We use a prototype version of the new-generation GFDL chemistry–climate model, GFDL-AM4.1 (Zhao et al., 2018a, b; Horowitz et al., 2020). A detailed description of the physics and dynamics in AM4.1 is provided by Zhao et al. (2018a, b). The version of AM4.1 with full interactive chemistry used in this work is described by Schnell et al. (2018). In its standard form, this model setup consists of a cubed-sphere finite-volume dynamical core with a horizontal resolution of ∼100 km and 49 vertical levels extending from the surface up to ∼80 km. The model's lowermost level is approximately 30 m thick. The chemistry and aerosol physics in this model have been updated from the previous version (GFDL-AM3; Naik et al., 2013a), as described by Mao et al. (2013a, b) and Paulot et al. (2016). There are a total of 102 advected gas tracers and 18 aerosol tracers, 44 photolysis reactions, and 205 gas-phase reactions included in the chemical mechanism in this version of AM4.1 to represent tropospheric and stratospheric chemistry.

The standard AM4.1 configuration uses global annual-mean methane concentrations as a lower boundary condition to simulate the atmospheric distribution of methane. This modeling framework does not allow for the full characterization of the drivers of methane trends and variability, nor does it capture latitudinal or seasonal variations in methane. To overcome this issue, we updated AM4.1 to be driven by methane emissions. Table 1 provides information on the methane emission datasets used in this work. Our initial estimates of surface emissions from anthropogenic sources – including agriculture (AGR), energy production (ENE), industry (IND), road transportation (TRA), residential, commercial, and other sectors (RCO), waste (WST), and international shipping (SHP) – are from the Community Emissions Data System (CEDS, version 2017-05-18; Hoesly et al., 2018) developed in support of the Coupled Model Intercomparison Project Phase 6 (CMIP6) for 1980–2014. Emissions for 2015–2017 are from a middle-of-the-road scenario of Shared Socioeconomic Pathways targeting a forcing level of 4.5 Wm-2 (SSP2–4.5), developed in support of the ScenarioMIP experiment within CMIP6 (Gidden et al., 2019). Biomass burning (BMB) emissions are from van Marle et al. (2017) for 1980–2014 and from SSP2–4.5 for 2015–2017, and they are vertically distributed over seven ecosystem-dependent altitude levels between the surface and 6 km above the surface, following the methodology of Dentener et al. (2006). Anthropogenic and biomass burning emissions are represented by monthly gridded emissions including seasonal and interannual variability. Natural emissions include wetland (WET) emissions from the WetCHARTs version 1.0 inventory (Bloom et al., 2017), ocean (OCN) emissions from Brasseur et al. (1998) with nearshore methane fluxes from Lambert and Schmidt (1993) and Patra et al. (2011), termites (TMI) from Fung et al. (1991), and mud volcanoes (VOL) from Etiope and Milkov (2004) and Patra et al. (2011). Wetland emissions and ocean emissions are climatological monthly means without interannual variability. The remaining natural emissions are based on a climatological annual mean (repeated every month without seasonal variability). Time series of the total emissions and emissions from major sectors over 1980–2017 are shown in Fig. 1. Trends in total emissions are primarily driven by trends in ENE, AGR, and WST emissions. Although wetlands are in reality a major contributor to interannual variability in methane emissions (Bousquet et al., 2006; Kirschke et al., 2013), our use of climatological wetland emissions causes the interannual variability in our methane emissions to be dominated by BMB emissions. Anthropogenic and biomass burning emissions of other short-lived species also follow the CEDS and SSP2–4.5 inventories. Natural emissions of other short-lived species are from Naik et al. (2013a). Biogenic isoprene emissions are calculated interactively following Guenther et al. (2006).

The methane sinks considered in AM4.1 include oxidation by OH, Cl, and O(1D) and dry deposition. Since the model does not represent tropospheric halogen chemistry, it does not consider removal of methane by Cl in the troposphere, a sink that remains poorly constrained (Hossaini et al., 2016; Gromov et al., 2018; Wang et al., 2019). The dry deposition flux of methane is estimated based on a monthly climatology of deposition velocities (Horowitz et al., 2003) calculated by a resistance-in-series scheme (Wesely, 1989; Hess et al., 2000) and used to mimic methane loss by soil uptake, which accounts for about 5 % of the total methane sink (Kirschke et al., 2013; Saunois et al., 2016).

