The impact of internal climate variability on OH trends between 2005 and 2014

The hydroxyl radical (OH) lies at the nexus of climate and air quality as the primary oxidant for both reactive greenhouse gases and many hazardous air pollutants. To better understand the role of climate variability on spatiotemporal patterns of OH, we utilize a 13-member ensemble of the Community Earth System Model version 2-Whole Atmosphere Community Climate Model version 6 (CESM2-WACCM6), a fully coupled chemistry-climate model, spanning the years 1950–2014. Ensemble members vary only in their initial conditions of the climate state in 1950. We focus on the final decade of the simulation, 2005–2014, when prior studies disagree on the signs of the global OH trends. The ensemble mean global airmass-weighted mean tropospheric column OH ( ΩTOH ), which is an estimate of the forced signal, increases by 0.06%/year between 2005 and 2014 while regional ΩTOH trends range from −0.56%/year over Southern Europe to +0.64%/year over South America. We show that ten-year ΩTOH trends are strongly affected by internal climate variability, as the spread of ΩTOH trends across the ensemble varies between 0.23%/year in Asia and 1.53%/year in South America. We train a fully connected neural network to emulate the ΩTOH simulated by the CESM2-WACCM6 model and combine it with satellite observations to interpret the role of OH chemical proxies. While the OH chemical proxies are subject to internal variability, the impact of internal variability on ΩTOH trends is primarily due to the meteorological parameters except for South America. Forced trends in global mean ΩTOH do not unambiguously emerge from trends driven by internal variability over the 2005–2014 period. The observation-constrained ΩTOH presents opposite trends due to climate variability, resulting in varying conclusions on the attribution of OH to CH4 trends.


Introduction
The hydroxyl radical (OH) determines the oxidative capacity of the troposphere (Levy 1971).OH influences air quality and climate by reacting with various trace gases, including nitrogen oxides (NO x ), non-methane volatile organic compounds (VOCs), carbon monoxide (CO), some ozone-depleting substances, and methane (CH 4 ).CH 4 stands out as a significant anthropogenic greenhouse gas, estimated to account for a third of the global temperature rise since the preindustrial era (Masson-Delmotte et al 2021).
OH has high reactivity and a correspondingly brief lifetime (∼1 s or less) (Lelieveld et al 2016), and is determined by a complex set of chemical and meteorological conditions.The primary production of OH is closely coupled to both water vapor and ultraviolet radiation.Globally, an increase in NO x leads to an increase in OH while the increases in VOCs and CO lead to a decrease in OH (e.g.Lelieveld et al 2002, Zhao et al 2023, Fiore et al 2024).While direct OH observations are not possible at the temporal and spatial scales relevant for understanding CH 4 trends, global OH can be inferred using inverse modeling approaches, which rely on the observation of one or more long-lived species where the OH reaction serves as the primary loss pathway, including methyl chloroform (MCF) (Prinn et al 1992, 2001, McNorton et al 2015, Rigby et al 2017, Turner et al 2017, Naus et al 2019, Patra et al 2021) and hydrofluorocarbon (HFC) species (Thompson et al 2024).OH can also be simulated in chemistry-climate models.Murray et al (2021) showed that these models disagree by ±30% in global mean OH and in the sign of its changes from the preindustrial to late 21st century, even when forced with identical anthropogenic emissions.Stevenson et al (2020) analyzed the global mean OH trends in the Coupled Model Intercomparison Project Phase 6 (CMIP6)/Aerosols and Chemistry Model Intercomparison Project (AerChemMIP) simulations and demonstrated that the model-derived OH trends from 1980 to 2005 are broadly consistent with trends inferred from different inversion methods.However, from 2005 to 2014, the models suggest a continued slight upward or near zero trend, contrasting with the declining OH trends reported in some inversion studies (Rigby et al 2017, Turner et al 2017).
The discrepancy in OH trends spanning 2005 and 2014 between forward and inverse model simulations is unresolved.Both approaches embed nonnegligible uncertainties.In the 'top-down' inverse modeling, the uncertainties in estimated OH trends arise from surface sampling bias and the choice of parameters prescribed in the inversions, including the inter-hemispheric exchange ratio and stratospheric loss rate.For instance, Naus et al (2019) accounted for each of the parameter biases in a two-box model OH inversion framework and found that there was no consistent decline OH trend as observed in Rigby et al (2017) andTurner et al (2017).In the 'bottomup' chemistry-climate model simulations, each parameter relevant to OH chemistry may be biased and any bias can propagate and affect the OH simulations.For instance, Zhao et al (2023) showed that the atmospheric chemistry models overestimate the global tropospheric airmass-weighted OH, primarily due to model underestimates of carbon monoxide and total ozone columns, and overestimated nitrogen dioxide.Nicely et al (2018) also suggested that the simulated OH can change with climatic conditions, including rising tropospheric water vapor, rising temperature, and widening of the climatological tropics due to expansion of the Hadley cell.
The OH trend between 2005 and 2014 can arise from changes in emissions of chemical parameters related to OH chemistry, changes in climatic conditions, as well as from internal variability.Unlike a chemical transport model, the chemistry-climate model simulates its own meteorology along with chemistry, which allows us to separate the roles of anthropogenic forcing versus climate variability since each ensemble member is forced with the same anthropogenic emissions, boundary conditions, and land-use change, but produces a unique meteorology consistent with those forcings.In this study, we utilize an initial-condition ensemble generated with a chemistry-climate model to explore the OH trends due to external forcing (i.e.anthropogenic emission change) and the sensitivity of OH trends to internal climate variability.We also investigate the impact of internal climate variability on three OH chemical proxies, including the tropospheric columns of nitrogen dioxide (Ω TNO2 ), formaldehyde (Ω THCHO ), and carbon monoxide (Ω TCO ).We then leverage machine learning (ML) techniques to emulate the OH chemistry in the model and to interpret the role of these three OH chemical proxies in the context of their responses to forcings versus internal variability, as well as the direct influence of meteorological variability on OH trends.

