The offline FLEXible PARTicle (FLEXPART) stochastic dispersion model is currently a community model used by many scientists. Here, an alternative FLEXPART model version has been developed and tailored to use with the meteorological output data generated by the CMIP5-version of the Norwegian Earth System Model (NorESM1-M). The atmospheric component of the NorESM1-M is based on the Community Atmosphere Model (CAM4); hence, this FLEXPART version could be widely applicable and it provides a new advanced tool to directly analyse and diagnose atmospheric transport properties of the state-of-the-art climate model NorESM in a reliable way. The adaptation of FLEXPART to NorESM required new routines to read meteorological fields, new post-processing routines to obtain the vertical velocity in the FLEXPART coordinate system, and other changes. These are described in detail in this paper. To validate the model, several tests were performed that offered the possibility to investigate some aspects of offline global dispersion modelling. First, a comprehensive comparison was made between the tracer transport from several point sources around the globe calculated online by the transport scheme embedded in CAM4 and the FLEXPART model applied offline on output data. The comparison allowed investigating several aspects of the transport schemes including the approximation introduced by using an offline dispersion model with the need to transform the vertical coordinate system, the influence on the model results of the sub-grid-scale parameterisations of convection and boundary layer height and the possible advantage entailed in using a numerically non-diffusive Lagrangian particle solver. Subsequently, a comparison between the reference FLEXPART model and the FLEXPART–NorESM/CAM version was performed to compare the well-mixed state of the atmosphere in a 1-year global simulation. The two model versions use different methods to obtain the vertical velocity but no significant difference in the results was found. However, for both model versions there was some degradation in the well-mixed state after 1 year of simulation with the build-up of a bias and an increased scatter. Finally, the capability of the new combined modelling system in producing realistic, backward-in-time transport statistics was evaluated calculating the average footprint over a 5-year period for several measurement locations and by comparing the results with those obtained with the reference FLEXPART model driven by re-analysis fields. This comparison confirmed the effectiveness of the combined modelling system FLEXPART with NorESM in producing realistic transport statistics.

Transport in the atmosphere can be simulated with grid-based methods or Lagrangian particle methods, and both modelling methods have their advantages and disadvantages. An advantage of Lagrangian particle models is that they are essentially free of numerical diffusion errors (except for errors associated with interpolations), whereas Eulerian and semi-Lagrangian grid-based methods normally suffer from numerical diffusion (see, e.g., Reithmeier and Sausen, 2002; Sofiev et al. 2015) one exception being the semi-Lagrangian method of Galperin (Sofiev et al., 2015). This limits, for instance, the capabilities of Eulerian models to simulate intercontinental pollution transport (Rastigejev et al., 2010). Purely Lagrangian transport schemes become an especially attractive option when the focus is on tracer transport rather than on atmospheric chemistry. In this case, the Lagrangian scheme naturally simulates only the domain of interest achieving a high computational efficiency with no compromises in the spatial accuracy. Lagrangian particle models presently used for atmospheric transport are mostly based on stochastic approaches to describe unresolved fluctuating motions of the particles in the atmosphere. This class of models is referred to as Lagrangian stochastic (LS) models (see, e.g., Thomson, 1987; Stohl et al., 1998; Draxler, 2016; Luhar and Hurley, 2003; Lin et al., 2003; Jones et al., 2007; Rossi and Maurizi, 2014). A nice feature of the offline Gaussian LS models is that they can be run backward in time without any model changes other than the sign changes of wind components (e.g. Thomson, 1987; Flesch et al., 1995; Stohl et al., 2003; Seibert and Frank 2004). Furthermore, atmospheric turbulence (in the boundary layer) can be treated more accurately in LS particle models than in grid-based dispersion models. All three features – minimal numerical diffusivity, possibility of time-reversed transport, and accurate turbulence description – are particularly attractive for atmospheric inversion studies, where sources of emissions (e.g. of greenhouse gases) are determined by combining information from atmospheric measurements and dispersion models. Not surprisingly, LS models are popular tools for such studies (e.g. Gerbig et al., 2003; Thompson and Stohl, 2014; Henne et al., 2016).

Therefore, and for many other purposes, many different offline Lagrangian models have been developed, probably the most popular being the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model (Draxler, 2016), the Stochastic Time-Inverted Lagrangian Transport (STILT) model (Lin et al., 2003), and the FLEXible PARTicle (FLEXPART) stochastic dispersion model (Stohl et al., 1998). In this paper, we concentrate on the FLEXPART model. The reference version of FLEXPART (Stohl et al., 1998, 2005; Stohl and Thomson, 1999) uses global meteorological data from the European Centre for Medium-Range Weather Forecasts (ECMWF) or the National Center of Environmental Prediction (NCEP). It is a versatile tool that has been applied in many different fields of atmospheric research ranging from classical pollution dispersion modelling to measurement data interpretation studies, inverse modelling, or studies of the hydrological cycle.

