An Overview of the HYSPLIT_4 Modelling System

for Trajectories, Dispersion, and Deposition

Roland R. Draxler

NOAA Air Resources Laboratory

Silver Spring, Maryland, U.S.A.

and

G.D. Hess

Bureau of Meteorology Research Centre

Melbourne, Australia

Australian Meteorological Magazine, 47 (1998) 295-308

(Manuscript received June 1997; revised January 1998)

 An Overview of the HYSPLIT_4 Modelling System
for
Trajectories, Dispersion, and Deposition

Abstract
The HYSPLIT_4 (HYbrid Single-Particle Lagrangian Integrated Trajectory) model is designed
for quick response to atmospheric emergencies, diagnostic case studies, or climatological analyses
using previously gridded meteorological data. Calculations may be performed sequentially on
multiple meteorological grids, going from fine to coarse resolution using either archive or forecast
data fields. Air concentration calculations associate the mass of the pollutant species with the
release of either puffs, particles, or a combination of both. The dispersion rate is calculated from
the vertical diffusivity profile, wind shear, and horizontal deformation of the wind field. Air
concentrations are calculated at a specific grid point for puffs and as cell-average concentrations
for particles. The model results are evaluated against ACE balloon trajectories, air concentrations
from the ANATEX tracer experiment, radiological deposition from the Chernobyl accident, and
satellite photographs of the Rabaul volcanic eruption. One common feature of the model results
was their sensitivity to the vertical atmospheric structure; trajectories in terms of their height
when near ground-level due to the strong gradients of wind speed and direction, air
concentrations with respect to the rate of vertical mixing, and deposition as a result of the vertical
distribution of the pollutant.

Introduction
Atmospheric dispersion modelling, the use of numerical methods to compute the time
history of air pollutant concentrations, is usually divided into two broad categories. Eulerian
models, which solve the advection-diffusion equation on a fixed grid, and Lagrangian models, in
which the advection and diffusion components are calculated independently. Eulerian methods
are typically applied when complex emission scenarios are considered, requiring solutions at all
grid points. Lagrangian methods are typically favored when single-point-source emissions restrict
computations to a few grid points. Further, Eulerian models generally require emissions to be
defined on a scale comparable to the model’s computational grid while Lagrangian models can
define the emissions at any resolution. These considerations aside, a review of the literature will
show that both methods have been successfully applied to a variety of different scenarios. The
intent in this paper is to review the details of a Lagrangian dispersion model (HYSPLIT_4 -
HYbrid Single-Particle Lagrangian Integrated Trajectories Version 4) designed to support the
Bureau of Meteorology (BoM) in its need to respond to atmospheric emergencies ranging in
character from accidental radiological releases to the hazards presented to aircraft operations from
volcanic ash eruptions.

The model calculation method is a hybrid between Eulerian and Lagrangian approaches.
Advection and diffusion calculations are made in a Lagrangian framework while concentrations
are calculated on a fixed grid. The model has evolved over several stages during the last decade
and a half. The initial version of the model (Draxler and Taylor, 1982) used only rawinsonde
observations and the dispersion was assumed to consist of uniform mixing during the daytime and
no mixing at night. Dispersion due to wind shear was introduced by splitting up the daytime
mixed layer into smaller layers each night. In the next revision (Draxler and Stunder, 1988),
variable strength mixing was introduced based upon a temporally and spatially varying diffusivity
profile. In the last revision (HYSPLIT_3 - Draxler, 1990; 1992), the use of rawinsonde data was
replaced by gridded meteorological data from either analyses or short-term forecasts from routine
numerical weather prediction models. It was shown (Draxler, 1991) that these routine products
provided calculations that were as accurate as those made using rawinsondes, even when special
off-time (0600 and 1800 UTC) soundings were included in the calculations.

