Neural networks are generated to predict the monthly mean OH mixing ratio for a given model following the method outlined in Nicely et al. (2017). Briefly, four NNs are trained for one model, each for one simulation month. To reduce the computational demands of NN training, we only establish NNs for four months, one for each season: January, April, July, and October. Separate NNs are trained for the SD (main text) and free-running (Supplement) simulations, and all training is performed with output from the year 2000. Each model grid box located below the tropopause (thermal, following the WMO definition, for all models except GEOS Replay, which uses a “blended” tropopause calculation combining thermal and potential vorticity definitions) is a single sample, so sample sizes are determined by a model's vertical and spatial native resolution. The number of tropospheric model grid points, and thus the training dataset sample size, is indicated for each model in Table S1 in the Supplement and always exceeds 100 000. Because separate NNs are trained for each month, and monthly mean output from each model simulation is used as input and training data, the dataset does not represent diurnal variations in OH chemistry.

The training process adjusts weighting factors such that mixing ratios of OH are predicted accurately when 3-D fields of the following variables are input to the NN: pressure, latitude, temperature (T), ozone (O3), specific humidity (H2O), methane (CH4), the sum of nitrogen oxide and nitrogen dioxide (NOx≡NO+NO2), the ratio NO:NOx, carbon monoxide (CO), isoprene (ISOP = C5H8), formaldehyde (HCHO), the photolysis frequency of NO2 (JNO2), the photolysis frequency of ozone to excited state O(1D) (JO1D), and stratospheric ozone column (O3 COL). Note that many of the inputs covary with one another depending on the chemical regime or meteorological conditions. A strength of the NN approach is that the inputs chosen need not be independent of each other. The NOx and NO:NOx inputs are calculated using monthly mean NO and NO2 fields. All chemical species are input to the NN as unitless mixing ratios, except for methane, which is normalized by the maximum tropospheric value and indicated by the notation CH4NORM. This normalization enables direct comparison of methane distributions between models, despite the fact that the use of boundary conditions sometimes results in substantially different amounts of methane between models. (While the CCMI models generally used roughly consistent boundary conditions, the additional simulations that were not formally part of CCMI exhibit methane concentrations outside the ranges of those in the CCMI models.) Pressure is provided in units of hPa, temperature in K, photolysis frequencies in s-1, and O3 COL in Dobson units (DU). Three of the inputs – HCHO, NO:NOx, and O3 COL – have been introduced to this analysis after the work of Nicely et al. (2017), due to availability of output from all models and to the added information they encompass that may be relevant for OH chemistry. For instance, having knowledge of the partitioning of NOx likely enables one to more accurately predict OH quantities compared to knowing just the total abundance of NOx. Likewise, the introduction of O3 COL is somewhat redundant when its primary effect on OH is through attenuation of ultraviolet (UV) flux to the troposphere, which is already encompassed by the input JO1D. However, JO1D is also altered by other factors such as clouds, which cannot as easily be included as an input for this analysis (some models provide 2-D cloud fraction fields, others output 3-D fields, and still others do not give any metric regarding clouds). Whether strong differences in JO1D are caused by clouds or overhead ozone should be clarified by inclusion of O3 COL as an input.

The neural network architecture is consistent with that of Nicely et al. (2017) and is shown in Fig. 1. However, the number of computational nodes was doubled from 15 to 30 given the availability of more powerful computing resources. Two hidden layers each containing 30 nodes provided strong performance of the NN in reproducing the OH mixing ratios from a given model. For training, the model output is randomly split 80 %/10 %/10 % into training, validation, and test datasets. During that process, the data from the training set are used to actively adjust weighting factors, and the validation set is evaluated to determine a training stopping point. When errors in predicting the validation data grow after adjusting weighting factors some number of iterations in a row, it is determined that the NN model prior to the growth in errors likely reached a local minimum in its cost function. This manner of “early stopping” helps to prevent overfitting, though application of the NNs to alternative years is not immune to overfitting, an issue discussed further in Sect. 4.3.1. For further application of this method across varying timescales, we would recommend a more methodical approach to sampling model output in time as well as in space. The final 10 % of data is then used to independently test the resulting NN and compare between different training iterations. A total of five trainings were performed for each NN, and the NN with best performance (evaluated by the correlation coefficient from comparison of NN-calculated and model-simulated OH values) was chosen as the NN to be used in further analysis. Further details of the training process and evaluation metrics can be found in Nicely et al. (2017).