In this work, we included 12 additional methane tracers tagged by source sector to attribute methane from agriculture (CH4AGR), energy (CH4ENE), industry (CH4IND), transportation (CH4TRA), residents (CH4RCO), waste (CH4WST), shipping (CH4SHP), biomass burning (CH4BMB), ocean (CH4OCN), wetland (CH4WET), termites (CH4TMI), and mud volcanoes (CH4VOL). The tracers are emitted from corresponding sources and undergo the same chemical and transport pathways as the full CH4 tracer. For analysis, we combine CH4IND, CH4TRA, CH4RCO, and CH4SHP as other anthropogenic tracers (i.e., CH4OAT), and we combine CH4OCN, CH4TMI, and CH4VOL as other natural tracers (i.e., CH4ONA).

Initially the model was spun up in a 50-year run with repeating 1979 emissions driven by 1979 sea surface temperatures and sea ice until stable atmospheric burdens of methane and tagged tracers were obtained. After spin-up, several sets of simulations were conducted for 1980–2017 to quantify the methane budget and investigate the impacts of changes in methane sources and sinks on methane abundance (see Sect. 2.3). All model simulations are forced with interannually varying sea surface temperatures and sea ice from Taylor et al. (2000), prepared in support of the CMIP6 Atmospheric Model Intercomparison Project (AMIP) simulations. Horizontal winds are nudged to the National Centers for Environmental Prediction (NCEP) reanalysis (Kalnay et al., 1996) using a pressure-dependent nudging technique (Lin et al., 2012).

We evaluate the simulated methane dry-air mole fraction (DMF) against a suite of ground-based and aircraft observations to thoroughly evaluate the model-simulated spatial and temporal distribution of methane. To evaluate surface CH4, we use measurements from a globally distributed network of air sampling sites maintained by the Global Monitoring Division (GMD) of the Earth System Research Laboratory at the National Oceanic and Atmospheric Administration (NOAA) (Dlugokencky et al., 2018). The global estimates are based on spatial and temporal smoothing of CH4 measurements from 45 surface marine boundary layer (MBL) sites. Locations of the MBL sites are shown in Fig. S1 in the Supplement, and information for each MBL site is listed in Table S1 in the Supplement. First, the average trend and seasonal cycle are approximated for each sampling site by fitting a second-order polynomial and four harmonics to the data. We characterize deviations from this average behavior by transforming the residuals to frequency domain, then multiplying by a low-pass filter (Thoning et al., 1989; Thoning, 2019). Zonal and global averages are determined by extracting values at synchronized times steps from the smoothed fits to the data, then fitting another curve as a function of latitude (Tans et al., 1989). We divide these fits into sine (latitude) =0.05 intervals, which define a matrix of zonally averaged CH4 as a function of time and latitude. The same sampling and processing approach (Thoning et al., 1989; Tans et al., 1989) is applied to the simulated monthly mean methane DMF to calculate global and zonal averages to facilitate consistent model–observation comparison. In addition to the comparison with global estimates from MBL sites, we also evaluate model performance at individual GMD sites to investigate regional emission representation. For site-specific evaluation, we sample the model grid cell at the location of the corresponding site and at the model layer with height closest to the altitude of the corresponding site.

To investigate background tropospheric methane variability, we compare the simulated vertical profiles with aircraft measurements from the High-performance Instrumented Airborne Platform for Environmental Research (HIAPER) Pole-to-Pole observation (HIPPO) campaigns from January 2009 to September 2011 (Wofsy et al., 2011, 2012). A total of 787 profiles were flown during five campaigns with continuous profiling between approximately 150 and 8500 m altitudes, but also including many profiles up to 14 km altitude. For each HIPPO mission, we spatially sample the model consistent with the observations and average the model for the months of the campaign to create climatological monthly means.

