FLEXPART

Lagrangian particle models compute trajectories of a large number of so-

called particles (not necessarily representing real particles, but
infinitesimally small air parcels) to describe the transport and diffusion of
tracers in the atmosphere. The main advantage of Lagrangian models is
that, unlike in Eulerian models, there is no numerical diffusion.
Numerical diffusion is a difficulty with computer simulations of continua
(such as fluids) wherein the simulated medium exhibits a higher
diffusivity than the true medium.

Advantages of Langranian Models:

 Can be computationally very efficient (depending on size of plume): only the
fraction covered with particles is simulated.
 Turbulent processes are included in a more natural way unlike Eulerian models
 There is no numerical diffusion due to a computational grid
 Grid and/or kernels are used only for output purpose therefore no artificial
diffusion is due to the averaging process
 Model is “self-adjoint” – can run backward in time, too.
 Many first order processes can be easily included with a prescribed rate:
radioactive decay, dry deposition, washout, etc.
 One particle can carry more than one species
 Gravitational settling is easily included (as long as particles carry a single species)
However: it is quite difficult and computationally expensive to include non-linear
chemical reactions and the process of gridding the output make as well loose some
of the advantages of Lagrangian modelling

FLEXPART
FLEXPART is a Lagrangian particle dispersion model that simulates the long-range and
mesoscale transport, diffusion, dry and wet deposition, and
radioactive decay of tracers released from point, line, area or volume sources.

FLEXPART can be used forward in time to simulate the dispersion of tracers from
their sources, or backward in time to determine potential source contributions for
given receptors. FLEXPART is coded following the Fortran 95 standard and tested with
several compilers (gfortran, Absoft, Portland Group) under a number of operating
systems (Linux, Solaris, Mac OS X, etc.). FLEXPART’s source code and a manual are
freely available from the internet page http://flexpart.eu. According to a user survey,
at least 34 groups from 17 countries are currently using FLEXPART. Different versions
of FLEXPART available are based on model level data of the numerical weather
prediction model of the European Centre for Medium-Range Weather Forecasts
(ECMWF), global National Centers of Environmental Prediction (NCEP) model data on
pressure levels, MM5, COSMO etc.

FLEXPART WORKING IN A NUTSHELL

The version of FLEXPART described here is based on model level data of the
numerical weather prediction model of the European Centre for Medium-Range
Weather Forecasts (ECMWF).

First:
 Correctly track the particles in a given velocity field.
Second:
Model the Sub-grid scale (SGS) unresolved physical processes that affect the
particles dispersion:
 Boundary Layer Turbulence
 Mesoscale Turbulence
 Cumulus turbulent convection
Third:
 Modify particles properties based on locally acting processes, e.g. radioactive
decay

Fourth:
 Count particles in a volume and extract concentration value

REQUIREMENTS:
FLEXPART needs five three-dimensional fields: horizontal and vertical wind
components, temperature and specific humidity. FLEXPART also needs the two-
dimensional fields: surface pressure, total cloud cover, 10 m horizontal wind
components, 2 m temperature and dew point temperature, large scale and
convective precipitation, sensible heat flux east/west and north/south surface stress,
topography, land sea-mask and subgrid standard deviation of topography.

DATA:
FLEXPART is an off-line model that uses meteorological fields (analyses or forecasts)
in Gridded Binary (GRIB) format in version 1 or 2 from the ECMWF numerical
weather prediction model (ECMWF, 1995) on a latitude/longitude grid and on native
ECMWF model levels as input. Optionally, GRIB data from NCEP’s GFS model,
available on pressure levels, can be used. The data can be global or only cover a
limited area. Furthermore, higher-resolution domains can be nested into a mother
domain.
The maximum dimensions of the meteorological field are specified by the
parameters nxmax, nymax, nuvzmax, nwzmax, nzmax in file includepar, for x, y, and
three z dimensions, respectively. All internal FLEXPART coordinates run from the
western/southern domain boundary with coordinates (0,0) to the eastern/northern
boundary with coordinates (nx-1,ny-1), where (nx,ny) are the mother grid
dimensions.
Input data must be on ECMWF model (i.e. ɳ) levels which are defined by a hybrid
coordinate system.
The conversion from ɳ to pressure coordinates is given by pk =Ak + Bkps and the
heights of the ɳ surfaces are defined by
ɳk = Ak/p0+Bk,where:
ɳk is the value of ɳ at the kth model
ps is the surface pressure and p0 =101325 Pa.
Ak and Bk are coefficients, chosen such that the levels closest to the ground follow
the topography,while the highest levels coincide with pressure surfaces; intermediate
levels transition between the two.
Velocity field Component
FLEXPART needs the vertical velocity in m s-1 because of the parametrized random
velocities
The conversion of vertical wind speeds from eta cordinate system to terrain following
system is given as:

