Description and Evaluation of a Secondary Organic Aerosol and New Particle Formation Scheme within TM5-MP v1.1

We have implemented and evaluated a secondary organic aerosol scheme within the chemistry transport model TM5-MP in this work. In earlier versions of TM5-MP the secondary organic aerosol was emitted as Aitken sized particle mass emulating the condensation. In the current scheme we simulate the formation of SOA from oxidation of isoprene and monoterpenes by ozone and hydroxyl radicals which produce semi-volatile organic compounds and extremely low-volatility compounds. Subsequently, SVOC and ELVOC can condense on particles. Furthermore, we have introduced a new particle for5 mation mechanism depending on the concentration of ELVOCs. For evaluation purposes, we have simulated the year 2010 with the old and new scheme, where we see an increase in simulated production of SOA from 39.9 Tgy−1 with the old scheme to 52.5 Tgy−1 with the new scheme. For more detailed analysis, the particle mass and number concentrations and their influence on the simulated aerosol optical depth are compared to observations. Phenomenologically, the new particle formation scheme implemented here is able to reproduce the occurrence of observed particle formation events. However, the concentrations of 10 formed particles are clearly lower as is the subsequent growth to larger sizes. Compared to the old scheme, the new scheme is increasing the number concentrations across the observation stations while still underestimating the observations. The total aerosol mass concentrations in the US show a much better seasonal cycle and removal of a clear overestimation of concentrations. In Europe the mass concentrations are lowered leading to a larger underestimation of observations. Aerosol optical depth is generally slightly increased except in the northern high latitudes. This brings the simulated annual global mean AOD 15 closer to observational estimate. However, as the increase is rather uniform, biases tend to be reduced only in regions where the model underestimates the AOD. Furthermore, the correlation against satellite retrievals and ground-based sun-photometer observations are improved. Although the process based approach to SOA formation causes reduction in model performance in some areas, overall the new scheme improves the simulated aerosol fields. 1 https://doi.org/10.5194/gmd-2021-49 Preprint. Discussion started: 22 March 2021 c © Author(s) 2021. CC BY 4.0 License.


20
Aerosols have a pronounced influence on the climate Boucher et al., 2013) and air quality (Isaksen et al., 2009;Monks et al., 2009). Particulate organic matter, also known as organic aerosol (OA), contributes between 20 % and 90 % of total aerosol mass (Kanakidou et al., 2005). This ubiquitous OA is a major component of the atmospheric aerosols across the globe (Zhang et al., 2007). It has two main sources, which are separated due to their formation mechanism. On the one hand, organic mass is emitted directly to the atmosphere. This component is often called primary organic aerosol (POA). On the other 25 hand, OA is formed in the atmosphere by oxidation of gaseous organic compounds. This part is known as secondary organic aerosol (SOA). POA sources include fossil fuel combustion, biofuel burning or wildfires, while organic gases are released into the atmosphere from both natural and anthropogenic sources producing volatile organic compounds (VOC). These VOCs can undergo chemical reactions in the atmosphere producing organic compounds with lower volatilities. Furthermore, some of them can take part in new particle formation (NPF) and condense onto existing particles, thereby creating SOA. Depending on 30 their volatility and their contribution to these two processes, these low-volatility products are often grouped into semi-volatile VOC (SVOC), low-volatity VOC (LVOC) and extremely low-volatility VOC (ELVOC).
Natural sources of VOCs account for approximately 85 % of total VOC emissions (Guenther et al., 2012;Lamarque et al., 2010). Important natural sources of VOC include e.g., terrestrial vegetat589527ion or marine phytoplankton (Guenther et al., 2006(Guenther et al., , 2012Meskhidze and Nenes, 2006;Gantt et al., 2009;Yassaa et al., 2008;Shaw et al., 2003). Due to their biogenic origin 35 these VOCs are often referred to as biogenic VOCs (BVOC). Emissions of BVOCs are dominated by isoprene and terpenes (Guenther et al., 2012;Glasius and Goldstein, 2016). Their contribution to OA production is significant and highlights the importance of the interactions between biosphere and atmosphere within the Earth system Paasonen et al., 2013). Emissions of BVOCs are a rather large source of VOCs, but due to complex chemistry the actual processes and the total amount of SOA formation is rather uncertain (Tsigaridis et al., 2014). The estimates of the total annual production of 40 SOA from bottom-up and top-down methods range between 12 and 1820 Tg yr −1 (Goldstein and Galbally, 2007;Hallquist et al., 2009). With deficiencies in understanding and complex pathways the descriptions of SOA formation in global models are often rudimentary. In the study by Tsigaridis et al. (2014) many models still treated SOA simply by emitting it as OA with prescribed SOA mass yields. These models with prescribed yields produce SOA amounts near the lower limit of the estimated source strength. In addition, models that treat SOA formation with a simple chemistry estimate similarly low SOA production. 2012). Due to recent advances, new parameterisations for NPF in the boundary layer were developed, e.g. involving sulfuric acid (Sihto et al., 2006), ammonia (Dunne et al., 2016) and VOCs in general (Paasonen et al., 2010;Riccobono et al., 2014;Bergman et al., 2015). After growth due to condensation or coagulation, newly formed particles can contribute to the global CCN number budget (e.g. Merikanto et al., 2009;Makkonen et al., 2012;Dunne et al., 2016;Gordon et al., 2017;Kerminen et al., 2018). 70 In conclusion, SOA formation can affect both NPF and condensational growth of existing aerosol particles. The two processes can promote the growth of particles into sizes relevant for CCN and therefore affect cloud properties. However, this effect is highly non-linear as the two processes distribute SOA differently on the aerosol size spectrum. Whereas NPF initially produces a large number of small particles, condensational growth increases the size of existing particles. Whether the addition of new particles due to SOA formation increases or decreases the number of CCN therefore depends on the share of SOA mass 75 between NPF and condensation.
In this work we present the implementation of a VOC oxidation scheme to calculate the production of ELVOCs and SVOCs to describe the formation of SOA within TM5-MP. Additionally, to improve the description of NPF in the boundary layer, we implemented a NPF scheme following Riccobono et al. (2014), which describes the production of new particles in the presence of ELVOCs. The new SOA and NPF scheme is part of the TM5 version of EC-Earth3-AerChem (Döscher et al.,in preparation;80 van Noije et al., submitted to Geoscientific Model Development), which is used in AerChemMIP (Aerosol Chemistry Model Intercomparison Project; Collins et al., 2017) of the Coupled Model Intercomparison Project phase 6 (CMIP6; Eyring et al., 2016). In this paper we describe and evaluate the new SOA and NPF scheme. TM5 simulations with and without the new scheme are compared and evaluated against in-situ and remote sensing datasets. The performance of the new SOA treatment is evaluated for key variables, such as SOA budget, total organic aerosol mass, aerosol optical depth (AOD) and aerosol number 85 concentrations in the surface layer.
In Section 2 we describe the new SOA formation and NPF schemes along with the observational data used for evaluation.
In Section 3 we present an evaluation of the simulation against observations. In Section 4 we give the conclusions. Oxidation by OH and O3 produces two organic surrogate species (ELVOC, SVOC). The lower volatility product ELVOC participates in new particle formation and can condense on the existing particles according to particle surface area. The semivolatile product SVOC can condense on existing particles according to their total OA mass. This freshly formed SOA mass was distributed near the surface with 80 % into heights below 30 m and 20 % in the height ranging from 30 to 100 m . The SOA mass was added to organic aerosol (OA) as additional mass into the soluble (65 % of the total) and insoluble (35 % of the total) Aitken modes with no increase in particle numbers.