Model
We utilize a 13-member initial-condition ensemble of the Community Earth System Model version 2-Whole Atmosphere Community Climate Model version 6 (CESM2-WACCM6) for our study.The model simulation covers the historical time period from 1950 to 2014, with a horizontal resolution of 1.25 • × 0.95 • and 70 vertical layers.Both anthropogenic and biomass burning emissions are obtained from the CMIP6 (Van Marle et al 2017).
As described in Fiore et al (2022), these 13 model realizations differ only in their initial climate states while all forcings (e.g.greenhouse gases and air pollutant emissions) were identical except for any natural emissions that are tied to meteorology (Gettelman et al 2019, Danabasoglu et al 2020, Emmons et al 2020).We use the monthly average 3D OH fields from the model ensemble and integrate OH from the surface to the tropopause to calculate the air mass-weighted tropospheric column OH (referred to as Ω TOH hereinafter).Therefore, the ensemble average of Ω TOH is taken to represent the signals due to anthropogenic and natural forcings (e.g.soil NO x emissions and volcanic eruption which is identical in all ensemble members), while the range across the ensemble provides a measure of internal climate variability.

Neural network (NN)
Previous studies have applied ML techniques to predict OH, including gradient-boosted trees (Anderson et al 2022, 2023, Zhu et al 2022a, 2022b) and NNs (Nicely et al 2017, 2020).Following Nicely et al (2020), we train a fully connected NN to emulate the CESM2-WACCM6 Ω TOH at the native model resolution.Here we use 11 monthly average variables as the input features, including 2 physical parameters (latitude, month), 4 meteorological parameters (lightning flash rates, cloud cover, air mass-weighted tropospheric column of water vapor and temperature), 2 photolysis parameters (airmass-weighted NO 2 and O 3 photolysis rates, J(NO 2 ) and J(O 1 D)), and 3 chemical parameters (Ω TNO2 , Ω THCHO , and Ω TCO ).
The architecture of the NN is distinct from Nicely et al (2020).The NN consists of four hidden layers, each containing 256, 256, 128, and 64 nodes, respectively.At each node, a ReLU (Rectified Linear Unit) function is performed on the linear combination of each input multiplied by a unique weighting.The ADAM algorithm is used to adjust NN weights during training (Kingma and Ba 2014), based on the second derivatives of the errors of the simulated Ω TOH calculated using the validation dataset.The NN configuration is described in further detail in section S1.For training and validation, we use the simulations from one ensemble member of the CESM2-WACCM6 model between 1950 and 2000, and randomly split 80% for training and the remaining 20% for validation.For testing and analysis, we use this trained NN to predict the gridded Ω TOH between 2005 and 2014 for each of the 13 ensemble members.We reprocess the satellite products to align them with the monthly average tropospheric column densities archived from CESM2-WACCM6.To account for the vertical sensitivity of the satellite instrument, we recalculate the daily air mass factor by replacing the a priori profiles with the vertical profiles from the ensemble average of the CESM2-WACCM6 model.Additionally, the daily overpass time of OMI is 13:30 local time, and the daily overpass time of MOPITT is 10:30 local time.We use the MERRA2-GMI model simulations (https://acd-ext.gsfc.nasa.gov/Projects/GEOSCCM/MERRA2GMI/) to derive scaling factors to convert from the snapshot at the overpass time to daily average column densities.We then average the daily column densities to the monthly scale and re-grid them to match the horizontal resolution of the CESM2-WACCM6 model simulation.