Several significant new features have been added to the current version of the model. In
particular, the advection algorithms have been updated to include temporal interpolation and
although the dispersion of a pollutant can still be calculated by assuming the release of a single
puff, the puff can now be defined with either a Gaussian or top-hat horizontal distribution. As in
previous versions, a single released puff will expand until its size exceeds the meteorological grid
cell spacing, and then it will split into several new smaller puffs. Further, a three-dimensional
particle dispersion routine has been added that computes air concentrations from the dispersal of
an initial fixed number of particles. The particle approach includes a modification, following
Hurley (1994), that combines both puff and particle methods by assuming a puff distribution in the
horizontal and particle dispersion in the vertical direction. This hybrid calculation may be started
with a single particle. In this way, the greater accuracy of the vertical dispersion parameterization
of the particle model is combined with the advantage of having an expanding number of puffs
represent the pollutant distribution as the spatial coverage of the pollutant increases. Perhaps
more significantly, all the equations used to compute the strength of the vertical mixing have been
revised based upon the more recent literature, and the rate of horizontal dispersion is no longer
assumed to be constant but can vary according to the deformation of the wind field. In general
only the features of the model that have changed significantly from the previous HYSPLIT
version will be discussed in any detail in the following sections, however a complete technical
description of the model, including all equations that have been omitted here, can be found in
Draxler and Hess (1997).

Trajectories
The basis of any Lagrangian model is that the dispersion is computed following the
particle or puff. That is, the advection of a particle is computed independently of the dispersion
calculation. The time integrated advection of each particle can be viewed as a simple trajectory
which only requires the three-dimensional velocity field.

Meteorological data

Meteorological data for trajectories and dispersion calculations are obtained as output

3
fields from meteorological models. Usually these fields cannot be directly used by HYSPLIT
without some pre-processing. Although most meteorological models use some variation of a
terrain-following coordinate system, the data fields are usually interpolated to a variety of
different vertical coordinate systems prior to output. To maintain a larger degree of flexibility
within HYSPLIT’s internal structure, i.e. the ability to use different meteorological data sources
for input, the profiles of the meteorological data at each horizontal grid point are linearly
interpolated to a terrain-following (F) coordinate system,

F = 1 - z/Ztop , (1)

where all the heights are expressed relative to terrain and where Ztop is the top of HYSPLIT’s
coordinate system. The model is capable of handling meteorological data fields provided on four
different vertical coordinate systems: pressure-sigma, absolute-pressure, terrain-sigma (typically
used mesoscale models covering a limited spatial domain), or a hybrid absolute-pressure-sigma
coordinate system (ECMWF, 1995; which consists of an absolute pressure added to the
pressure-sigma level).

The dispersion model’s horizontal grid system is always identical to that of the
meteorological input data. Three different conformal map projections are supported, Polar
Stereographic, Mercator, and Lambert Conformal, using a set of universal mapping
transformation routines (Taylor, 1997). A simulation may use several meteorological data files,
each on different projections. For instance, a calculation may start on a Lambert Conformal
regional grid and then switch to the Polar Stereographic global grid. Gridded fields of
meteorological variables are also required at regular temporal intervals which should be constant
for each defined grid, i. e. the regional grid may be available at 3 h intervals while the coarser
global grid may be available only every 6 h.

At a minimum HYSPLIT requires U, V (the horizontal wind components), T

(temperature), Z (height) or P (pressure), and the pressure at the surface, P0. If wet deposition
processes are to be included, then the model also requires the rainfall field. Normally ground-
level (#10m) winds are available from most meteorological models. In the rare case when these
fields are not provided, these winds are estimated using a logarithmic profile for neutral
conditions.

In most circumstances the meteorological data will contain a vertical motion field, usually
in pressure units, and regardless upon which vertical coordinate system these data are provided,
the vertical velocity field is almost always relative to the meteorological model’s native terrain-
following coordinate system. The trajectory and dispersion model calculations can use these data
fields directly because the model’s internal coordinate system is terrain-following. This is one of
the primary reasons why the meteorological data for the dispersion model are always re-mapped
internally to a common terrain-following vertical coordinate system.