120
The online SOA scheme described in this paper calculates the production of SOA from isoprene and monoterpenes (see Fig. 1). To track the mass of SOA in the atmosphere we have expanded M7 by including a new particulate mass tracer into all soluble modes and the insoluble Aitken mode. Despite the production of SOA being different from primary organic aerosol, we assume otherwise the same characteristics (such as density and refractive index) for SOA as for the primary organic aerosols.
The scheme is kept minimal in detail and computationally light to allow long, centennial scale integrations. In the following 125 we detail the microphysical processes of production of gas-phase SOA precursors, their condensation and the ELVOC-induced new particle formation.

Emissions of SOA precursors
Plants emit isoprene and monoterpenes depending on ambient conditions creating a diurnal cycle with highest emissions during daytime (Funk et al., 2003;Holzke et al., 2006). Biogenic VOC emissions calculated by the MEGANv2.1 model depend on radiation, temperature, leaf area, leaf age, and CO 2 (for detailed description of the model see Guenther et al., 2012). The emissions are shown in Fig. 2 (see section 2.4 for the explanation of the different simulations). Funk et al. (2003) show that the isoprene emissions from plants follow a diurnal cycle, but due to conflicting results accurate modelling of its variation requires more research. Due to lack of a more realistic emission treatment and our dependence on emission inventories, we assume the release of isoprene during the day to vary as the cosine of the solar zenith angle with 135 zero emissions during nighttime (Huijnen et al., 2010). The monoterpenes are emitted from storage pools mainly as a function of leaf temperature, but also other factors affect the emission such as temperature and drought (Holzke et al., 2006). These variables are not available in the model and therefore for monoterpene emissions we follow a similar approach to the one used for isoprene. However, we use a sinusoidal function which has a minimum at night and peak around noon to emulate the emission from storage pools and from photosynthesis.

140
In this work we employ monthly mean isoprene (572.3 Tg yr −1 ) and monoterpene (95.5 Tg yr −1 ) emissions from an inventory derived from MEGANv2.1 (Sindelarova et al., 2014;Guenther et al., 2012, ;geographical distribution from our simulation setup shown in Fig. 2a,c). In the old scheme the calculation of SOA production is based on monoterpene emissions from an earlier MEGANv1 with an annual total of 144.2 Tg yr −1 (Fig. 2b). The difference to the new scheme is shown in Table 1. Rate coefficients (Atkinson et al., 2006) and molar yields used to calculate production of ELVOC and SVOC in reactions between OH and O3 and monoterpene and isoprene. are monthly emissions without diurnal variations as defined in the emission inventory (van Marle et al., 2017). At present the model does not include oceanic isoprene or monoterpene emissions, mainly due to low strength of isoprene emissions (Arnold et al., 2009) and high uncertainties of monoterpene emissions (Arnold et al., 2009;Yassaa et al., 2008).

Production of extremely low volatility and semi-volatile organic compounds
Jokinen et al. (2015) assumed the molar yield of SOA precursors from monoterpene and isoprene oxidation to be 15 % and 5 150 %, respectively, which are divided into ELVOC and SVOC. Furthermore, they determined experimentally the ELVOC molar yields from oxidation of monoterpenes and isoprene by ozone (O 3 ) and the hydroxyl radical (OH). In our implementation ELVOC and SVOC are formed in reactions of isoprene and monoterpenes with O 3 and OH. The reaction rate coefficients and yields for SVOC and ELVOC can be seen in Table 1. We assume two products from these reactions: extremely low-volatile organic compounds (C 10 H 16 O 7 ; ELVOC) and semi-volatile compounds (C 10 H 16 O 6 ; SVOC), which have different volatilities.