Trend analysis
We consider four sets of Ω TOH between 2005 and 2014.The first set originates directly from the 13 ensemble members of CESM2-WACCM6.The second set comprises the NN predictions ('ML-NN') with the input features from each of the CESM2-WACCM6 members.We replace three OH chemical proxies from each ensemble member with ensemble averages from CESM2-WACCM6 to yield a third set ('ML-NN-ChemAvg').For the fourth set, we replace the chemical parameters (Ω TNO2 , Ω THCHO , and Ω CO ) from the model simulations with the re-processed satellite observations (section 2.3).We then combine them with other parameters from the CESM2-WACCM6 simulations to yield the partially observationally constrained NN prediction ('ML-NN-Sat').The observed record includes both the forced signal and some unknown influence of climate variability over the 2005-2014 period.We infer biased model simulations for any case where the satellitederived trend falls outside the range simulated by the model ensemble.Differences between 'ML-NN' and the original CESM2-WACCM6 ensemble members provide an estimate of the NN accuracy.Differences between 'ML-NN' and 'ML-NN-ChemAvg' show the influence of climate variability operating solely through the OH chemical proxies.Because 'ML-NN-Sat' includes both forcing and climate variability, the difference from 'ML-NN-ChemAvg' reflects the combined influence of model biases in the forcings, biases in the model responses to those forcings, plus any (unknown) deviations due to different influences from internal climate variability.
In this study, we focus specifically on the Ω TOH trend between 2005 and 2014 rather than Ω TOH in each year.We compute the annual Ω TOH normalized to the 2005 mean (ratio of annual Ω TOH to 2005 mean).As in Chua et al (2023), we calculate the Ω TOH trends defined as the slope from the linear regression of normalized Ω TOH versus year using the Theil-Sen method.In addition, we calculate the trends of Ω TNO2 , Ω THCHO , and Ω CO from both CESM2-WACCM6 model simulations and satellite observations (section 3.2).

Results and discussion
Even though our ensemble is too small to capture the full possible range of climate variability compared to other studies that include 40-100 initialcondition ensemble members (e.g.Deser et al 2012, 2020a, 2020b), we calculate the trend of the ensemble average of Ω TOH , which offers the best estimate of the changes occurring in response to anthropogenic and natural forcing.We then use the range across the ensemble to assess the sensitivity of the trend of Ω TOH to internal climate variability.Conversely, negative Ω TOH trends are simulated in the Eastern region of North America and the majority of Southern Europe.At the continental scale, the trends of ensemble average Ω TOH in CESM2-WACCM6, exhibit positive trends in Asia (0.25%/year) and South America (0.64%/year), and negative trends in North America (−0.26%/year) and Southern Europe (−0.56%/year).

