Multi-model simulation of CO and HCHO in the Southern Hemisphere : comparison with observations and impact of biogenic emissions

We investigate the impact of biogenic emissions on carbon monoxide (CO) and formaldehyde (HCHO) in the Southern Hemisphere (SH), with simulations using two different biogenic emission inventories for isoprene and monoterpenes. Results from four atmospheric chemistry models are compared to continuous long-term groundbased CO and HCHO column measurements at the SH Network for the Detection of Atmospheric Composition Change (NDACC) sites, the satellite measurement of tropospheric CO columns from the Measurement of Pollution in the Troposphere (MOPITT), and in situ surface CO measurements from across the SH, representing a subset of the National Oceanic and Atmospheric Administration’s Global Monitoring Division (NOAA GMD) network. Simulated mean model CO using the Model of Emissions of Gases and Aerosols from Nature (v2.1) computed in the frame work of the Land Community Model (CLM-MEGANv2.1) inventory is in better agreement with both column and surface observations than simulations adopting the emission inventory generated from the LPJ-GUESS dynamical vegetation model framework, which markedly underestimate measured column and surface CO at most sites. Differences in biogenic emissions cause large differences in CO in the source regions which propagate to the remote SH. Significant inter-model differences exist in modelled column and surface CO, and secondary production of CO dominates these inter-model differences, due mainly to differences in the models’ oxidation schemes for volatile organic compounds, predominantly isoprene oxidation. While biogenic emissions are a significant factor in modelling SH CO, inter-model differences pose an additional challenge to constrain these emissions. Corresponding comparisons of HCHO columns at two SH mid-latitude sites reveal that all models significantly underestimate the observed values by approximately a factor of 2. There is a much smaller impact on HCHO of the significantly different biogenic emissions in remote regions, compared to the source regions. Decreased biogenic emissions cause decreased CO export to remote regions, which leads to increased OH; this in turn results in increased HCHO production through methane oxidation. In agreement with earlier studies, we corroborate that significant HCHO sources are likely missing in the models in the remote SH.


G. Zeng et al.: Multimodel simulation of SH CO and HCHO
and non-methane volatile organic compounds (NMVOCs) (e.g. Duncan et al., 2007). Its principal sink is the reaction with the hydroxyl radical (OH); hence, CO plays a key role in controlling the oxidizing capacity in the atmosphere (e.g. Levy, 1971). The oxidation of methane and NMVOCs, such as isoprene (C 5 H 8 ), monoterpenes (C 10 H 16 ), acetone (CH 3 COCH 3 ) and higher aldehydes, leads to the formation of formaldehyde (HCHO), which, through photolysis and reaction with OH, is the major chemical source of CO (Atkinson, 2000). Once formed, CO has a relatively long lifetime of around 1-2 months, and therefore it is often used as a chemical marker for characterizing the long-range transport of air pollutants away from important source regions (e.g. Staudt et al., 2001;Heald et al., 2003;Liang et al., 2004;Fisher et al., 2010).
Due to a lack of strong regional emission sources, the Southern Hemisphere (SH) acts as a global sink for many of the polluting trace species emitted in the tropics, where polluted plumes are transported away out over the relatively clean ocean becoming subject to chemical processing. The relatively low population density, and thus low anthropogenic activity, in the SH means that direct emission sources of CO are principally limited to biomass burning (BB) and direct biogenic processes (e.g. Swinnerton et al., 1970;Watson et al., 1990;Fishman et al., 1991). Satellite and ground-based observations of CO in the SH have been used to identify the effect of BB and its footprint through longrange transport in the SH, which dominates the CO seasonal cycle there (e.g. Rinsland et al., 2005;Edwards et al., 2006;Gloudemans et al., 2006;Morgenstern et al., 2012). Global distributions of HCHO are much more inhomogeneous than CO, due to the much shorter lifetime of HCHO (on the order of a few hours), and the concentration of HCHO drops off sharply away from the source regions. Observations of HCHO are commonly used to constrain isoprene emissions in high-emission regions, because it is a high-yield product of isoprene oxidation (e.g. Palmer et al., 2003;Shim et al., 2005;Barkley et al., 2008).
Global chemical models have been extensively used to estimate the sources and sinks of CO (e.g. Holloway et al., 2000;Duncan et al., 2007). However, systematic discrepancies between modelled and observed CO still exist, with models generally underestimating CO in the more polluted Northern Hemisphere (NH) and overestimating CO in the SH (e.g. Shindell et al. , 2006;Naik et al., 2013;Stein et al., 2014). In the remote SH, however, the extremely low HCHO concentrations are expected to further complicate the comparisons of model results with observations.
In contrast to the anthropogenic emissions dominating CO sources in the more polluted NH, biogenic volatile organic compounds (VOCs) are important sources of CO and HCHO in the SH, and isoprene oxidation contributes significantly to the regional CO and HCHO abundances in this region (Pfister et al., 2008). However, large uncertainties exist in biogenic emissions inventories, in particular for surface fluxes of iso-prene and monoterpenes (Arneth et al., 2008). Bottom-up estimates of annual isoprene emissions vary between 400 and 600 Tg C yr −1 (Arneth et al., 2008), and the typical range of annual total isoprene emissions implemented in global atmospheric chemistry models is ∼ 200-600 Tg C yr −1 . The effect of such uncertainties in biogenic emissions on SH composition, such as CO and HCHO, has not been adequately assessed. Moreover, the sparsity of the ground-based CO and HCHO measurements in the SH also limits our ability to constrain these biogenic emissions.
In this study, we perform a number of simulations using an ensemble of chemical transport models (CTMs) and chemistry-climate models (CCMs) as part of the Southern Hemisphere Model Intercomparison Project (SHMIP), to compare modelled CO and HCHO to observations and to investigate the factors that influence the distributions of CO and HCHO in the SH. Given the relatively low anthropogenic emissions in the SH and the dominance of biogenic emissions of VOCs (mainly isoprene), we determine the influence that different emission inventories of isoprene and monoterpenes have regarding their effects on modelled CO and HCHO columns in the SH. Satellite observations of SH CO usually are in good agreement with groundbased observations; however, the data quality of the satellite data deteriorates towards the poles . Morgenstern et al. (2012) found that CO columns exhibit a large-scale mode of variability in the remote SH that does not exist in the NH. For our purposes, we make use of high-precision ground-based Fourier transform infrared spectroscopy (FTIR) measurements of CO columns from four SH sites that have not previously been fully exploited for model evaluations, namely Darwin (12.43 • S, 130.89 • E) and Wollongong (34.41 • S, 150.88 • E) in Australia, Lauder (45.04 • S, 169.69 • E) in New Zealand, and Arrival Heights (77.82 • S, 166.65 • E) in Antarctica. We also compare the modelled HCHO columns to those observed by the FTIR instruments at Wollongong and Lauder. In a companion paper, Fisher et al. (2015) have evaluated the vertical gradients of CO from the SHMIP models in the vicinity of Cape Grim, Australia, which is representative of SH mid-latitude background air, using multi-year aircraft measurements available from the Cape Grim Overflight Programme (Langenfelds et al., 1996). The influence of both chemistry and transport on the modelled vertical gradients of CO are addressed. Although there are biases of various magnitudes across the different models, the seasonal variability and extent of the gradients in tropospheric CO are shown to be captured reasonably well, especially during the tropical BB season.
In this paper we address the sensitivity of CO and HCHO distributions in the SH to biogenic emissions of isoprene and monoterpenes as provided by the LPJ_GUESS emission inventory (Arneth et al., 2007a, b;Schurgers et al., 2009) and the MEGANv2.1 model (Guenther et al., 2012) across the models included in SHMIP. In Sect. 2 we provide model descriptions, the common emission inventories used to drive the models, and the observations used in the study. In Sect. 3 we compare results for the period 2004-2008 and show comparisons between modelled CO and HCHO columns and the FTIR measurements at the four SH sites mentioned above, modelled and MOPITT CO columns, as well as comparisons between modelled and observed surface CO. In Sect. 4 we analyse differences in the models' abilities to reproduce SH CO and HCHO columns, and the underlying differences in the models' chemistry and transport. In Sect. 5 we further analyse the chemical production and loss terms to address differences in models' NMVOC oxidation mechanisms. In Sect. 6 we assess the sensitivity of modelled CO and HCHO to changes in biogenic emissions and the effect of such changes on the oxidizing capacity in the clean SH. Finally, in Sect. 7 we present our conclusions.

Model simulations and observations
The SHMIP intercomparison uses four global models, including two CTMs (Tracer Model 5 (TM5), GEOS-Chem) and two CCMs (the Community Atmosphere Model with chemistry (CAM-chem), NIWA-UKCA -the National Institute of Water and Atmospheric Research -UK Chemistry and Aerosols Model). In this section we provide the description of the simulations performed, the common emission inventories employed, a brief description of each model, the meteorological drivers, and the observations used for evaluating the performance of the models.