FLEXPART has become a community model. Scientists from several countries
contribute to its development and share model versions and model branches on
the website (

The current paper describes a new branch of FLEXPART that uses output data from the Norwegian Earth System Model (NorESM1-M; Bentsen et al., 2013; Iversen et al., 2013), which is based on the Community Climate System Model (CCSM4; Gent et al., 2011; Vertenstein et al., 2010). In NorESM, the Community Atmosphere Model (CAM4; Neale et al., 2010) is modified to include the aerosol module developed for NorESM (CAM4-Oslo; Seland et al., 2008; Kirkevåg et al., 2013). This version of FLEXPART is named FLEXPART–NorESM/CAM, and is tailored particularly to climate applications and adjusted to use the NorESM1-M and CAM4 output coordinate system and data formats. The combination of the NorESM1-M climate model and FLEXPART allows for Lagrangian transport analysis for periods where no re-analysis data might be available, e.g. paleoclimate analysis and future climate projections.

In Sect. 2, we introduce FLEXPART–NorESM/CAM while a description of relevant aspects of NorESM1-M and CAM4-Oslo is reported in Appendix A and some technical details of the model coupling are reported in Appendix B. In Sect. 3, we validate FLEXPART–NorESM/CAM as a tool for transport diagnostics, by comparing its Lagrangian offline tracer dispersion calculations with the finite volume online tracer calculations used in CAM4. The comparison includes tracer releases from several point sources around the globe, and it allows for evaluating (i) the correctness of this FLEXPART version, (ii) the differences introduced by the need to transform the vertical coordinate system and obtain an appropriate vertical velocity, (iii) the use of two methods to obtain the vertical velocity in FLEXPART–NorESM/CAM, (iv) the influence on the models results of different sub-grid-scale (SGS) parameterisations for convection and boundary layer height, and (v) the possible advantages entailed in using a numerically non-diffusive Lagrangian particle solver instead of a grid-based solver. The maintenance of the well-mixed state (e.g. Thomson, 1987) of the particles is also investigated in Sect. 3, comparing the results of the reference FLEXPART model and the new version in a 1-year global simulation of the whole atmosphere up to about 20 km aboveground. Such a global scale well-mixed test was not previously done, to our knowledge, for an offline Lagrangian stochastic dispersion model, and it is relevant in case there is a need to simulate long-term evolution of the whole atmosphere and eventually include chemistry and mixing in a Lagrangian framework (e.g. Collins et al., 1997; Reithmeier and Sausen, 2002; Stenke et al., 2009). In Sect. 4, we compare climatological FLEXPART–NorESM/CAM calculations with equivalent calculations done with the reference FLEXPART model version driven with atmospheric re-analysis data to test the effectiveness of the combined modelling system FLEXPART–NorESM in producing realistic transport statistics. Finally, in Sect. 5, we draw conclusions.

FLEXPART–NorESM/CAM has been developed on the basis of FLEXPART version 9.1,
which can be used with meteorological data from ECMWF or NCEP. No detailed
separate documentation of FLEXPART version 9.1 exists but the code is
available from

The FLEXPART model is coded in Fortran 95. The physics of the FLEXPART model is described in detail in Stohl et al. (2005), although since then many improvements have been introduced. FLEXPART is a Lagrangian stochastic particle model and uses the well-mixed criteria of Thomson (1987) and diffusion coefficients to define the motions of notional fluid particles. The stochastic components simulate the effects of the planetary boundary layer (PBL) turbulence and unresolved mesoscale motions. We underline that the model for the PBL accounts for different stabilities, including the possibility of skewed turbulence in convective conditions, and for the vertical air density gradient (Stohl and Thomson, 1999; Cassiani et al., 2015). More details of the PBL turbulence parameterisations can be found in the FLEXPART user's manual. FLEXPART simulates deep moist convective exchanges using a non-local transilient transport matrix constructed consistently with the scheme of Emanuel and Živković-Rothman (1999) (see Forster et al., 2007), and includes dry/wet deposition processes and linear chemical and radioactive loss processes. Overall, the Lagrangian representation used in the model does not have significant numerical diffusivity, and well preserves the structures generated during the advection and dispersion processes and this will also be shown in detail below and compared to the finite volume solver of NorESM. A grid or weighting kernel is used only to extract statistically averaged information from the particles at the required spatial resolution. The model can run both in forward or backward mode to study dispersion from a source or receptor respectively (e.g. Thomson, 1987; Flesch et al., 1995; Stohl et al., 2003).

