Top-down constraints on atmospheric mercury emissions and implications for global biogeochemical cycling

. We perform global-scale inverse modeling to constrain present-day atmospheric mercury emissions and relevant physiochemical parameters in the GEOS-Chem chemical transport model. We use Bayesian inversion meth-ods combining simulations with GEOS-Chem and ground-based Hg 0 observations from regional monitoring networks and individual sites in recent years. Using optimized emissions/parameters, GEOS-Chem better reproduces these ground-based observations and also matches regional over-water Hg 0 and wet deposition measurements. The optimized global mercury emission to the atmosphere is ∼ 5.8 Gg yr − 1 . The ocean accounts for 3.2 Gg yr − 1 (55 % of the total), and the terrestrial ecosystem is neither a net source nor a net sink of Hg 0 . The optimized Asian anthropogenic emission of Hg 0 (gas elemental mercury) is 650–1770 Mg yr − 1 , higher than its bottom-up estimates (550–800 Mg yr − 1 ). The ocean parameter inversions suggest that dark oxidation of aqueous elemental mercury is faster, and less mercury is removed from the mixed layer through particle sinking, when compared with current simulations. Parameter changes affect the simulated global ocean mercury budget, particularly mass exchange between the mixed layer and subsurface waters. Based on our inversion results, we re-evaluate the long-term global biogeochemical cycle of mercury, and show that legacy mercury becomes more likely to reside in the terrestrial ecosystem than in the ocean. We estimate that primary anthropogenic mercury contributes up to 23 % of present-day atmospheric deposition.


Introduction
Mercury (Hg) is a ubiquitous trace metal that cycles between the atmosphere, ocean, land, and biosphere (Selin, 2009).Atmospheric mercury transports globally (Driscoll et al., 2013) and, in aquatic systems, can be converted to methylmercury, a bioaccumulative toxic compound (Mergler et al., 2007).Human activities have strongly affected the mercury global cycle by both unintentional and intentional releases (Streets et al., 2011).Since mercury deposited to terrestrial and ocean surfaces can remobilize, the atmosphere continues to be affected by its historical releases (Lindberg et al., 2007;Amos et al., 2013).Atmosphere-surface fluxes of mercury are still poorly constrained, limiting our ability to fully understand timescales of its global biogeochemical cycle (Pirrone et al., 2010;Mason et al., 2012).A better knowledge of these fluxes is important for assessing its impacts on humans and evaluating the effectiveness of policy actions (Selin, 2014).
Current estimates of mercury fluxes to the atmosphere are mainly built on a bottom-up approach.Anthropogenic inventories are based on emission factors, activity levels, and abatement efficiency (Pacyna et al., 2010;S. Wang et al., 2014;Muntean et al., 2014).Flux estimates from ocean and terrestrial surfaces extrapolate limited direct measurements to larger scales and use simplified process models (Mason, 2009;Kuss et al., 2011).The top-down or inverse approach, combining observations and atmospheric modeling, has been widely used to derive sources and sinks of greenhouse gases and ozone-depleting substances (Gurney et al., 2002;Xiao et al., 2010).Inverse studies have addressed mercury at a regional scale (Roustan and Bocquet, 2006;Krüger et al., 1999).For example, a hybrid inversion combining back trajectories and a regional chemical transport model (CTM) identified Hg 0 emission using year-long urban observations (de Foy et al., 2012).This scheme was expanded to estimate sources of oxidized Hg (de Foy et al., 2014).
In this paper, we apply a top-down approach at global scale to quantitatively estimate present-day mercury emission sources (emission inversion) as well as key parameters in a CTM (parameter inversion), in order to better constrain the global biogeochemical cycle of mercury.Section 2 describes the overall methodology.We combine groundbased observations of atmospheric Hg 0 (Sect.2.1) and simulations with the GEOS-Chem global CTM (Sect.2.2).Reference (also known as a priori) emissions are from GEOS-Chem parameterizations and agree well with bottom-up estimates (Sect.2.3).We adopt a Bayesian inversion method (Sect.2.4) to obtain the optimized (a posteriori) emissions, with a monthly time step, taking into account uncertainties associated with both reference emissions and groundbased observations (Sect.2.6).Section 3 presents results and discussion.Comparisons of observations and model outputs are given in Sect.3.1.The optimized emissions from ocean and terrestrial surfaces and from anthropogenic sources are shown in Sect.3.2.We use results of the emission inversion to identify key uncertain model parameters, and optimize them in the parameter inversion (Sects.2.5 and 3.3).Finally, we discuss implications of our inversion results for the global biogeochemical mercury cycle (Sect.3.4) and summarize our conclusions (Sect.4).

Atmospheric mercury observations
Tropospheric mercury exists mainly as gaseous elemental mercury (GEM) but also as two operationally defined species, gaseous oxidized mercury (GOM) and particlebound mercury (PBM) (Valente et al., 2007).Manual methods of measuring GEM or total gaseous mercury (TGM = GEM + GOM) were applied in the 1970s (Slemr et al., 1981).High-frequency measurements (time resolution < 1 h, e.g., using Tekran automated ambient air analyzers) became available in the 1990s and have substantially replaced manual sampling (time resolution of about several hours).We only use GEM and TGM observations in this study because we are not able to quantify the uncertainty in GOM and PBM measurements (Jaffe et al., 2014;McClure et al., 2014).We identify high-frequency observations of GEM and TGM concentration for our inversions using two criteria.First, we choose sites in rural/remote areas not strongly affected by local emissions.Second, we require that observations at different sites are minimally correlated (Brunner et al., 2012).Data sets are drawn from the Atmospheric Mercury Network (AMNet) (Gay et al., 2013), the Canadian Measurement Networks (including the Canadian Air and Precipitation Monitoring Network (CAPMoN) and other sites sponsored by Environment Canada) (Cole et al., 2014), and the European Monitoring and Evaluation Programme (EMEP) (Tørseth et al., 2012).We use data from 2009 to 2011, when all these networks were active.To expand spatial coverage of observations, we also collected data from individual sites for recent years (2007)(2008)(2009)(2010)(2011)(2012)(2013).Some sites are included in the Global Mercury Observation System (GMOS) (Pirrone et al., 2013).All sites use Tekran analyzers, operated in sampling intervals of 5-30 min.We calculate Pearson's correlation coefficients between each pair of sites using hourly data.Several sites are excluded due to strong correlations within each other, as shown in Table S1 in the Supplement.Table 1 shows the names, locations, and affiliated networks of the 27 ground-based sites used in our inversion.Site locations are also plotted in Fig. 1.For most of these sites GEM data are used, and for a few sites where GEM data are not available we use TGM data (see Table 1).The concentration difference between measured GEM and TGM concentrations in remote near-surface air is usually < 1 % (Lan et al., 2012;Fu et al., 2012a;Weigelt et al., 2013;Steffen et al., 2014) and thus we do not distinguish between measured GEM and TGM concentrations and use Hg 0 to represent them in the paper.These sites are all uncorrelated or only weakly correlated (−0.3 < r < 0.4, n = 10 3 -10 4 ) (see Table S2 in the Supplement).