Simulations
We perform simulations covering the period of 2004 to 2008 using a 1-year spin-up for 2003. The two CTMs are driven by the meteorological analysis for the same period from their respective sources, whereas NIWA-UKCA uses observed sea surface temperature and sea ice data sets. CAM-chem runs in the specified-dynamics mode, using meteorological fields from the reanalysis data. Two simulations are performed in all models with identical emission inventories for the anthropogenic and BB components, but different inventories are adopted for biogenic isoprene and monoterpene emissions. We also include passive CO tracers in the simulation defined as having the same global primary, surface emission sources as CO, but with one having a fixed lifetime of 25 days and a second having the lifetime determined by OH distribution in each respective model. These tracers allow for the differentiation of the inter-model variability with respect to transport of CO to the SH from the main source regions.
Although we have been careful to harmonize the emissions used across models, differences in the chemical mechanisms which are employed result in the aggregated emissions of the NMVOCs being somewhat different across the models. For anthropogenic emissions, we adopt the yearly specific Monitoring Atmospheric Composition and Climate (MACC)/CITYZEN (MACCity) global emission estimates Granier et al., 2011) nested with the Regional Emission Inventory in Asia (REASv2.1) for the East Asian region (Kurokawa et al., 2013). Interannually varying monthly mean BB emissions are taken from the Global Fire Emissions Database version 3 (GFEDv3) (van der Werf et al., 2010). For lightning-NO x emissions, each model adopts individual parameterizations, which interact with the models' convection schemes. Natural emissions of soil NO x and CO from the ocean are taken from the Precursors of Ozone and their Effects in the Troposphere (POET) database (http://eccad.sedoo.fr). The annual total emission fluxes for key species are listed in Table 1 for the simulation period of 2004-2008. Biogenic emissions for isoprene, monoterpenes, CO, methanol, and acetone are based on MEGANv2.1 (Guenther et al., 2012) and are calculated offline using the Community Land Model (CLM4.0; Lawrence et al., 2011), driven by Climate Research Unit-National Centers for Environmental Prediction (CRUNCEP) reanalyses (http://dods.extra.cea. fr/data/p529viov/cruncep/readme.html) for each year. We refer to this data set as Model of Emissions of Gases and Aerosols from Nature (v2.1) computed in the frame work of the Land Community Model (CLM-MEGANv2.1) hereafter. We then replace the MEGANv2.1 emissions for isoprene and monoterpenes with the "GUESS-ES" emissions that were calculated by the LPJ-GUESS model (http://eccad. sedoo.fr) in our second set of simulations for comparison (hereafter referred to as LPJ-GUESS simulations); this emission data set is generated using the dynamical vegetation model LPJ-GUESS driven with Climate Research Unit Timeseries (CRU TS) 3.1 climate data (Arneth et al., 2007a;Schurgers et al., 2009). None of the models currently include any higher terpenes. The yearly varying annual global total emissions of isoprene from CLM-MEGANv2.1 (462-508 Tg yr −1 ) are markedly larger than the LPJ-GUESS emissions (431-450 Tg yr −1 ), i.e. ∼ 5-10 % relative to CLM-MEGANv2.1 (Table 1). The corresponding differences are much larger for the SH between these two inventories, i.e.
∼ 10-20 % relative to CLM-MEGANv2.1, over the same period. For monoterpenes, the annual total CLM-MEGANv2.1 emissions are substantially larger than the LPJ-GUESS emissions, i.e. a factor of 4 and 6 larger globally and in the SH, respectively. The two data sets have been generated from the respective land surface models driven by similar meteorological fields, as specified in the above references. Here, we do not harmonize the model meteorology to those used in generating the biogenic emissions. Instead, we prescribe the monthly mean biogenic emissions in the models to ensure the consistency. Figure 1 shows the SH and regional monthly total isoprene emission fluxes from LPJ-GUESS and CLM-MEGANv2.1 for Australia and part of Indonesia (0-44 • S, 94-156 • E), southern Africa (0-37 • S, 9-44 • E), and South America (0-57 • S, 34-84 • W), the regions with high isoprene emissions. The largest monthly emissions oc-  cur in austral summer in both data sets when the differences between these two data sets are also largest. Overall, the integrated CLM-MEGANv2.1 isoprene emissions (especially the summer maxima) are substantially higher than LPJ-GUESS emissions with the exception of southern Africa, where LPJ-GUESS emissions are larger. Figure 2 shows the spatial distribution of the fluxes for both isoprene and monoterpenes from the two emission inventories for January 2005. In Amazonia, tropical Africa, and Australia emissions are visibly larger in CLM-MEGANv2.1 than in LPJ-GUESS.

NIWA-UKCA
NIWA-UKCA stands for the version of the UK Chemistry and Aerosols Model (UKCA) that is used and further developed at the National Institute of Water and Atmospheric Research (NIWA). The background climate model is similar to HADGEM3-A (Hewitt et al., 2011) with a horizontal resolution of 3.75 • × 2.5 • and 60 vertical levels extending from the surface to 84 km. The physical processes in NIWA-UKCA, including interactive dry and wet deposition of the trace gases and the FAST-JX photolysis scheme, have been described in detail by Morgenstern et al. (2013) and O'Connor et al. (2014). Unlike the version described by O'Connor et al. (2014), here we now apply dry deposition following Wesely (1989) only to the bottom model layer rather than through- out the boundary layer. The model setup used here comprises a coupled stratosphere-troposphere chemistry scheme. The stratospheric reactions are the same as in Morgenstern et al. (2009) and include explicit chlorine and bromine chemistry. We have updated the NIWA-UKCA chemical mechanism from that described in Morgenstern et al. (2013) to account for emissions and degradations of ethene (C 2 H 4 ), propene (C 3 H 6 ), methanol (CH 3 OH), isoprene, a representative monoterpene, and a lumped species that accounts for missing NMVOCs in the model similar to the approach taken in the IMAGES model (Müller and Brasseur, 1995), with primary emission sources. In addition to the above, the NIWA-UKCA chemical scheme includes nitrogen oxides (NO x ), CO, ethane (C 2 H 6 ), propane (C 3 H 8 ), HCHO, acetone (CH 3 COCH 3 ), and acetaldehyde (CH 3 CHO) as primarily emitted species (O'Connor et al., 2014). The isoprene oxidation scheme is the mechanism described by Pöschl et al. (2000), as previously used by Zeng et al. (2008), but with rate coefficients of reactions between OH and isoprene nitrates and between NO and isoprene peroxy radicals updated following Paulot et al. (2009a, b). A diurnal cycle is imposed on isoprene emissions as a function of the solar zenith angle. We adopt a set of monoterpene oxidation reactions initiated by OH, O 3 , and NO 3 , described by Brasseur et al. (1998). Methane mixing ratios are prescribed at the surface with a constant value of 1813 ppbv north of the Equator and a constant value of 1707 ppbv south of the Equator, and are the same for each year. Surface emissions are as specified in Sect. 2.1. Lightning-NO emissions are based on the parameterization of Rind (1992, 1994), as a function of convection. The model uses prescribed sea surface conditions following the Atmospheric Model Intercomparison Project (AMIP) II (http://www-pcmdi.llnl.gov). The chemistry is run in a semi-offline mode; i.e. chemistry feedbacks to meteorology and hydrology are deactivated. Table 2 summarizes key model properties for all models.