155
However, in contrast to Jokinen et al. (2015) we assume total production of SOA precursors from isoprene to be 1 % instead of 5 % due to very high production of SOA precursors in our inital tests with higher yields. This is in line with Kroll et al. (2005) who report 0.9-3.3 % mass yields.

Gas-particle partitioning of ELVOC and SVOC
Organic condensation in large scale models is generally calculated using either partitioning theory (Pankow, 1994) or kinetic 160 condensation on the surface of aerosols . The former assumes that organic vapor molecules find equilibrium instantly with the aerosol, the latter assumes that vapors are non-volatile and condensation depends on the surface area.
We apply both methods following the work of Jokinen et al. (2015). SVOCs are assumed to partition among different types of particles according to the existing total OA mass in each mode (equilibrium model) whereas ELVOCs condense according 165 to particle surface area of each mode (kinetic approach). The change in SOA mass by condensation of ELVOC and SVOC in a where i is the log-normal mode index. The first term on the right-hand side describes the condensation of ELVOC, where CS i is the condensation sink (Pirjola et al., 1999) of a single mode i which is proportional to the surface area of the mode. The 170 condensation for ELVOC is applied for all soluble modes and the insoluble Aitken mode where SOA and/or POA is present.
∆M ELVOC is the mass of gas-phase ELVOCs available for condensation within one timestep.
The second term on the right-hand side describes the increase in SOA mass by condensation of SVOC on particles of mode i, where ∆M SVOC is the available gas-phase SVOC mass within one timestep, M i,POA+SOA is the total organic aerosol mass in mode i. Here we follow the original M7 modal assumption for SOA, meaning that SVOC can condense to soluble and 175 insoluble Aitken mode, soluble accumulation mode and soluble coarse mode. Similarly to Jokinen et al. (2015) we assume that both of these compounds have a volatility low enough that all of them will be condensed onto the existing particles within a time step of the model. Thereby we reduce the computational cost without the need to calculate the transport of gas-phase SOA precursors.

180
During new particle formation (NPF) low volatility gases form small molecular clusters which can transform into stable particles. This process is ubiquitous and mainly driven by sulphuric acid . Therefore, global models often parameterise NPF as function of sulphuric acid concentration (Vehkamäki et al., 2002;Sihto et al., 2006;Kulmala et al., 2006;Laakso et al., 2004). However, recent research has shown that a variety of other compounds participate in the process e.g.

185
In previous TM5 versions, NPF is described using the parameterisation of Vehkamäki et al. (2002) (classical nucleation theory). They calculate the nucleation rate and particle size depending on the concentration of water and sulphuric acid using a fitted formulation. In this work, we introduce an additional NPF parameterisation which takes into account gas-phase organics (Riccobono et al., 2014). This scheme formulates the formation of particles of 1.7 nm in diameter as a function of the concentrations of sulphuric acid (H 2 SO 4 ) and oxidised biogenic compounds (BioOxOrg), which is based on measurements done in 190 the CLOUD chamber (Cosmics Leaving OUtside Droplets) at CERN (Kirkby et al., 2011). In our implementation we represent the BioOxOrg as ELVOC following Eq. 2 in Riccobono et al. (2014) with p=2 and q=1, as where K m = 3.27 × 10 −21 cm 6 s −1 is an empirical factor and [H 2 SO 4 ] and [ELVOC] are the gas-phase concentrations of sulfuric acid and ELVOCs, respectively.

195
The new scheme combines two different parameterisations to calculate the NPF: the binary homogeneous water -sulphuric acid nucleation (Vehkamäki et al., 2002, referred to as BHN in the following) and the semi-empirical parameterisation by Riccobono et al. (2014) (referred to as RICCO in the following). Calculation of the parameterised growth by Kerminen and Kulmala (2002) to 1.7 nm, Phase 3: Calculation of semi-empirical particle formation rate JRICCO considering ELVOC, Phase 4: Calculation of the parameterised growth by (Kerminen and Kulmala, 2002) from 1.7 nm to 5 nm.

Parameterisation of particle growth to 5 nm
The growth of small particles in M7 is hindered by the modal structure of the model (Korhola et al., 2014). Therefore, we 200 calculate the parameterised NPF rate for particles of 5 nm in diameter using the formulation of Kerminen and Kulmala (2002, KK in the following) in four phases (see schematic in Fig. 3). First, we calculate the BHN rate (J BHN ) and associated diameter of the formed particles. Second, to combine the formation rate of the BHN and the RICCO schemes the formation rate of the BHN scheme is required for particles of the same size (i.e. 1.7 nm). KK is then used to calculate the growth of particles by ELVOC and sulfuric acid vapors from BHN formation size to 1.7 nm diameter particles. In the third step, the formation of 205 1.7 nm particles according to RICCO is calculated and added to the one from BHN. Finally, we use KK again to calculate the growth from 1.7 nm to 5 nm due to condensation of ELVOC and sulfuric acid. A more detailed description can be found in Appendix A.