Atmos
Original observational data are converted into hourly averages and then into monthly averages (Fig. S1 in the Supplement).We require > 30 min data to derive an hourly average and > 10-day data to derive a monthly average.Where full data are available, median values are used to suppress the influence of high Hg 0 due to local or regional pollution events (Weigelt et al., 2013;Jaffe et al., 2005) or occasional low Hg 0 due to non-polar depletion events (Brunke et al., 2010).For a few individual sites (see Table 1), the original data are not available and monthly arithmetic means are used.Finally, multiple-year averages are calculated.Hg 0 concentrations are given in nanograms per cubic meter at standard temperature and pressure.
Four polar sites are included (ALT, ZEP, and ADY in the Arctic and TRS in Antarctica, see Table 1).Episodically low Hg 0 is observed at these sites in polar spring (Cole et al., 2013;Pfaffhuber et al., 2012).These atmospheric mercury depletion events (AMDEs) result from rapid Hg 0 oxidation and deposition driven by halogens (Steffen et al., 2008).Volatilization of the deposited Hg and the large quantities of imported mercury from circumpolar rivers to the Arctic Ocean are hypothesized to contribute to the observed summer Hg 0 peak in the Arctic region (Dastoor and Durnford, 2013;Fisher et al., 2012).The lack of understanding of the above physical and chemical processes limits GEOS-Chem's ability to reproduce Hg 0 in the polar spring and summer.For these reasons we remove Hg 0 data at polar sites for this period (i.e., March-September in the Arctic and October-March in Antarctica).
We also include three mountaintop sites (LUL, MBO, and MLO, see Table 1).These sites are affected by upslope surface air during the day and downslope air from the free troposphere at night (Sheu et al., 2010;Fu et al., 2010).The downslope air usually contains higher levels of GOM than the upslope air due to oxidation of Hg 0 to GOM in the free troposphere (Timonen et al., 2013).Therefore, Hg 0 at mountaintop sites peaks in the afternoon whereas GOM peaks between midnight and early morning (Fig. S2 in the Supplement), showing an opposite diurnal pattern to most lowelevation sites (Lan et al., 2012).The minimum hourly Hg 0 at night is calculated to be ∼ 90 % of the all-day average.Thus, to represent Hg 0 modeled at a vertical layer in the free troposphere (this layer is obtained by matching observed air pressure), the observed mountaintop Hg 0 data are multiplied by 0.9.
We do not use over-water Hg 0 observations (i.e., from ship cruises) in the inversion because they are very limited and usually cover large areas, making their observational errors difficult to estimate.Instead, we use over-water observations as an independent check of our inversion results.The North Atlantic Ocean is the most densely sampled ocean basin.Soerensen et al. (2012) assembled Hg 0 measurements from 18 ship cruises in this region during 1990-2009 and found a statistically significant decrease of −0.046 ± 0.010 ng m −3 yr −1 .However, previous GEOS-Chem simulations of Hg 0 concentration did not take this multidecadal trend into account in evaluating its seasonal variability (Soerensen et al., 2010a 2006) for MBO, Fu et al. (2012b) for CBS, Fu et al. (2012a) for WLG, Zhang et al. (2015) for SGR, MOEJ (2013) for OKN, Sheu et al. (2010) for LUL, Müller et al. (2012) for NWN, Slemr et al. (2011) for CPT, Angot et al. (2014) for AMS, and Slemr et al. (2015) for the Southern Hemispheric sites.e Unit for errors is picograms per cubic meter.f Equation of NRMSE (quantity without unit) is given in Sect.3.1.all 19 ship cruises to Hg 0 levels consistent with year 2009 based on a fitted decline trend (Table S3 and Fig. S3 in the Supplement).Seasonal variation is estimated by dividing the normalized Hg 0 (Hg 0 nor ) by month of measurement.As shown in Fig. 2, Hg 0 nor are smaller and show less seasonal variability compared to Hg 0 obs .

GEOS-Chem model
GEOS-Chem (v9-02) is a CTM driven by assimilated meteorological fields from the NASA Goddard Earth Observing System (Bey et al., 2001).The original GEOS-5 has a resolution of 1/2 • × 2/3 • and is degraded to 2 • × 2.5 • for input into our simulations.The GEOS-Chem global mercury simulation was described and evaluated in Selin et al. (2007) and Strode et al. (2007), with updates by Selin et al. (2008), Holmes et al. (2010), Soerensen et al. (2010b), and Amos et al. (2012).It couples a three-dimensional atmosphere, a twodimensional mixed layer slab ocean, and a two-dimensional terrestrial reservoir.For consistency with most ground-based observations, we use meteorological years 2009-2011 for analysis after a spin-up period of 4 years.
Three mercury tracers (representing GEM, GOM, and PBM) are simulated in the atmosphere in GEOS-Chem.Models have assumed that Hg 0 is oxidized by OH, ozone, and/or halogens (Lei et al., 2013;De Simone et al., 2014;Travnikov and Ilyin, 2009;Durnford et al., 2010;Grant et al., 2014).Some studies suggested the gas-phase reaction with Br was the most important Hg 0 oxidation process globally (Seigneur and Lohman, 2008;Hynes et al., 2009), and here we use Br as the only oxidant of Hg 0 (Holmes et al., 2010;Goodsite et al., 2012).Tropospheric Br fields are archived from a full chemistry GEOS-Chem simulation (Parrella et al., 2012).Models also hypothesize gas-and/or aqueousphase reductions of oxidized Hg and scale their kinetics to match atmospheric observations (Holmes et al., 2010;Pongprueksa et al., 2011;Selin et al., 2007).However, an accurate determination of potential pathways is lacking (Subir et al., 2011(Subir et al., , 2012)), and their atmospheric relevance is unknown (Gårdfeldt and Jonsson, 2003).Thus, we do not include atmospheric reduction of oxidized Hg in our simulations.

Emission inversion: reference emissions
For our reference emissions, we use parameterizations in GEOS-Chem with improvements from recent literature.As shown in Table 2, the global mercury emission is estimated as 6.0 Gg yr −1 , with an uncertainty range of 0.4-12.2Gg yr −1 .Mercury released via natural processes is assumed to be entirely Hg 0 (Stein et al., 1996), while a small fraction of anthropogenic mercury is in oxidized forms.Anthropogenic emission is unidirectional, but air-surface exchange is bidirectional (emission and deposition) (Xu et al., 1999;Gustin et al., 2008).A positive net emission from a surface means it is a net source of Hg 0 , whereas a negative value means it is a net sink.We describe below our reference emissions for individual sources.