TM5
TM5 is a global 3-D CTM driven by meteorological fields from the European Centre for Medium-Range Weather Forecast (ECMWF) ERA-interim re-analysis (Dee et al., 2011) using an update frequency of 3 h. Interpolated fields are used for the interstitial time periods. The version used here is identical to that described by Williams et al. (2014) and uses the modified CB05 chemical mechanism (Williams et al., 2013) for describing the chemistry which occurs in the troposphere, along with online photolysis rates. Details relating to the convection, advection, and deposition processes employed are given by Huijnen et al. (2010). TM5 includes a full description of HO x and NO x reactions, as well as explicitly treating all C1 to C3 organic species in addition to ammonia (NH 3 ), sulfur dioxide (SO 2 ), and dimethyl sulfide (DMS). For this study a horizontal resolution of 3 • × 2 • is chosen adopting 34 vertical layers from the surface up to 0.5 hPa. The isoprene and monoterpene oxidation schemes are based on the mechanisms developed by Yarwood et al. (2005), with modifications to both the oxidation rate of organic hydroperoxide (ROOH) and the production efficiency of HO 2 from the OH-initiated oxidation of isoprene following recommendations by Archibald et al. (2010). Also in TM5, a diurnal cycle is applied to the monthly mean isoprene emissions. Methane emissions are included and the simulated surface concentrations are then nudged towards a latitudinally and monthly varying climatology based on surface observations; a detailed description of the approach is given by Williams et al. (2013). Mean surface methane mixing ratios for the year 2004 are listed in Table 2, and their interannual variation is small.
Here we use the v9-01-03 coupled O 3 -NO x -HO x -VOCaerosol simulation with the Caltech isoprene mechanism (Paulot et al., 2009a, b), which includes 57 transported species (both gas phase and aerosol) and an additional 49 species that interact chemically but are not transported (The detailed chemical mechanism used in this study can be found at http://wiki.seas.harvard.edu/geos-chem/index.php/ New_isoprene_scheme_prelim). As monoterpenes are not included as an explicit chemical species, their emissions are used to produce CO (assumed 20 % molar yield; Hatakeyama et al., 1991) and acetone (assume a 12 % molar yield; Jacob et al., 2002). CH 4 mixing ratios are prescribed annu-  (Morgenstern et al., 2013) Modified CB05 chemical mechanism, 60 species (Williams et al., 2013) tropospheric chemistry, 121 species, 106 transported 150 species, MOZART scheme  Isoprene oxidation mechanism Mainz Isoprene Mechanism (Pöschl et al., 2000) with update OH and NO initiation rates (Paulot et al., 2009a, b) CB05 (Yarwood et al., 2005); Modified HO 2 yields (Archibald et al., 2010) Caltech Isoprene Mechanism (Paulot et al., 2009a, b) MOZART scheme ) * GEOS-Chem version v9-01-03 is used in this study. * * Surface CH 4 mixing ratios shown here are for year 2004; NIWA-UKCA uses the same values for each year and the interannual variation is small in other models.
ally and latitudinally based on the National Oceanic and Atmospheric Administration's Global Monitoring Division (NOAA GMD) surface observations and are listed in Table 2. Interactive chemistry is computed in the troposphere only, with stratosphere production and loss rates for most species taken from the NASA Global Modeling Initiative (GMI) Combo CTM Aura4 model (Murray et al., 2013). Stratospheric ozone is simulated using the Linoz linearized ozone scheme (McLinden et al., 2000). Lightning-NO emissions are based on the cloud top height parameterization of Price and Rind (1992) with regional correction to match lightning distributions from satellite, as described by Murray et al. (2012). Other processes in GEOS-Chem v9-01-03 including mixing and deposition are described in detail by Mao et al. (2010Mao et al. ( , 2013a. The version used here was modified from the standard v9-01-03 to include irreversible HO 2 uptake by aerosols with a gamma value of 0.2 (Mao et al., 2013b), and to include methanol as an interactive tracer based on the offline simulation of Millet et al. (2008). In the standard GEOS-Chem simulation, biogenic emissions are computed online using a coupled version of the MEGAN model. Here, to ensure consistency with the other SHMIP models, we used the pre-computed biogenic emissions described above (Sect. 2.1) and applied an imposed diurnal variability tied to solar zenith angle.

CAM-chem
CAM-chem is a component of the NCAR Community Earth System Model (CESM). The version used for this study is the same as that used for the Chemistry-Climate Model Initiative (CCMI) , and very similar to the CAM4 version described in Tilmes et al. (2015). For SHMIP, CAM-chem was run in the specified-dynamics mode, us-ing meteorological fields from the Modern-Era Retrospective Analysis For Research And Applications (MERRA) reanalysis product (http://gmao.gsfc.nasa.gov/merra/), regridded to the model horizontal resolution of 1.9 • × 2.5 • , using the lowest 56 levels. In this study, the internally derived meteorological fields (temperature, winds, surface heat, and water fluxes) are nudged at every time step (30 min) by 1 % towards the reanalysis fields (i.e., a 50 h Newtonian relaxation time). The chemical mechanism, based on MOZART-4 , includes both tropospheric and stratospheric chemistry, with 150 compounds and 400 photolysis and kinetic reactions, and a tropospheric bulk aerosol model (Lamarque et al., 2012). Heterogeneous reactions on aerosols are included as described in Emmons et al. (2010), including the uptake of HO 2 with a reaction probability of 0.2 producing H 2 O 2 . While the option of running with online MEGANv2.1 biogenic emissions is available in CAM-chem, this was not used and all surface emissions were based on those specified for this intercomparison, with diurnal variation imposed for isoprene and monoterpenes. Methane surface mixing ratios are specified for monthly zonal averages, as used for CCMI, based on RCP6.0 (Meinshausen et al., 2011). Lightning-NO emissions are determined according to the cloud height parameterization of Price and Rind (1992) and Price et al. (1997). The vertical distribution follows De-Caria et al. (2006) and the strengths of intra-cloud and cloudground strikes are assumed equal, as recommended by Ridley et al. (2005).

Observations of CO and HCHO in the SH
Long-term measurements of trace gases up to the upper troposphere in the remote SH are sparse. Continuous multi-year tropospheric columns of CO are observed at four SH sites: Darwin, Wollongong, Lauder, and Arrival Heights, with high spectral resolution FTIR spectrometers. In addition, HCHO columns have been retrieved at Wollongong and Lauder. The data records and retrieval methods have largely been presented before (Rinsland et al., 1998(Rinsland et al., , 2002Jones et al., 2001Jones et al., , 2009Paton-Walsh et al., 2010;de Laat et al., 2006;Zeng et al., 2012;Morgenstern et al., 2012) and therefore we only give a brief description here.
At Wollongong, Lauder, and Arrival Heights, mid-infrared (MIR) spectra from the FTIR measurements are used to retrieve CO columns, and these stations are part of the Network for the Detection of Atmospheric Composition Change (NDACC; http://www.ndacc.org). The retrieval of trace gas information from these recorded spectra was performed based on the SFIT2 profile retrieval algorithm using the 4.7 µm band, and is similar to that described by Rinsland et al. (1998Rinsland et al. ( , 2002. At Arrival Heights, there are no measurements during the polar nights which last 4-5 months per year. CO total columns have been observed at Darwin since 2005 with solar remote sensing using FTIR measurements in the near infrared (NIR), as part of the Total Column Carbon Observing Network (TCCON) Wunch et al., 2011). The spectra used for CO retrieval are analysed with the GFIT spectral fitting algorithm (Washenfelder et al., 2006) for total column CO. Details of the retrieval method and data are described in Paton-Walsh et al. (2010). Daily averaged time series of CO columns from 2004 to 2008 are used for comparison with the models. Due to the very small "smoothing" error for CO retrievals, which indicates the difference between realistic and retrieved CO columns, averaging kernels are not applied when comparing with the modelled CO data (de Laat et al., 2006;Zeng et al., 2012). Comparisons are made against daily-mean output from each model.
Total columns of HCHO were retrieved at Wollongong and Lauder from the mid-infrared spectra using the SFIT2 inversion algorithm (Jones et al., 2009). HCHO is a very weak absorber in the mid-infrared spectral region. Due to its large "smoothing" errors, the averaging kernels and a priori applied in the retrieval were also applied to the modelled data for a like-with-like comparison between the modelled and retrieved HCHO columns following the method described by Zeng et al. (2012) and references therein.
In order to provide comparisons on a larger spatial scale, we also perform multi-year comparisons for surface CO against flask measurements available from the NOAA Global Monitoring Division network (Novelli et al., 1998). The selected sites are all situated in the SH and cover an extensive latitudinal range. They are typically located away from regions which exhibit strong local emissions of CO. The sites shown are Mahe Island (4.7 • S, 55.5 • E), Ascension Island all sites used in this study are displayed in Fig. S1 in the Supplement.
To illustrate how the models perform on the global scale in general, we also show comparisons between modelled CO and observed CO columns made by the Measurements Of Pollution in the Troposphere (MOPITT) satellite instrument (https://www2.acd.ucar.edu/mopitt). We use the MO-PITT version 6 level 3 thermal-infrared product, and the data are monthly averages. A description of the data and the retrieval method is given by Deeter et al. (2003Deeter et al. ( , 2014. Here the daytime monthly mean MOPITT CO columns for January and September 2005 are used for comparison. Model outputs are monthly averaged and have been interpolated to the MOPITT horizontal grid of 1 • × 1 • and 10 vertical levels with a 100 hPa spacing. The MOPITT CO averaging kernel and a priori data are applied to the calculation of the modelled CO columns, as described by Morgenstern et al. (2012) and the references therein; such an approach is generally recommended when comparing modelled data to data from satellite remote-sensing instruments (Rodgers et al., 2003). Figure 3 shows the direct comparison between modelled and FTIR daily-mean CO columns for the CLM-MEGANv2.1 simulation. Here, we use the tropospheric FTIR partial columns, for the reasons that not all models have wellresolved stratospheric chemistry in the model, there is a significant contribution of CO from the mesosphere during polar spring (Velazco et al., 2007), and that all models lack or have deficiencies in handling the mesospheric chemistry. FTIR CO partial columns (0-12 km) are used for comparison at Arrival Heights, Lauder, and Wollongong instead of total columns. However, the partial columns of CO at Darwin are not available so we use total columns for comparison. Note that the contribution of mesospheric CO to the total column is expected to be minimal at Darwin given its tropical location (Velazco et al., 2007); therefore, the differences between the partial and the total columns are expected to be small. The model data at all locations have been interpolated to the dates when the measurements were carried out. Figure 3 shows that CO seasonal cycles are well reproduced by all four models at all locations. Models accurately reproduce the total columns of CO at Darwin with very small inter-model differences. The Darwin measurement site is the closest to the tropical source regions; this indicates that the emissions in this area are well represented in the models. Inter-model differences are notably larger at other sites which are located further from the source regions, with consistent overestimation by TM5 and underestimation by CAM-chem at both Arrival Heights and Lauder. Such differences are possibly associated with both differences in the oxidative capacities in these two models and differences in transport (discussed in Sect. 4). All models underestimate CO columns at Wollongong, especially during the peak BB season; this may be due to its proximity to large forested areas and/or the cities of Sydney and Wollongong whose direct emissions may be underestimated in the MACCity inventory. Note that due to the coastal location of Wollongong, model grid boxes may not be representative of the measurement site.