Simulations
We have run two simulations with TM5 to evaluate the impact of the new SOA scheme on the organic mass and particle number 210 concentrations. The first simulation, called OLDSOA, is done with the old formulation where all SOA mass is added to the primary OA in Aitken mode as explained in the beginning of section 2.3. This simulation uses prescribed SOA production based on monoterpene emissions from the older MEGANv1 (Guenther et al., 1995). The second simulation, called NEWSOA, utilises the SOA description of this paper. Biogenic monoterpene and isoprene emissions for SOA production in NEWSOA use MEGANv2.1 (Sindelarova et al., 2014;Guenther et al., 2012). For all other biogenic emissions we use MEGANv2.1 215 (Sindelarova et al., 2014;Guenther et al., 2012)  Other natural emissions are prescribed as in van Noije et al. (2014). For the emissions of gases and particulate matter from anthropogenic sources and biomass burning we use the CMIP6 input4MIPs inventory (Hoesly et al., 2018;van Marle et al., 2017). In this work, a horizontal resolution of 3 • longitude by 2 • latitude and 34 hybrid-sigma levels are used.

Observational data used in model evaluation
In order to evaluate the impact of the new SOA scheme on the simulated aerosol properties, the model is compared against observations of organic mass concentrations, number concentrations and AOD. The observational data from surface measurements, in-situ remote sensing and satellite retrievals are described below.

Organic mass concentrations on the surface 230
We evaluate the model performance of simulating OA concentrations at the surface by comparing to two freely available observational networks, the United States' Interagency Monitoring of Protected Visual Environments (IMPROVE; http://vista.cira.colostate.edu/improve/ last access 11.1.2018; Malm et al., 1994) and European monitoring and evaluation project (EMEP; http://www.emep.int; last access 27.7.2017; Tørseth et al., 2012). For the IMPROVE network we use the organic mass in PM2.5 particles from 175 stations (see Table S1 for a list of stations) and for EMEP we use PM2.5 or PM10 235 depending on the station for 15 stations (see Table S2 for a list of stations). The sum of simulated primary and secondary organic aerosol concentrations in the lowest model layer has been collocated with the location, and time of the observations.
However, here we report aggregated monthly and yearly means of the observed and simulated values.
Both EMEP and IMPROVE networks measure particulate organic carbon (OC) instead of total organic mass in the particles. Therefore, the carbon content is usually converted to organic mass with a constant factor. For the whole IMPROVE 240 network the suggested ratio between carbon and particulate organic matter in PM2.5 particles is 1.8 (Pitchford et al., 2007, http://vista.cira.colostate.edu/Improve/the-improve-algorithm/), which is used in our analysis also. The conversion factor from OC to OA at the European sites is commonly assumed to be 1.4 (Putaud et al., 2004;Sillanpää et al., 2005). However, since Yttri et al. (2007) show that usually the ratio between OC and OA should be higher than that, we follow the IMPROVE network implementation and assume a factor of 1.8 also for the EMEP stations. It has to be noted that in the primary emitted carbon for 245 POA is converted to total mass with constant factor of 1.6.  et al., 2012). The stations there provide condensation particle counter (CPC) observations around the globe, but the coverage is very sparse. In this dataset, USA and Europe are overpresented while most of Asia is lacking observations. Nevertheless, 250 we have collocated the simulated concentrations of particles with diameter larger than 10 nm in time and space to 27 stations (see Table S3 for a list of stations), where number concentration data was available for 2010 to compare the annual mean from NEWSOA and OLDSOA simulations to the observations (Fig. 8).

Remote sensing data
The comparison to satellite retrievals on aerosol optical depth (AOD) provides the opportunity to evaluate the new SOA 255 description globally. Here we use the AOD products from two different satellite instruments: the MODerate resolution Imaging Spectroradiometer (MODIS) retrieval and the Advanced Along Track Scanning Radiometer (AATSR) retrieval.
MODIS is located aboard two satellites Aqua and Terra providing good coverage in the morning and afternoon (King et al., 1999). We use here the combined product from Deep Blue and Dark Target  The sun-photometer network AERosol RObotic NETwork (AERONET; Holben et al., 1998) provides global coverage of in-265 situ measurements of AOD, although in many areas the network is rather sparse. Nonetheless, these observations are considered as a ground truth in aerosol optical depth (AOD) retrievals. Since AOD at 550 nm is not available from many AERONET sites, their instantaneous aerosol optical depths at 550 nm (AOD 550 ) are derived using Ångström power law, Ångström exponent (440-675 nm;α) and level 2 AERONET AOD at 500 nm (AOD 500 ): We have collocated the simulated hourly AOD at 550nm with the AERONET AOD in space and time to provide us with the best possible evaluation of the model performance. In our comparison we have included all 299 stations which provide data for the year 2010. After the collocation we show the aggregated monthly and yearly mean data.

