by a fraction of the air parcels but not by so small and infrequent a fraction as to challenge the ability of current CTMs to simulate these observations and thus be used to study the oxidation budgets. In comparing ATom results with modeled climatologies, we find a systematic ATom–model difference: models show a large relative drop in O3 production below 2 km over the tropical oceans, but ATom shows an increase (C.Pac.), no change (E.Pac.), or a much lesser drop (Atl.). We traced this result to the lack of NOx at 20–60 ppt levels in the models below 4 km and believe it provides a clear challenge in modeling ozone.

Building our chemical statistics (PDs) from the ATom 10 s air parcels on a scale of 2 km by 80 m, we can identify the fundamental scales of spatial heterogeneity in tropospheric chemistry. Although heterogeneity occurs at the finest scales (such as seen in some 1 s observations), the majority of variability in terms of the O3 and CH4 budgets occurs across scales larger than neighboring 2 km parcels. The PDs measured in ATom can be largely captured by a global model's 100 km by 200 m grid cells in the lower troposphere. This surprising result is evident by comparing the ATom 1D PDs – both species and reactivities – with those from the models' climatologies (Fig. 5). These comparisons show that the modeled PDs are consistent with the innate chemical heterogeneity of the troposphere as measured by the 10 s parcels in ATom. A related conclusion for biomass burning smoke particles is found by Schill et al. (2020), where most of the smoke appears in the background rather than in pollution plumes, and therefore much of the variability occurs on synoptic scales resolved by global models (see their Fig. 1 compared with Fig. 2 here).

As a quick look at the opportunities provided by the ATom data, we present an example based on the Wolfe et al. (2019) study, which used the F0AM model and semi-analytical arguments to show that troposphere HCHO columns (measurable by satellite and ATom) are related to OH columns (measured by ATom) and thus to CH4 loss. Figure 6 extends the Wolfe et al. (2019) study using the individual air parcels and plotting L-CH4 (ppb d-1) versus HCHO (ppt) for the three tropical regions where most of the CH4 loss occurs. The relationship is linear but with a lot of scatter and has slopes ranging from 3.5 to 4.4 per day over the three tropical regions, but for the largest reactivities (0–4 km, 1–3 ppb d-1), L-CH4 is not so well correlated with HCHO.

As is usual with new model intercomparison projects, we have an opportunity to identify model “features” and identify errors. In the UCI model, an error in the lumped alkane formulation (averaging alkanes C3H8 and higher) did not show up in P2018, where UCI supplied all the species, but when the ATom data were used, the UCI model became an outlier. Once found, this problem was readily fixed (hence the current UCIZ model version). Inclusion of the F0AM model with its extensive hydrocarbon oxidation mechanism provided an interesting contrast with the simpler chemistry in the global CCM/CTMs. For a better comparison of the chemical mechanisms, we should have F0AM use 5 d of photolysis fields from one of the CTMs. The anomalous GISS results have been examined by a co-author, but no clear causes have been identified as of this publication. The problem goes beyond just the implementation of the RDS protocol, as it shows up in the model climatology (Figs. 4 and 5, also in P2017).

Decadal-scale shifts in the budgets of O3 and CH4 are likely to be evident through the statistical patterns of the key species, rather than simply via average profiles. The underlying design of ATom was to collect enough data to develop such a multivariate chemical climatology. As a quick look across the four deployments, we show the joint 2D PDs on a logarithmic scale as in P2017 for HOOH versus NOx in Fig. 7. The patterns for the tropical central Pacific are quite similar for the four seasons of ATom deployments, and the fitted ellipses are almost identical for ATom 2, 3, and 4. Thus, for these species in the central Pacific, we believe that ATom provides a benchmark of the 2016–2018 chemical state, one that can be revisited with an aircraft mission in a decade to detect changes in not only chemical composition, but also reactivity.

ATom identifies which “highly reactive” spatial or chemical environments could be targeted in future campaigns for process studies or to provide a better link between satellite observations and photochemical reactivity (e.g., eastern Pacific mid-troposphere in August, Fig. 2). The many corollary species measured by ATom (not directly involved in CH4 and O3 chemistry) can provide clues to the origin or chemical processing of these environments. We hope to engage a wider modeling community beyond the ATom science team, as in H2018, in the calculation of photochemical processes, budgets, and feedbacks based on all four ATom deployments.