We conduct several sets of hindcast simulations for 1980–2017, as listed in Table 2, to quantify the methane budget and investigate the contributions of sources and sinks to the trend and variability of methane. The model simulation using the initial methane emissions inventory (Einit) described in Sect. 2.1 was found to largely underestimate the methane DMF by 126 ppb (see Figs. S2 and S3 in the Supplement). Assuming that this mismatch is due to a bias in the simulated methane budget, we can either increase methane sources or decrease methane sinks to match the observations. We perform several optimization simulations that explore the sensitivity of methane to uncertainties in emissions of methane and levels of OH, the dominant sink for methane. Because OH trends and variability depend on a number of factors, including temperature, water vapor, O3, and emissions of nitrogen oxide (NOx), carbon monoxide (CO), and volatile organic compounds (VOCs), it is not straightforward to perturb OH. Previous work has shown that interannual variability of global OH is highly correlated with NOx from lightning (Fiore et al., 2006; Murray et al., 2013). Therefore, we apply scaling factors to lightning NOx (LNOx) emissions to indirectly adjust OH levels without influencing its variability. The LNOx emissions are calculated interactively as described by Horowitz et al. (2003) as a function of subgrid convection parameterized in the model. The climatological global mean LNOx emission simulated by standard AM4.1 is about 3.6 TgNyr-1, within the range of 2–8 TgNyr-1 estimated by previous studies (e.g., Schumann and Huntrieser, 2007). We additionally apply scaling factors (e.g., 0.5 and 2.0) to LNOx emissions, producing LNOx at the lower and upper limits of the estimated range for sensitivity simulations described below.

We test the sensitivity of simulated methane to changes in OH using (1) standard OH levels simulated by AM4.1 (referred to as “S0”), (2) low OH levels via application of a scaling factor of 0.5 to the default LNOx emission calculations (referred to as “S1”), and (3) high OH levels via application of a factor of 2 to the default LNOx emission calculation (referred to as “S2”). For each OH option, we begin with initial methane emissions and then optimize global total emissions as described below to match simulated methane with surface observations. Different OH levels lead to different estimations of the optimized total emissions, which provide a measure of uncertainties in our optimized total methane emissions.

The estimates of optimized emissions are based on comparison of simulated surface methane with NOAA GMD MBL observations. We apply a simple mass balance approach to optimize global total methane emissions, following the methodology of Ghosh et al. (2015). In this approach, we calculate an increment ΔE, by which global emissions need to be modified for each year. We do so by converting the differences in surface methane DMFs between observations and model estimates to the differences in methane burden growth rate and in total methane loss. We iterate the optimization process a couple of times to account for the methane–OH feedback until the simulated surface methane DMF matches the observations. Unlike inverse modeling studies (Houweling et al., 2017), we do not optimize emissions for each grid cell. Instead, we uniformly scale emissions for particular sectors (as described below) globally for each year by the rate of the optimized emission total (Eopt=Einit+ΔE) to the initial emissions (Einit). We assume that the spatial distribution of methane emissions from the initial emission inventories is the best available information we have. Considering the large uncertainties in the anthropogenic and wetland emissions, we perform two simulations for the standard (S0) LNOx scenario, in which we achieve the optimized emission totals by scaling either anthropogenic and biomass burning sources only (referred to as “Aopt”) or the wetland sector only (referred to as “Wopt”). The purpose of conducting these simulations is to investigate the impact of optimizing emissions from different sectors on methane predictions. For the Aopt case, eight anthropogenic sectors (i.e., AGR, ENE, IND, TRA, RCO, WST, SHP, and BMB) are uniformly scaled by the ratio of ΔE to total anthropogenic emissions (ΔE/Eanthro.), keeping the fractions of individual sources unchanged. For the Wopt case, wetland emissions are rescaled to increase this source by ΔE. For S1 and S2 scenarios, we scale the wetland sector only. The total Eopt emissions are the same for both Aopt and Wopt cases. Time series of optimized total emissions and emissions from major sectors from S0Aopt and S0Wopt over 1980–2017 are shown in Fig. 1. As shown in Fig. 1, the emission optimization to match observations resulted in higher interannual variability in total emissions than in the initial emissions. Although the interannual variability of methane emissions is mainly dominated by that from wetland and biomass burning, it could also exist in anthropogenic emissions due to the dependence of microbial methane sources, such as rice paddies, on soil temperature and precipitation (e.g., Knox et al., 2016). Because the purpose of S0Aopt is to investigate the role of changes in total anthropogenic emissions (including BMB) rather than individual sectors, we applied this interannual variability to all anthropogenic sectors, which we acknowledge introduces some unrealistic interannual variability in the anthropogenic emissions. We chose this experimental construct to limit the number of sensitivity simulations.