When the vertical motion fields are missing, or perhaps there are some special conditions

4
required for a simulation, HYSPLIT has an option to replace these fields with a vertical velocity
calculated based upon an assumption that the pollutant parcel is transported on some other
surface. The velocity (W0 ) to maintain a parcel on the selected (0) surface is computed from,

W0 = (-M0/Mt -u M0/Mx -v M0/My) / (M0/Mz), (2)

given the slope of the surface and its local rate of change. The surfaces, 0, may be defined as
isobaric (p), isosigma (F), isopycnic (D), or isentropic (2).

Advection calculation

Once the basic (U,V,W) meteorological data have been processed and interpolated to the
internal model grid, trajectories can be computed to test the advection components of the model.
The key improvement in the advection calculation over previous versions of the model is that time
interpolation of the meteorological fields is directly incorporated into the advection algorithm.
The advection of a particle or puff is computed from the average of the three-dimensional
velocity vectors for the initial-position P(t) and the first-guess position P'(t+)t). The velocity
vectors are linearly interpolated in both space and time. The first guess position is

P'(t+)t) = P(t) + V(P,t) )t (3)

and the final position is

P(t+)t) = P(t) + 0.5 [ V(P,t) + V(P',t+)t) ] )t, (4)

Trajectories are terminated if they exit the model top, but advection continues along the surface if
trajectories intersect the ground. The integration time step ()t) can vary during the simulation. It
is computed from the requirement that the advection distance per time-step should be less than
0.75 of the meteorological grid spacing.

The integration method (e.g. Kreyszig, 1968) is common and has been used for trajectory
analysis (Petterssen, 1940) for quite some time. Higher-order integration methods were
investigated but as long as the data observations are linearly interpolated from the grid to the
integration point, higher-order methods will not yield greater precision.

Trajectory example - ACE

Following a balloon path provides an excellent test of a model’s advection calculation

(Draxler, 1996). A sample trajectory calculation was made for the time and place from which
three constant-level balloons were released on 8 December 1995 during the first Aerosol
Characterization Experiment (ACE - S.T. Siems, personal communication; Bates et al. (1998))
south of Tasmania. The meteorological data were obtained from NOAA’s global AVN model

5
(Peterson and Stackpole, 1989) at an output resolution of 190 km on mandatory pressure surfaces
at 12-h intervals, and the BoM global GASP model (Bourke et al., 1995) at an output resolution
of 285 km (2.50 ) on 19 vertical levels at 12-h intervals. Data from 6-h forecasts from the AVN
and GASP models were used in the period
between analyses. The balloons were tracked for
24 hours, each released about 1 hour apart, and
each had a slightly more westerly path. The
calculated and observed trajectories for the first
balloon are shown in Fig. 1, where the calculated
trajectories are labelled as “GASP” and “NOAA”
according to the meteorological data used for the
calculation. The trajectory calculation was
started from the 0000 UTC balloon position and
height (about 200 m AGL - above ground level).
Although the calculated tracks are close to the
balloon’s path, the measured balloon track ends
slightly to the west of all the calculations. After
24-h, the observed differences of 10% to 20% of
the transport distance are consistent with
Figure 1. ACE balloon path and corresponding previous studies (Draxler, 1996, 1991) of
calculated trajectories using GASP and NOAA trajectory accuracy in an inhomogeneous
meteorological data for a starting time of 0000 UTC environment. It is seen that the two models
8 December 1995. Latitude and longitudes are produce results that are in close agreement with
indicated at 2 degree intervals. each other, although the GASP model exhibits
faster winds and its trajectory travels further
south on this occasion.