FLEXPART uses a terrain-following vertical coordinate system

From the original NorESM and CAM model output fields, a few more fields necessary to run the model, are obtained by online post-processing: the vertical velocity, the 10 m wind, and the dew point.

The vertical velocity provided by NorESM and CAM output is

From the definition of total derivative in the hybrid

From this definition, the quantity

Equation (4) has been discretised with a simple finite difference scheme in
space and time,

The second method uses explicitly the hydrostatic approximation, i.e.

The vertical velocity

Other minor modification necessary to use FLEXPART with the wind fields
generated by NorESM include a procedure to obtain the 10 m wind components
and the dew point. The 10 m wind components are obtained from the 10 m wind
velocity module and the surface stresses as follows:

In Eq. (7) the prime denotes a fluctuation from the resolved mean quantity
and the overbar is the averaging operator; therefore,

The dew point is obtained as (see, e.g., Campbell and Norman, 1998, p. 43)

Despite this model version being similar to the reference FLEXPART model, several changes have been introduced. These changes needed testing to verify the correctness and performance of the new model and created also the opportunity for models comparisons. First, the results obtained by the online advection and diffusion equation solver in NorESM and the offline Lagrangian particle tracking and diffusion used in FLEXPART–NorESM/CAM were compared in numerical tracer experiments. In doing this, we investigated the effects on the results of the change of vertical coordinate system, SGS convection scheme, and PBL depth parameterisation, which are different between the offline FLEXPART–NorESM/CAM and the online transport scheme of NorESM. Furthermore, with the aim of validating the procedure used to obtain the vertical velocity, and the possible numerical errors introduced by this procedure, a comparison between the vertical well-mixed state in FLEXPART–NorESM/CAM and the reference FLEXPART–ECMWF has been performed for a 1-year global domain-filling simulation of the atmosphere. To our knowledge such quantitative and comparative tests on the vertical well-mixed state were never before presented in the literature for a global-scale offline Lagrangian particle model over a 1-year simulation.

In this section we compare the results for scalar tracer transport obtained using the NorESM online grid-based advection and diffusion equation solver (see Appendix A) and the results obtained offline using the fully Lagrangian FLEXPART–NorESM/CAM. An integrated solver is generally more consistent with the dynamics of the model than a post-processor (see, e.g., Byun, 1999; Ngan et al., 2015), which, moreover, in the present case works with a different coordinate system. Furthermore, online tracer advection and diffusion in NorESM is calculated with the model's 30 min time step, whereas FLEXPART–NorESM/CAM is driven offline with 3-hourly NorESM output fields. FLEXPART uses here a 5 min time step for the integration of the stochastic differential equations for particle velocities with a refinement to 75 s for the fluctuating vertical velocity component in the PBL. However, resolved scale processes, numerical diffusion and parameterised processes all influence the modelled dispersion of a tracer in the atmosphere. In this respect Lagrangian particle tracking has some advantages compared to a grid-based solver including negligible numerical diffusivity and a much better framework for including turbulence parameterisations especially in the PBL (e.g. Thomson, 1987; Stohl and Thomson, 1999; Cassiani et al., 2015). SGS parameterisations of PBL turbulence and depth, topographical effects on PBL depth, mesoscale horizontal motions and moist convection are all important for tracer dispersion, and they are different in the online NorESM transport scheme and the offline FLEXPART–NorESM/CAM. Thus, while it may not be clear which transport scheme performs better overall, it is important to ensure the overall consistency of the two schemes.

For comparisons of the two transport schemes, tracer dispersion from 15
sources on the globe was calculated using both the NorESM online transport
scheme and FLEXPART–NorESM. Passive tracers were released from
1.89

List of point source locations used in the performed model test.