Anthropogenic sources
We use the anthropogenic emission inventory based on activity data for year 2010, developed by AMAP/UNEP (2013).As shown in Table 2, the total anthropogenic emission is 1960 Mg yr −1 , with an uncertainty range of 1010-4070 Mg yr −1 (AMAP/UNEP, 2013).We do not optimize oxidized mercury emissions (accounting for 19 % of the total anthropogenic sources) because this form has a short atmospheric lifetime (days to weeks) and may not significantly contribute to observed TGM concentrations.The geospatial distribution for emissions from contaminated sites (Kocman et al., 2013) is not available for this inventory, and we distribute this small source (80 Mg yr −1 ) based on the locations of mercury mines (Selin et al., 2007).We do not consider in-plume reduction of oxidized Hg emitted from coal-fired power plants (Y.Zhang, et al., 2012).About 50 % of global emissions are from Asia (defined as 65-146 • E, 9 • S-60 • N), and a small fraction are from Europe and North America (together < 10 %).For other regions like Africa and South America, there is no effective observational site to constrain emissions (Fig. 1).Thus, only anthropogenic emissions from Asia are optimized in the inversion, but we still include other regions' anthropogenic emissions in the GEOS-Chem simulations.
www.atmos-chem-phys.net/15/7103/2015/Atmos.Chem.Phys., 15, 7103-7125, 2015 a Flux values in parentheses indicate estimated uncertainty ranges.For sources included in the inversion, "average ± SD" is shown.The uncertainty ranges of contaminated sites and geogenic emissions are from AMAP/UNEP (2013) and Mason (2009), respectively.If the uncertainty range of a source is not available, we assume that its SD is a half of its best estimate.b Only selected mercury emission sources are included in the inversion, see Sect.2.3.4.c Oxidized Hg emissions from anthropogenic sources are not included in the inversion."Asia" and "Other regions" (except Asia) refer to emissions of Hg 0 .d Because air-terrestrial interactions are bi-directional, we assume that uncertainties of prompt re-emission and Hg 0 deposition have been covered by that of soil emission.e Total mercury emissions are the sum of anthropogenic, net ocean, and net terrestrial emissions.

Ocean
The mixed layer (ML) slab ocean model in GEOS-Chem is described in Soerensen et al. (2010b).Net Hg 0 emission from ocean surfaces is determined by the supersaturation of Hg 0 aq in the ML relative to the atmosphere and the air-sea exchange rate.Hg 0 aq in the ML is mainly produced by the net photolytic and biotic reduction of Hg 2+ aq .Atmospheric deposition accounts for most Hg 2+ aq inputs into the ML, but subsurface waters also contribute a considerable fraction.The ML interacts with subsurface waters through entrainment/detrainment of the ML and wind-driven Ekman pumping.
We improve several parameterizations in GEOS-Chem based on recent findings.(1) Basin-specific subsurface water mercury concentrations are updated according to new measurements (Lamborg et al., 2012;Munson, 2014), as shown in the Supplement, Fig. S4.(2) Soerensen et al. (2010b) used the Wilke-Chang method for estimating the Hg 0 aq diffusion coefficient (D Hg ) (Wilke and Chang, 1955), but this estimate was believed to be too high (Loux, 2004).We adopt a revised D Hg derived by molecular dynamics (MD) simulation (Kuss et al., 2009).As shown in the Supplement, Fig. S5, compared to the Wilke-Chang method, the MD simulation obtains a D Hg that agrees much better with laboratory results (Kuss, 2014).(3) Particulate mercury (Hg P aq ) sinking from the ML is estimated by linking the organic carbon export (biological pump) and Hg P aq : C ratios.Soerensen et al. (2010b) used the model of Antia et al. (2001) for estimating carbon export fluxes, giving a global total of 23 Gt C yr −1 .However, this estimate is mainly based on the flux measurement data from much deeper depths and may not well represent carbon export from the ML.Different models suggest global carbon export fluxes ranging from 5 to 20 Gt C yr −1 with a best estimate of 11 Gt C yr −1 (Sanders et al., 2014;Henson et al., 2011).Thus, we multiply carbon export fluxes in GEOS-Chem by a factor of 0.47 (11 Gt C yr −1 /23 Gt C yr −1 ) to match this best estimate.
Net global ocean emission of 2990 Mg yr −1 from the improved GEOS-Chem (considered as reference emission, shown in Table 2) compares favorably with best estimates of 2680 Mg yr −1 using a bottom-up approach (Pirrone et al., 2010;Mason, 2009).Due to their different seasonal characteristics, we divide the global ocean into the NH (Northern Hemisphere) and SH (Southern Hemisphere) oceans and optimize their emissions separately.

Terrestrial ecosystem
Although atmosphere-terrestrial Hg 0 exchange is bidirectional, only recently developed exchange models have coupled deposition (downward) and emission (upward) fluxes and dynamically estimated net fluxes by gradients be- 15, 7103-7125, 2015 www.atmos-chem-phys.net/15/7103/2015/tween air Hg 0 and "compensation points" inferred from surface characteristics (Bash, 2010;Bash et al., 2007).Because their complex parameterizations lack field data for verification (X.Wang et al., 2014), such exchange models have not been incorporated into current global CTMs.As described in Selin et al. (2008) and Holmes et al. (2010), GEOS-Chem treats emission and deposition fluxes of Hg 0 separately.Only dry deposition is considered for Hg 0 due to its low Henry's law constant (Lin and Pehkonen, 1999).Net emission from terrestrial surfaces (E net ) represents the sum of these processes: volatilization from soil (E soil ), prompt re-emission of deposited Hg (E pr ), geogenic activity (E gg ), biomass burning (E bb ), and dry deposition to surfaces (E ddHg 0 ).
Soil emission (E soil ) is specified as a function of solar radiation and soil Hg concentration: where C soil is soil Hg concentration (ng g −1 ) and R g is the solar radiation flux at the ground (W m −2 ).GEOS-Chem assumes a global average soil concentration of 43 ng g −1 for preindustrial conditions and derives its spatial distribution from the local equilibrium between emission and deposition.The scaling factor β (1.2 × 10 −2 g m −2 h −1 ) is obtained from the global mass balance of the preindustrial simulation.Selin et al. (2008) assumed that present-day soil mercury reservoir and emission have both increased by 15 % compared to the preindustrial period and distributed this global average increase according to the present-day deposition pattern of anthropogenic emission.However, by linking soil mercury with organic carbon pools, Smith-Downey et al. ( 2010) estimated that present-day Hg storage in organic soils has increased by 20 % while soil emission by 190 %. Mason and Sheu (2002) suggested doubled soil emissions compared to preindustrial times.Thus, following Smith-Downey et al. ( 2010), we assume a 190 % global increase in the present day, and distribute this increase according to the anthropogenic emission deposition pattern.The present-day reference soil emission is calculated to be 1680 Mg yr −1 .An additional 520 Mg yr −1 is emitted from the soil, vegetation, and snow (E pr ) through rapid photoreduction of recently deposited oxidized Hg (Fisher et al., 2012).Geogenic emission (E gg ) is set as 90 Mg yr −1 , consistent with its best bottom-up estimate (Mason, 2009;Bagnato et al., 2014).Biomass burning (E bb ) of 210 Mg yr −1 is estimated using the Global Fire Emissions Database version 3 of CO (van der Werf et al., 2010) and a Hg : CO ratio of 100 nmol mol −1 (Holmes et al., 2010).This amount falls at the lower end of bottom-up estimates (Friedli et al., 2009).Dry deposition of Hg 0 is estimated using a resistance-in-series scheme (Wesely, 1989) and has a downward flux of 1430 Mg yr −1 .Using Eq. ( 1), net emission of Hg 0 from terrestrial surfaces is calculated to be 1070 Mg yr −1 in GEOS-Chem (Table 2), at the lower end of the bottom-up estimates (1140-5280 Mg yr −1 ) (Mason, 2009;Pirrone et al., 2010) and also lower than 1910 Mg yr −1 by Kikuchi et al. ( 2013) using a different empirical mechanism (Lin et al., 2010).