FTIR CO columns
We performed a second set of simulations using LPJ-GUESS isoprene and monoterpene emissions (see Fig. S2 in the Supplement); the models visibly underestimate the observed FTIR tropospheric CO columns at all sites. The deviation of model ensemble-mean CO columns from the observed FTIR columns are shown at the four measurement sites (Fig. 4) in comparison with the simulation using CLM-MEGANv2.1 biogenic emissions. The differences between these two simulations are also shown (i.e. CLM-MEGANv2.1 minus LPJ-GUESS). It appears that a larger negative bias exists when adopting the LPJ-GUESS emissions for all of the column measurement stations (i.e. CLM-MEGANv2.1 results in better agreement with the FTIR observations). The deviations of both simulations from the ob-  served CO columns exhibit large seasonal variations but seasonal and interannual variations are consistent between these two simulations. Figure 5 shows differences between the modelled and observed FTIR CO columns at the four measurement sites from the multi-annual ensemble-mean data for both CLM-MEGANv2.1 and LPJ-GUESS simulations. As in Fig. 4, the seasonal variations of the biases from these two sets of simulations follow a very similar pattern, implying that the effect of different biogenic emissions is reflected in the differences in the background CO columns in the SH. The biases shown in the ensemble model means from both simulations are largest during the SH tropical BB season of September,  October, and November (SON), although at Darwin the negative biases are also high in July and increase from October to December. For Wollongong, Lauder, and Arrival Heights the largest negative biases are in October, November, and December, respectively; this suggests an underestimation of SH BB sources in GFEDv3 and the subsequent effect on CO columns at SH remote locations through long-range transport. At Darwin, CO columns are more likely influenced by local or nearby BB sources which may have a different seasonality. The annually averaged biases of the model ensemble means for each site are shown in Table 3; the lowest biases are at Arrival Heights for both simulations, followed by those for Darwin, Lauder, and Wollongong. Note that the low ensemble bias at Arrival Heights is largely the result of cancellation of a positive bias in TM5 with a negative bias from CAM-chem with a similar magnitude. The large spread between the models indicates that substantial differences exist in other physical and/or chemical processes which are unrelated to emissions. The individual model biases are also shown in Fig. 5. For both simulations, inter-model variability is notably larger during months that lie outside the seasons when most intensive BB occurs, i.e. typically in austral summer and autumn (covering December and January to May). Such a seasonal dependence of inter-model variability is consistent with that described by Fisher et al. (2015) who compare modelled vertical CO gradients in the Cape Grim region using the same simulations, and is due to the difference in chemistry that controls CO chemical production and loss processes in the seasons other than the peak biomass burning season. Intermodel variability is generally larger in CLM-MEGANv2.1 than in LPJ-GUESS for all seasons and locations, primarily due to the larger response of modelled CO to its higher precursor emissions.

MOPITT CO columns
MOPITT CO columns and the individual model biases for January and September 2005 are shown in Fig. S3 in the Supplement. The model data are monthly means convolved with the MOPITT averaging kernels and a priori data. MOPITT data exhibit a lot of gaps over the Amazon region in January and over Africa in September, due to persistent cloud cover. All models underestimate MOPITT CO columns in the NH (with the exception of East Asia), and overestimate CO in the plumes, particularly over the tropical Atlantic in January and the Pacific in September. These plumes originate in tropical Africa and South America, respectively, indicating an overestimation of biomass burning emissions in these regions.
Comparing the models, TM5 shows higher CO columns throughout the SH and CAM-chem the lowest, in agreement with the comparison of FTIR CO columns. Figure 6 shows the percentage differences between the ensemble-mean modelled and measured columns for both the CLM-MEGANv2.1 and the LPJ-GUESS simulations. There is a general underestimation of CO columns by both ensembles in the NH by up to ∼ −25 %. Both ensembles overestimate CO in the source regions, with up to +30 % over tropical Africa in January and over Amazonia in September, i.e. during the months of peak biomass burning. In the SH, away from the CO plumes, the CLM-MEGANv2.1 ensemble clearly compares better with MOPITT CO than the LPJ-GUESS ensemble, with biases typically between −10 and +10 % in January and Septem-  ber, whereas errors typically are in the range −20 to −5 % from the LPJ-GUESS ensemble. Both ensembles also underestimate CO columns over Australia in September, which suggests an underestimation of biomass burning in GFEDv3. This is also reflected in the comparison between modelled and FTIR CO columns in the four SH locations shown above, which generally show negative biases in modelled CO. The two ensembles are fairly similar in the NH with regard to their CO columns but exhibit significant differences in the extratropical SH. This is consistent with a larger relative role of biogenic emissions in the SH versus the NH.

Surface CO
To assess the models' ability to capture both the seasonality and interannual variability of CO at the surface over the simulation period, we show in Fig. 7 comparisons between the CLM-MEGANv2.1 simulations and monthly mean CO values observed at the eight surface sites listed in Sect. 2.3. Consistent with the FTIR column comparisons, all models capture the seasonal cycles of observed surface CO at each location. In line with Fig. 3 in Fisher et al. (2015), TM5 typically exhibits a high bias and CAM-chem exhibits a low bias of the order of 5-10 ppbv. Large variations exist in seasonal cycles at both Mahe Island and Ascension Island, but the timing of the peaks are different. At Mahe Island, surface CO peaks in January and February due to the influence of anthropogenic emissions from India (Wai et al., 2014), whereas at Ascension Island, the seasonal cycle is principally driven by CO which originates from BB in southern Africa during June-July-August (e.g. Williams et al., 2012;Wai et al., 2014). The interannual variability and timing in peak mixing ratios is not captured well at Ascension Island, especially for GEOS-Chem and TM5; this is possibly related to too strong westerly transport out of southern Africa and too weak an oxidative capacity, especially in TM5. For the more southerly oceanic sites, the seasonal cycles and amplitudes are remarkably similar, indicating that the variability in background CO is rather low at the surface in the SH remote locations. In general, NIWA-UKCA and GEOS-Chem display a better agreement with the observations in the remote SH, indicating that their oxidative capacities are more realistic. The consistent high and low biases in TM5 and CAM-chem, respectively, are related to the oxidizing capacity in these models; this is discussed in Sect. 4.
We quantify the differences between the multi-annual ensemble means for surface CO and the corresponding values derived from the observations for both the CLM-MEGANv2.1 and the LPJ-GUESS simulations (Fig. 8). As was seen in the model comparisons to the FTIR and MO-PITT CO data, the observed distributions of surface CO in the SH are better reproduced by CLM-MEGANv2.1 for most of the chosen sites. A comparison of sites shows that the seasonal biases are more variable for the tropical sites which are affected by the interannual variability in tropical BB. For the mid-to high-latitude sites, the CLM-MEGANv2.1 ensemble mean accurately reproduces the observations in most cases, whereas the LPJ-GUESS ensemble is consistently biased low. The individual model biases (shown only for the CLM-MEGANv2.1 simulation) are up to ±20 %, and are much larger than the differences between the two ensemble means (∼10 %). The generally better agreement between modelled and observed surface CO, relative to the agreement between modelled and FTIR CO columns in the remote SH, reflects that there may be some deficiencies in the models' vertical transport of either CO and/or its precursors. This generally underestimation of observed vertical gradients of CO by the models in the remote SH was shown by Fisher et al. (2015) for the Cape Grim region.