We begin our analysis with the results obtained for a continuous release at the ALP location. Here, by switching on and off the parameterisation, SGS parameterisation of topography was tested to have a minor role and the moist convection scheme was found as well to have some but not a dominant influence. Therefore, resolved scale motions dominate this plume dispersion. The topography at the source is also intermediate among the examined cases. Figure 1 shows the vertically integrated and meridionally integrated tracer concentrations after 144 h of simulation. For FLEXPART–NorESM/CAM results obtained with both the methods to obtain the vertical velocity are shown. First, we note that the two methods to obtain the vertical velocity produce almost indistinguishable results. The tracer fields by NorESM and FLEXPART–NorESM/CAM are also generally very similar, and the major differences can be attributed to the higher numerical diffusivity of the online finite volume-based solver, a point that will be further investigated in Sect. 3.1.4. The purely Lagrangian particle solver shows the characteristic ability to maintain a sharper gradient (e.g. Reithmeier and Sausen, 2002) and preserve filamentary structures that are generated by chaotic advection (Ottino, 1989), which grid-based models have difficulties to capture. The results shown in Fig. 1 can be considered typical, i.e. with no extreme topography and without substantial influence from SGS parameterisations. Below, we will investigate cases that allow for highlighting particular aspects of tracer transport.

Tracer dispersion for a continuous release from the ALP source, with
the NorESM online tracer solver

Here we examine the results obtained for a continuous release in the SUM location. This release point is over high topography where the transformation of the vertical velocity has a significant role and can potentially create larger errors than over flat and low terrain. Moreover, this location was selected since the roles of SGS parameterisation of topography and convection were found (by switching on and off the parameterisations in separate tests; not shown) to be completely negligible in the FLEXPART–NorESM/CAM simulations. Although the same test was not performed for the NorESM native solver, it is reasonable to assume that the role of the convection scheme was also minor in that case. In support of this assumption, we note that Tost et al. (2010) found that both the Zhang and McFarlane (1995) and the Emanuel and Živković-Rothman (1999) convection schemes hardly calculate any convection in polar regions. Therefore, for this case the SGS physical parameterisations can be considered negligible while at the same time the transformation of vertical coordinate is very important. In Fig. 2, the vertically integrated and meridionally integrated tracer concentrations are shown after 144 h of simulation. It is possible to see again that the two methods to obtain the vertical velocity in FLEXPART–NorESM/CAM produce almost indistinguishable results. The tracer fields simulated by NorESM and FLEXPART–NorESM/CAM are very similar, the difference being even smaller than in the previous (ALP) case examined, and this shows that in general the different model parameterisations introduce some deviations in the results, while the differences induced by the transformation of vertical velocity are in comparison of minor importance. The effects of parameterisations will be further investigated below in Sect. 3.1.2 and 3.1.3. In this SUM case the effect of the higher numerical diffusivity of the grid-based online solver is even clearer than in the ALP case.

As in Fig. 1 but for SUM.

Tracer dispersion for a continuous release from the HIM source, with the NorESM online tracer solver (left panels) vs. Lagrangian tracking with FLEXPART–NorESM/CAM, with vertical velocity obtained with method one (right panels). Upper panels show maps of vertically integrated tracer concentrations, the lower panels show meridionally integrated tracer concentrations as a function of longitude and altitude, 144 h after the start of the simulation.

A further test of the correctness of the transformation of vertical velocity over topographic slope regards the HIM release point, which is over the highest NorESM grid point of the whole domain. By switching on and off the convection scheme in FLEXPART–NorESM/CAM, we have found that the convection scheme in this case has a minor but not completely negligible influence, similar to the ALP case investigated above. In any case, the agreement between NorESM and FLEXPART–NorESM/CAM is again very good, with differences mainly generated by numerical diffusivity. The results of FLEXPART–NorESM/CAM when using the two different methods to obtain the vertical velocity were again almost indistinguishable hence in Fig. 3, and in the remainder of the paper, only results obtained using the first method are shown.

Overall, for both these cases of high topography the agreement between the online and offline transport schemes is remarkable. These initial tests grant confidence that both the procedures used to obtain the vertical velocity in the FLEXPART terrain-following coordinate are appropriate and do not add any significant errors and that the use of an offline transport scheme is adequate. Moreover, the Lagrangian solver is able to better maintain the filamentary structure generated by chaotic advection.

Continuous release from the EQA source. FLEXPART–NorESM/CAM results without convective scheme (right panels) and with convective scheme (central panels) vs. NorESM integrated solver results (left panels) are shown. The upper panels show maps of vertically integrated tracer concentrations, the lower panels show meridionally integrated tracer concentrations as a function of longitude and altitude, 144 hours after the start of the simulation.

In model runs with horizontal resolution of
1.89

As can be seen in Fig. 4, comparing right and central panels, the deep convection scheme in this case totally alters the vertical (and consequently also the horizontal) distribution of the tracer by lifting it up to 20 km compared with an uplifting of less than 5 km without it. The results of FLEXPART–NorESM/CAM with the convection activated and NorESM are quite similar, despite the different convection schemes used. However, the convection scheme in FLEXPART transports a larger mass fraction above about 7.5 km, which may be due to simulation of overshooting convection with the Emanuel and Živković-Rothman (1999) parameterisation. This in turn generates a faster horizontal advection with higher total column integrated concentration extending over the central region of the African continent.