Sources included in emission inversion
Because of limitations in both observations and the CTM, only anthropogenic emission from Asia, ocean evasion (separated into the NH and SH), and soil emission are optimized in the emission inversion (see Table 2).The remaining sources are still included in the simulation but not inverted because they are too diffusely distributed, their magnitude is small, and/or observations are not sensitive to them (Chen and Prinn, 2006).The seasonal sources (the NH ocean, SH ocean, and soil) usually have strong spatiotemporal variations and the inversion optimizes their monthly magnitudes and uncertainties.For the aseasonal Asian anthropogenic emission, the inversion optimizes its annual magnitude and uncertainty.

Bayesian inversion method
We use a Bayesian method to invert emissions and parameters with a weighted least-squares technique (Ulrych et al., 2001).The unknowns (correction factors for reference emissions and parameters) are contained in a state vector x and their a priori errors (uncertainties in reference emissions and parameters) in a matrix P. In the emission inversion, as we include one aseasonal source (Asian anthropogenic emission) and three monthly sources (the NH ocean, SH ocean, and soil), the vector x contains 37 elements.P is a 37 × 37 diagonal matrix with each diagonal element equal to the square of 1σ a priori error of the corresponding element in x (see Sect. 2.6.1).
Our inversion method assumes a linear relationship between the observation vector y obs and x, as shown in the measurement equation: where y ref contains monthly Hg 0 concentrations modeled by GEOS-Chem using the reference emissions and parameters.The vectors y obs and y ref both have 12 (number of months per year) × 27 (number of observational sites) = 324 elements.ε represents the model and observational errors which will be discussed in detail in Sect.2.6.The state vector x is related to monthly Hg 0 concentrations by the sensitivity matrix H, in which the elements are written as where i and j are indices for the observational and state vectors, respectively.H describes how monthly Hg 0 concentrations at different observational sites respond to changes in the  S6).
The GEOS-Chem CTM acts as a mathematical operator relating the emissions/parameters to monthly Hg 0 concentrations.For the emission inversion, sensitivities for the seasonal and aseasonal sources are generated by two different types of simulations.The aseasonal Asian anthropogenic emission is perturbed above the reference level by 50 %, and we run the GEOS-Chem CTM until steady state is reached.
For the seasonal sources (e.g., the NH ocean emission from March), a 1-month pulse of Hg 0 is emitted, and we track modeled Hg 0 concentrations by GEOS-Chem for the next 3 years.After this, we assume that the perturbed concentrations at all observational sites will exponentially decrease (Saikawa et al., 2012).
The objective function J with respect to x is where R, a diagonal 324 × 324 matrix, represents errors related to observations and the CTM and will be described in detail in Sect.2.6.By minimizing J, we obtain the expression for the optimal estimate of the state x: where the matrix Q contains the a posteriori errors of x.The size of Q is the same as the matrix P. Each diagonal element in Q is the square of 1σ a posteriori error of the corresponding element in x.A detailed mathematical derivation of the above equations can be found in Wunsch (2006).As shown in Eqs. ( 6) and ( 7), several vectors and matrices need to be calculated during the optimization procedure, including the observational vector y obs and its error matrix R, the error matrix P of the a priori state, the sensitivity matrix H, and the vector y ref which is obtained from the reference simulation of the GEOS-Chem CTM.

Parameter inversion
As described in Sect.3.2.1, based on results of ocean evasion in our emission inversion and sensitivity tests of model parameters, we identify two ocean parameters in GEOS-Chem for improvement: the rate constant of dark oxidation of Hg 0 aq (denoted as K OX2 , following notations in Soerensen et al., 2010b) and the partition coefficient between Hg 2+ aq and Hg P aq (denoted as K D ).For simplicity they are expressed in decimal logarithms (−log K OX2 and log K D ).
A −log K OX2 (s −1 ) of 7.0 is specified in GEOS-Chem (Soerensen et al., 2010b).From a survey of laboratory studies (see details in the Supplement) (Amyot et al., 1997;Lalonde et al., 2001Lalonde et al., , 2004;;Qureshi et al., 2010), we suggest that this value is too low and that a more appropriate range of −log K OX2 is 4.0-6.0.The chemical mechanisms for dark oxidation of Hg 0 aq remain unclear.OH generated from photochemically produced H 2 O 2 via the Fenton reaction may oxidize Hg 0 aq in dark conditions (Zhang and Lindberg, 2001;Zepp et al., 1992).Light irradiation before a dark period is needed, and dark oxidation kinetics depend on intensity and duration of light (Qureshi et al., 2010;Batrakova et al., 2014).Future work could include a more mechanistic representation of this process as laboratory studies become available.
K D (= C s /C d C SPM ) describes the affinity of aqueous Hg 2+ for suspended particulate matter (SPM), where C s , C d , and C SPM are the concentrations of Hg P aq , Hg 2+ aq , and SPM, respectively.GEOS-Chem uses a log K D (L kg −1 ) of 5.5 based on measurements in the North Pacific and North Atlantic oceans (Mason and Fitzgerald, 1993;Mason et al., 1998).
In the parameter inversion, we attempt to constrain these two ocean model parameters using the Bayesian approach described in Sect.2.4.For consistency with sources in the emission inversion, two other parameters are included, i.e., emission ratios for soil (ER Soil ) and Asian anthropogenic sources (ER Asia ).It is noted that the emission inversion and parameter inversion are carried out separately.Because the responses of Hg 0 concentrations to changes in ocean parameters are nonlinear, as shown in the Supplement Fig. S7, we use a two-step iterative inversion method (Prinn et al., 2011).At each iteration step, the sensitivity matrix H is estimated by linearizing the nonlinear function around the current parameter estimate.In the parameter inversion, the state vector x contains four elements (corresponding to the four parameters), and P and Q are 4 × 4 matrices.

Error representation
Successful estimation of x (Eq.6) and its uncertainty Q (Eq.7) depends on reasonable representations of all relevant errors, including the a priori errors associated with reference emissions/parameters (contained in P) and errors related to Hg 0 observations and the CTM (contained in R).R consists of three parts: observational errors, model-observation mismatch errors, and model errors.

Errors in reference emission and parameters
For the emission inversion, we set the 1σ errors in reference emissions as 50 % in order to match uncertainties in their estimates using bottom-up approaches (see Table 2).For example, the reference emissions and 1σ errors for the NH and SH oceans are 1230 ± 630 and 1760 ± 880 Mg yr −1 , respectively.The uncertainty range of reference emission from the global ocean is estimated as 470-5510 Mg yr −1 , comparing very well with 780-5280 Mg yr −1 from bottom-up estimates (Mason, 2009;Pirrone et al., 2010).For the parameter inversion, the a priori estimates of two ocean model parameters −log K OX2 (5.0 ± 1.0) and log K D (5.3 ± 0.4).The a priori uncertainties of ER Soil and ER Asia are chosen as 50 %, the same as in the emission inversion.
The instrumental precision (σ IP ) of high-frequency Hg 0 measurements using the Tekran instrument is ∼ 2 % (Poissant et al., 2005).Here an intercomparison error (σ IC ) is used to represent the comparability of Hg 0 concentrations measured by different research groups using the Tekran instrument.In principle, it includes several inaccuracies during the measurement process (e.g., the instrument's flow control and the permeation source rate for the automated calibration) and also arises from the different data management and quality control protocols taken by different research groups (Steffen et al., 2012).Its value has been assessed during several field intercomparisons (Temme et al., 2006;Aspmo et al., 2005;Munthe et al., 2001;Ebinghaus et al., 1999;Schroeder et al., 1995).Hg 0 concentrations measured by different groups have a relative SD of reproducibility of 1-9 %, and we choose a generous uniform intercomparison error of 10 %.Sampling frequency error (σ SF ) reflects the ability of each site to capture the overall variability of Hg 0 concentration in 1 month and is calculated as the monthly SD divided by the square root of the number of valid hourly data points in this month (Rigby et al., 2012).Table 1 shows observational errors at each site, averaged over 2009-2011.The total observational errors are dominated by intercomparison errors.The other two types of errors have small contributions.