The sensitivity of the calculated trajectory to initial position can be evaluated by starting
simultaneous trajectories at different heights and levels. Trajectories started at ±0.5o about the
origin measure the homogeniety of the wind field and the results using the AVN meteorological
data are shown in Fig. 2. Clearly all the trajectories are parallel to each other and slight errors in
horizontal position would have made little difference in any accumulated trajectory error.
However when trajectories are started at 3 different heights (10, 200, and 750 m AGL) the
resulting differences shown in Fig. 3 are quite substantial. The actual balloon track did not follow
a constant altitude and at times travelled much closer to the surface that the 200 m trajectory
would indicate and as expected the measured horizontal balloon path falls between the 10 m and
200 m calculated trajectories. This is only a single illustrative case and a more formal evaluation
would require many more cases before quantitative conclusions could be drawn from the results.
However, due to the larger variations of wind speed and direction near the ground than at higher
altitudes, it is clear that proper specification of the atmosphere’s vertical structure is essential for
the accurate computation of boundary layer pollutant transport.

6
Figure 2 Calculated trajectories using NOAA Figure 3 ACE balloon path and calculated
meteorological data starting at four different trajectories using NOAA meteorological data
positions 0.5 degrees about the initial ACE balloon starting at three different heights, 10 m, 200 m, and
position. Latitude and longitudes are indicated at 2 750 m AGL. Latitude and longitudes are indicated
degree intervals and trajectory positions are marked at 2 degree intervals, trajectory positions are marked
every 6 h. every 6 h, and the balloon heights are indicated at
the end of each trajectory in pressure units.

Air Concentrations
A Lagrangian model can compute air concentrations through either of two assumptions.
In a puff model, the source is simulated by releasing pollutant puffs at regular intervals over the
duration of the release. Each puff contains the appropriate fraction of the pollutant mass. The
puff is advected according to the trajectory of its center position while the size of the puff (both
horizontally and vertically) expands in time to account for the dispersive nature of a turbulent
atmosphere. Air concentrations are then calculated at specific points (or nodes on a grid) by
assuming that the concentrations within the puff have a defined spatial distribution. In a
Lagrangian particle model, the source can be simulated by releasing many particles over the
duration of the release. In addition to the advective motion of each particle, a random component
to the motion is added at each step according to the atmospheric turbulence at that time. In this
way a cluster of particles released at the same point will expand in space and time simulating the
dispersive nature of the atmosphere. Air concentrations are calculated by summing the mass of all
the particles in a grid cell. In a homogeneous environment the size of the puff (in terms of its
standard deviation) at any particular time should correspond to the second moment of the particle
positions.

Although a particle approach can more realistically represent dispersion in an

inhomogeneous atmosphere, the puff methods used in HYSPLIT permit the puffs to split into
smaller components when they exceed certain dimensions, usually the dimensions of the

7
meteorological grid. In this way, subgrid turbulence processes are modeled by the dispersion
parameterizations while grid-scale processes are simulated by the puff-splitting process.
However, one limitation of puff-splitting is that it can generate too many puffs, especially in the
vertical direction if the mixing is strong. A novel approach, developed by Hurley (1994), uses
particle dispersion in the vertical direction and puff dispersion in the horizontal, is incorporated
into this version of the model. This permits the more accurate particle representation to be used
in the vertical, where discontinuities can be large, and where the puff-splitting approach has its
limitations, while using the puff representation in the horizontal to limit the number of particles
that are required to adequately represent the horizontal distribution.

Although the model can be configured to compute dispersion using three-dimensional puff
methods (similar to the previous version) or three-dimensional particle dispersion, only the
combination (particle-vertical, puff-horizontal) approach will be described in detail here. See
Draxler and Hess (1997) for a complete description of the other options. Regardless of which
dispersion approach is selected for a simulation, it is first necessary to compute the atmospheric
stability, the corresponding atmospheric mixing, and the rate of pollutant dispersion. Each of
these processes will be addressed in sequence.

Stability

There are two ways to estimate the boundary layer stability. The preferred method is to
use the fluxes of heat and momentum provided by the meteorological model, if available.
Otherwise the vertical temperature and wind gradients at each grid-point are used to estimate
stability. This latter situation may not be ideal if meteorological data aloft are only provided on
mandatory surfaces with large distances between levels near the ground. In either situation the
friction velocity and temperature are estimated first, from which the Obukhov length is calculated.