Based on evidence from δ13CH4, recent studies suggest increasing wetland emissions may be responsible for the renewed growth of methane (Dlugokencky et al., 2009; Nisbet et al., 2016). We perform two additional sensitivity simulations to test the possibility of wetland emissions driving the renewed methane growth during 2006–2014. One simulation is driven by repeating 2006 S0Aopt anthropogenic and biomass burning emissions for 2006–2014 but adjusting wetland emissions to ensure that the total methane emissions are the same as in S0Wopt (or S0Aopt), which would imply that the increases in methane emissions are only due to the increases in wetland emissions. This sensitivity simulation is referred to as “S0A06”. Another sensitivity simulation is driven by a combination of emissions for S0Aopt and S0Wopt as follows: S0Aopt emissions for 1980–2005 and S0Wopt emissions for 2006–2014. This simulation is referred to as “S0Comb”.

The detailed model evaluation for S0Aopt and S0Wopt is discussed below. We first evaluate the mean climatological spatial distribution and seasonal variability simulated by the model and then evaluate the trends and variability.

Figure 5 shows time series of optimized total emissions, global sink, and global burden of methane based on S0Wopt. Since global totals in the S0Aopt and S0Wopt simulations are very similar, we only show the budget for S0Wopt. As depicted in Fig. 5, the simulated global methane burden steadily increases from 1980 to 1992, with a growth rate of 39 Tgyr-1. During 1993–1998, the global methane burden growth slows with a growth rate of 16 Tgyr-1. The simulated growth rate in global methane burden decreases to 4 Tgyr-1 during 1999–2006 while it increases to 16 Tgyr-1 during 2007–2017 and reaches over 20 Tgyr-1 during 2014–2016. The changes in the global burdens are due to the imbalance between methane emissions and sinks. As shown in Fig. 5, the optimized emissions in general increase during 1980–2017, with an annual mean of 580±34 Tgyr-1 (mean±standard deviation), and show much larger interannual variability during 1991–1993 and 1997–2000, which is likely due to the strong El Niño events during 1991–1992 and 1997–1998 as well as the Mt. Pinatubo eruption in 1991 (Dlugokencky et al., 1996; Bousquet et al., 2006; Bândă et al., 2016). Although there is an overall increasing trend in total global emissions, growth in annual mean emissions has increased from the 1980s (with an annual emission growth rate of 3.9 Tgyr-1) to the 1990s (4.4 Tgyr-1), but decreased to 0.3 Tgyr-1 during 2000–2006, and increased again to 2.3 Tgyr-1 during 2007–2017. Interannual variability of the optimized emissions mainly results from interannual variability in simulated OH levels during emission optimization. Uncertainties in the interannual variability of simulated OH levels and therefore methane sinks could lead to uncertainties in the interannual variability of the optimized emissions.

Unlike methane emissions, the methane sink increases during 1980–2007, with relative stabilization during 2008–2014 but a resumed increase during 2015–2017. The annual mean methane sink during 1980–2017 is 560±44 Tgyr-1 (mean±standard deviation). The trends in methane sink are affected by the changes in both methane and OH levels (assuming that other sinks are minor) and temperature. Figure 6 shows the tropospheric OH anomalies with respect to 1998–2007. An interesting finding is that AM4.1 predicts higher OH levels during 2008–2014 than 1998–2007 by 3.1 %, whereas previous studies applying multispecies inversion with a box-model framework (e.g., Rigby et al., 2017; Turner et al., 2017) suggest a decline in OH levels after 2007. However, a recent study by Naus et al. (2019) found a shift to a positive OH trend over 1994–2015 after applying bias corrections based on a 3-D chemical transport model to a similar box-model setup. In addition, OH levels simulated by AM4.1 decrease from 2013 to 2015 but increase again afterwards, leading to an increase in methane sinks during 2015–2017. As shown in Fig. 5, larger methane emissions than sinks during 1980–1998 lead to an increase in methane burden. A relative balance between methane sources and sinks during 1999–2006 leads to the methane stabilization. Compared to 1999–2006, both methane sources and sinks are greater during 2007–2017, but methane emissions outweigh sinks after 2007, leading to renewed methane growth.