HCHO columns
Here we examine the models' ability to reproduce observed HCHO columns at the SH mid-latitude sites Lauder and Wollongong. Figure 9 shows comparisons between modelled daily-mean HCHO columns (from the CLM-MEGANv2.1 simulation) convolved with FTIR a priori data and averaging kernels, and observed daily-mean HCHO columns from the FTIR measurements. The seasonal cycles are generally well reproduced across the entire model ensemble, with the seasonal maxima in austral summer and the minima in winter, but all models significantly underestimate observed columns in all seasons. Inter-model differences in modelled HCHO columns are larger at Lauder than at Wollongong; the highest HCHO columns are produced in GEOS-Chem, whereas the lowest are from TM5. Such variations between the models indicate that the differences in the models' chemistry are the driving factor, in particular at the sites that are further away from the emission sources. Significant and persistent low bias across all models cannot be reconciled by considering the diurnal cycle in HCHO; for testing purposes, we also calculated HCHO columns by replacing daily-mean HCHO data shown in Fig. 9 with the daily maximum of the 3 hourly data from one of the ensemble members (CAM-chem). This resulted in small overall changes, with ∼ 10-15 % increases that occur in some summer months, and the increases were not sufficient to close the gap between the models and the observations. Therefore, we are confident that using dailymean modelled HCHO columns for comparing to columns from FTIR observations that occur during the daylight is satisfactory. Figure 10 shows the multi-annual monthly mean FTIR HCHO columns and model ensemble means averaged for the same years with both CLM-MEGANv2.1 and LPJ-GUESS emissions for isoprene and monoterpenes. Overall, the models underestimate the observed HCHO columns by approximately 50 %. Differences in biogenic emissions do not appreciably affect this.
In the case of Wollongong, proximity to Sydney and the influence of episodic BB events in the vicinity (Williamson et al., 2013) could introduce local direct and indirect sources of HCHO and chemical precursors which are unaccounted for and might have contributed to the low bias simulated in the models, particularly for the seasonal peaks. However, at Lauder there are no known significant local sources of HCHO. We therefore assume that the underestimation of observed FTIR HCHO columns by the models is very likely related to missing emissions of precursors.
The underestimation of measured HCHO by the models at the remote SH locations had been shown in some previous studies, and in those studies various assumptions about missing processes have been explored (e.g. Ayers et al., 1997;Jones et al., 2009;Vigouroux et al., 2009). Ayers et al. (1997) used a box model to simulate the measured surface HCHO at Cape Grim and were unable to capture the magnitude of the observed mixing ratios of HCHO by including a set of standard methane oxidation reactions in the model. Among the major HCHO production channels, Ayers et al. (1997) assumed a 100 % yield of CH 3 OOH from CH 3 O 2 + HO 2 . Ayers et al. (1997) then experimented with an alternative oxidation pathway that involved the direct production of HCHO (40 %) from CH 3 O 2 + HO 2 , which resulted in a much im-proved comparison. We have not applied such high direct yield of HCHO in our models. However, following the recent recommendation of the International Union of Pure and Applied Chemistry (IUPAC) (Atkinson et al., 2006), a 10 % direct yield of HCHO has been adopted by NIWA-UKCA but no direct yield has been applied in the other three models. The recent IUPAC recommendations (Atkinson et al., 2006) assume a temperature-dependent branching ratio for the direct HCHO production channel (i.e. 0.09 to 0.29 for temperatures ranging from 298 to 218 K). Adopting this recommendation, an additional test was performed in TM5, showing some modest increases in HCHO in the extratropics of up to ∼ 10 %. However, this is not sufficient to explain the large bias shown here. Another hypothesis suggested by Ayers et al. (1997) is the possibility of a small marine biological source of isoprene (e.g. Bonsang et al., 1992). Recently, Lawson et al. (2015) found relatively abundant HCHO precursors (dicarbonyls) in two regions of the southwest Pacific, corroborating the hypothesis that marine biological activity might be responsible for the measured HCHO abundance. However, spatial sampling and understanding of the underlying biological processes remain poor.
The HCHO column data set we use here is an extension of the 1992-2005 data set described by Jones et al. (2009), retrieved using the same algorithm. They also derived HCHO mixing ratios at a coarse vertical resolution. Jones et al. (2009) performed a box model simulation based on subsets of the Master Chemical Mechanism (MCM) (Saunders et al., 2003) including the isoprene oxidation scheme of the MCM. They found that high-HCHO mixing ratios retrieved at Lauder cannot be explained by methane oxidation alone and that additional local sources, possibly isoprene, are needed to explain the observed near-surface HCHO mixing ratios at Lauder. A recent study by Vigouroux et al. (2009) compared modelled and observed FTIR HCHO columns at Réunion Island, using the global chemical transport model IMAGESv2; they also underestimate the observed FTIR HCHO columns albeit with a smaller magnitude than that shown in this study. The time series shown by Vigouroux et al. (2009) are for August to November 2004 and for May to November 2007, respectively, and the differences between modelled and observed HCHO columns are around 30 and 25 %, respectively. The isoprene mechanism used by Vigouroux et al. (2009) is based on the MCM and is described by (Müller and Stavrakou, 2005). The isoprene emissions used by Vigouroux et al. (2009) are from the MEGAN-ECMWF inventory (Müller et al., 2008), and the yearly totals averaged over 2004 to 2006 are around 10 % lower than the CLM-MEGANv2.1 inventory used here. To investigate the possible causes for the low bias in modelled HCHO, Vigouroux et al. (2009) include methane oxidation by tropospheric chlorine, but the impact of this process on HCHO columns is only about 1-2 % and therefore cannot explain the underestimation. They also experimented using a different OH climatology; this increase of OH abundance results in better agreement between observed and modelled HCHO columns but cannot fully reconcile the substantial differences, and a more probable explanation is an underestimation of HCHO precursors transported from Madagascar to Réunion Island. This finding, together with our finding here, suggests that the underestimation of HCHO columns is persistent throughout the SH. Observations of HCHO in the remote SH regions are extremely sparse, and it is impossible to fully constrain modelled HCHO. Note that in both studies (Jones et al., 2009;Vigouroux et al., 2009), FTIR HCHO columns compare well with satellite measurements, and with both satellite and MAX-DOAS measurements, respectively. This again suggests that the FTIR HCHO retrieval is robust at all sites, and that the likely cause for model-observation differences is missing sources of HCHO and/or its precursors in the models.