For all the tropical sources, the convection scheme dominates the dispersion and the relative behaviour of the two models in all these cases is similar, but not perfectly consistent. For example, Fig. 5 shows the comparison for a short (30 min) release in the GUI location after 144 h of dispersion. As in the EQA case a higher overall plume vertical extension in FLEXPART–NorESM/CAM with respect to NorESM can be observed. However, in the GUI release the mass fraction transported above about 7.5 km in FLEXPART–NorESM/CAM is lower than that of NorESM. The overall agreement is good, considering that different convection schemes are used, but some discrepancies can be observed. Near the east coast of Australia and north-west of Madagascar FLEXPART–NorESM/CAM simulates significantly higher column integrated concentrations than the NorESM integrated solver, while NorESM simulates higher column integrated concentrations close to the north-west coast of Australia and over Malaysia.

Puff dispersion after 144 h from the GUI source. NorESM integrated solver results (left panels) and FLEXPART–NorESM/CAM results (right panels) are shown. The upper panels show maps of vertically integrated tracer concentrations, the lower panels show meridionally integrated tracer concentrations as a function of longitude and altitude.

Tracer dispersion for a 30 min release from the ALE point after 144 h of simulated time, with the NorESM online tracer solver (left panels) vs. Lagrangian tracking with FLEXPART–NorESM/CAM. FLEXPART results are shown with (middle) and without (right) the SGS topography PBL scheme.

These comparisons show that the convection scheme can be fundamental in large-scale dispersion and may overshadow the influence of other possible sources of discrepancy between the offline and online transport schemes including numerical diffusivity.

Boundary-layer depths after 21 h of simulation for FLEXPART–NorESM/CAM without parameterisation of SGS topography (right), with parameterisation of SGS topography (middle), and for NorESM (left). The red dot indicates the source location.

FLEXPART and NorESM have a similar formulation for the definition of the boundary layer depth, both based on Vogelezang and Holtslag (1996), which uses the critical Richardson's number concept. In all the simulations, but one, of those listed in Table 1, the differences between the scalar transport in the two models could be explained in terms of the difference in numerical diffusivity and/or deep convection scheme. In this section, we will investigate the exception where small differences in the height of the boundary layer top triggered visible differences in the scalar dispersion. FLEXPART has the option to activate an additional SGS parameterisation of the topography effects that increases the boundary layer depth in presence of significant unresolved topography (see FLEXPART user guide). The role that the SGS parameterisation of topography may have on the scalar dispersion in FLEXPART–NorESM/CAM is shown in Fig. 6 for a short duration (30 min) release from the ALE source. The right and middle panels show the result from FLEXPART–NorESM/CAM without and with the direct additional accounting of SGS topography in PBL height respectively, while the left panel shows the results for NorESM. Comparing the right and central panels, it is clear that in this case the SGS topography effect is substantial. It generates a higher boundary layer allowing the tracer to reach an altitude with different horizontal transport. The better consistency between NorESM and FLEXPART–NorESM/CAM with the SGS topographic effects included is a bit surprising given that NorESM does not have a direct additional accounting of SGS orography in the calculation of the PBL height.

However, Fig. 7 shows the PBL depths in FLEXPART calculated when the dispersion of the puff, with and without the SGS parameterisation of topography, initially start departing. The PBL depths seems in better agreement with those in NorESM if the SGS parameterisation is used. This is visible along the coast of Greenland including the north-east of Greenland, where the puffs tracer distribution originally started deviating. This could be related to the coarser time resolution (3 h) of the meteorological field used in FLEXPART–NorESM/CAM compared to the native resolution (30 min) of NorESM. The FLEXPART SGS parameterisation of topography is indeed intended to also compensate the coarser time resolution. However, we note that in general the PBL depths calculated by FLEXPART with the additional SGS parameterisation of topography activated are larger than those calculated in NorESM.

Results for the 30 min release from ALE (puff, left panels) and the continuous release from SUM (plume, right panels) as simulated with FLEXPART–NorESM/CAM with a 10-fold increase in the horizontal diffusivity coefficients. Compare with NorESM results in Figs. 6 and 2.