We note that alternative machine learning algorithms have seen increased application to problems within atmospheric science in the last few years and may be equally or even better suited than neural networks for studying non-linear chemical systems. In particular, random forest regressions and gradient-boosting techniques offer greater computational efficiency and, in the case of random forests, have the capability to quickly identify which inputs are most strongly influencing the calculated output, known as “feature importance” (Hu et al., 2017; Grange et al., 2018; Liu et al., 2018; Keller and Evans, 2019). Additionally, linear regression algorithms such as Ridge and Lasso regression may be beneficial in curbing issues related to extrapolation. We also do not intend to suggest that our chosen NN input list, architecture, and general method are the best approach; input variables were largely determined by available output, and architecture testing was conducted on the computing resources available at the time of the study. It is possible that a single NN could suffice for predicting OH variations throughout an entire year, rather than for just a single month, following methodical subsampling methods to create the initial training dataset. As such, we encourage exploration of modifications to this method as well as additional algorithms for future machine learning applications to atmospheric chemistry.

Once NNs are established for each model, an analysis is conducted to quantify the OH and τCH4 differences attributable to individual input terms. To accomplish this, each model, A, is paired with another model, B, such that one input to the NN of model A is substituted with the same field from model B. All other inputs are held fixed, using fields from model A for the year 2000. Fields are interpolated to the resolution of the native model, A in this case, bilinearly across latitude and longitude, and linearly in log(pressure) space for the vertical coordinate. Any resulting changes in OH can then be directly attributed to the substituted variable.

The “swaps” that are performed in the manner described above undergo a process we refer to as “extrapolation control,” which restricts the substituted variable from leaving the range of values over which the native model's NN was trained. For example, if O3 is being substituted from CAM4-Chem into the GMI NN, we not only check that a given CAM4-Chem O3 value lies within the minimum and maximum GMI tropospheric O3 values but also that the GMI value of CO at that grid point can be associated with the new CAM4-Chem O3 value. This check is performed across all variables and essentially prevents the substitutions from venturing too far outside of the chemical regimes simulated within the native model. In the case that a swapped variable exceeds the acceptable range of values, it is revised up or down accordingly. For reference, we tally the number of instances in which extrapolation control is invoked for two categories: coarse adjustments, when a NN input value from another model falls entirely outside the range of the NN input values from the native model, and fine adjustment, when a value from another model must be tweaked to preserve the native model's chemical regimes. On average, coarse adjustments are incurred for 3.5 % of all swapped data points, while fine adjustments are made to 18.8 % of swapped values. We find that extrapolation control is critical to achieve meaningful results with the NN intermodel comparison method, though it necessarily forces the attributed changes in OH and τCH4 to be conservative estimates.

Metrics used to evaluate the results of variable swaps include tropospheric OH integrated columns for visualization and changes in τCH4 for a globally summed quantity. Tropospheric columns are integrated vertically and weighted by the mass of methane and the temperature-dependent rate constant of reaction between OH and methane. The global mean lifetime of methane is found using Eq. (1): 1τCH4=∑Mair×χCH4∑[OH]×kOH+CH4×Mair×χCH4, where Mair is the mass of air within a grid box, brackets denote number density, χ denotes mixing ratio, kOH+CH4 is the reaction rate constant for the OH+CH4 reaction calculated for each grid box temperature, and summations are performed over all tropospheric model grid boxes. This formulation is equivalent to the standard lifetime calculation of burden divided by loss rate, adapted to the quantities most directly related to model outputs available (Chipperfield et al., 2014). Again, we note that this is strictly the atmospheric lifetime of methane with respect to loss by tropospheric OH. If one additionally includes all stratospheric grid boxes within the above summation, annual average lifetimes of almost all models consistently increase by ∼1.2 years.