Model differences in chemistry and transport
Although the four models are constrained by the same emissions, there are significant differences in the models' abilities to reproduce observed CO columns and surface CO in the remote SH, as shown above. Here we explore the underlying factors driving these differences. To diagnose the extent of differences in transport between the models, we ex- amine the two passive CO tracers defined in Sect. 2: one with a fixed lifetime of 25 days (referred to as CO 25 ) and the second with first-order loss via model calculated OH (referred to as CO OH ). Both tracers are subjected to the same surface emissions as the full simulations, but not subjected to any secondary production of CO from methane and NMVOC oxidation. Dry deposition of CO is not included for either of the additional CO tracers as it is considered a minor loss channel for the SH. The global tropospheric CO columns from all models for January and September are shown in Fig. 11. January and September represent the seasonal maxima of biogenic and biomass burning emissions in the SH, respectively. Here, we define the tropospheric columns as the columns below the chemical tropopause marked by the 150 ppbv O 3 isopleth in each model (monthly mean O 3 used here is averaged over [2004][2005][2006][2007][2008]. Although here we focus on the SH, we note that the inter-model differences apparent in the SH are consistent with those occurring in the NH, namely, the lowest CO columns occur in CAM-chem, followed by NIWA-UKCA, with higher CO columns from GEOS-Chem and TM5 for both hemispheres, indicating systematic differences between the models. Comparing the seasonal variations, CO columns are generally higher in September than in January in the SH, primarily due to the timing of the most intensive tropical BB events in austral spring. Of the four models, CAM-chem simulates the lowest CO columns in both the source regions and in the remote mid-to high latitudes. Examining the distributions of the tropospheric columns of CO 25 shown in Fig. 12, CO 25 exhibits similar distributions among the models for both seasons in source regions as those shown in Fig. 11. The differences become more obvious in the extratropics, with NIWA-UKCA showing slightly weaker transport towards the poles, whereas GEOS-Chem shows somewhat stronger export of CO 25 out of the source regions and towards the poles. Overall, despite some differences, the magnitude and distribution of CO 25 are very similar among the models. However, such differences and similarities in transport among the models are not reflected in the differences in CO columns shown in Fig. 11 in which TM5 simulates the highest CO columns and CAM-chem the lowest in both the source and remote regions. CO OH (Fig. S4 in the Supplement) is a more realistic proxy of CO, reflecting the influences of the models' variable OH concentrations. Like CO 25 , the magnitudes and distributions of CO OH are similar across the models; hence, the main driver of the model differences in total CO cannot be attributed to primary emissions.
Next, we quantify the roles of transport and chemistry in determining the inter-model variability in CO columns in the SH. We examine three zonal bands defined as 0-30, 30-60, and 60-90 • S. These latitude bands capture the main tropical source regions and mid-and high latitudes, respectively. Figure 13 shows the monthly mean tropospheric columns of CO as well as ratios of CO 25 /CO and CO OH /CO columns, averaged across each of these zones for each model. CO sec = CO − CO OH is an estimate of the fraction of CO that is produced by oxidation of CH 4 and NMVOCs; the ratio of CO sec to CO is also shown in Fig. 13. These ratios define the contributions of CO 25 , CO OH , and CO sec to the total CO columns in each model. Figure 13 shows that CO columns decrease towards the high latitudes and the seasonal maxima are during the September/October BB season in all zones. Although CO 25 is an idealized tracer designed to diagnose differences in the long-range transport simulated in each model, CO OH should be a more realistic measure of how much primary emissions of CO contribute to the CO columns because CO OH reflects the locally varying lifetime of CO due to the spatial variability of OH. The ratio of CO 25 to CO drops sharply from the tropics to the pole for all models (from ∼ 20 to ∼ 5 % in the annual average), as would be expected from the hemispheric distribution of emissions and the timescales for meridional transport. By contrast CO OH /CO reduces only from ∼ 30 to ∼ 25 % in the yearly average. This reflects that the lifetime of CO is considerably longer outside of the source region due to lower background O 3 levels (and therefore lower OH levels) in the more pristine environment away from strong NO x sources. CO sec /CO shows a moderate increase from 70 to 75 % from the tropical zone to the high latitudes. Overall, primarily emitted CO makes up ∼ 25-45 % of total tropospheric CO in the source region and ∼ 20-40 % in the polar region, depending on season, while the secondary CO makes up the remainder of the tropospheric columns, i.e. ∼ 55-75 % in the source region and ∼ 60-80 % in the polar region. Regarding seasonal variability, CO 25 and CO OH have proportionally larger contributions in austral spring when BB dominates the CO emissions, whereas CO sec shows larger contributions in austral summer/autumn. Of all the models, NIWA-UKCA shows the smallest contribution from primary CO to the columns and the largest contribution from the secondary CO, relative to the other three models.
Inter-model differences in CO columns and the additional CO tracers are expressed as the ratio of individual model columns vs. the multi-model mean columns for each zone, shown in Fig. 14. For CO columns, the inter-model differences are smallest in the tropical zone and gradually increase towards the pole, with the highest CO columns from TM5 and the lowest from CAM-chem, in agreement with the FTIR comparisons and the surface comparisons shown earlier. Examining the inter-model differences in CO 25 , the model spread increases substantially towards the polar zone, and is characterized by the strongest transport out of the source region from GEOS-Chem and the weakest from NIWA-UKCA (also shown in Fig. 12). Note that this behaviour is not reflected in the model spread of CO columns (i.e. the highest CO occurs in TM5 and the lowest in CAM-chem). By contrast, the patterns of model spread in CO sec and to a lesser degree in CO OH are consistent with that seen in the CO columns, indicating that the inter-model differences in modelled CO columns are strongly influenced by the differences in CO sec , which is dependent on the oxidizing capacity in the model that also drives the loss of primary-emitted CO by OH. Considering also the absolute contributions of both primary CO sources and secondary CO production to the SH CO columns (these being ∼ 35 and ∼ 65 %, respectively), we can deduce that inter-model differences in CO columns attribute about one-third to primary and two-thirds to the secondary CO production in the SH. Note that here we only take into account the accumulated effects of primary and secondary contributions to CO; we do not differentiate or individually identify the separate influences, e.g. of transport and chemistry. For example, the large CO columns in TM5 can be the result of combined effects of slower chemical loss of CO due to lower OH levels in the model and a faster secondary CO production in the source region, as reflected in higher ratios of CO sec to CO shown in Fig. 15. In contrast, GEOS-Chem CO has faster loss by OH than TM5 (but slower than the others), but this is outweighed by a stronger transport resulting in higher CO compared to that in NIWA-UKCA and in CAM-chem. For CAM-chem, moderately slow transport of CO out of the source region combined with slower secondary CO production results in the lowest CO columns. More quantitative analyses of differences in chemistry are carried out in Sect. 5. To further probe the differences between the models, we show vertical profiles of modelled key species, namely CO, HCHO, O 3 , and OH mixing ratios from each model in Fig. 15. We display data for January 2005 from the CLM-MEGANv2.1 simulation, because in austral summer the chemical production maximizes due to stronger photochemistry and higher biogenic emissions, and absolute inter-model differences in CO columns are also larger than in other seasons. TM5 is characterized by consistently high CO throughout the SH. The CO values in NIWA-UKCA and in GEOS-Chem are very close in all three zones, exhibiting differences of ∼ 5-10 %, although CO in NIWA-UKCA is slightly higher than that in GEOS-Chem in the tropics but becomes lower towards remote regions. This may reflect slower meridional transport in NIWA-UKCA (shown in CO 25 ) combined with larger chemical production in the source region. The HCHO mixing ratios decrease sharply with altitude due to the dominant chemical precursors residing in the boundary layer and the efficient photo-dissociation, but the vertical gradient becomes smaller away from the source region, particularly in TM5, due to depletion of the biogenic precursor emissions in the remote SH. HCHO abundances in the four models correlate with OH to some extent; i.e. both OH and HCHO are relatively large in GEOS-Chem, whereas both are relatively small in TM5; this reflects the approximate linear-ity between the modelled HCHO abundance and methane oxidation via OH in the remote SH. However, there is no simple linear relationship between HCHO and OH; OH is involved in both the loss and the production of HCHO, and HCHO is one of the OH sources. The modelled OH profiles do not seem to be closely related to O 3 (the primary source of OH) in that TM5 has the lowest OH but its O 3 values lie in the middle of the model range; this is likely due to differences in photolysis schemes. Water vapour fields are very similar among the models.