Extracting forced signals in
However, Ω TOH trends vary across the ensemble due to internal climate variability.Individual ensemble members simulate opposite-signal Ω TOH trends in most regions.Figure 1(b) shows the spatial distribution of the standard deviation of Ω TOH trends between 2005 and 2014, which ranges from 0%/year to 1.67%/year.Two regions stand out for their sensitivities of Ω TOH trends to internal climate variability (large standard deviation).The first region, covering the Pacific Ocean between 25 • S and 10 • N, is influenced by the El Niño-Southern Oscillation (ENSO), exhibiting a large standard deviation of OH trend at 1.32%/year.ENSO is a dominant mode of naturally occurring climate variability and it is known to impact OH (e.g.Turner et al 2018, Anderson et al 2021).The second sensitive region is the Amazon rainforest in South America, where the standard deviation of Ω TOH trends is as high as 1.67% per year.At the continental scale (figure 2), the spread (maximum-minimum) of Ω TOH trends across the ensemble is as large as 1.53%/year in South America.Asia experiences weaker influences from climate variability, as the spread of Ω TOH trend is 0.27%/year across the ensemble.

Interpreting the impact of forced signals and internal variability in the trends of OH chemical
proxies, including Ω TNO2 , Ω TCO and Ω THCHO While internal climate variability directly leads to varying meteorological parameters, it also affects the chemical parameters related to OH chemistry because they respond to changes in meteorology.Here, we select Ω TNO2 , Ω TCO and Ω THCHO , as OH chemical proxies for NO x , CO and VOCs.In figure 3, we investigate the sensitivity of forced signals and internal variability to the trends of these three OH chemical proxies.Besides, the impact of OH chemical proxies on OH trends can be biased as the OH chemical proxies are not accurately represented in the model.Therefore, we also examine whether or not the observed trend from satellite observations falls into the range of the modeled trends to evaluate the model skill in representing these OH chemical proxies.The modeled grids are sampled only where satellite observations are available.As the sunlight-dependent OMI observations are missing during winter in the higher latitudes of both hemispheres, the comparison is confined to 40 • S to 45 • N (figure S1).The geographical extent of continents within these boundaries is shown in figure S2.The comparison of these OH chemical proxies at native model resolution is described in section S2 and summarized below.
The modeled Ω TNO2 trends (figure 3(a)) mainly reflect the emissions trends (figure 3(b)), implying they are forced, with little spread across the ensemble (ranging from 0.32%/year in Asia to 1.72%/year in Australia).The ensemble average successfully reproduces the Ω TNO2 trends derived from satellite observations in Asia and North America, demonstrating the model's capability to represent the trends of anthropogenic NO x emissions in both continents.The most significant discrepancy in the Ω TNO2 trends between the model and observation is in Southern Europe, where the model simulates a larger declining trend of −2.59%/year in NO 2 column, consistent with the trend in emissions but in contrast to the satellite observations of −0.80%/year.It indicates that the anthropogenic NO x emissions in this region are biased, or weather-sensitive soil NO x emissions could be offsetting the decline in the anthropogenic emissions in a way not captured by the model.
In contrast, the ensemble average of modeled Ω TCO trends are not consistent with the trends of anthropogenic CO emissions (figures 3(c) and (d)).Both modeled Ω TCO and Ω THCHO trends are more strongly influenced by internal climate variability, as indicated by the spread across the individual circles in figures 3(c) and (e).The spreads (maximumminimum) in both Ω THCHO and Ω TCO trends across the ensemble are most pronounced in South America, with the min-max range of 2.35%/year for Ω THCHO and 1.22%/year for Ω TCO .The impact of climate variability on Ω THCHO and Ω TCO trends arises from climate-sensitive biogenic emissions (figure 3(f)), as both HCHO and CO are chemically produced by biogenic VOC oxidation (Wolfe et al 2016, Worden et al 2019).
The CESM2-WACCM6 model yields a range of modeled Ω TCO trends covering the satellite-derived trends in North America, Southern Europe, South America, and Australia, yet it fails to represent the declining Ω TCO trends in Asia and Africa observed by MOPITT.The satellite-derived Ω THCHO trends are within the range of modeled trends in Asia, South America, and Australia.However, the modeled Ω THCHO diverges in North America and Africa, with a trend of 0.71%/year and 0.44%/year, respectively.The CESM2-WACCM6 model also fails to capture the satellite-derived Ω THCHO trends in Southern Europe, yielding a near-zero trend.However, due to the substantial influence of internal variability, the mismatch between model simulations and satellite observations can not be unambiguously attributed to an error in emissions, as additional constraints are needed on the inter-annual variations in weather-dependent regional biogenic sources over the decade.