285
The total SOA production here is roughly double to that predicted by Jokinen et al. (2015). In addition, the shares of ELVOC and SVOC as well as the shares of precursors are quite different. In their model the ELVOC production is only 0.9 Tg yr −1 , while in our implementation it is 7.3 Tg yr −1 . Thereby, also the ELVOC fraction is higher in our case, with 14 % of total production being ELVOC compared to 3.2 % by Jokinen et al. (2015). For the precursors they show 12 % of SOA originating from isoprene while in our NEWSOA simulation it is 47 %. Therefore, even though both models employ very similar SOA 290 schemes, different emissions and chemistry cause distinct differences in the global annual mean production of SOA.
When comparing the regional differences in SOA production between our two simulations in Figs. 4a and 4b, the emission regions remain mostly the same. Strongest production is found in the rain forests of South America and Africa and the lowest production over the deserts, most notably in the Sahara and Gobi deserts. However, there are some changes in the main production regions in South America and Africa. Additionally, in Australia the production increases strongly in NEWSOA 295 compared to OLDSOA resulting from inclusion of isoprene in NEWSOA (Fig. 4c). Another interesting change is in the northern high latitudes, where the SOA production in NEWSOA is lower than in OLDSOA throughout the year (Fig. S4), with clearly lower production in boreal regions in Scandinavia, northern Canada and northern Russia.
Mainly the differences are caused by using SOA production calculated from older Guenther et al. (1995)    it is reduced by 10 %-100 %, which is expected since the production in the northern high latitudes has decreased significantly as noted above. In the Southern hemisphere (SH) the increase in SOA burden is more than 50 % almost everywhere, which 310 is inline with the strong increase in production in the SH. A lower increase of 25-50 % can be seen in sub-Saharan Africa and eastern Brazil. In contrast, over Australia a strong increase in SOA burden can be observed (more than 100 %), which is expected due to strong isoprene emissions (Guenther et al., 2006) and subsequent SOA formation from isoprene, which is not accounted for in OLDSOA.
The zonal annual mean concentrations of the SOA in simulations OLDSOA and NEWSOA and their difference are shown 315 in Figure 6. On average the concentrations of SOA are higher near the Earth's surface in OLDSOA than in NEWSOA, which is expected, because in OLDSOA all of the produced SOA mass is distributed into the lowest 100 m of the boundary layer. In the Southern Hemisphere the concentrations aloft are as much as 100 % higher in NEWSOA compared to OLDSOA as can be seen in Fig. 6c. Higher SOA concentrations aloft in NEWSOA are expected due to the SOA production occurring throughout the atmosphere from oxidation products of monoterpenes and isoperene instead of emitting it into the boundary layer below 320 100 m as in OLDSOA. In contrast, the figure shows clearly that the concentrations are lower in NEWSOA in the northern high latitudes (50-75 % lower) due to the lower production of SOA in the northern Hemisphere in NEWSOA. This results mainly from spatially different emissions of monoterpenes in MEGANv1 compared to MEGANv2.1 as stated earlier (see Fig. 2).
The removal mechanism of SOA is almost exclusively wet deposition (98.9 %). It has increased by 12.8 Tg yr −1 (38.5 %) in NEWSOA compared to the earlier scheme mainly due to higher production in NEWSOA. Although a high fraction of wet 325 deposition is expected (Flossmann and Pruppacher, 1988), the removal fraction by wet deposition is clearly higher than the contribution of 85 % observed in the AeroCom multi-model ensemble (Tsigaridis et al., 2014). The bottom row shows the concentration of particles in the nucleation mode in runs NEWSOA (c) and OLDSOA (d).
The lifetime of SOA in Table 2 is calculated as a ratio between burden of SOA over its annual mean removal. It has increased in NEWSOA by about one day from 7.8 to 9.1 days. The new lifetime is higher compared to the multi-model mean of about 8 for 12 models in Tsigaridis et al. (2014). The increase is expected due to calculating SOA production ubiquitously in the 330 atmosphere when SOA precursors are available. It is, however, slightly counter intuitive. Due to a larger portion of SOA mass in the accumulation mode in OLDSOA, a longer lifetime of SOA could be expected (see Schutgens and Stier, 2014, for modewise lifetimes in M7). However, as SOA concentration aloft in NEWSOA is increased compared to OLDSOA, a longer lifetime is reasonable.
Although wet deposition, burden and lifetime fall within the multi-model ranges of Tsigaridis et al. (2014), the dry deposition 335 is clearly lower than either their multi-model mean or multi-model range. Therefore, it would be worthwhile to examine the wet and dry deposition processes of SOA in TM5 in more detail in a dedicated study.

Changes in new particle formation (NPF)
As explained in Sect. 2.3.4, in the previous version of TM5 the NPF is calculated using the parameterisation of Vehkamäki et al.
(2002), which is not able to produce the observed particle number concentrations in the boundary layer (Spracklen et al., 2006;340 Makkonen et al., 2009). Here we show how the particle formation and number concentrations change using the organically enhanced new particle formation by Riccobono et al. (2014). Figure 7a,b shows the annual mean new particle formation rate at the surface for NEWSOA and OLDSOA. The annual mean formation rate at the surface is 0.021 cm −3 s −1 and 0.00021 cm −3 s −1 for NEWSOA and OLDSOA, respectively. The binary homogeneous nucleation produces very few particles near the surface except in Siberia and Antarctica. Since the RICCO NPF 345 parameterisation depends on the ELVOC and sulfuric acid concentrations, formation rates over the oceans are negligible (below 1 × 10 −4 cm −3 s −1 ). Over land areas of highest formation rates mostly coincide with locations of high ELVOC production.
However, the formation rates in the Amazon and sub-Saharan Africa are very low although ELVOC production rate is high.
These locations have low sulfuric acid concentrations, which in turn limits the new particle formation here. Contrarily, the Arabian peninsula shows high formation rates due to high sulfuric acid concentrations although ELVOC concentrations are 350 relatively low. In general the updated SOA and new NPF parameterisation caused an increase in particle number concentrations, but at two stations the concentrations were decreased. This can be caused by two confounding factors. Firstly, in the new SOA scheme with low concentrations of ELVOC, particle formation is low. Secondly, with higher production of SOA and its condensation