Analysis of chemical production and loss rates
To quantify the effects of differences in model chemistry, we analyse chemical production (CP) and chemical loss (CL) rates of CO and of HCHO, as listed in Tables 4 and 5 for both simulations (i.e. with CLM-MEGANv2.1 and LPJ-GUESS emissions, respectively). The budget terms displayed are for year 2004 and for the whole globe, the SH, and the three SH latitudinal bands defined above. The corresponding burdens of CO, HCHO, and OH are shown in Table 6. We define the tropopause of each model as the 150 ppbv O 3 isopleth in each model, as in Sect. 4.
Examining the CO budget terms shown in Table 4 main contribution to the SH CPs come from the 0-30 • S latitude band; production decreases sharply towards the southern polar region. In general, chemical production and loss rates of CO are larger in the CLM-MEGANv2.1 simulations, indicating larger biogenic emissions leading to larger CO production. However, the CO production from methane oxidation is generally larger in LPJ-GUESS as a result of increased OH (shown in Table 6) due to the reduction of biogenic emissions. In all models, the ratio of CP of CO to surface emissions is markedly larger in the SH than the global values; this results in larger inter-model differences in the SH due to the differences in the models' underlying chemistry. NIWA-UKCA shows the largest total CPs for all domains, followed by GEOS-Chem, TM5, and CAM-chem, and the differences in total CPs are dominated by the differences in the oxidation of NMVOCs. Methane oxidation is more constrained among the models, and the differences in methane oxidation are mainly driven by differences in OH (shown in Table 6). Note that we do not calculate CO production rates from NMVOC oxidation explicitly in the models; instead they are deduced from the total CP and methane oxidation terms, assuming a 100 % yield of CO from methane oxidation; this is only for diagnostic purposes and we do not make such assumptions in the actual mechanisms. Dry deposition of CO is a small loss term, particularly for the SH. TM5 and NIWA-UKCA have comparable dry deposition loss rates. Note that CO loss through dry deposition is not included in GEOS-Chem and is not provided for CAM-chem. The seasonal variation of CP and CL of CO in the SH are shown in Fig. 16. The surface emission of CO peaks in September in the SH which is dominated by biomass burning emissions. CPs and CLs maximise in austral summer and minimise in austral winter. CAM-chem shows much lower total CP and NMVOC oxidation compared to the other three models, in particular during the summer months, indicating below-average oxidizing capacity in that model. Examining the contribution to the total CPs, methane oxidation and oxidation from NMVOCs are nearly equal in all models except CAM-chem where NMVOC oxidation is significantly lower than methane oxidation. Methane oxidation is largest in GEOS-chem, followed by NIWA-UKCA reflecting the higher OH in these two models. The peak chemical loss shown in all four models in October is in response to the peak of surface emissions of CO. We also display the ratio of CO production from NMVOCs to the total CP, showing that TM5 has the highest ratio, indicating a fast conversion of NMVOCs to CO in TM5. In comparison, CAM-chem has a substantially lower NMVOC oxidation to total CP ratio, indicating a slower NMVOC to CO conversion; this is the primary cause for low CO in this model. We have not explicitly quantified the CO production from isoprene oxidation but assume that isoprene oxidation is the dominant contributor to NMVOC oxidation in the SH (Pfister et al., 2008). We therefore suggest that the different isoprene oxidation schemes used in the models are responsible for the differences in the chemical production rates of CO. Without a detailed comparison of the chemical mechanisms used in the models, we cannot identify which processes and/or parameters that make up the mechanisms are responsible for the differences in the models employed here, and such tests would be more suitably done in a box model in which parameters can be more straightforwardly controlled (e.g. Archibald et al., 2010). Four different isoprene oxidation mechanisms are included in the models presented here. They vary in complexity and also in the approaches to treat degradation products. The isoprene oxidation mechanism in NIWA-UKCA is based on a smaller mechanism (Mainz Isoprene Mechanism (MIM); Pöschl et al., 2000) than those used in GEOS-Chem (Paulot et al., 2009a, b) and in CAM-chem . NIWA-UKCA contains some recently updated rate coefficients of reactions between NO and peroxy radicals from the OH-initiated isoprene oxidation reactions, and reactions between OH and isoprene nitrate (Paulot et al., 2009a, b). The isoprene oxidation scheme in TM5 is based on the CB05 chemical mechanism (Yarwood et al., 2005) with modifications made to both the oxidation rates of peroxides and the production efficiency of HO 2 from the OHinitiated oxidation of isoprene based on recommendations by Archibald et al. (2010). Our results here show that the rates of NMVOC oxidation are substantially faster in TM5 (shown in Table 4 and Fig. 16) than in the other models, and such faster NMVOC oxidation rates are largely driven by the isoprene oxidation scheme in that model, which, together with the lower OH (shown in Table 6), lead to higher CO than in the other models. M: CLM-MEGANv2.1 emissions; G: LPJ-GUESS emissions; CP: chemical production; units in Tg yr −1 ; * A conversion factor of 1.0 from methane oxidation is assumed here for diagnostic purposes; * * NMVOC oxidation is derived from total chemical production and methane oxidation, i.e. CP NMVOCs = CP Total − CP CH 4 .
HCHO budget terms from NIWA-UKCA and TM5 are listed in Table 5 (These terms were not saved in the other models). The surface emissions of HCHO are small compared to the in situ chemical production and loss terms. The global total CP in NIWA-UKCA is slightly larger than in TM5 for both simulations, but the amounts are comparable for the SH. Methane oxidation rates are higher in NIWA-UKCA for all regions due to the higher OH in that model, and NMVOC oxidation rates are significantly larger in TM5. Examining the chemical losses, HCHO loss through the reaction with OH is much higher in NIWA-UKCA; however, HCHO losses through photolysis are comparable between these two models. Together with the smaller burden of HCHO in TM5, this implies that HCHO photolysis rates are larger in TM5 than in NIWA-UKCA. (This diagnostic is not directly available for either model.) The much larger wet deposition of HCHO in TM5 (i.e., ∼ 10 % of the to-tal loss terms), compared to that in NIWA-UKCA (∼ 3 %), could explain the lower HCHO burden/columns in TM5. An additional hydration of HCHO is applied in TM5 (but not in the other models), which further enhances the effective solubility of HCHO in aqueous solution (Huijnen et al., 2010). This may have resulted in an additional loss of HCHO to wet deposition in TM5 which is however still substantially smaller than the gas-phase loss processes. The Henry's law coefficients, governing gas-and liquid-phase partitioning of HCHO, applied in the other models are comparable.   ( Table 3). Here we further quantify the changes in CO and HCHO columns in response to changes in biogenic emissions at the hemispheric scale, and also highlight associated changes in the corresponding OH columns. Figures 17-19 display the monthly mean global distributions of relative differences in CO, HCHO, and OH columns between simulations using CLM-MEGANv2.1 and LPJ-GUESS, respectively, for January and July (averaged over 2004-2008). The differences calculated for all species are expressed as the percentage change relative to the CLM-MEGANv2.1 simulation. Here, we show results for the January and July months in order to contrast the seasonal features in oxidizing capacity. For all models, applying LPJ-GUESS emissions results in significant decreases in CO columns throughout the SH, with the largest decreases in the South American and Australian source regions (Fig. 17), in response to the smaller emission fluxes of isoprene and monoterpenes from LPJ-GUESS (the accumulated peak isoprene emissions in CLM-MEGANv2.1 are 25 % higher than in LPJ-GUESS during the peak season of the austral summer months, shown in Fig. 1, and the biggest differences are in South America). Away from these source regions, the differences are largely homogeneous in the mid-to high latitudes. The models' responses to changes in biogenic emissions vary considerably, with TM5 having the largest sensitivity of CO columns change to changes in biogenic emissions, namely ∼ 35 % in January and ∼ 25 % in July in the source regions and 10-15 % over the remote SH. GEOS-Chem has the lowest sensitivity with 15-20 % changes in January and 10-15 % in July in the source regions and less than 10 % in remote regions in response to the same emission changes of isoprene and monoterpenes. For corresponding changes in tropospheric HCHO columns (Fig. 18), substantial decreases (up to ∼ 50-60 %) occur in the source regions of South America and Australia in response to smaller emission fluxes in LPJ-GUESS, relative to CLM-MEGANv2.1. These reductions in HCHO columns propagate to the sub-tropical remote oceans where the magnitude of the decreases is greatly reduced. There are some increases in HCHO columns over southern Africa, which are responses to the higher isoprene emissions in LPJ-GUESS. However, there is a consistent increase of up to 5 % over large areas of the mid-to high latitudes which is apparently not directly caused by reduced biogenic emissions. We find that changes in both CO and HCHO are associated with changes in OH (Fig. 19); the tropospheric OH columns exhibit substantial increases in the source regions as a result of reduced isoprene and monoterpene emissions; qualitatively these effects follow the differences in the geographical distributions of the emissions, and are of opposite sign to both the CO and the HCHO columns changes there. OH increases in remote regions are largely positive, and are opposite in sign to the CO changes; i.e. reduced loss rates of CO cause increases in OH. However, increases in OH columns away from the source regions correlate with HCHO changes; this implies that increases in HCHO in remote regions under LPJ-GUESS emissions are due to strengthened methane oxidation through increases in OH. The inter-model differences in HCHO changes are generally small in remote regions; TM5 shows the largest sensitivity over the source regions in both OH and HCHO, due primarily to the faster isoprene oxidation processes in that model. Note that the large relative differences in both HCHO and OH in July at high latitudes shown in CAM-chem are not significant because the background abundances of both species in the polar region are extremely small.
In Table 7 we summarise hemispheric changes in chemical production and loss rates of tropospheric CO and HCHO, in response to the differences in biogenic emissions. Values expressed are percentage changes (i.e. LPJ-GUESS minus CLM-MEGANv2.1 relative to CLM-MEGANv2.1), and are given for both hemispheres to assess the hemispheric impact of biogenic emissions. In the SH, the changes in all terms are negative, except for the rates of chemical production of both CO and HCHO from methane oxidation; this is generally the result of increased OH in the LPJ-GUESS simulation, in response to reduced biogenic emissions in that inventory. For all models, relative reductions in NMVOC oxidation rates (−17.8 to −35.6 %) are substantially larger than relative increases in CP from methane oxidation (3.2 to 8.0 %), in response to changes in biogenic emissions. Therefore, NMVOC oxidation (mainly of isoprene) is the driving factor for model differences in in situ CO and HCHO production. The burden changes are closely related to the changes in total CP; i.e. TM5 has the largest changes in both bur-  den and the CP, and GEOS-Chem has the smallest terms for both. For all models, relative responses in the SH are much larger than in the NH, emphasizing the importance of biogenic emissions for CO and HCHO formation in the SH.
Complementing the comparison of columns, we here compare the seasonal differences in vertical profiles of CO mixing ratios between CLM-MEGANv2.1 and LPJ-GUESS simulations, averaged zonally and over 2004 to 2008. Figure S5 in the Supplement shows large reductions in CO over the SH tropics in all simulations using LPJ-GUESS emissions of isoprene and monoterpenes, relative to those using CLM-MEGANv2.1, and these reductions propagate to the upper SH tropical troposphere and spread throughout the middle and high latitudes. This shows that the CO column changes in the extratropics are dominated by the changes in the free and upper troposphere, where CO has a relatively long lifetime. Overall, the impact of biogenic emissions on CO are more significant in the SH than the NH. In the SH, throughout the depth of the troposphere, the LPJ-GUESS simulations have reduced CO, which is linked to much reduced CO in the tropics. This effect maximizes during austral winter and spring.