Model-observation mismatch errors
The mismatch error (σ MM ) exists because an observation is made at a single point in space, but its corresponding grid box in model represents a large volume of air.We estimate σ MM as the SD of monthly Hg 0 concentrations in the eight surrounding grid boxes (at the same vertical layer) from the reference simulation (Chen and Prinn, 2006).As shown in Table 1, σ MM values are larger over strongly emitting continental areas (e.g., SGR and WLG) and smaller over remote marine areas (e.g., CPT and AMS).

Model errors
All existing CTMs including GEOS-Chem are imperfect, due to both errors in meteorological data driving the CTMs and errors induced by their parameterizations of physical and chemical processes.The former type of model errors is termed "forcing errors" and the latter "process errors" (Locatelli et al., 2013).Physical processes consist of horizontal/vertical resolution, advection/convection, turbulence, planetary boundary layer mixing, etc.The CTM for Hg is subject to large process errors due to highly uncertain atmospheric chemistry.Recent studies have shown that Br concentration may be significantly underestimated in GEOS-Chem (Parrella et al., 2012;Gratz et al., 2015) and that current Br-initiated oxidation mechanisms are incomplete in describing all possible radical reactions (Dibble et al., 2012;F. Wang, et al., 2014).In order to provide a preliminary assessment of the effect of Br oxidation chemistry on our inversion, we perform an additional parameter inversion including six new elements in the state vector x, and each of them represents Br columns in a 30 • latitudinal band (see results in Sect.3.3 and Fig. S8 in the Supplement).
Quantifying model errors requires incorporating many CTMs which are driven by different meteorology and which contain different parameterizations (Prinn, 2000).Multi-CTM intercomparison studies have been performed for CO 2 and CH 4 (Gurney et al., 2002;Baker et al., 2006;Locatelli et al., 2013), suggesting that model errors can impact inverted emissions.Few other global CTMs exist for Hg (Bullock et al., 2008(Bullock et al., , 2009)).Due to our inability to quantify model errors using a single CTM, model errors are not incorporated in our inversion, like many other inverse studies (Huang et al., 2008;Xiao et al., 2010;Rigby et al., 2012).As a result, R in Eq. ( 5) only includes observational errors and modelobservation mismatch errors.

Emission inversion: model-observation comparison
We first test whether the comparison between ground-based Hg 0 observations and model outputs improves when using optimized emissions, compared to reference emissions.Figure 3 shows the modeled and observed Hg 0 concentrations at all 27 sites.To quantify model performance, we calculate the normalized root mean square error (NRMSE) for each site: where X obs,i and X mod,i are the observed and modeled Hg 0 concentrations at the ith month (n in total), respectively.As shown in Table 1, an average NRMSE of 0.13 is obtained for the emission inversion, smaller than that of 0.16 for the reference simulation, indicating that the emission inversion can better reproduce ground-based observations.While this is a relatively small uncertainty reduction (−0.03), we do not expect better performance for our inversion.This is because errors in Hg 0 observations (as described above, and in Table 1) are roughly 13 %, which constrain the optimization.
Our inversion brings the average NRMSE within the observation error.The NRMSEs are not reduced for all 27 sites (see Table 1).For three Nordic sites (ZEP, ADY, and BKN) and four Asia-Pacific sites (WLG, SGR, LUL, and MLO), the NRMSEs increase.Hg 0 concentrations are ∼ 1.8 ng m −3 at the three Nordic sites, higher than the modeled values (Fig. 3) from both reference simulation and emission inversion, and also higher than those measured at many background sites Atmos.Chem.Phys., 15, 7103-7125, 2015 www.atmos-chem-phys.net/15/7103/2015/ in Europe (Ebinghaus et al., 2011;Kentisbeer et al., 2014;Weigelt et al., 2013).Part of the differences may be explained by a positive bias in the instrumentation of these Nordic observations when compared to other laboratories (Temme et al., 2006).It is also possible that GEOS-Chem cannot sufficiently capture local meteorology and/or emissions at these sites.For the Asia-Pacific sites, the reference simulation underestimates Hg 0 at SGR (−32 %, calculated as (y ref /y obs − 1) × 100 %, hereinafter the same) and WLG (−19 %) and predicts comparable values at MLO (+2 %) and LUL (+0 %).Such discrepancies likely arise from unknown intercomparison errors and are influenced by local emission and meteorology factors not captured by the CTM (Fu et al., 2012b;Wan et al., 2009).These sites are operated by three different laboratories but, to the best of our knowledge, no field intercomparisons have been conducted among these laboratories.
Figure 4 compares monthly Hg 0 observations with model simulations for sites aggregated into four regions: Asia-Pacific, North America, Europe, and Southern Hemisphere.The emission inversion significantly improves the comparison for the SH sites (CPT, AMS, and TRS, see Table 1).In the reference simulation, Hg 0 concentrations at the SH sites vary seasonally, with a high in austral winter (∼ 1.3 ng m −3 ) and a low in austral summer (∼ 0.9 ng m −3 ).However, observed Hg 0 shows little seasonal variation with monthly concentrations of ∼ 1.0 ng m −3 .The emission inversion reduces the Hg 0 concentration in austral winter and fits the observations much better (the average NRMSE decreases from 0.19 to 0.10).As shown in Fig. 3, all three SH sites show improvement after optimization.
The emission inversion also improves the comparison for sites in North America (the average NRMSE decreases from 0.13 to 0.08).Hg 0 data at a total of 11 sites are available, including five coastal sites (ALT, SAT, KEJ, SCZ, and GRB), five inland sites (BRL, EGB, HTW, ATS, and YKV), and one mountaintop site (MBO) (see Fig. 1 and Table 1).Hg 0 at the coastal and inland sites are observed to be 1.41 ± 0.04 and 1.29 ± 0.06 ng m −3 , respectively.This coastal-inland difference in observation is consistent with results of Cheng et al. (2014), who found that air masses from open ocean at the site KEJ had 0.06 ng m −3 higher Hg 0 concentrations than those originating over land.The reference simulation and emission inversion both obtain comparable Hg 0 concentrations at the coastal sites (1.43 ± 0.06 and 1.38 ± 0.07 ng m −3 ).At the inland sites, the emission inversion predicts Hg 0 concentrations (1.38 ± 0.03 ng m −3 ) closer to observations than the reference simulation (1.50 ± 0.06 ng m −3 ).
Over-water Hg 0 observations serve as an independent test of the emission inversion.As shown in Fig. 2, Hg 0 concentrations over the North Atlantic Ocean from both the reference simulation and the emission inversion fall within 1σ uncertainty ranges of Hg 0 nor .The NRMSEs for the reference simulation and the emission inversion are 0.09 and 0.10, respec-tively.Thus, using Hg 0 emissions constrained by groundbased observations, GEOS-Chem still matches these regional over-water observations.
We additionally test the performance of the inversion by comparison with regional wet deposition data.Since most oxidized Hg is formed from the oxidation of Hg 0 , changing Hg 0 emissions may have an effect on modeled oxidized Hg and its subsequent deposition.We compare model results to the observed wet deposition fluxes from NADP/MDN (2012), as shown in the Supplement, Fig. S9.We use the monitoring sites active in 2009-2011 (n = 126).Both the reference simulation and the emission inversion fit observations well (R ≈ 0.7, NRMSE ≈ 0.3).Accordingly, the effect of the inversion on the NADP/MDN (National Atmospheric Deposition Program/Mercury Deposition Program) wet deposition fluxes is insignificant.