In the sections above we have discussed the importance of the convection scheme and a case where the use of the SGS parameterisation of topography played a role. However, in most cases the differences in the results between FLEXPART–NorESM/CAM and NorESM can be attributed to their different diffusivity. FLEXPART, as any Lagrangian particle model, is only minimally affected by numerical diffusivity (see, e.g., Reithmeier and Sausen, 2002). Therefore, in FLEXPART horizontal motions unresolved in the driving meteorological fields need to be parameterised as an additional horizontal diffusivity (see FLEXPART user guide). NorESM, on the other hand, does not have any parameterisation of these effects and the horizontal diffusivity of the scalar is purely numerical. In the following test the puff release from the ALE point (see Fig. 6) and the plume from the SUM point (see Fig. 2, right panels) have been re-run with a 10-fold increase in the horizontal diffusivity coefficients in FLEXPART. The comparison shows that the results with the increased diffusivity in FLEXPART (Fig. 8) are considerably more similar to those of NorESM (Figs. 2 and 6, left panels). Albeit simple, this test shows that, for the resolution and numerical scheme used in NorESM, the horizontal diffusion in NorESM is about 10 times stronger than the parameterised diffusion of unresolved scales used in FLEXPART. These results agree, at least qualitatively, with the findings of Reithmeier and Sausen (2002) where, for the ECHAM model, a better behaviour was found for their online Lagrangian transport scheme compared to the semi-Lagrangian scheme (grid based) due to an excessive numerical diffusion in the latter one. From the point of view of model inter-comparison (and FLEXPART–NorESM/CAM validation), it is reassuring that a large part of the difference between the two models can be explained in terms of just horizontal diffusivity. This comparison also shows the added value of the offline Lagrangian model, as the numerical diffusion in NorESM is likely much too strong and leads to a too rapid destruction of filamentary tracer structures. By running FLEXPART–NorESM/CAM with NorESM output, such structures can be recovered.

In FLEXPART–NorESM the vertical velocity in the terrain-following
coordinates is obtained from the vertical velocity in pressure coordinates
(omega) available in the NorESM output files. As shown previously, this does
not introduce significant errors in short-term simulations. However, in
longer time simulations it may introduce errors in the vertical mass
distribution that in Lagrangian stochastic modelling is often described as a
perturbation of the vertical well-mixed state of the model (e.g. Thomson,
1987). In order to test this, the well-mixed state in FLEXPART–NorESM/CAM
and the standard FLEXPART–ECMWF model are compared. FLEXPART–ECMWF uses a
different method to calculate the required vertical velocity in the
terrain-following coordinates, which is based on the velocity in the original

The well-mixedness of the tracer in physical space requires that the
probability density function (pdf) of particle positions is proportional to
the mean air density at any time if this was initially so (e.g. Thomson,
1987, 1995; Pope, 1987; Stohl and Thomson, 1999; Cassiani et al., 2015). For
a single atmospheric column with surface area

Here

Scatter plot of volume-averaged air density calculated from meteorological data of the driving model (NorESM and ECMWF), and using particle counting in FLEXPART according to Eq. (10). The linear regressions are shown as red lines.

Figure 9 shows a scatter plot between the volume-averaged air density
calculated as volume average of the driving model (using random samples) and
using particle counts as reported in Eq. (10). The results are for vertical
layers up to about 20 km. At the beginning of the simulation, the particles
are initialised according to air density and any disagreement between air and
particle density is due only to discretisation and statistical noise. This
is, fairly well, confirmed by the linear regression in Fig. 9, which gives

Conserved tracer-averaged residence time in the lowest 1 km of the
atmosphere for 21-days retroplume, for four stations. Results are averaged
over 5 years (1995–2000) for the months JJA. The values are expressed as
the logarithm (base 10) of the residence time in seconds divided for the
surface of the grid cell in km

The previous comparisons have shown that FLEXPART–NorESM/CAM is technically working as expected. However, to test whether the combined modelling system FLEXPART with NorESM can also provide realistic transport climatologies, a further model inter-comparison between FLEXPART–NorESM/CAM and FLEXPART–ECMWF driven by ERA-Interim meteorology (Dee et al., 2011) was performed. This test also indirectly compares the climatologies of meteorological variables in NorESM and ERA-Interim that are important for driving FLEXPART (especially the wind data). For our comparison, we chose one of the most common applications of global-scale Lagrangian particle models, the modelling of retroplumes (Stohl et al., 2003). Retroplumes from a measurement site, for instance, are often used to establish transport climatologies, to identify pollution sources, and to quantify source contributions to pollution at the site. If sources are at the surface, typically the FLEXPART output for the lowest model layer (so-called footprint) is of greatest interest. Therefore, the goal of this comparison is to gain confidence in the performance of FLEXPART–NorESM/CAM for generating footprint climatologies (see, e.g., Hirdman et al., 2010) by applying it to recent historical periods for which ERA-Interim data are also available.