A new element of this analysis applies the already-established NNs of each model to examine the time evolution of OH over several decades of simulation. For this, we focus on the REF-C1SD simulation set, as it contains the most realistic representation of historical emissions and meteorological conditions and thus is most likely to resemble true OH variations. All models that provided SD-type simulations as described in Sect. 2.2 and 2.3 are included, with the exception of GEOS-Chem and OsloCTM, both of which only provided output for the year 2000. Using a similar swapping technique as described in Sect. 3.2, the NN for a given model is used to quantify the effect of substituting individual inputs from different years. No intermodel substitutions are conducted; instead, a single input is taken from the various years of the simulation (1980–2015) while all other inputs are fixed to their 2000 values. Because all swaps are performed on an intra-model basis, extrapolation control is largely unnecessary, since that model's chemical regimes do not vary drastically from the original year 2000 training output. However, we do see some instances, noted in Sect. 4.3, of anomalous behaviour in the τCH4 results because some variables undergo significant changes, particularly between the 1980s and the training year (2000). Overall, the NN technique should be sufficiently generalizable to provide meaningful results even when using inputs lying modestly outside of the range of training values. Robustness of the results is demonstrated by the emergence of several consistent features between the eight models examined, as discussed in Sect. 4.

Figure 2 shows values of τCH4 found for all models that produced SD-type simulations. Annually and globally averaged lifetimes vary from 6.59 years (OsloCTM) to 8.41 years (GMI). All models exhibit the expected seasonal variation in τCH4, with minimum values in the Northern Hemisphere (NH) summer months due to higher OH at this time of year. Specifically, the seasonal variation in the global mean is a result of greater anthropogenic influence in the NH and resulting increases in concentration of two OH precursors: ozone and NOx.

An example of NN performance is shown for the January WACCM model in Fig. 3, relative to the native model OH fields. Tropospheric OH columns are shown for the model and NN alongside the absolute value of the difference between the two. In general, the NNs from all models show similar magnitudes and spatial patterns in their calculated OH field, with errors somewhat randomly scattered and maximizing locally to values of ∼10 % of the total column value. Figures S1–S4 in the Supplement show the performance of all NNs, for each of 10 SD-type model simulations and for each of the four months, while Table S2 provides further statistics on all NNs used here. Performance of all model NNs for the year 2000 is strong, with values of τCH4 calculated from the NN-generated OH field within 0.006 years of the parent model's τCH4 on average. The maximum error in τCH4, an overestimate by 0.012 years, occurs for the MRI-ESM1r1 model in the month of January. Performance is generally poorest in boreal winter, with average offsets in τCH4 of 0.007 years, and strongest in boreal summer, for which the mean bias is only 0.004 years.

The intermodel comparison component of this analysis can be understood fundamentally by the OH and τCH4 differences generated by substituting input fields between models. An example of the OH column and τCH4 changes that are calculated through individual variable swaps is shown in Fig. 4. The two models with the highest and lowest values of τCH4, GMI and OsloCTM, respectively, are chosen for this example. Swaps performed between the two models for the month of January reveal that local O3, JO1D, HCHO, and NOx account for the largest differences in τCH4 for this particular model pairing. A complete budgeting of the changes in τCH4 attributable to all inputs for GMI and OsloCTM is shown in Table 1. Note that the values of τCH4 shown in Table 1 correspond to lifetimes for the month of January rather than annual averages and so will differ from the lifetimes noted at the beginning of Sect. 4.1.