Emission inversion: optimized emissions
The annual reference and optimized emissions of mercury are shown in Table 2 where n = 12 months and σ t is monthly error, is used to compute the annual uncertainty for seasonal processes (Chen and Prinn, 2006).The uncertainty of the aseasonal source (annual Asian anthropogenic emission) is obtained directly from Eq. ( 7).The global optimized mercury emission is ∼ 5.8 Gg yr −1 , with an uncertainty range of 1.7-10.3Gg yr −1 .Compared to our reference emission of ∼ 6.0 Gg yr −1 (uncertainty range: 0.4-12.2Gg yr −1 ), the emission inversion results in a slightly smaller value and also reduces its uncertainty range.The optimized value is smaller than previous estimates of 7.5 Gg yr −1 by Pirrone et al. ( 2010) using a bottom-up approach.The emission inversion increases emissions from anthropogenic sources and ocean surfaces but decreases those from terrestrial surfaces.The ocean accounts for more than half (55 %) of the total, while the terrestrial surface contributes only a small fraction (6 %).

Ocean
Net Hg 0 evasion from the global ocean is optimized by the emission inversion as 3160 Mg yr −1 , with an uncertainty range of 1160-5160 Mg yr −1 (Table 2).The NH and SH oceans contribute similar amounts to the total but, on an area basis, evasion from the NH ocean is higher since it is 30 % smaller.We are able to reduce ocean evasion uncertainty from 50 to 40 % by using top-down constraints.
Figure 5 shows the monthly reference and optimized emissions of seasonal sources.We find, for both hemispheres, that the emission inversion generally results in increased ocean emissions in summer and decreased emissions in winter, compared to the reference simulation.As a result, we hypothesize that one or more ocean processes that affect the www.atmos-chem-phys.net/15/7103/2015/Atmos.Chem.Phys., 15, 7103-7125, 2015 seasonal behavior of aqueous mercury and its evasion are not well-represented in GEOS-Chem.We therefore conduct a series of sensitivity studies of model parameters to test their potential effects on the seasonal pattern of ocean emission.
We also compare the parameter values used in GEOS-Chem with their possible ranges in a recent review (Batrakova et al., 2014).The tested model parameters in GEOS-Chem include rates of redox chemical reactions and physical processes in the ML and subsurface mercury concentrations affecting physical exchange between the ML and subsurface waters.Through these sensitivity tests and literature review, we identify two processes as candidates for improvement, the rate constant of dark oxidation of Hg 0 aq (K OX2 ) and the partition coefficient between Hg 2+ aq and Hg P aq (K D ).We optimize these two ocean model parameters in the parameter inversion, as described in Sect.2.5.

Terrestrial ecosystem
As shown in Table 2, the emission inversion reduces soil emissions of Hg 0 by about 50 %, from 1680 ± 840 to 860 ± 440 Mg yr −1 .Using Eq. ( 1), the optimized net emission flux from terrestrial surfaces (E net ) is 340 Mg yr −1 .If we do not consider geogenic activities (90 Mg yr −1 ) and biomass burning (210 Mg yr −1 ), the E net2 (calculated as E soil + E pr − E ddHg 0 and representing net emissions from soils/vegetation) is almost zero after optimization.Thus, terrestrial surfaces are neither a net source nor a net sink of Hg 0 .This is in contrast to bottom-up estimates that the terrestrial surface is a net source of about 2000 Mg yr −1 (Pirrone et al., 2010;Mason, 2009).
Vegetation is now believed to serve as a net sink of atmospheric Hg 0 through foliar uptake and sequestration (Gustin et al., 2008;Stamenkovic and Gustin, 2009;X. Wang et al., 2014).Although its size has not been well quantified, we suggest that this sink is important in global mass balance since litterfall transfers 2400-6000 Mg Hg yr −1 to terrestrial surfaces (Gustin et al., 2008).Air-soil flux measurements show that Hg 0 emissions from background soils generally dominate over dry deposition (Obrist et al., 2014;Edwards and Howard, 2013;Park et al., 2013;Denkenberger et al., 2012;Ericksen et al., 2006).Our result of a smaller soil Hg source is consistent with a study by Obrist et al. (2014), which suggested that Hg was unlikely to be re-emitted once incorporated into soils and that terrestrial Hg emission was restricted to surface layers (Demers et al., 2013).Our result is also in agreement with estimates of terrestrial fluxes of southern Africa using Hg 0 correlations with 222 Rn, a radioactive gas of predominantly terrestrial origin (Slemr et al., 2013).Considering that soil is a smaller source while vegetation a sink of Hg 0 , our result that the terrestrial ecosystem is neither a net source nor a net sink of Hg 0 is reasonable, implying that the magnitudes of soil emission and dry deposition of Hg 0 (primarily to vegetation) are similar.We evaluate dry deposition fluxes modeled by GEOS-Chem against data in L. Zhang et al. (2012), which estimated fluxes at sites in North America and obtained good agreements with surrogate surface and litterfall measurements (Graydon et al., 2008;Lyman et al., 2007).As shown in the Supplement, Fig. S10, there is no bias in the average dry deposition flux at eight background sites, indicating that ∼ 1400 Mg yr −1 (modeled by GEOS-Chem) may be reasonable estimates for both emission and dry deposition of Hg 0 .

Anthropogenic emission from Asia
Table 3 summarizes Asian emissions of Hg 0 (only GEM) estimated by several recent bottom-up emission inventories and modeling studies.These inventories reported Asian anthro- pogenic emissions ranging from 550 to 800 Mg yr −1 .In our model simulations, the reference emission of 770 Mg yr −1 follows AMAP/UNEP (2013).The emission inversion using all 27 sites increases this value to 1060 ± 110 Mg yr −1 .Uncertainty in Asian anthropogenic emission should be larger than that obtained using our inversion method, because emission estimates are sensitive to the Asia-Pacific sites used in the inversion.As discussed above, model performance at several Asia-Pacific sites is affected by unknown intercomparison errors and local emission and meteorological factors not captured by GEOS-Chem.To obtain a more accurate estimate of uncertainty, we perform seven emission inversions, each including only one Asia-Pacific site.