Particle size distributions at selected sites
Considering the small amount of continuous long-term comprehensive observations of the aerosol-chemical system, it remains extremely difficult to constrain how distinct aerosol dynamical processes perturb the particle size distribution, and how those processes are influenced by regional and large scale physical and chemical conditions. From a modelling perspective, a large set of simulations with varying parameter values (Perturbed Parameter Ensemble, PPE) would allow to assess the sensitivity of the simulated size distributions to underlying parameter uncertainties, although limited to a single model framework (e.g. Sengupta et al., 2020). Such experiment is outside the scope of this paper, and we limit ourselves to presenting selected cases 380 which highlight the size distribution properties simulated by the improved TM5 model.  9b), since observed N20 (particles with particle diameter d p > 20 nm) reaches even 5000 cm −3 which is higher than total condensation nucleii (CN) peak in TM5 during that day. Nevertheless, TM5 simulates aerosol growth from nucleation to 385 Aitken and accumulation modes, resulting in cloud concensation nucleii (CCN) peak on 25th of April. During 24.-29.9.2010 ( Fig. 9a) TM5 simulates the evolution of both CN and CCN except for peak concentrations. In Vavihill, during the first week of July (Fig 9c), TM5 successfully predicts nucleation events but seems to underestimate nucleation or subsequent growth to the Aitken mode, or both. During 2nd of July, nucleation and resulting N3 (particles with d p > 3 nm) concentrations match the observations and clear growth to Aitken mode is simulated. Nevertheless, simulated growth to accumulation mode remains 390 slower and less efficient, rendering N70 (particles with d p > 70 nm) concentrations significantly lower than observed values.
During a few days in mid-June, TM5 simulates nucleation and growth events in Harwell (Fig. 9d). While the observed Aitken-mode growth is not visible in the simulated distribution, N20 peak concentrations are well simulated, although peak times deviate from the observed one. In Waldhof (Fig. 9e), aerosol formation rates are clearly underestimated even though the observed size distribution starts at 20 nm. Only a minor contribution from nucleation to N20 and N70 is simulated, and the 395 increased trend in CCN is due to a shift in airmass trajectories. Both simulations and observations in Mace Head show a distinct aerosol size distribution for marine and continental airmasses. Fig. 9f shows an example of a simulated nucleation event (15th October 2010) with visible growth in the Aitken mode. The sustained growth leads to an increase in N70 during the course of 12 hours.
To summarize, TM5 with modal microphysics including improved nucleation and SOA mechanisms is able to capture 400 nucleation and growth events but effective formation of CCN-sized particles from new particle formation events might remain limited due to numerical challenges (Whitby et al., 2002;Wan et al., 2013;Korhola et al., 2014), underestimated nucleation rates, and underestimated concentrations of vapours available for sub-CCN growth. However, it should be recognized that we do not expect a one-to-one match between a coarse-grid global model and local aerosol observations .

Organic mass concentrations at the surface 405
In this section we compare the simulated surface concentrations of organic mass in PM2.5 and PM10 particles to EMEP and IMPROVE network observations which are described in Section 2.5. or NEWSOA. This peak is caused by the addition of all SOA mass in the Aitken mode without any increase in particle number.

Evaluation at IMPROVE stations
Therefore the particles in Aitken mode grow quickly to accumulation mode, which has a lower deposition rate (Seinfeld and Pandis, 2006). In NEWSOA, however, the SOA mass is condensed more realistically across the particle size distribution using either kinetic or thermodynamic assumptions (see Section 2.3.3). Therefore, in NEWSOA we see behaviour similar to the observations. This improvement in the representation of organic mass in NEWSOA reduces the deviation between 415 observations and model substantially (NMB from 184 % in OLDSOA to -18 % in NEWSOA, see Table 3) and results in a correlation coefficient R=0.59. The strong overestimation (NMB=184 %) of the OLDSOA run, also seen in the scatter plot ( Fig. 10d), is removed and the correlation with observations is improved notably.
The new formulation is now underestimating the OA concentration slightly. However, it does not include any anthropogenic SOA nor the production from oxidation by NO 3 . In addition, the conversion factor for primary emitted carbon to organic mass 420 (1.6) in the model is lower than than the conversion factor used by the IMPROVE network (1.8), which has a contribution to the underestimation. For the annual mean the situation is mostly similar with NEWSOA underestimating the surface concentrations of PM2.5 435 and PM10 (NMB=-24 % and NMB=-25 %, respectively) more strongly than OLDSOA (NMB=-4 % and NMB=-14 %, respectively). The root mean square error is slightly increased in the NEWSOA simulation (from 1.72 to 1.83) for PM2.5 while it has decreased for PM10 (from 2.09 to 2.02). However, as for IMPROVE it is expected that the total OA is underestimated, since the model doesn't account for antropogenic SOA and other biogenic SOA sources (e.g. VOC oxidation by NO 3 ). In addition, the conversion factor of POA carbon to total organic mass (1.6) in the model is lower than than the one used for the 440 station data (1.8), which has a small contribution to the underestimation. Furthermore, the good local agreement in OLDSOA with prescribed SOA formation can be accidental, while online oxidation in NEWSOA takes into account the oxidant levels.

Evaluation at EMEP stations
Therefore the missing sources may lead to underestimations in certain areas.