It is worth discussing several features that are evident in the visualized OH changes shown in Fig. 4. First is the spatial distribution of the OH variations. Depending on how the sink or source term undergoing the swap affects OH chemistry, the strongest impacts may occur in localized areas or may distribute evenly over the globe. For instance, varying local ozone and NOx (Fig. 4a, b and g, h, respectively) exert the greatest influence on OH over the climatological tropics, with maximum impacts over land but extending over the oceans as well. This is likely a result of the anthropogenic or biomass burning emissions sources, which limit the largest differences in ozone and NOx between the two models to areas proximate to the South American, African, and Indonesian source regions for the month of January. The OH changes resulting from substitutions of the inputs JO1D and HCHO, however, are distributed over oceans as well as over land masses and, in the case of HCHO, are strongest in remote marine regions. This pattern is common for species that influence OH chemistry through mechanisms that are largely independent of local emissions. In the case of HCHO, its role as a secondary source of OH through methane oxidation is relatively more important in the absence of large VOC concentrations; thus, its stronger influence is seen away from terrestrial vegetation.

The second feature to note in Fig. 4 is the symmetry between input swaps in opposing directions. In other words, the swap of an input from OsloCTM into the GMI NN generally yields OH column and τCH4 changes that are equal but opposite to the changes resulting from use of a GMI input in the OsloCTM NN. With few exceptions, almost all regions of OH increase (red) in one model's NN are matched by OH decreases (blue) in the other model's NN in Fig. 4. The changes in τCH4 are correspondingly similar in magnitude but opposite in sign. This behaviour is expected because a swap that may, e.g. increase an OH precursor and subsequently cause an increase in OH for one model will manifest as a decrease in that same precursor when the substitution occurs in the NN of the other model. While this pattern occurs for the vast majority of cases across all model pairings and swaps performed for this analysis, there are instances when symmetry is not maintained. This could happen for two reasons.

First, the sensitivities of the two models to a particular change in an OH precursor or sink could differ. For example, one model may be sensitive to an increase in isoprene, causing OH concentrations to drop in response. Another model may incorporate buffering effects, such as reactions involving oxidized volatile organic compounds (Taraborrelli et al., 2012; Lelieveld et al., 2016) that allow OH to be recycled following its reaction with isoprene, causing it to be less sensitive to the same change in methane. We refer to these variations in model sensitivities as chemical mechanism differences, as they are most likely a result of the chemical reactions, species representations, or reaction rates implemented within a model's chemical mechanism.

The second explanation for lack of symmetry in the OH response to a model swap is a forced asymmetry in the swapped inputs themselves, imposed by the extrapolation control technique described in Sect. 3.2. It is possible that the swap of an input in one direction, i.e. from Model A to Model B, could proceed with no alteration to the substituted variable, while the swap in the other direction, i.e. from Model B to A, results in the variable lying outside the trained range of Model A. The extrapolation control process will revise the substitute variable field from Model B, such that the difference between it and the native field from Model A is lessened. As such, the first swap into the NN of Model B will yield a larger magnitude change in the input as compared to the swap into the NN of Model A. The impact of these factors is indirectly quantified through a remainder term that falls out of a full budgeting analysis, described below.

A third consideration in interpreting the information presented in Fig. 4 is the conditions that must be met in order for a large change in OH to manifest through this analysis. First, the two models between which a swap is conducted must exhibit differences in the parameter of interest. Should the two models exhibit, e.g. very similar ozone fields, then swapping one model's O3 with the other's will produce little difference in the NN-calculated OH. Second, the model must have some OH sensitivity to the variable being swapped. If a model is insensitive to changes in methane, swapping in a drastically different methane field may not cause a perceivable difference in OH. Therefore, the absence of an OH response does not necessarily mean that input fields are similar between models. Conversely, the existence of large OH changes indicates that differences in the swapped input field exist between the two models and that the native model demonstrates a dependence of OH on that input variable.