Atmos
As shown in Table 3, these inversions result in Asian anthropogenic emissions of Hg 0 ranging from 650 to 1770 Mg yr −1 .Comparing this range to its bottom-up inventory estimates of 550-800 Mg yr −1 , we suggest that it is very likely to be underestimated.We estimate total (anthropogenic + natural + legacy) Hg 0 emission in Asia as 1180-2030 Mg yr −1 .Our uncertainty ranges cover those in Strode et al. (2008), which estimated total Asian emission of 1260-1450 Mg yr −1 with 890-990 Mg yr −1 from anthropogenic sources, by comparing GEOS-Chem to the observed Hg : CO ratio at sites OKN and MBO.Pan et al. (2007) assimilated aircraft observations into a regional CTM and estimated total Hg 0 emission in east Asia as 2270 Mg yr −1 , at the upper end of our range.Fu et al. (2015) obtained a total Hg 0 emission in Asia of 1590-1870 Mg yr −1 , which compared well with our range, using the Hg 0 : CO and Hg 0 : CO 2 slopes observed at ground-based sites and inventories of CO and  2008) estimated natural terrestrial emission in east Asia was about 710 Mg yr −1 , much higher than our 0-230 Mg yr −1 in a larger domain.The difference is due to their larger estimation of vegetation evapotranspiration (630 Mg yr −1 ).

Parameter inversion
Results of the parameter inversion are presented in Table 4.The a posteriori K OX2 of 6 × 10 −6 s −1 is much larger than its current value (1 × 10 −7 s −1 ) in GEOS-Chem, suggesting that Hg 0 aq dark oxidation in the ML is more important than previously thought.The a posteriori log K D of 4.2 is lower than seawater values in the literature (Fitzgerald et al., 2007;Batrakova et al., 2014) but agrees with the lower end of freshwater measurements (Amos et al., 2014).We attribute this discrepancy to several simplifying assumptions in GEOS-Chem.K D is linked to the estimates of SPM concentrations in the ML and organic carbon export.As described above, the amount of organic carbon export is very uncertain (5-20 Gt C yr −1 ).A smaller organic carbon export may correspond to a larger log K D .The uncertain spatial and seasonal variations of carbon export may also affect the estimate of log K D .In addition, there are no available global data sets of SPM in the ML.GEOS-Chem derives SPM concentrations from MODIS satellite chlorophyll a and C : Chl a ratios (Soerensen et al., 2010b).Thus, the uncertain SPM fields may also affect log K D .As for the other two parameters (ER Soil and ER Asia ), the parameter inversion decreases soil emission but increases Asian anthropogenic emission, consistent with the emission inversion (see Table 4).
Similar to our model-observation comparison for the emission inversion, we run GEOS-Chem using optimized parameters and calculate the NRMSEs for all ground-based sites (Table 1).A smaller average NRMSE of 0.14 for the parameter inversion than that of 0.16 for the reference simulation shows improvement in model performance.GEOS-Chem simulations using optimized parameters also match regional over-water Hg 0 (NRMSE = 0.10, Fig. 2) and wet deposition measurements (Fig. S9 in the Supplement).In addition, we evaluate the optimized model against recent surface ocean measurements of total aqueous mercury (Hg T aq ), Hg 0 aq , and Hg P aq (Table 5).For Hg T aq , 50 and 75 % (6 and 8 out of 12) of the modeled data from the reference and optimized simulations, respectively, are within measurement ranges.For Hg 0 aq , 60 % (6 out of 10) of the modeled data from both simulations are within measurement ranges.For Hg P aq , the reference simulation predicts a higher value while the parameter inversion predicts a lower value than the only measurement data.These results suggest that the parameter inversion is comparable or potentially better than the reference simulation with regard to modeling surface ocean mercury.
Optimizing the two ocean model parameters, −log K OX2 and log K D , changes the global ocean Hg budget in GEOS-Chem, as shown in Fig. 6.Sources of Hg aq in the ML in- clude deposition of oxidized Hg and physical transport from subsurface waters.They are balanced by Hg 0 evasion and Hg P aq sinking.In the reference simulation, although deposition (20.2 Mmol yr −1 ) accounts for most ML Hg aq inputs, the two physical transport processes, entrainment/detrainment of the ML and Ekman pumping, together supply a considerable amount (F INT : 6.1 Mmol yr −1 ) from subsurface waters.This upward flux is a result of the gradient in Hg T aq between the ML (0.8 pM) and subsurface waters (1.1 pM).Hg 0 evasion and Hg P aq sinking remove 14.9 and 11.4 Mmol yr −1 from the ML, respectively.The combined effect of the larger K OX2 and smaller K D in the parameter inversion is, in the ML, that Hg 2+ aq increases from 0.69 to 0.95 pM, Hg P aq decreases from 0.05 to 0.004 pM, and Hg 0 aq remains 0.06 pM.Hg P aq sinking becomes a smaller sink (1.7 Mmol yr −1 ) due to the lower K D .Physical transport contributes a downward flux (−1.5 Mmol yr −1 ) since the gradient of Hg T aq between the ML (1.0 pM) and subsurface waters (1.1 pM) is diminished.
Physical transport and Hg P aq sinking affect seasonal variations of simulated Hg 0 evasion from the ocean (Soerensen et al., 2010b).In summer, enhanced biological productivity increases Hg P aq sinking and decreases Hg 0 evasion by shifting speciated Hg aq equilibrium in the ML towards Hg 0 aq loss.During winter months, the ML deepens and Hg aq in subsurface waters invade the ML by entrainment; additionally, Hg 0 evasion will be enhanced if subsurface waters contain higher Hg T aq .In the parameter inversion, physical transport and Hg P aq sinking are both weakened, as described above.As a result, the parameter inversion overturns seasonality of simulated ocean evasions in both hemispheres (Fig. 5), agreeing with results from the emission inversion.
1.0 ± 0.5 0.37 ± 0.08 0.37 ± 0.19 0.24 ± 0.1 (soil emission decreases by 76 %) ER Asia 1.0 ± 0.5 1.7 ± 0.1 1.7 ± 0.9 1.9 ± 0.1 (Asian anthropogenic emission increases by 90 %) * For the second iteration, we use the best estimates derived from the first iteration, but larger parameter uncertainties.The uncertainty of 1.0 for −log K OX2 is the same as that for the a priori estimate.The uncertainties for ER Soil and ER Asia are chosen as 50 % of their best estimates, consistent with the emission inversion.The uncertainty for log K D is chosen as 0.2 because it is approaching the lower end (4.2) of the possible values in the literature survey.columns in different latitudinal bands.As shown in the Supplement, Fig. S8, −log K OX2 is found to be strongly correlated with Br columns at 30-60 • N, 30

Implications for the Hg biogeochemical cycle
We use the box model developed by Amos et al. (2013Amos et al. ( , 2014) ) to explore the long-term impact of our inverted emissions and parameters on the global biogeochemical cycling of mercury.This seven-box model dynamically couples the atmosphere, three terrestrial reservoirs (fast, slow, and armored), and three ocean reservoirs (surface, subsurface, and deep).All rate coefficients of Hg mass between reservoirs are assumed to be of the first order.The simulation is initialized with geogenic emissions to represent the natural mercury cycle and, after reaching steady state, it is driven by historical anthropogenic emissions (Streets et al., 2011;Horowitz et al., 2014).Two box-model simulations are performed.The first uses rate coefficients from the present-day global budget in the reference simulation.The second uses those from our emission and parameter inversions and has higher anthropogenic emissions, lower re-emission from terrestrial surfaces, and less sinking out of the surface ocean than the first one does (Table S4 in the Supplement).The second simulation obtains larger terrestrial mercury reservoirs, highlighting their important role in sequestering legacy mercury.The oceans are a smaller mercury reservoir of ∼ 1700 Mmol in the second simulation, compared to that of ∼ 2000 Mmol in the first simulation.The former number is more consistent with the estimates of about 1300-1400 Mmol by Lamborg et al. (2014) and Zhang et al. (2014).The first box-model simulation shows that 18 % of present-day atmospheric deposition is from primary anthropogenic emissions, 76 % is legacy, and 6 % is natural (i.e., geogenic emissions).Applying our inversion results into the box model, the second simulation suggests that primary anthropogenic emissions account for a larger fraction (18-23 %) of present-day atmospheric deposition.Legacy releases of mercury contribute a smaller proportion (72-76 %) but still play a major role.