Summary of surface organic mass concentrations
In NEWSOA the OA burden has been reduced in the Northern Hemisphere and increased in the Southern Hemisphere as 445 shown in Fig. 12a. At the surface OA concentration (Fig. 12b) is mostly increased in the Southern Hemisphere. However, in South America and Africa the surface concentrations, even with roughly similar production of SOA (see Fig. 4), are lower due to the NEWSOA production occurring more aloft while OLDSOA produces all SOA within 100 m of the surface. The decreases in the Northern Hemisphere are largely due to lower monoterpene emissions. Also notable is the increase in the surface concentration and burden of SOA around Australia due to the addition of isoprene as a precursor of SOA.
The new SOA scheme shows an improvement in simulating the mean organic mass concentrations at the IMPROVE sites, although there is a small underestimation of the annual mean organic mass. For the EMEP sites, the performance is slightly worse, but since the process description in OLDSOA is clearly flawed, the reduced performance is more likely to result from lack of other sources. It is likely that the reason for the reduction lies in the missing production of SOA from anthropogenic sources, reactions with nitrate radical, other oxidant concentrations or even problems due to resolution.

455
The annual cycle at the IMPROVE stations is more realistic than before with a flat distribution comparable to the observations. At the EMEP stations the annual cycle is largely unchanged, but the underestimation of concentrations has increased.
Regardless, for IMPROVE the new treatment removes a clear overestimation during the summer improving the overall SOA description.
The observations of organic mass concentrations depicted in Fig better agreement with observations. Therefore, the implementation of a VBS or another more sophisticated SOA scheme into TM5 could be investigated in the future.

Satellite and ground-based remote sensing
In this section we compare the modelled AOD to the remote sensing observations which are described in Section 2.5.

MODIS
470 Figure 13a shows the difference in the annual mean collocated AOD from the NEWSOA simulation modelled AOD fields compared to MODIS. The global annual mean AOD is underestimated for both SOA schemes. However, in NEWSOA the global annual mean AOD is improved by 0.01 from 0.12 to 0.13, which is still 0.02 lower than the mean MODIS value (0.15), but regionally both undestimates and overestimates occur. The area weighted normalized mean bias is improved in NEWSOA to -16 % compared to -20 % in OLDSOA. The simulated AOD is underestimated in large areas over oceans in both model runs, 475 however the SOA scheme was not expected to impact these regions. Over land AOD is mainly increasing causing regions of increased overestimation and decreased underestimation (e.g. China and Australia). Only in some regions of Canada, Finland and Russia AOD is decreasing in an area with existing low biases (see Fig. S13 for change from OLDSOA to NEWSOA). In Russia the underestimation east of Finland is reduced but a negative bias remains. This is expected since the wildfire emissions in this area in summer of 2010 are difficult to reproduce even with high resolution model simulations 480 2018). The underestimation in central Africa is reduced, but it still remains large. The strong increase in production and burden of SOA north of the Congo region changes the underestimation to a slight overestimation. In the outflow region from Africa towards South America the underestimation is decreased, but not removed. However, this is more related to dust AOD than SOA. In the Amazon region AOD is improved with NEWSOA compared to OLDSOA. Since this underestimation is mostly caused by biomass burning emissions in the Southern Hemisphere dry season (September-October-November) we expect this 485 bias to remain with the improved SOA scheme. To summarize, the annual mean modelled AOD collocated to MODIS has increased relatively evenly. The global mean AOD is improved. However, already existing overestimations are increased in NEWSOA. Figure 13b shows the collocated annual mean AOD difference between NEWSOA and AATSR. The collocated annual mean is underestimated compared to AATSR. Furthermore, the AOD compared to AATSR over South America is underestimated as in the comparison to MODIS, although the underestimation is slightly more pronounced. However, AATSR overestimates over bright surfaces (Che et al., 2018), which means that an underestimate is to be expected in such areas, e.g. Saharan Desert, Australia.

AERONET
500 Figure 14 shows the comparison of AOD measured by AERONET and modelled AOD in NEWSOA (a), and OLDSOA (b).
Furthermore, Fig. 14d presents a map showing the stations and their regional grouping for the statistics in Table 4. Both simulations have similar behaviour overestimating the low AOD and underestimating the high AOD. However, compared to OLDSOA the absolute bias across all stations increases in NEWSOA by 0.007 to a low bias of 0.003. Additionally, there is a slight increase of 0.009 in correlation coefficient (R; see last row of Table 4 for the values across stations). Visually both 505 simulations show a very similar deviation from the observations. Figure 14c shows the seasonal cycle of the mean AOD at the AERONET stations comparing the measurements and the two simulations. In general both simulations reproduce the observed annual cycle reasonably well. It is evident that the new SOA scheme is increasing the AOD throughout the year. However, the observed AOD has a peak in March, but in both simulations this peak seems to be delayed by one month. The local minimum in May is not seen at all in the simulations. NEWSOA 510 reproduces the AOD from June and July almost exactly. The peak in August obtained with NEWSOA is closer to the observed value but is still a bit too low. The reduction during fall (Sep-Oct-Nov), the modelled AOD is well produced in both simulations, and only the December AOD is clearly overestimated. Figure 15a shows the change in AOD between OLDSOA and NEWSOA as collocated to AERONET AOD. The AOD at many stations is still underestimated with NEWSOA as shown in Fig. 15b. It shows that the AOD is increasing at most stations 515 causing reductions in the bias at some locations while at other locations with overestimation existing biases have increased.
All but few stations show a larger AOD in NEWSOA than in OLDSOA, even in the boreal region where the production of SOA is reduced compared to OLDSOA. This supports the finding that the new SOA scheme is affecting particles that are more relevant for AOD. This implies that in the boreal region regardless of the lower SOA production we see an increase in AOD when the SOA is distributed more realistically onto the particles. It is noteworthy that the AOD in the Amazon has increased, 520 but some bias remains. In Canada the decrease in burden of SOA does lead to reduction of the AOD, but the AOD is still overestimated. In central and northern Europe however, the simulated AOD is increased and is near the observed AOD, even though the burden of SOA in the boreal region is reduced.
The regional statistics between modelled and observed AOD at AERONET stations are presented in Table 4. in SOA concentration at the surface in the US is not reflected in the AOD. Actually the AOD has increased in all analyzed regions, which suggests that the increase in concentrations of smaller particles are more relevant for the AOD. This is probably due to reduced growth to accumulation mode and more even distribution of the SOA mass in NEWSOA.