Interpreting the role of chemical proxies on OH trends
We interpret the impact of internal climate variability on continental-scale OH trends and tease out the role of OH chemical proxies using ML.In figure 2, besides continental-scale Ω TOH trends from the CESM2-WACCM6 model, we show NN predictions using input features from CESM2-WACCM6 ('ML-NN'), and predictions using the ensemble average of the OH chemical proxies from CESM2-WACCM6 ('ML-NN-ChemAvg') and satellite observations ('ML-NN-Sat') as inputs.
The NN is trained to emulate the OH chemistry.As shown in figure 2, we apply the NN separately to each ensemble member to calculate the continental-scale Ω TOH trends from NN predictions, reporting both the ensemble average and the variation across the ensemble members.The ensemble average Ω TOH trends from the NN predictions ('ML-NN') capture the continental variations in the CESM2model except in Asia and South America.The NN underpredicts the ensemble average Ω TOH trend by 0.24%/year in Asia and by 0.48%/year in South America.It indicates that some of the Ω TOH variations in these two regions cannot be captured by the input features.Despite the shifts in the ensemble average Ω TOH trend, the NN reproduces the variation in Ω TOH trends due to internal climate variability across the ensemble at each continent, which arises from both direct impacts through meteorological parameters (temperature, water vapor, lightning flash rates, and cloud cover) and indirect impacts through the OH chemical proxies (Ω TNO2 , Ω THCHO and Ω TCO ).The evaluation of the NN performance is further described in section S3.
We then replace the OH chemical proxies from each ensemble member with the ensemble average from CESM2-WACCM6 and with satellite observations, both of which are used as inputs to NN to yield 'ML-NN-ChemAvg' and 'ML-NN-Sat' Ω TOH trends, respectively.The 'ML-NN-ChemAvg' Ω TOH trends at the continental scale include the climate variability in meteorological parameters only and the 'ML-NN-Sat' Ω TOH trends further correct any model bias relative to the satellite-derived trends in representing these OH chemical proxies including any differences in the real-world forcing, the chemical response to that forcing, and the particular trajectory of climate variability over the decade.When we keep the OH chemical proxies constant across the ensemble members by switching to the ensemble average, the spread of Ω TOH trends in the 'ML-NN-ChemAvg' predictions is comparable to the spread in 'ML-NN' predictions.It suggests that in most continents, the impact of internal climate variability on Ω TOH trends is primarily due to the meteorological parameters.The exception is South America, where the variation in Ω TOH trends is partially attributed to the varying OH chemical proxies (e.g.biogenic emissions) due to internal climate variability.
While the trends of OH chemical proxies are different in CESM2-WACCM6 compared to the satellite observations (section 3.2), the changes in Ω TOH trends are small when we switch from the OH chemical proxies from CESM2-WACCM6 to the satellite observations (figure 2).By removing interensemble variation in Ω TNO2 , Ω THCHO and Ω TCO , both 'ML-NN-ChemAvg' and 'ML-NN-Sat' Ω TOH trends exhibit smaller variations over every continent.In North America, the trend of ensemble average Ω TOH constrained by satellite observations shifts from −0.22%/year to 0.01%/year (figure 2), which is most likely due to the difference in the Ω THCHO trends between CESM2-WACCM6 and satellite observations (figure 3).An increase in the 'ML-NN-Sat' ensemble average Ω TOH trend also occurs in Southern Europe, changing from −0.46%/year to −0.31%/year.