Table 4 provides a summary of the decadal mean methane budget for 1980–2017. Compared to Kirschke et al. (2013) and Saunois et al. (2020), total natural emissions from the initial emission inventories (203 Tgyr-1) are at the lower range of top–down estimates during this period, except for the 1990s, when they are slightly greater than top–down estimates but still much lower than the bottom–up estimates. Since there is no observational constraint on bottom–up estimates, total natural emissions are simply summed over independent individual sources, which could be overestimated in the bottom–up approach considering the relatively large uncertainties in each individual source. In addition, in the bottom–up estimate from Kirschke et al. (2013) and Saunois et al. (2016), some other natural sources, such as freshwater, are not included in the initial emission inventories in this work; however, they are likely double counted in the bottom–up estimates (e.g., high-latitude inland waters are likely also considered as wetland areas) as pointed out in Saunois et al. (2020). The natural emissions estimated in this work (e.g., 203–297 Tgyr-1) are much more comparable to previous top–down estimates (e.g., 150–273 Tgyr-1 as shown in Kirschke et al., 2013), which demonstrates confidence in the natural source estimates. Total anthropogenic emissions from the initial emission inventories are overall within the range of top–down or bottom–up estimates, except for 1980–1989, when they are less than the estimates in Kirschke et al. (2013). The low values in the 1980s result mainly from low estimated emissions from agriculture and waste sectors in the CEDS inventory. With the optimized global total emissions, the total sources used in this work and the total sinks estimated by AM4.1 are in the range of either top–down or bottom–up estimates by previous studies. As a result, the imbalance between total emissions and total sinks estimated in this work is, overall, within the range of estimates by previous studies, although we find a smaller imbalance than previous estimates for the 2000s and afterwards. The atmospheric growth rates simulated by the model (sampled identically as for observations) are also comparable to the observed atmospheric growth rates.

In this section, we apply the Mann–Kendall (M–K) test to estimate the linear trend of global mean source-tagged tracers and total methane for 1983–1998, 1999–2006, and 2007–2017 to investigate possible drivers of total methane trends. Figure 7 compares the trends of source-tagged tracers and total methane from S0Aopt and S0Wopt during each time period. For S0Aopt, total methane increases strongly at 10.5 ppbyr-1 during 1983–1998. The tagged anthropogenic tracers all show increasing trends during 1983–1998 despite the increases in OH levels, with dominant increasing trends by CH4AGR and CH4WST consistent with emission trends. Since wetland emissions and other natural emissions are kept constant every year in S0Aopt, with increases in OH levels during 1983–1998, all tagged natural tracers show a weak decreasing trend. During 1999–2006, total methane shows a small increasing trend of 1.0 ppbyr-1, due to the increasing trends of CH4ENE and CH4WST compensated by the decreasing trends of other source-tagged tracers. The increasing trends of CH4ENE and CH4WST are mainly driven by the increases in the emissions from energy and waste sectors in S0Aopt, whereas the decreasing trends of other source-tagged tracers are mainly driven by the increases in OH levels. Compared to the rapid growth during 1983–1998, only CH4ENE growth rate shows a small increase during 1999–2006 (2.6 vs. 2.2 ppbyr-1 in 1983–1998), with all other tracers showing a decrease in their growth rates. Despite higher anthropogenic emissions during 1999–2006 than previous periods, the sinks are also higher, leading to a relative stabilization. During 2007–2017, total methane shows a renewed increasing trend of 5.3 ppbyr-1, dominated by a strong increasing trend of CH4ENE (5.9 ppbyr-1) and smaller increasing trends of CH4AGR and CH4WST. Compared to 1999–2006, there is a significant increase in CH4ENE growth rate with smaller increases in CH4AGR growth rate during 2007–2017. Although the CH4WST growth rate decreased in 2007–2017, the continued increasing trend of CH4WST together with those of CH4AGR and CH4ENE contribute to the renewed growth in methane. The results from S0Aopt suggest that globally anthropogenic tracers dominate total methane trends during the entire simulation period. During the 1980s and 1990s, emissions from agriculture, energy, and waste sectors are the major contributors to the methane increase. During 1999–2006, when atmospheric methane stabilizes, increases in methane sinks and methane sources alternatively dominate the trend for different tracers. Compared to 1999–2006, higher emissions from agriculture, energy, and waste sectors during 2007–2017 are the major drivers for the renewed growth in methane, with the energy sector as the largest contributor.