When surface fluxes are available from the meteorological model, the friction velocity u*
is computed from the momentum flux and the friction temperature T* is computed from the
sensible heat flux. The stability parameter z/L is then computed from the friction values, such
that

z/L = Z2 k g T* (u*2 T2)-1, -2 # z/L # 10, (5)

and where Z2 indicates height of 2nd model level, assumed to be the depth of the surface layer.

When no fluxes are provided by the meteorological model, a typical situation when using
older meteorological data archives, z/L is estimated from wind and temperature profiles. The
meteorological sounding is used first to compute the bulk Richardson Number,

Rb = g )2 )z {212 [()U)2+()V)2]}-1, (6)

where ) indicates a gradient between levels 1-2 and 212 is the layer-average virtual potential

8
temperature. Then the stability parameter z/L is estimated from Rb. This is done by using
empirical interpolation formulas fit to Monin-Obukhov profile functions explicitly derived for a
surface layer height of 75 m (similar equations have been derived by Launiainen (1995),
Choudhury, et al. (1986), and Abdella and McFarlane (1996)). The friction velocity and
temperature are then given by

u* = k z2 )U (Nm )z)-1, (7)

T* = k z2 )2 (Nh )z)-1, (8)

where k is von Kármán's constant (k . 0.4), and the stability dependent normalized profiles (N)
for heat (h) and momentum (m) are from Beljaars and Holtslag (1991) for a stable surface layer
(0# z/L#10), and from Betchov and Yaglom (1971) and Kadar and Perepelkin (1989) in an
unstable surface layer (-2 # z/L # 0).

Horizontal and vertical mixing

First the boundary layer depth (Zi) is computed at each grid point. It is assumed to equal
the height at which the potential temperature first exceeds the value at the ground by 2K. The
temperature profile is analyzed from the top down to determine the boundary layer depth. The
top-down approach reduces the influence of shallow stable layers near the ground. In addition, a
minimum depth of 250 m is assumed for all hours. The height was chosen to correspond with the
minimum height resolution typical of the meteorological input data.

The pollutant vertical mixing coefficient (Kz ) is assumed to follow the coefficients for
heat. Within the Boundary Layer (BL), vertical mixing coefficients are computed following Troen
and Mahrt (1986) and Holtslag and Boville (1993), where

Kz = k wh z (1 - z/Zi)2, (9)

and wh is a function of the Prandtl number, u*, Nh, Nm , and the convective velocity scale w*.
Following Beljaars and Betts (1993), mixing through the inversion layer at z = Zi during
convective conditions is computed based upon the surface flux parameters and the strength of the
inversion, where

Kz = - Cs u* T* (M2/Mz)-1, (10)

and Cs = 0.4. During stable conditions mixing through the interface is calculated from Eq. 11.
Above the BL, pollutant mixing in the remainder of the atmosphere is defined by the vertical
diffusivity for heat using mixing length theory where

Kz = R2 * MV/Mz * Nh(R/Lo)-1, (11)

where R is a Blackadar-type mixing length proportional to height above ground and Lo is the local

9
Obukhov length.

Once the Kz profile has been computed, a single average value for the entire BL is
computed from the profile and that value replaces all the values within the BL. Each horizontal
grid point will have a different value. The resulting mixing profile, an adaptation of the Troen-
Mahrt profiles, was determined empirically to give improved results in both the near and far fields.
Although near-field concentrations depend more on the details of the mixing profile while the far-
field results depend more upon the average mixing over the BL, the model is intended to be
applied for long-range simulations, suggesting a unified approach in averaging the BL mixing.

The subgrid-scale horizontal mixing coefficient (Kh) is computed from the velocity
deformation (Smagorinsky, 1963; Deardorff, 1973),

Kh = 2-0.5 (c O)2 [ (MV/Mx + MU/My)2 + (MU/Mx - MV/My)2 ]0.5, (12)

where O is the meteorological data grid size, and c is 0.14. Grid scale dispersion is simulated by
the particles and puffs horizontally spreading into different wind regimes.