As in Fig. 10 but for DJF.

Comparison between the FLEXPART–NorESM/CAM and the FLEXPART–ECMWF
(driven by ERA-Interim) results for the 5-year-averaged residence time in the
lowest 1000 m of the atmosphere. Correlation (Corr) and fraction of data
within a factor of 2 (FA2) for JJA and DJF are reported for both the
conserved and depositing (black carbon) tracer. The

In this section backward trajectories have been calculated with
FLEXPART–NorESM/CAM and FLEXPART–ECMWF driven with ERA-Interim data for six
observatories: ALE, BAR, SUM, TRO, ZEP (listed in Table 1), and Birkenes (BIR;
58

FLEXPART has the possibility to treat physical loss processes of a tracer,
such as wet and dry deposition. Therefore, we simulated both a passive tracer
and a species resembling the behaviour of black carbon, which is subject to
wet and dry deposition. Figures 10 and 11 show the results obtained for the
conserved tracer and for four stations for JJA and DJF respectively. The
residence time (in seconds) in the grid cell has been scaled by the grid
surface area (in km

The residence times obtained for the black-carbon-like tracer subject to wet
and dry deposition (not shown here) are marginally less consistent between
the two models. This is to be expected since differences in the precipitation
patterns, especially, lead to differences in wet deposition and thus tracer
patterns. However, the overall picture is fundamentally unchanged and the
level of consistency between the results of the two models is still high.
This can be seen in Table 2, which reports a statistical comparison between
the two models for both DJF and JJA and for both conserved and depositing
tracer. A threshold value of 1 s for the averaged residence time in the grid
cell is used. The Pearson correlation coefficient (Corr) and the fraction of
data within a factor of 2 (FA2) are shown. For the conserved tracer the
correlation varies between 0.84 and 0.99 and FA2 values are between 68 and
96 %. For the depositing tracer, the correlation varies between 0.85 and
0.98 and FA2 values between 56 and 91 %. The

We have developed a version of FLEXPART, FLEXPART–NorESM/CAM, that can ingest input data from the NorESM1-M/CAM4-Oslo model. This provides a new advanced tool to directly analyse and diagnose atmospheric transport properties of the climate model NorESM in a reliable way and can be used for climatological studies. To validate this newly developed FLEXPART model version, we performed multiple comparisons both with the online transport scheme embedded in NorESM and with the reference FLEXPART version. From these comparisons, we can draw the following conclusions.

Comparison between online tracer calculations with the grid-based (vertically semi-Lagrangian) advection scheme built into NorESM1-M and the offline Lagrangian particle tracer calculations in FLEXPART–NorESM/CAM showed very good agreement, even for releases over high terrain where errors introduced by the transformation of the vertical coordinate system are largest. In fact, in most cases the largest differences between the offline Lagrangian and online grid-based transport schemes were attributed to the higher numerical diffusion in the online finite volume transport scheme. This was proven by artificially enhancing the horizontal diffusion coefficients in the Lagrangian model calculations by 1 order of magnitude, which resulted in much closer agreement between the two methods. This proves a possible added value of dispersion calculations for non-reactive tracers done with FLEXPART–NorESM/CAM compared to online tracer calculations.

NorESM1-M/CAM4-Oslo and FLEXPART–NorESM/CAM use different convection schemes. Nevertheless, the tracer transport is similar both with the online and LS offline transport schemes, even for release locations strongly affected by deep convection. Conversely, switching off the convection scheme in FLEXPART–NorESM/CAM leads to very large differences from both online and LS transport using the convection scheme.

Tests have also been performed on the vertical well-mixedness of particles in longer-term (1 year) simulations. While some deviations from well-mixedness occur, these are of a similar magnitude as in FLEXPART–ECMWF, confirming that the vertical coordinate transformation in FLEXPART–NorESM/CAM does not introduce additional errors. However, the deviations from well-mixed state may be relevant for some applications.

A further model inter comparison was done for one of the most common applications of offline Lagrangian models: backward in time calculations of the retroplume and the footprint emission sensitivity for specific measurement stations. The inter-comparison between FLEXPART–NorESM/CAM and FLEXPART–ECMWF driven by ERA-Interim meteorology showed that FLEXPART–NorESM/CAM provided realistic transport climatologies for the years 1995–2000. This lends confidence to using this combined tool for climatological analyses.