The source-tagged tracers behave slightly differently in S0Wopt. For S0Wopt, total methane shows an increasing trend similar to that of S0Aopt. During 1983–1998, the tagged anthropogenic tracers all show increasing trends except CH4ENE, with overall smaller increasing trends than those in S0Aopt. CH4WET shows a strong increasing trend (7.0 ppbyr-1), dominating the total methane trend. This is mainly because wetland emission growth is larger than anthropogenic emission growth due to the emission optimization in S0Wopt during this period. During 1999–2006, similar to S0Aopt, the total methane trend results from the increasing trends of CH4ENE and CH4WST compensated by the decreasing trends of other source-tagged tracers. During this time, CH4WET shows a slightly decreasing trend (-0.8 ppbyr-1), reflecting the slightly greater CH4WET sinks (226 Tgyr-1) than emissions (223 Tgyr-1). Similar to S0Aopt, only CH4ENE shows an increase in its growth rate during this time period compared to previous time periods. During 2007–2017, the total methane trend is dominated by the increasing trends of CH4AGR, CH4ENE, and CH4WST, with CH4ENE as the largest contributor, similar to S0Aopt. On the other hand, CH4WET shows a significant decreasing trend during this period, with CH4WET sinks (217 Tgyr-1) larger than emissions (206 Tgyr-1). Compared to 1999–2006, there is a significant increase in CH4ENE growth rate with a noticeable increase in CH4AGR growth rate during 2007–2017. Although the CH4WST growth rate also decreased in 2007–2017, similar to S0Aopt, the continued increasing trend of CH4WST together with those of CH4AGR and CH4ENE contribute to the renewed growth in methane. On the other hand, CH4WET shows a significant decrease in its growth rate during this time period compared to 1999–2006, mainly due to lower emissions (206 Tgyr-1 in 2007–2017 vs. 223 Tgyr-1 in 1999–2006) imposed in this scenario. Compared to the S0Aopt results, S0Wopt suggests CH4WET as the largest contributor for the methane trends during the 1980s and 1990s, mainly due to emission optimization of different sectors. However, both scenarios suggest CH4AGR, CH4WST, and CH4ENE are the major contributors to the renewed growth in methane, with CH4ENE as the largest contributor.

As shown in Figs. 5 and 6, OH levels slightly decrease and methane sinks are relatively stable during 2007–2013, but large interannual variability exists during 2013–2017. Decreasing OH levels could lead to increases in methane lifetime and therefore methane buildup. Combined with increases in the emissions, methane starts to increase again during this period. However, it is difficult to separate the contributions from methane emissions and sinks as optimized methane emissions are based on methane mass balance (e.g., changes in methane loss would act as a feedback on estimates of optimized total emissions). Nevertheless, it is clear that the decrease in OH levels alone (e.g., if emissions are kept constant) would not be enough to reproduce the renewed growth. The remaining question is then as follows: which emission sectors are the major contributors to the renewed growth from 2007 to 2017? Both S0Wopt and S0Aopt suggest that the agriculture, waste, and energy sectors are the major contributors to renewed methane growth. However, both cases depend largely on the initial emission inventory and the scaling methods chosen. For example, S0Wopt relies on the emission growth of other sectors from the initial emission inventory, which means if the emission growth of a certain sector is overestimated or underestimated in the initial emission inventory, it would lead to different results. Therefore, we conducted two additional sensitivity simulations (i.e., S0A06 and S0Comb as described in Sect. 2.3) with different emission growths for anthropogenic and wetland sectors as in S0Aopt and S0Wopt for 2006–2014.