A fourth issue is the fact that NNs can exhibit some degree of random behaviour based on how they were trained and initialized. Our method involved training five NNs and selecting from those the one that performed best when compared to the independent test dataset. That single NN was used in all subsequent analysis. However, it is a useful exercise to evaluate the role of NN randomness in our results. We show, in Figs. S5 and S6, the left and right panels of Fig. 4, reproduced for the alternate NN trainings of the GMI and OsloCTM models, respectively. A visual comparison of tropospheric OH column differences among the five trainings of each model's NN reveals markedly similar spatial distributions and magnitudes. The values of calculated τCH4 changes (ΔτCH4) do differ somewhat between the training instances, with larger effects on some variable swaps than for others. For instance, the standard deviation of the values of ΔτCH4 calculated for all five trainings of the GMI NN is about 0.2 years for the J(O1D) and HCHO swaps but less than 0.05 years for O3 and NOx. We note, though, that some of the NNs displayed in Figs. S5 and S6 exhibit worse performance than the one ultimately chosen for subsequent use. As a result of this exercise, the uncertainties resulting from this analysis method may be considered, at most, to be ∼0.2 years.

The final point of interest in Fig. 4 is the general consistency in the signs of OH and ΔτCH4 for each model. The substitutions of all four variables generally cause an increase in OH within the GMI NN (and corresponding decrease in τCH4) and a decrease in OH (increase in τCH4) within the OsloCTM. This feature is most pronounced for this particular pair of models due to our reasoning for choosing them: they exhibit the largest difference in τCH4 among our group of 10 models. Because the native GMI model has a longer τCH4 value compared to OsloCTM, it makes sense that incorporation of OsloCTM's various OH precursor and sink fields into the GMI NN will tend to decrease the GMI τCH4, bringing it into closer agreement with that of OsloCTM. This characteristic points to the utility of this analysis as a budgeting tool for quantifying the cause of the difference in τCH4 between two models. The τCH4 accounting for the GMI and OsloCTM set of swaps conducted for January is shown in Table 1. When considering all 12 variable swaps that were performed, the NN analysis more than explains the original gap in τCH4 between the two models. The GMI January lifetime of 9.24 years is decreased to 6.71 years (τORIG+Δτ) after summing all Δτ values, while the OsloCTM lifetime is increased from 7.18 years to 9.48. This budgeting rarely provides a perfect accounting of the τCH4 gap due to the same reasons that give rise to asymmetric OH responses to a given swap: chemical mechanism differences and asymmetric swaps of inputs due to extrapolation control. As a result, a remainder term, found as the difference between the other model's τORIG and the present model's τORIG+Δτ, is attributed to these factors. This term is listed in the last row of Table 1 with the label “Mech.”

Results from analysing individual model pairs reveal a multitude of insights regarding idiosyncrasies in emissions of, global distributions of, and OH sensitivities to the various input parameters. These results, available in the archived dataset described in the data availability section, may be especially useful to the reader with an interest in a particular species or model. However, with over 4000 plots (12 species × 10 models × 9 submodels × 4 months = 4320) and 180 τCH4 budget tables generated, it is beyond the scope of this paper to highlight and explain every interesting feature. Instead, we aggregate the results across all models to identify some primary conclusions. Figure 5 shows the change in τCH4 for a specific model and substituted input variable, averaged over all nine pairings. For example, the data point shown for CAM4-Chem JO1D is calculated from the nine ΔτCH4 values obtained when swapping the JO1D fields from the other nine models into the CAM4-Chem NN. The circular point represents the mean of those nine values, while the whiskers indicate 1 standard deviation about the mean. Aggregate results shown in this manner are compiled both for individual months (available in the archived dataset noted above) as well as for annually averaged output. The latter is calculated as the average of the four monthly mean and standard deviation values, and is shown in Fig. 5.