Particle and puff dispersion

Both the horizontal puff and vertical particle dispersion equations are formulated in terms
of the turbulent velocity components. These velocity components are a function of the
diffusivities computed in the previous section. In the particle model, the dispersion process is
represented by adding a turbulent component to the mean velocity obtained from the
meteorological data. The approach used follows the one described by Fay et al. (1995) such that
after computation of the new position at a time step due to the mean advection of the wind, a
turbulent component is added to the mean particle positions (Z),

Zfinal(t+)t) = Zmean(t+)t) + W' (t+)t) )t. (13)

The contribution of the vertical turbulent wind component (W') is added to the "mean" position
(from Eq. 4) to give a final position from which the advection at the next time step is computed.
Full reflection is assumed for particles that intersect the ground or model-top.

The turbulent velocity component at the current time W'(t+)t) is computed from the
turbulent velocity component at the previous time W'(t), an autocorrelation coefficient (R) that
depends upon the exponential of the time step (exp [-)t / TLw ]), the Lagrangian time scale (TLw),
and a computer-generated random component (W"), such that

W' (t+)t) = R()t) W' (t) + W" ( 1-R()t)2 )0.5 + TLw (1-R()t)) MFw2/Mz. (14)

The last term is applied to prevent accumulation of particles in low turbulence regions (Legg and
Raupach, 1982), the importance of which is somewhat reduced due to the averaging of the

10
diffusivity profile within the BL. The Lagrangian time scale is assumed to be a constant 100 s
for convenience. This value results in a random walk (R. 0) vertical dispersion for most of the
longer time steps. The random component W" is the product of a computer-generated Gaussian
random number and the standard deviation of the turbulent velocity,

Fw = (Kz / TLw)0.5, (15)

which is computed from the previously calculated vertical diffusivity (Eqs. 9-11).

Puff dispersion is treated in two domains, when the puff is smaller than the meteorological
model grid size and when it is larger. In the latter case it is assumed that the meteorological
model is capable of resolving turbulent motions on that scale. Gaussian and "top-hat" puffs are
treated almost identically. When puffs have dimensions less than the meteorological grid spacing,
the rate of puff horizontal growth,

dFh/dt = o2 Fu, (16)

where the horizontal turbulent velocity standard deviation (Fu) is computed using an equation
comparable to (15) but with the Lagrangian time scale equal to 10800 s (-1/f) and the diffusivity
is from (12). Additional growth is achieved through the puff-splitting process as the puff grows
into different mixing regions. When a puff expands to cover several meteorological grid points, a
top-hat puff splits horizontally into four puffs, each with 25% of the mass, while a large Gaussian
puff splits into five smaller puffs, with an additional puff at the center position with 60% of the
mass while the outside 4 puffs each get 10% of the mass.

Pollutant air concentration

For each puff, concentrations are summed at each time step to all grid points that fall
within the puff defined for top-hat distributions as ±1.54 Fh or ±3.0 Fh for Gaussian distributions.
The incremental concentration contribution by each top-hat puff of mass m to a grid point is,

)c = m (B r2 )z)-1, (17)

where the horizontal radius r = 1.54 Fh. All grid-nodes within the puff receive the same )c. The
incremental concentration contribution for a Gaussian puff is,

)c = m (2 B Fh2 )z)-1 exp(-0.5 x2/Fh2), (18)

where x is the distance from the puff center to grid-node and with )z defined as grid-cell height.
Note that in the vertical particle calculations are summed into a grid-cell rather than computed at
a grid-point. A cell is defined at the center of the node and has an area corresponding to the half-
way distance to adjacent nodes. The final average concentration is the incremental sum divided
by the number of time steps in the concentration averaging period.