The trends for source-tagged tracers and total methane by S0A06 and S0Comb are shown in Fig. 7. Interestingly, in S0A06, where anthropogenic and biomass burning emissions are kept constant every year for 2006–2014, anthropogenic tracers, particularly CH4ENE and CH4WST, still show increasing trends during 2007–2014, whereas CH4WET shows a small decreasing trend despite rising emissions. As OH levels slightly decrease during this time (but are still higher than 1999–2006), with constant emissions except for wetland, one might expect possible increasing trends in all tagged tracers except CH4WET. In fact, major anthropogenic tracers such as CH4AGR, CH4ENE, CH4WST, and CH4BMB increase over 2007–2014 in S0A06, but at a slower rate than in S0Wopt (and S0Aopt) due to no emission growth for these tracers. On the other hand, the decreasing OH levels (Fig. 6) would lead to a smaller methane sink and therefore higher methane concentrations. Since methane loss is proportional to the product of OH levels and methane concentrations, and concentrations of CH4WET are much greater than other source-tagged tracers, the loss of CH4WET is also much greater than other tracers. Higher CH4WET loss (223 Tgyr-1) than CH4WET emissions (222 Tgyr-1) leads to a slightly decreasing trend in CH4WET. In other words, despite the increasing source contributions from wetlands to total methane emissions, the relative contributions of wetland tracer to total methane abundance are declining. Compared to 1999–2006, there are major increases in the growth rates of CH4ENE and CH4BMB, with a smaller increase in CH4AGR and CH4OAT growth rates, which drives the renewed methane growth. Meanwhile, CH4WET is still declining during 2007–2014 (-1.1 ppbyr-1), but at a larger decrease rate than for 1999–2006 (-0.8 ppbyr-1). Nevertheless, S0A06 results suggest that the renewed growth during 2007–2014 is contributed by the increased growth rates of CH4ENE, CH4BMB, and CH4AGR as well as increasing trend of CH4WST, mainly due to higher anthropogenic emissions than 1999–2006 and decreases in OH levels during 2008–2014. The results also suggest OH trends play an important role in determining the increasing trend of total methane since emissions of the energy and waste sectors are kept constant in this sensitivity simulation. In addition, increases in wetland emissions alone are not able to drive increases in CH4WET over this period, as CH4WET sinks are equally important for determining the trend in CH4WET. Our analysis also suggests that increased emissions from other microbial sources (e.g., agriculture and waste) are needed to match the observed negative trend in δ13CH4 since 2007 (Nisbet et al., 2019).

The trends for source-tagged tracers and total methane behave differently in S0Comb, where we combine S0Aopt emissions for 1980–2005 and S0Wopt emissions for 2006–2014. During 2007–2014, all anthropogenic tracers show decreasing trends except CH4ENE (2.8 ppbyr-1), whereas CH4WET shows a significant increasing trend (5.9 ppbyr-1) and dominates the total methane trend. This is mainly due to lower anthropogenic emissions during this period than previous periods, allowing sinks of anthropogenic methane tracers to start to take over their trends except for CH4ENE. At the same time, significantly higher wetland emissions during this period than previous periods dominate the increasing trend of CH4WET. Interestingly, even with the same wetland emissions in S0Wopt and S0Comb for 2006–2014, CH4WET shows different trends. This is mainly because the CH4WET concentrations at the beginning of 2006 are much lower in S0Comb than in S0Wopt. Therefore, CH4WET loss is much lower in S0Comb (190 Tgyr-1) compared to S0Wopt (220 Tgyr-1) over this time, leading to an increasing CH4WET trend in S0Comb, but a decreasing trend in S0Wopt. Compared to 1999–2006, there is a significant increase in CH4WET growth rate with a minor increase in CH4ENE growth rates during 2007–2014, which drives the renewed growth in methane. S0Comb results suggest the need for a sharp increase in wetland emissions with a concomitant sharp decrease in anthropogenic emissions in 2006 to drive the methane growth by wetland tracer. It is a likely scenario for a sharp increase in wetland emissions considering the interannual variability. However, it is a less likely scenario for a concomitant sharp decrease in anthropogenic emissions as both top–down and bottom–up inventories indicate anthropogenic emissions increasing over the 2000s. A more likely scenario is that both anthropogenic and wetland emissions increase (i.e., higher during 2007–2014 than 1999–2006). However, in that case, the dominance of wetland emissions in driving the total methane trend would decrease based on our analysis.