As with the individual OH tropospheric column change plots (Fig. 4), numerous conclusions can be drawn by studying the aggregated results in Fig. 5. The method for reading the data in Fig. 5 is demonstrated in the following example. The mean ΔτCH4 value attributable to JO1D for the WACCM model is +0.99 years. This indicates that use of JO1D fields from other models causes τCH4 to increase by ∼1 year, meaning the native JO1D field from WACCM imparts a low bias to τCH4 of 1 year, relative to the other models. A low τCH4 would result from OH concentrations being too high. Since OH and JO1D are positively correlated (i.e. JO1D can be thought of as a source for OH), the too-high OH is an indication of too-high JO1D. In general, positive values of ΔτCH4 correspond to relative high biases in input parameters that are source terms for OH and to low biases for species that instead serve as sinks. This reasoning is less straightforward for species such as HCHO, which can both produce and consume OH, while it is also produced by OH-initiated oxidation. We stress that these comparisons are strictly relative to other models, not to any observation or other indication of truth. So, points that appear as outliers in Fig. 5 should not necessarily be interpreted as an erroneous result but rather considered an area for further examination.

The ordering of variables along the x axis of Fig. 5 denotes the average magnitude of ΔτCH4 values across all models, with parameters on the left accounting for the largest τCH4 differences. As such, JO1D is the largest driver of OH differences in the CCMI SD model simulations, followed by local O3 and NOx. The subsequent variables (H2O, CO, the NO:NOx ratio, and HCHO) cause moderate variations in tropospheric OH, while ISOP, CH4, JNO2, O3 COL, and T are not responsible for intermodel spread in τCH4. We note that T differences between the SD simulations are likely limited due the meteorological constraints imposed on the models. However, examination of the free-running simulations, discussed in the Supplement, also shows practically no impact of T on OH. Thus, we conclude that the effect of temperature on OH chemistry is likely indirect, acting through pathways embodied by other variables, such as H2O and species that exhibit strongly temperature-dependent reaction rates. Finally, the Mech. term, described in the discussion of Table 1, appears on the far right, indicating its origins as a remainder term from the budget analysis of individual model pairs. The magnitudes of ΔτCH4 values attributed to chemical mechanism differences and asymmetric swaps between models are large enough to consistently rank the Mech. term third, between O3 and NOx, in terms of importance for OH in this analysis. Especially in model simulations conducted with common emissions inventories (though inventories can be implemented very differently among models, as demonstrated by Young et al., 2013), we expect some of the disparity in a short-lived species like OH to emerge from differences in chemical mechanism implementations. In other words, when responses in OH to a given change in a source or sink term differ between two models, the remainder term (or term labelled “Mech.” in Table 1) will increase, representing variations in the sensitivity of OH that presumably arise due to the two different implementations of the chemical mechanism. It is possible that other factors are represented by this term; e.g. other chemical species that influence OH chemistry but are not considered in the NN analysis could contribute to the Mech. term. However, previous analysis using a 0-D chemical box model as a “standard” mechanism in Nicely et al. (2017) suggested a correlation between actual biases in OH imparted by a given model's chemical mechanism and the remainder term resulting from the NN analysis. Therefore, we have some confidence that the Mech. term is meaningful, though significant further study would be required to parse the actual mechanistic differences responsible for imparting bias in OH calculations.

Significant intermodel differences in the largest driver of τCH4 spread, JO1D, could arise from two possible sources. The amount of solar UV light penetrating down to the troposphere is largely dictated by the stratospheric column ozone amount. However, the differences in total ozone column are generally small and insufficient to cause the variations in JO1D seen among the CCMI models. Rather, JO1D likely varies to a great extent due to differences in cloud cover, and dissimilar treatments of clouds within model photolysis codes. Figure S7 highlights this effect by showing the ratio of JO1D at the surface to JO1D in the upper troposphere (UT) for each model. The relatively small column amounts of ozone within the troposphere should account for very little absorbed UV light, making it much more likely that deviations in this ratio from 1.0 are driven by scattering due to clouds and possibly aerosols. The fact that models show large spatial differences in this ratio is a strong indication that clouds underlie the model differences in JO1D.