535
We have implemented a new scheme for online production of SOA from oxidation of monoterpene and isoprene together with a new particle formation mechanism depending on the ELVOC concentration. We have run the model with 1-hourly output for one year with emissions set to 2010 levels using the old and new schemes. The two simulations have been compared to each other as well as to in-situ and remote sensing observations. Surface particle number concentrations were compared to observations available from the EBAS database. Surface concentrations of total organic aerosol mass were evaluated against 540 measurements from the EMEP and IMPROVE surface networks. Aerosol optical depths were compared to retrievals from the AERONET sun-photometer network, as well as satellite observations from the AATSR and MODIS instruments.
The global production of SOA mass was increased by 35 % to 52.5 Tg(SOA) with an increase by 1.3 days in SOA lifetime (to 9.1 days), mainly due to increased concentrations of SOA in the upper troposphere. This is caused by more SOA being produced aloft in the atmosphere and a more realistic description of condensation of SOA precursors onto particles. The new 545 lifetime is similar to the average from the AeroCom model ensemble (Tsigaridis et al., 2014). The SOA production and burden increase is rather uniform except for northern high latitudes where the production and burden are decreased compared to the old scheme. This is caused mainly by lower emissions of monoterpenes in this area in MEGAN v2.1. in conversion factor from carbon to organic mass.
The global annual mean AOD collocated to MODIS and AATSR increases from OLDSOA to NEWSOA by 0.01, which brings the simulated annual mean AOD closer to observations in both cases. However, since the increase is quite uniform, regionally the change is improving areas with underestimations and degrading areas with overestimation. Since we see an increase in AOD also at high latitudes, where the SOA production and OA burdens are decreased, the influence of particles 565 affecting radiation is stronger. The increase could be attributed to humidity effects, but both simulations use the same humidity while OA does not affect water uptake. Therefore, the changes result mainly from changes in particle number concentrations.
Similarly satellite instruments, the simulated AOD collocated to AERONET measurements increases by 0.007. Even regionally the changes are small, with the strongest improvement in South America, where the bias changes from -32.7 % to -22.5 %, but also a deterioration in Australia from -4.0 % to 23.3 %. The RMSE and correlation across AERONET network improve 570 slightly while regionally there is improvement and deterioration.
Although the model shows an improvement in many aspects, especially regarding organic mass concentrations in the US, for example at EMEP sites there is a decrease in performance. As it is commonly the case, a more detailed process description can initially lead to reduced performance due to error compensation in the simpler description. Further development is needed targeting the identified model deficiencies. Some future lines of study could focus on 1) missing anthropogenic SOA sources, 575 2) additional precursor compounds (e.g. sesquiterpenes and anthropogenic gases) and oxidants (e.g. NO 3 ), 3) transport and removal of gas-phase SOA precursors (e.g. ELVOCs and SVOCs), 4) interactive emission of precursor gases from a the dynamic vegetation model. Appendix A: Parametrisation of growth from nucleation to 5 nm As shown by Kerminen and Kulmala (2002) in Eq. 21, the growth rate (in nmh −1 ) due to gas-phase compound i can be Here the nuclei density ρ nuc is assumed to be 1 gcm −3 . The molecular speed c i is calculated online depending on the environment and gas-phase properties of either SA or ELVOC. Molecular weight M i for ELVOC is 248 gmol −1 and for SA 98 gmol −1 . The gas-phase concentration C i for both gases is calculated online.
By combining Eq. 11 and 13 from (Kerminen and Kulmala, 2002) we can calculate the fraction F of particles surviving 600 to diameter d p using the growth rates (GR SA ,GR ELV OC ), reduced condensation sink CS' of existing aerosol population and semi-empirical proportionality factor γ. Here we calculate CS' following Kulmala et al. (2001) and γ following Kerminen and Kulmala (2002).
This fraction is calculated and used in two phases during the NPF parameterisation. First, to calculate the fraction survinving 605 to 1.7 nm from binary nucleation. Secondly, to calcluate the fraction surviving to 5 nm from combined J 1.7 from binary nucleation and Riccobono new particle formation.
Volume fractions (V f r) for SA and ELVOC (j) in the produced 5 nm particles are assumed to be proportional to the cube of their growth rates.
Which are then normed to account for all mass in the particles: Rest of the volume is then assigned to the remaining compound In case of insufficient amount of compound i we recalculate the volume fractions so that the other gas phase compound is 615 assumed to produce the remaining growth to 5nm.
Here C i , M i and ρ i are the concentration, molecular weight and density of i, respectively. R a is the Avogadro number, v tot is the volume of the 5 nm particle. ∆t is the length of the timestep. The volume fraction for the other compound is then calculated using A5 620 Furthermore, if both gas-phase compounds are too low to uphold the formation rate at 5 nm recalculate the formation rate based on the available gas-phase compound concentrations.