As described in Sect. 2.3, we perform two additional simulations for low and high OH levels (i.e., S1 and S2) for 1980–2017 to investigate the sensitivity of methane predictions to different OH levels. For both OH cases, the interannual variations in OH levels are the same as in S0 because the simulations are driven by the same meteorology. Figure 8a and b show global tropospheric OH concentrations, OH-driven methane loss, and tropospheric methane lifetime for the three cases (i.e., S0, S1, and S2) in which wetland emissions are optimized (Wopt; Aopt has a very similar global OH trend to Wopt). Compared to S0, scaling LNOx production in the model by a factor of 0.5 leads to a reduction in simulated annual global mean OH levels by -6.4 % and an increase in methane lifetime by 0.5 years in S1 over 1980–2017; scaling by a factor of 2 leads to an increase in simulated annual global mean OH by +9.1 % and a decrease in methane lifetime by 0.7 years in S2. The global mean OH levels increase from 1980 to 2008 (by 3.6 %, with respect to the 1980 level), decrease from 2008 to 2015 (by 2.3 %, with respect to the 2008 level), and increase from 2015 to 2017 (by 4.6 %, with respect to the 2015 level). However, compared to the 1998–2007 average, OH levels during 2008–2015 and 2015–2017 are still greater by 2.5 % and 1.3 %, respectively. Changes in OH levels depend on a number of factors (e.g., temperature, water vapor, O3, NOx, CO, and VOCs). Therefore, OH is influenced by the specific chemistry and forcing data used in the model. Nevertheless, our estimates in OH trends and variabilities from all three cases are quite comparable to the those estimated by the Chemistry Climate Model Initiative (CCMI) models (e.g., Zhao et al., 2019). Since emission optimization is also based on methane sinks, the total optimized emissions in S1 are lower than those in S0 by about 4.1 % (with an annual mean of -24 Tgyr-1), and the total optimized emissions in S2 are higher than those in S0 by about 5.8 % (or 33 Tgyr-1). This indicates that a 1 % change in OH levels could lead to about 4 Tgyr-1 difference in the optimized emissions. Increasing methane loss due to OH is simulated for 1980–2007 in the three cases due to increases in OH and methane concentrations (except over the stabilization period when methane was not increasing but OH was increasing). During 2007–2013, the simulated decrease in OH levels combined with increasing methane concentrations leads to relative stabilization in OH-driven methane loss in the three cases. The large interannual variability in OH levels during 2013–2017 dominates the interannual variability in methane OH loss despite the continued increases in methane.

All three simulations show a similar trend for tropospheric methane lifetime, with a decreasing trend from 1980 to 2007 (-0.04 yryr-1 in S0, -0.05 yryr-1 in S1, and -0.03 yryr-1 in S2), a clear increasing trend during 2011–2015 (0.08 yryr-1 in all three simulations), and a decreasing trend during 2015–2017 (-0.2 yryr-1 in all three simulations). The mean tropospheric methane lifetime due to OH loss for 1980–2017 is 9.9±0.4 years in S0Wopt, which is about 0.5 years lower than S1Wopt (10.4±0.5 years) and about 0.7 years higher than S2Wopt (9.2±0.3 years), due to different OH levels and therefore methane sinks, but with similar methane burdens. This indicates that a 1 % change in OH levels could lead to about a 0.08-year difference in the tropospheric methane lifetime. The mean tropospheric methane lifetime simulated by the three simulations is within the uncertainty range of model estimates of 9.3±0.9–9.8±1.6 years (Voulgarakis et al., 2013; Naik et al., 2013b) and in general comparable to the observation-derived estimates of 9.1±0.9 years for the present day (Prather et al., 2012), with a slightly higher estimate by S1Wopt. All simulations show an increase in methane lifetime during 2011–2015, which could be a signal of the methane feedback on its lifetime (Holmes, 2018) in the model. Continued increases in methane emissions (Fig. 5) during this time, along with decreases in tropospheric OH concentrations (Fig. 8), lengthen the lifetime of methane and therefore amplify methane's response to emission changes. If methane emissions continue to increase, we can expect stronger increases in atmospheric methane due to the amplifying effect of the methane–OH feedback as occurred in the significant increases in methane growth rates during 2014 and 2015.