verttransform.f
verttransform_gfs.f

Boundary layer
PBL HEIGHT
Spatial and temporal variations of ABL heights play an important role in
determining the thickness of the layer over which tracer is effectively
mixed.
The height of the convective ABL reaches its maximum value (say 1500
m) in the afternoon (say, at 1700 local time (LT)), before a much
shallower stable ABL forms. Now, if meteorological data are available
only at 1200 and 1800 LT and the ABL heights at those times are, say,
1200 m and 200 m, and linear interpolation is used, the ABL height at
1700 LT is significantly underestimated (370 m instead of 1500 m). If
tracer is released at the surface shortly before the breakdown of the
convective ABL, this would lead to a serious overestimation of the
surface concentrations (a factor of four in the above example).
Similar arguments hold for spatial variations of ABL heights due to
complex topography and variability in landuse or soil wetness.
• somewhat arbitrary parameterization is used to avoid a significant
bias in the tracer cloud thickness and the surface tracer
concentrations- Envelope ABL Height

• Here, Z is the standard deviation of the ECMWF model subgrid

topography, c is a constant (here: 2.0), V is the wind speed at
height hmix, and N is the Brunt-Vaisala frequency.
• The envelope ABL height is, thus, the diagnosed ABL height plus
the subgrid topography
• (assuming that the ABL height over the hill tops effectively
determines the dilution of a tracer cloud located in a
convectiveABL).

Transport and Diffusion

Gaussian turbulence is assumed in FLEXPART, which is strictly valid
only for stable and neutral conditions. Under convective conditions, when
turbulence is skewed and larger areas are occupied by downdrafts than by
updrafts, this assumption is violated, but for transport distances where
particles are rather well mixed throughout the ABL, the error is minor.

Mesoscale Fluctuations
Mesoscale motions eg mountain waves, pulsating drainage flows, wake
vortices, are neither resolved by the ECMWF data nor covered by the
turbulence parameterization.
This unresolved spectral interval needs to be taken into account at least
in an approximate way, since mesoscale motions can significantly
accelerate the growth of a dispersing plume.
For this, we use a similar method as Maryon (1998), namely to solve an
independent Langevin equation for the mesoscale wind velocity
fluctuations (“meandering” in Maryon’s terms). Assuming that the
variance of the wind at the grid scale provides some information on its
subgrid variance.the wind velocity standard deviation used for the
Stohl et al.: FLEXPART v9.3 description mesoscale Langevin equation is
set to turbmesoscale (set in file includepar) times the standard deviation
of the grid points surrounding the particle’s position.
The corresponding time scale is taken as half the wind interval.
REMOVAL PROCESSES
RESEARCH PAPER
Using Bayesian optimization method and FLEXPART tracer model to
evaluate CO emission in East China in springtime
ABSTRACT
This study attempted to evaluate CO emission in East China using the
analytical Bayesian inverse method and observations at Mount Hua in
springtime. The mixing ratio of CO at the receptor was calculated using 5-day
source-receptor relationship (SRR) simulated by a Lagrangian Particle
Dispersion Model (FLEXPART) and CO emission flux. The stability of the
inversion solution was evaluated on the basis of repeated random sampling
simulations. The inversion results demonstrated that there were two city cluster
regions (the Beijing–Tianjin–Hebei region and the low reaches of the Yangtze
River Delta) where the difference between a priori (Intercontinental Chemical
Transport Experiment-Phase B, INTEX-B) and a posteriori was statistically
significant and the a priori might underestimate the CO emission flux by 37 %.
A correction factor (a posteriori/a priori) of 1.26 was suggested for CO
emission in China in spring. The spatial distribution and magnitude of the CO
emission flux were comparable to the latest regional emission inventory in
Asia(REAS2.0)
CONTENT
Inverse modeling is an alternative approach which constrains a priori
information by optimizing the mismatch between simulated and observed
concentrations. This method can avoid dealing with large emissions datasets
(fuel consumption, emission factors, etc.) from a variety of sources and
categories. Inverse modeling was introduced by a book to solve linear
problems and then was used more recently to investigate emission problems in
East Asia. Satellite onboard measurements were recently incorporated in
inversing modeling, and the authors suggested that CO emission in China
might be significantly underestimatedCO emission inventory (updated to 2006)
in East Asia has been adopted by numbers of modeling studies; however, it was
still not satisfied because of large variability between simulated results and
observations.
To address this need, an analytical inversion study was carried out based on
Lagrangian dispersion model simulation and CO observations at a remote
mountain site at the westernmost point of the North China Plain, which was
identified as a potential source region for carbonaceous particle-rich plumes. A
better understanding of CO emission there would also help to improve
inventories of other combustion-related aerosols such as black carbon. This
analytical inversion in this study will mainly address the emission in springtime
and the possible adjustment of spatial distribution of CO emission in East
China.
Analytical inversion calculations
Description of the dispersion model
In this research, the spatial distribution of source–receptor relationship (SRR) was
determined by a Lagrangian Particle Dispersion Model (FLEXPART v6.2). SRR was
in unit of seconds per kilogram in particular grid cell and proportional to the particle
residence time in that cell and essentially used to simulate CO mixing ratio at the
receptor. FLEXPART model was driven by the high-resolution Weather, Research,
and Forecasting Model (WRFv3.3). Initial meteorological field data (NCEP FNL
analysis) for the WRF calculations was obtained from CISL/RDA/NCAR. The run of
model was set to a horizontal resolution of 75 by 75 km with vertical level of 45. This
model was capable of providing refined meteorological field in studying the transport
of pollutions in regional scale . In present study, 5-day SRR were determined by
releasing 10,000 particles from the observation site and mixing ratio of CO was
simulated by the spatial integration of SRR and emission flux .
A period of 5 days was proved sufficient to explain most of the CO variability at the
receptor. As known, emission strength of CO from both surface and high-altitude
point sources (chimney of power plants, industries, etc.) vary temporally. In the
present study, the inversion was tested with different footprint heights (37, 66, 110,
and 267 m). Statistical results demonstrated that the analytical inversion was
insensitive to this variation, and a footprint layer of 0–110 m was used in the analysis
described below.
The principle of inversion is to optimize the emission flux density by minimizing the
mismatch between observed and simulated concentrations (Stohl et al. 2009). The
basic equation of the inversion approach can be written as