The current FLEXPART–NorESM/CAM model is based on the FLEXPART 9.1 model, which is a purely serial code. Lagrangian one-particle models, such as FLEXPART, are well suited for trivial parallelisation by running multiple instances of the same simulation with a different independent random number string. However, a parallel (MPI-based) version of FLEXPART has been recently developed, which can automatise parallelisation. Therefore, we aim to include the parallelisation in the FLEXPART–NorESM/CAM model branch in the near future.

The FLEXPART–NorESM/CAM model branch can be downloaded from

This study uses the CMIP5-version of NorESM used for concentration-driven greenhouse gas experiments, NorESM1-M, which is thoroughly presented by Bentsen et al. (2013) and Iversen et al. (2013). It belongs to the family of state-of-science global climate and Earth system models that contributed to the Coupled Model Intercomparison Project Phase 5 (CMIP5; Taylor et al., 2012). The basis for NorESM1-M is the Community Climate System Model (CCSM4; Gent et al., 2011). The ocean component in CCSM4 is replaced by a different ocean model (NorESM-O), which is an elaborated version of the Miami Isopycnic coordinate Ocean Model (MICOM; Bleck et al., 1992) adapted for multi-century simulations in coupled mode by Assmann et al. (2010) and Otterå et al. (2009). Further extensions are described by Bentsen et al. (2013). The atmospheric component is CAM4-Oslo and includes an advanced aerosol–cloud–chemistry representation (Seland et al., 2008; Kirkevåg et al., 2013) into the version of CAM4 included in CCSM4 (Neale et al., 2010). The land and sea ice components – the Community Land Model (CLM4; Lawrence et al., 2011) and the Los Alamos Sea Ice Model (CICE4; Holland et al., 2012) – are the same as in CCSM4, except that aerosol deposition on snow (Flanner and Zender, 2006) and sea ice are treated prognostically in NorESM1-M, and minor adjustments of parameters for the thermodynamic properties of snow on sea ice are made.

With respect to physical processes in the atmosphere not directly associated
with the aerosols, CAM4-Oslo applies the standard configuration of CAM4. This
includes the Rasch and Kristjansson (1998) scheme for stratiform cloud
processes and the CAM-RT radiation scheme. Deep convective clouds are
parameterised following Zhang and McFarlane (1995) extended with the plume
dilution and convective momentum transport also used in CCSM4 (Richter and
Rasch, 2008; Neale et al., 2008). Shallow convection follows a
parameterisation by Hack (1994). The turbulence parameterisation includes
computation of diffusivities for the free atmosphere, based on the gradient
Richardson number, and an explicit, non-local atmospheric boundary layer
parameterisation (Holtslag and Boville, 1993). The finite volume (FV)
dynamical core with a semi-Lagrangian approach with re-interpolation to the
grid for the vertical (Rasch et al., 2006; Lin, 2004) is used for transport
calculations. The time integration within the FV dynamics is fully explicit,
with sub-cycling within the 2-D Lagrangian dynamics. However, the transport
for tracers can take a larger time step (e.g. 30 min) as for the physics.
The horizontal resolution is 1.89

Two types of simulations have been used for adapting and testing FLEXPART
with NorESM1-M and CAM4-Oslo. In a first experiment, the period 1990–2070
has been simulated (scenario RCP6.0 from year 2005 onward) using the
fully-coupled NorESM1-M. The time-evolution of global mean atmospheric
concentrations of the greenhouse gases CO

We emphasise that FLEXPART–NorESM/CAM does not depend on any NorESM1-M-specific features and can be driven with output generated by either NorESM or the Community Earth System Model (CESM; Lindsay et al., 2014), the successor of CCSM4. However, CCSM4 itself lacks the capability to output 10 m wind speed, and a code modification is therefore needed in order to produce output suitable for FLEXPART–NorESM/CAM. The following instantaneous 3-hourly fields need to be specified in CAMs input namelist: PS:I, U:I, V:I, OMEGA:I, T:I, Q:I, CLDTOT:I, U10:I, TREFHT:I, PRECL:I, PRECC:I, SHFLX:I, TAUX:I, TAUY:I, QREFHT:I, SNOWHLND:I, FSDS:I. Here “:I” means instantaneous field and a full description of the field is give in Appendix B, Table B1.

This work has been supported by the Research Council of Norway through the projects EarthClim and EVA (grant no. 229771), by Nordforsk through the Nordic Centers of Excellence ESTICC (grant no. 57001) and the project CRAICC, and through the European Commission FP7 projects PEGASOS (FP7-ENV-2010-265148), and ACCESS (FP7-ENV-2010-265863). Computational and storage resources for NorESM simulations have been provided by NOTUR (nn2345k) and NorStore (ns2345k). Edited by: V. Grewe Reviewed by: two anonymous referees