Summary and conclusion
Here, we perform global-scale inverse modeling combining ground-based Hg 0 observations and GEOS-Chem mercury simulations.Using Bayesian inversion methods, we are able to constrain present-day mercury emission fluxes from major sources (emission inversion) and relevant key parameters in GEOS-Chem (parameter inversion), and reduce uncertainties associated with these fluxes and parameters.
The emission inversion better reproduces the groundbased Hg 0 observations (particularly for sites in the Southern Hemisphere and North America) than the reference simulation and also matches measured Hg 0 over the North Atlantic Ocean and wet deposition fluxes in North America.We obtain a global Hg emission of 5.8 Gg yr −1 (uncertainty range: 1.7-10.3Gg yr −1 ), smaller than the estimate of 7.5 Gg yr −1 using a bottom-up approach (Pirrone et al., 2010).The global ocean accounts for 3.2 Gg yr −1 Hg (55 % of the total).The terrestrial ecosystem is neither a net source nor a net sink of atmospheric Hg 0 , in contrast to its bottomup estimate as a significant source (Pirrone et al., 2010).The optimized Asian anthropogenic emissions range from 650 to 1770 Mg yr −1 , suggesting that bottom-up inventories (550-800 Mg yr −1 ) may have underestimated their value.The total Asian Hg 0 emission (including anthropogenic, natural and legacy sources) is estimated as 1180-2030 Mg yr −1 , consistent with recent studies (Fu et al., 2015;Strode et al., 2008;Pan et al., 2007).
The emission inversion changes seasonal patterns of ocean emissions in both hemispheres.We identify and constrain two ocean model parameters in GEOS-Chem that can explain this seasonal pattern, the rate constant of dark oxidation of Hg 0 aq (K OX2 ) and the partition coefficient between Hg 2+ aq and Hg P aq (K D ).The a posteriori K OX2 (6 × 10 −6 s −1 ) is larger than its current value in GEOS-Chem (1 × 10 −7 s −1 ), suggesting that dark oxidation of Hg 0 aq is more important than previously thought.The a posteriori log K D (4.2) is smaller than its a priori (5.3), leading to less Hg P aq sinking out of the mixed layer.These changes in parameters affect the simulated global ocean mercury budget, especially mass exchange between the mixed layer and subsurface waters.The parameter inversion changes seasonality of ocean emissions in both hemispheres, agreeing with results from the emission inversion.
Our inversion results suggest changes in our understanding of the timescales of cycling between different mercury reservoirs.Based on these changes, the long-term biogeochemical box-model simulations result in larger estimated terrestrial mercury pools and smaller ocean mercury pools.Legacy mercury accounts for a smaller fraction of present-day atmospheric deposition than previous estimates, whereas the contribution of primary anthropogenic emissions becomes larger (up to 23 %).
Our inversion results identify specific knowledge gaps in mercury observation and modeling that currently limit our ability to constrain the biogeochemical cycle of mercury.First, and most important, effective inversions are hampered by the uncertain atmospheric Hg measurements, particularly the large intercomparison errors in measured GEM.Only a few experiments have been made to evaluate the comparability of mercury measurements (Gustin et al., 2013).Our results show that intercomparison errors (about 10 %) dominate the total observational errors and thus limit the uncertainty reduction possible by our inverse approach.Our inversions only lead to moderate reductions of the average NRMSE (Sect.3.1).Therefore, research aimed at quantifying and reducing the intercomparison errors should be given high priority by the mercury measurement community.Second, observational sites are sparse in some regions (e.g., the Southern Hemisphere).More sites in these regions are necessary to further constrain emissions.Third, the uncertainty in atmospheric mercury chemistry also affects our inversion re- Improving our understanding of atmospheric mercury chemistry at both global and regional scales (e.g., the polar regions) requires a combination of both measurement and modeling advances.

Atmos
The Supplement related to this article is available online at doi:10.5194/acp-15-7103-2015-supplement.

Figure 1 .
Figure 1.Locations of ground-based observational sites.

a
Observational sites without original data are MBO, CBS, WLG, NMC, SGR, LUL, and NWN.b Observational sites where we use TGM data are ALT, BRL, SAT, EGB, CBS, WLG, NMC, SGR, and NWN.For all other sites, we use GEM data.c Unit for altitude is meters.d Network affiliations: (1) Canadian networks, (2) EMEP, (3) AMNet, and (4) individual observational sites.More information about these individual sites can be found in Weiss-Penzias et al. (

Figure 2 .
Figure 2. Observed and modeled monthly Hg 0 concentrations over the North Atlantic Ocean.The observational data and related references are given in the Supplement.Hg 0 obs are the concentrations observed from 19 ship cruises during 1990-2009, whereas Hg 0 nor are the concentrations normalized to levels consistent with year 2009.The gray shaded region shows the 1σ error of Hg 0 nor , which is composed of the observational error, mismatch error, and regression error.

Figure 3 .
Figure 3. Monthly Hg 0 concentrations for all ground-based observational sites.Note different scales on vertical axes.Error bars correspond to the total errors described in Sect.2.6.The two numbers in parentheses after the name of each site are its latitude and longitude.For polar sites (ALT, ZEP, ADY, and TRS), the gray color shows the observed Hg 0 concentrations that are not used in our inversions due to AMDEs, as shown in Sect.2.1.

Figure 5 .
Figure 5. Monthly emissions for the three seasonal sources (NH ocean, SH ocean, and soil) from the reference simulation (blue solid lines), emission inversion (red solid lines), and parameter inversion (green dashed lines).The gray shaded regions and red error bars indicate 1σ uncertainties for the reference emissions and emission inversion, respectively.

Figure 6 .
Figure 6.Global ocean mercury budget modeled by GEOS-Chem.Blue color indicates the reference simulation and red color the parameter inversion.Fluxes are in megamoles per year.Notations in this figure follow Soerensen et al. (2010b).F INT denotes net fluxes from subsurface waters through entrainment/detrainment of the mixed layer and Ekman pumping.
, in constraining ocean model parameters).

Table 1 .
Information for ground-based observational sites of atmospheric mercury.

Table 2 .
Global mercury emissions into the atmosphere (Mg yr −1 ). a

Table 4 .
Evolution of the parameters' estimates in the parameter inversion.

Table 5 .
Recent surface ocean mercury measurements and simulated concentrations.a

7125, 2015 S. Song et al.: Top-down constraints on atmospheric mercury emissions
• S-0 • , and 60-30 • S. The other three factors, log K D , ER Soil , and ER Asia , have no or weak correlations with Br columns.Thus, we suggest that the inversion results of smaller terrestrial emissions and larger Asian anthropogenic emissions are not likely to be afwww.atmos-chem-phys.net/15/7103/2015/Atmos.Chem.Phys., 15, 7103-