Investigating the impact of OH chemical proxies on global Ω TOH trends and its implications for CH 4 trends
We also average the Ω TOH to the global scale between 2005 and 2014, and the corresponding Ω TOH trend is shown in figure 4. We compare Ω TOH trends derived here versus previous studies for other periods in table S1

Conclusion
We investigate Ω TOH trends in the initial-condition ensemble of the CESM2-WACCM6 chemistryclimate model.Even though the model neglects to include the role of internal variability in driving soil NO x or biomass burning emissions, we show continental-scale variations in Ω TOH trends, ranging from −0.56%/year over Southern Europe to +0.64%/year over South America.As these model ensemble members only differ in the initial climate state, the spread of Ω TOH trends across the ensemble members provides a measure of internal climate variability.We show that the Ω TOH trends in Asia present weaker climate variability than the Ω TOH trends in South America, implying differences in the dominant drivers of OH variation across continents that future work should explain further.
We interpret the sensitivity of forced signals and internal variability to the trends of three OH chemical proxies, including Ω TNO2 , Ω THCHO and Ω TCO in CESM2-WACCM6 model.We show that the trends of Ω TCO and Ω THCHO are more sensitive to internal variability than Ω TNO2 trends.Comparing against satellite observations, we show that the satellitederived Ω TNO2 trends fall within the range of modeled Ω TNO2 trends in the CESM2-WACCM6 model except for Southern Europe where the model suggests larger decreasing trends.Besides, the CESM2-WACCM6 model fails to capture the observed trends of Ω THCHO over North America, Southern Europe, and Africa, and Ω TCO over Asia and Africa, though we cannot rule out the possibility that our small ensemble does not fully capture the possible range of internal variability.
We train a NN as an emulator of the CESM2-WACCM6 model.The NN adeptly reproduces Ω TOH trends in the CESM2-WACCM6 model in most continents, albeit without completely capturing continental Ω TOH trends in Asia and South America.By combining the NN with OH chemical proxies from ensemble average and satellite observations, we investigate how these OH chemical proxies affect the impact of interval variability on Ω TOH trends.We show that the impact of internal climate variability on Ω TOH trends is primarily due to the meteorological parameters except for South America.Compared to the CESM2-WACCM6 model, the observationconstrained Ω TOH trends present a smaller variation between continents.The global Ω TOH trends derived from the observation-constrained Ω TOH predictions presents a wide spread due to the internal climate variability, ranging from −0.25%/year to 0.07%/year.Q Z was supported by the NOAA Climate & Global Change Postdoc Fellowship.A M F acknowledges the National Aeronautics and Space Administration (NASA) Grant (80NSSC23K0925).We also acknowledge the free use of QA4ECV tropospheric NO 2 and HCHO column data from the OMI sensor (www.temis.nl).The CESM project is supported primarily by the National Science Foundation.We would like to acknowledge high-performance computing support from Cheyenne (doi: https:// doi.org/10.5065/D6RX99HX)provided by NCAR's Computational and Information Systems Laboratory, sponsored by the National Science Foundation.
The three OH chemical proxies used in the NN, Ω TNO2 , Ω THCHO and Ω TCO , are routinely retrieved from satellite intruments.The CO mixing ratio vertical profiles are obtained from the spaceborne MOPITT instrument (Measurements Of Pollution In The Troposphere v9 TIR-NIR product; (Deeter et al 2022)).The daily tropospheric NO 2 vertical column density and tropospheric HCHO column density are acquired from the QA4ECV (Quality Assurance for Essential Climate Variables) OMI (Ozone Monitoring Instrument) retrieval (Boersma et al 2011, 2018, De Smedt et al 2015).Previous studies have validated the trends of Ω TNO2 , Ω THCHO and Ω TCO from these satellite observations by comparing them against surface observations and other satellite products (e.g.Worden et al 2013, Lamsal et al 2015, Wang et al 2022).

Figure 1 .
Figure 1.The Ω TOH trend is driven by both forcings (anthropogenic emissions) and internal variability.Spatial distribution of the ensemble average (a) and the standard deviation (b) of Ω TOH trend (%/year).Hatching in (a) represents regions where consistent signed OH trends occur in all 13 ensemble members, indicating a clear signal from anthropogenic forcing.
Ω TOH trends from internal climate variability in CESM2-WACCM6 Figure 1(a) shows the spatial distribution of the ensemble average Ω TOH trends at the native model resolution between 2005 and 2014.The Ω TOH trends are further integrated into the continental scale, as shown in figure 2. Positive Ω TOH trends are prominent in the Eastern and Southern regions of Asia, the tropical Pacific belt, and South America.