While the model differences in JO1D, O3, NOx, and chemical mechanisms appear to drive the bulk of the τCH4 spread among this group of CCMI models, we emphasize that individual models may not adhere to these conclusions. As such, any efforts to improve a particular model should instead focus on the results specific to that model. For instance, HCHO plays a very small role in describing intermodel differences in OH on average, but for the OsloCTM model, HCHO is a much more important factor. Thus, we refrain from offering an across-the-board solution for remedying the large model spread in τCH4 and instead suggest a more individualized approach of studying plots such as those shown in Fig. 4 for more spatially and temporally resolved information. Visualizations of all model swaps, for all months and species, are available in our archived dataset described in the data availability section for this purpose.

There are several other qualifications to note when considering the results of the intermodel comparison. One is the negating effect between the JO1D and tropospheric O3 variables. Many, but not all, model ΔτCH4 values for JO1D in Fig. 5 are opposite in sign to the ΔτCH4 values attributed to O3. Physically, photolysis of tropospheric ozone by light at wavelengths below 336 nm to form excited state O(1D) and subsequent reaction with H2O to form OH is a loss pathway for ozone. Therefore, more UV flux will tend to decrease tropospheric ozone concentrations while increasing OH, and vice versa. This physical mechanism, then, can explain the frequent cancellation of the ΔτCH4 values attributed to these two factors. Should a modeller attempt to alter a model's OH field by forcing adjustments in its JO1D, the opposing impact of tropospheric O3 may result in no change for the value of τCH4. However, this does not preclude the finding that both JO1D and tropospheric O3 are substantially different in the models for reasons we do not fully understand. Tropospheric ozone can also vary between models for reasons external to the radiative environment. For instance, differences in the stratosphere–troposphere exchange, wet and dry deposition, and lightning NOx emissions can each cause substantial variations in tropospheric ozone among models (Wild, 2007). Further parsing of the reasons for the ozone differences seen among the CCMI models is difficult without specialized output, including tracers such as ozone of stratospheric origin and NOx generated by lightning. We recommend a targeted study to address the underlying reasons for the variations in tropospheric ozone.

Another qualification concerns the issue of causation versus correlation. Machine learning techniques, and NNs in particular, are generally more adept at identifying the predictors of a certain phenomenon than traditional methods, such as multiple linear regression. However, it is still possible that an input that is tightly correlated with the output may be misidentified as a driver of variations in the output. This is particularly relevant to keep in mind for species that serve as sinks of OH, such as CO and methane. Whether a decline in OH initiates or results from an increase in its sinks is difficult to differentiate, even with advanced analysis methods. Therefore, descriptions of CO and methane as drivers of OH variations in this text may just as well be interpreted conversely, as downstream indicators of the change in oxidizing capacity.

A final qualification is this analysis constitutes a foundationally hypothetical experiment. It essentially addresses the following questions: what if we could switch the fields of just one chemical species between two global models? What would be the instantaneous impact on OH? On τCH4? This approach, then, necessarily neglects the roles of feedbacks in the atmospheric system (e.g. if the NOx field is perturbed, this will propagate to changes in ozone as well, with time). However, for the objective of teasing apart the influences on global OH abundance and τCH4 and explaining intermodel differences, a notoriously difficult task, we regard our approach as a valuable exercise.

The second half of our NN analysis interrogates temporal trends in OH and τCH4. Figure 6 shows the evolution of τCH4 in the SD-type simulations conducted for 1980–2010. Two models, GEOS-Chem and OsloCTM, only provided output for the year 2000 and thus only appear as single points in Fig. 6. In addition, some models provided output beyond the year 2010; output from years through the end of 2015 was included when available. The lifetimes all show a general downward trend over time, consistent with the upward trend in global mean tropospheric OH concentration shown by Zhao et al. (2019; their Fig. 4). Results concerning attribution of the τCH4 time series are presented in Sect. 4.3.1, while derivation and analysis of trends are shown in Sect. 4.3.2.