11
Air concentration example - ANATEX

Inert atmospheric tracers provide excellent benchmarks for testing atmospheric dispersion
models. Most long-range experiments have been conducted under very controlled conditions
(pre-selected meteorological conditions), leading to the question of how this practice might be
biasing the statistical evaluation of the dispersion models. This problem was addressed in the
Across North America Tracer Experiment (ANATEX - Draxler et al., 1991) in which two
different inert perfluorocyclohexane gases (PTCH and PDCH) were released for three hours from
two different locations, Glasgow, Montana (GGW, 48.4N, 106.5W) and St. Cloud, Minnesota
(STC, 45.6N, 94.2W) every 2½ days with about 75 samplers located at weather stations over the
whole eastern half of the U.S. The experiment lasted three months, from January 5th through
March 29th of 1987.

Dispersion model calculations were made using a consecutive series of forecast (2- to 12-
h) meteorological data fields from NOAA’s Nested Grid Model (NGM - Peterson and Stackpole,
1989) with output fields available every 2-h on a 90 km grid on the NGM’s sigma coordinate
system (levels at: 0.982, 0.943, 0.897, 0.844, ...). The lowest NGM data level is above the first
two dispersion model levels, however the NGM archive contains both heat and momentum fluxes,
which were used in the calculation of vertical mixing.

The HYSPLIT model was run for each dispersion combination independently for each
source location. The first half of the experimental period was selected for analysis. The model
was configured to run for 814 hours, starting with the first release on the 5th of January and
ending on the 8th of February. This period includes 14 tracer releases from each location. Only
simple statistical analyses were performed on the results. The daily measured and calculated
concentrations at each sampling station were averaged over the duration of the simulation. The
resulting correlation coefficient was computed from these pairs of measured and model calculated
concentrations and it represents the average spatial correlation. The model simulation results are
summarized in Table 1.

Table 1. Model performance during ANATEX for the period of 5 January 1987 through 8 February 1987
in terms of the spatial correlation coefficient (R) and the ratio of the mean calculated to mean measured
concentrations (C/M). Total computational time is indicated in the last column.
______________________________________________________________________________
Dispersion Model GGW STC
Horizontal / Vertical R C/M R C/M CPU (h)
______________________________________________________________________________
T/T 0.76 1.57 0.55 1.75 1.2
G/T 0.74 1.48 0.56 1.69 2.7
P/P 0.70 1.23 0.65 1.94 2.1
T/P 0.72 1.25 0.69 1.87 2.9
G/P 0.73 1.20 0.72 1.89 6.0
______________________________________________________________________________
T - top hat puff, G - Gaussian puff, P - Particle dispersion model
CPU - IBM RS6000/590.

12
 In general all model combinations show similar results with a slight over-calculation,
higher at STC than at GGW. This is believed to be due to the fact that the STC release was in the
middle of the sampling network, and hence there were many more samplers relatively (compared
with GGW) close to the release point and hence the statistical results give greater weight to the
over-calculation of the nearby samplers. The three-dimensional puff combinations, T/T and G/T,
comparable to the previous version of the model, show correlations that are lower at STC than
GGW, while the vertical particle simulations showed comparable correlations between GGW and
STC. This suggests that the particle method provides a more realistic representation of the
pollutant’s vertical structure. Gaussian and top-hat simulations produce almost identical results
because the horizontal dispersion is dominated by grid-scale processes rather than the subgrid-
scale parameterizations. In subsequent calculations, unless noted otherwise, the base calculation
is made using the horizontal top-hat, vertical particle combination of the model, because of its
simple horizontal configuration (requires less computer time than Gaussian versions) and its
potential to more realistically represent the pollutant’s vertical structure.

An illustration of the over-

calculation can be seen for the T/P
combination for GGW in Fig. 4 for
both measured and calculated
concentrations. The units are 10-12
g m-3 and the contours are drawn at
levels of 10, 20, 50, 100, and 200
units. The basic over-calculation is
evident from the larger areal
coverage of the calculated contours
as compared with the measured
contours, especially for the
samplers nearest to the release.
Because this is a winter period, the
strongly stable structure of the
atmosphere should increase the
model’s sensitivity to the vertical
mixing computation method. In
contrast, if the BL were well mixed,
it would matter little if the model
Figure 4 Average measured and calculated concentrations using the mixed pollutants twice as fast or
top-hat/particle model combination for the period of 5 January through slow as it should at the long source
8 February, 1987 for tracer releases from the Glasgow, Montana to receptor distances typical of the
location. Concentration units are 10-12 g m-3 and the contours are ANATEX experiment.
drawn at levels of 10, 20, 50, 100, and 200 units. Values less than 10
approach the detection limit of the analytical system.