RESULTS
CO measurements
CO measurements were performed at the summit of Mount Hua (110.07° E, 34.5° N,
2,065 m a.s.l.) from March to June2007 using a commercial gas filter nondispersive
infrared CO analyzer (Model 541, Kimoto Electric Co. Ltd.; precision, 10 ppbv). The
observatory was built at a bulk granitic rock and local sources of CO emission were
very limited in the 100-km area around the observation site because the region is
sparsely populated. The zero check was periodically performed in the first 20 min of
each hour, and span calibrations were performed by injecting standard span gas [1,040
ppbv] at the beginning and end of field measurements. Total uncertainty of the CO
measurement was estimated to be less than 5 %. Furthermore, CO spikes with
duration of 1 h were removed from the dataset because the short time peaks were not
representative of regional pollution transport processes. The total number of data
points involved in the inversion calculations was 2,153, about nine times greater than
the numbers of grid cells in the domain of interest.
Inversion results
The modeled CO mixing ratios at the receptor with a priori and a posteriori are shown
in Fig. 2. It is clear that the optimized CO emission could improve simulation
performance, especially for high-CO pollution episodes (March 10–25). Background
of CO determined by a posteriori was consistent with that by a priori since the SRR
region had weak CO emission flux. Figure 3b shows the spatial distribution of CO
emission flux for a posteriori values. The striking feature was that two important
regions (the Beijing–Tianjin–Hebei region and the lower reaches of the Yangtze
River) were identified as regions where the CO emission flux might be substantially
underestimated by a priori (Fig. 3c). CO emission of a posteriori in the central East
China Plain was slightly lower than the a priori values; nevertheless, the decrease was
statistically insignificant (within±1σ). The spatial features of the a posteriori values
were generally consistent with the latest regional emissions inventory in Asia
(REAS2.0), which is a technology-based bottom-up statistical inventory covering the
whole area of Asia (Fig. 3d). Figure 3e, f shows the adjustment for the a posteriori
(difference between a posteriori and a priori) and the adjustment for REAS2.0
(difference between REAS2.0 and a priori) for the regions where
Conclusions
In this study, Bayesian optimization method has been used to perform an inversion
analysis of CO emission in East China based on measurements at a high mountain site
(Mt. Hua) in spring 2007. The general features of the inversion results were consistent
with the latest REAS emission inventory (version 2.0). However, the present results
implied that the CO emission of two densely populated regions (the Beijing–Tianjin–
Hebei region and the low reaches of the Yangtze River Delta) might have been
underestimated significantly by bottom-up statistical methods. This study also
suggested use of a correction factor of
1.26 to adjust the CO emission in these regions in springtime.

REFERENCES
1. X. L. Pan & Y. Kanaya & Z. F. Wang & X. Tang & M. Takigawa & P. Pakpong
& F. Taketani & H. Akimoto, Using Bayesian optimization method and
FLEXPART tracer model to evaluate CO emission in East China in
springtime, Environ Sci Pollut Res (2014) 21:3873–3879

2. Stohl, A., C. Forster, A. Frank, P. Seibert, and G. Wotawa (2016):

Technical Note : The Lagrangian particle dispersion model
FLEXPART version 9.3. Atmos. Chem. Phys. 5, 2461-2474.
3. Stohl, A., M. Hittenberger, and G. Wotawa (1998): Validation of
the Lagrangian particle dispersion model FLEXPART against large
scale tracer experiments. Atmos. Environ. 32, 4245-4264