7 Summary and conclusions
We have compared modelled daily-mean CO and HCHO columns from a four-model ensemble with the observed daily-mean FTIR columns of these two species at SH sites including the tropical site Darwin, the mid-latitude sites Wollongong and Lauder, and the Antarctic site Arrival Heights for CO, and Wollongong and Lauder for HCHO. We use CLM-MEGANv2.1 biogenic emissions for the first set of simulations; for these simulations modelled and measured CO are in reasonable agreement, albeit with some low biases, in all models at most locations; annually averaged deviations relative to the observations are −3.2 % at Arrival Heights, −8.6 % at Lauder, −19.2 % at Wollongong, and −6.9 % at Darwin for the four-model mean. The largest discrepancies between modelled and observed CO columns occur at Wollongong which is heavily influenced by local urban and industrial sources and episodic nearby bush fires that are most likely unaccounted for in the emission inventories. Large inter-model differences exist at all locations for all seasons with the exception of austral spring at Darwin where the local biomass burning sources dominate the CO columns. We also compare the modelled surface CO to observations; significant inter-model differences exist although the ensemble mean exhibits good agreement with the observed values for most sites. The inter-model differences for modelled surface CO are markedly larger than the differences between the ensemble mean and observed surface CO. In agreement with previous modelling studies of HCHO in the remote SH (Ayers et al., 1997;Jones et al., 2009;Vigouroux et al., 2009), the models significantly underestimate observed HCHO columns at Wollongong and Lauder by more than a factor of 2, and the largest discrepancy occurs during austral summer. We cannot reconcile such significant differences between the modelled and observed HCHO columns over the remote SH with our current understanding. We hypothesize that missing local sources and/or missing chemical processes are the most likely causes. The fact that model differences are much smaller than the differences between the models and the observations indicate that the cause of such a large discrepancy probably goes beyond what the differences in chemical mechanisms can explain.
To determine the sensitivity of CO and HCHO distributions to biogenic emissions, we perform a second set of simulations with emissions of isoprene and monoterpenes from the LPJ-GUESS data set; results show that the LPJ-GUESS simulations exhibit systematically lower CO columns and lower surface CO than the CLM-MEGANv2.1 simulations, in response to an average of ∼9 % reduction in isoprene emissions globally and a ∼17 % reduction in the SH (monoterpene emissions are also substantially lower in LPJ-GUESS; see Table 1). Annually averaged relative differences between ensemble model mean and observed CO columns are −10.5 % at Arrival Heights, −17.1 % at Lauder, −27.5 % at Wollongong, and −19.9 % at Darwin. The dif-ferences in surface CO at remote monitoring sites between the two simulations are generally smaller than 5 %. At neither Wollongong nor Lauder do we find that differences in biogenic emissions have any significant impact on modelled HCHO columns.
Examining the response of CO and HCHO columns to differences in biogenic emissions of isoprene and monoterpenes on the hemispheric scale, we show that both species exhibit large sensitivity to emissions in the source regions, with 30-40 % reductions in CO and HCHO columns, as a direct consequence of the mainly reduced emissions of isoprene and monoterpenes in the LPJ-GUESS inventory, relative to CLM-MEGANv2.1 (i.e. ∼ 37 % reduction of isoprene emissions in Australia and Indonesia, ∼ 23 % reduction in South America, and ∼ 13 % overall increase in Africa, with both increases and reductions occurring in different regions), and these reductions in CO and HCHO are generally larger in summer than in winter. Away from the source regions and throughout the SH, decreases in CO columns are roughly half those occurring in the source regions, whereas there are moderate increases in HCHO columns (∼ 5 %) despite the significant decreases in and near the source regions for all models. We show that the increases in HCHO columns in the remote SH for LPJ-GUESS, relative to CLM-MEGANv2.1, are linked to the increases in OH columns through enhanced methane oxidation in the remote SH (see Tables 5  and 6). There are substantial increases in OH columns in the source regions in direct response to the reduced isoprene and monoterpene emissions in the LPJ-GUESS inventory, whereas the general increase (up to ∼ 5 % across the models) in the remote regions is the result of reductions in CO and possibly other longer-lived isoprene oxidation products. Significant inter-model differences exist in modelled CO columns; we quantify these differences in three latitudinal regions (SH tropics, mid-, and high latitudes). The ratios of individual model columns to the ensemble-mean columns (annually averaged and averaged across the three regions) are between 0.85 and 1.15 for the tropical region, and the range increases to between 0.7 and 1.2 at high latitudes. Using diagnostic tracers, we assess the impact of modelled transport (by CO 25 ), the contribution from primarily emitted CO (by CO OH ), and CO produced and transported from secondary CO production (CO sec = CO−CO OH ). The results reveal that the differences in transport are not sufficient to explain the differences in modelled CO columns. The modelled range of CO OH corresponds much better to the modelled CO columns than CO 25 but still cannot fully explain the intermodel differences in modelled CO columns. The differences in secondary CO production, i.e. CO sec , however, correspond well with those in modelled CO columns. TM5 exhibits the highest values in both variables, followed by GEOS-Chem, NIWA-UKCA, and CAM-chem in magnitude. We calculate that CO sec contributes around 65 % to CO in the tropics and around 75 % in the polar region in each model, and is responsible for two-thirds of the inter-model differences in mod-elled CO columns overall. This suggests that the models' differences in secondary CO production from methane and NMVOC oxidation play a major role in their ability to reproduce the CO columns in the SH, as also noted by Fisher et al. (2015).
We further quantify the models' differences in chemistry by examining the chemical production and loss terms of CO and HCHO in the models. Results show that large differences in chemical production between the models are largely attributed to differences in the rates of NMVOC oxidation, which are mainly driven by the differences in isoprene oxidation processes, which exhibit varying degrees of complexity in the models. We show the collective effects that different isoprene oxidation schemes have on the rates of chemical production of CO and HCHO but are not able to individually quantify which reactions/processes are responsible for the differences in modelled CO and HCHO. Among the four models, NIWA-UKCA has the highest total chemical production rates of CO, followed by GEOS-Chem, TM5, and CAM-chem which has the lowest chemical production rates. Methane oxidation rates are mainly driven by the OH abundance in the models with TM5 having the lowest OH hence the lowest methane oxidation rates. The fastest conversion rates from NMVOCs to CO occurs in TM5, and the slowest in CAM-chem, leading to respectively high and low CO in these two models. Modelled CO in NIWA-UKCA and GEOS-Chem both better matches the observations in general, irrespective of the different complexities of the isoprene oxidation schemes employed in these two models. Moreover, GEOS-Chem includes some recent advances in isoprene oxidation mechanisms, for example, OH formation of epoxide species which regenerate OH under low NO x conditions (Paulot et al., 2009b). Epoxides are not included in other models. We have not specifically tested how recent experimental evidence on isoprene oxidation mechanisms, e.g. OH regeneration in a low-NO x environment (Fuchs et al., 2013), will impact on modelled species. More detailed and targeted studies will be needed to clarify how individual approaches/processes making up isoprene oxidation schemes will impact chemical production of CO and HCHO in global models.
Production and loss terms of HCHO are assessed in NIWA-UKCA and TM5. We find that total chemical productions are comparable in the two models, with moderately larger chemical production and loss rates in NIWA-UKCA. Again, the production of HCHO from the oxidation of NMVOCs is faster in TM5 although this is partly offset, for HCHO production, by the slower methane oxidation rates due to lower OH. The markedly lower HCHO in TM5 than in NIWA-UKCA could be due to the substantially larger wet deposition loss rate of HCHO, and a faster photodissociation rate of HCHO in TM5. Despite the differences in rates of HCHO formation and loss, we cannot, based on these differences alone, explain the substantial low bias in modelled HCHO in all models compared to the observed HCHO columns at Lauder and Wollongong. We therefore suspect that missing local sources and/or HCHO precursors might contribute to the differences between modelled and observed HCHO.
We conclude that the uncertainty in biogenic emissions remains a significant problem in modelling both long-and short-lived species throughout the SH. Understanding the differences between isoprene oxidation mechanisms and the resultant differences in modelled CO and HCHO is critical, and might result in an improvement in these mechanisms, allowing for a more robust use of HCHO and CO columns to constrain biogenic emissions and reduce this uncertainty. Given that the differences between the two biogenic emissions inventories used here are moderate compared to the much larger uncertainties existing in the current estimates of isoprene and monoterpene emissions, the resultant uncertainty in modelled CO could be much larger. Although the ensemble model mean satisfactorily compares to observed CO in the SH, the large inter-model differences add more uncertainties in modelled CO and in constraining biogenic emissions. Note that in this paper, we do not separately quantify the effect from changes in monoterpene emissions. The emissions from monoterpenes are around 30 and 10 % of those of isoprenes in CLM-MEGANv2.1 and LPJ-GUESS inventories, respectively, which could have a significant impact on modelled CO. However, due to the large uncertainty in emissions and the varying degrees of complexity of the monoterpene degradation schemes included in each model, this will further complicate the interpretation of the impact from changing monoterpene emissions.
The Supplement related to this article is available online at doi:10.5194/acp-15-7217-2015-supplement.