Figure 2 .
Figure 2. The role of OH chemical proxies in interpreting the impact of internal climate variability on Ω TOH trends.Continental-scale OH trends from CESM2-WACCM6 model simulations in black, NN prediction ('ML-NN') in blue, the NN prediction with ensemble average OH chemical proxies ('ML-NN-ChemAvg') in purple, and the NN prediction constrained by satellite observations ('ML-NN-Sat') in red.The dots represent the OH trends from each ensemble member.The triangles represent the trends of the ensemble average OH.

Figure 3 .
Figure 3.The trends of OH chemical proxies are determined by both forcings (e.g.anthropogenic emissions of OH source and sink gases) and internal variability.The continental-scale trends of OH chemical proxies and emissions, including Ω TNO2 (a), Ω THCHO (c), and ΩCO (e), as well as anthropogenic NOx emission (b), anthropogenic CO emissions (d) and biogenic isoprene emissions (f).The blue dots represent the trends of each ensemble member.The blue triangles represent the trends of the ensemble average of each OH chemical proxy.The red triangles represent the trends of each chemical OH driver from the satellite observations.

Q
Figure 4. Global annual mean OH trend between 2005 and 2014 in our study and as reported in inversion studies constrained by methyl chloroform observations (Rigby et al 2017, Turner et al 2017, Naus et al 2019).The solid black line represents the ensemble average global annual OH from CESM2-WACCM6 model simulations, and the gray shadings represent the spread across ensemble members.The solid blue and red lines represent global OH trends from NN prediction ('ML-NN') and the NN prediction constrained by satellite observations ('ML-NN-Sat'), respectively.The corresponding spread across the ensemble members is denoted as blue and red shading.The dashed olive and red lines denote global OH trends from Rigby et al (2008), in which global OH is estimated using inversions constrained by methylchloroform measured at two observation networks, the Advanced Global Atmospheric Gases Experiment and NOAA.The dashed purple and green lines denote global OH trends from Turner et al (2017) and Naus et al (2019), respectively.
. The ensemble-mean global Ω TOH remains relatively stable between 2005 and 2014, with a trend of 0.06%/year.The global Ω TOH trends range from −0.16%/year to 0.34%/year across the ensemble members, highlighting the influence of internal climate variability.This global Ω TOH trend is consistent with the simulated Ω TOH trends from Stevenson et al (2020), despite differing from Ω TOH trends estimated from MCF inversions (e.g.Rigby et al 2017, Turner et al 2017, Naus et al 2019).The large disparity in Ω TOH trends from previous studies (e.g.Rigby et al 2017, Turner et al 2017, Naus et al 2019) reflects the intrinsic uncertainty in MCF inversions due to the choice of parameters, which highlights the necessity of constraining Ω TOH trends from other approaches.The NN generally captures the OH simulation in the CESM2-WACCM6 model, albeit with a slight underprediction, as the ensemble average global Ω TOH trend from NN predictions is −0.12%/year.However, this difference between CESM2-WACCM6 and NN is minor compared to the variation between modeled Ω TOH trends and those from MCF inversions.The NN predictions, constrained by OH chemical proxies from satellite observations, yield a near-global Ω TOH trend of −0.09%/year.Even though it only includes regions where satellite observations are available, we show that the near-global Ω TOH trend from 'ML-NN-Sat' predictions is representative of the global Ω TOH trend (section S4).The influence of internal climate variability leads to a spread of global Ω TOH trends from 'ML-NN-Sat' predictions, resulting in varying conclusions on the cause of the increasing CH 4 trend observed after 2006.The largest decline in Ω TOH trends from 'ML-NN-Sat' predictions, at −0.25% per year, explains some of the increasing CH 4 trends if CH4 emissions remain constant (Turner et al 2017).However, the upper bound in Ω TOH trends, at 0.07% per year, indicates that CH 4 declines due to OH changes.Therefore, the observed increasing CH 4 trend is likely driven by an increase in CH 4 emissions offsetting the declining trend due to increasing OH trends.