13
 One issue regarding the model’s sensitivity to the mixing can be addressed by turning on
or off various elements of the mixing parameterization. This test can only be done using the 3D
particle version of the model, as it is not possible to compute a concentration at a sampling node
for a puff that has no horizontal or vertical dimensions. The results are summarized in Table 2
and it is quite striking that the P/P-2 calculation, where the horizontal equivalent of Eq. 14 equals
zero, gives almost identical results to the base simulation (P/P) which included both horizontal
and vertical turbulence. Although horizontal dispersion is not explicitly included in the P/P-2
calculation, the vertical mixing of particles into regions of different wind speed and direction
results in a horizontal dispersion of particles that dominates the sub-grid scale dispersion
parameterization. The no dispersion (P/P-1) control case and the horizontal dispersion only (P/P-
3) simulation show substantial over-calculation of air concentration.

Table 2. Model sensitivity during ANATEX to horizontal and vertical dispersion for the three-
dimensional particle version of the model. Statistical results are the same as those defined in
Table 1.
_____________________________________________________________
GGW STC
Model Simulation R C/M R C/M
_____________________________________________________________
P/P Base - Table 1 0.70 1.23 0.65 1.94
P/P-1 No dispersion 0.36 67. 0.54 108.
P/P-2 Vertical only 0.71 1.22 0.64 1.81
P/P-3 Horizontal only 0.40 53. 0.60 128.
______________________________________________________________

Considering the importance of the vertical mixing to the model’s performance there are
several other simple tests that can be conducted. Varying the Lagrangian time scale by a factor of
2 above (T/P-2) and below (T/P-1) the base value of 100s results in concentration changes of
greater than a factor of 2 as shown in Table 3. As noted earlier, the BL vertical mixing
coefficients from (Eq. 9) are vertically averaged to reduce the sensitivity of long-range dispersion
results to the details of the mixing profile. The use of the actual (un-averaged) Kz profile in the
BL (T/P-3) resulted in less overall model bias (more of the tracer remained above the ground),
however, the correlation was less at STC than for the base case (T/P). In the final simulation
(T/P-4), z/L was computed from the bulk Richardson number (Eq. 6) rather than defined from
the heat and momentum fluxes of the meteorological model and those results were quite
comparable to the base case. This result may not hold true for other meteorological data with
coarser vertical resolution.

14
 Table 3. Model sensitivity during ANATEX to various vertical dispersion parameterizations for
the horizontal-puff, vertical-particle version of the model. Statistical results are the same as those
defined in Table 1.
_____________________________________________________________
GGW STC
Model Simulation R C/M R C/M
_____________________________________________________________
T/P Base - Table 1 0.72 1.25 0.69 1.87
T/P-1 TLv=50s 0.68 0.51 0.71 0.80
T/P-2 TLv=200s 0.74 2.55 0.65 3.65
T/P-3 No Kz averaging 0.69 0.95 0.57 1.24
T/P-4 No flux fields used 0.71 1.06 0.62 1.76
_____________________________________________________________

The last issue regarding the model’s vertical structure relates to the computation of the
vertical motion. The base calculation uses the vertical velocities provided in the NGM data fields.
However when the air concentration predictions are recalculated using each of the vertical motion
options using Eq. 2, the results as shown in Table 4 indicate that the isentropic option (T/P-5)
provides some improvement in both correlation and bias for both release sites. The same is true,
but to a lesser extent, for the isobaric (T/P-6) and isopycnic (T/P-8) options. However the
isosigma calculation (T/P-7) shows a clear and substantial degradation in the results. One can
speculate that in a winter stable environment, pollutants are more likely to follow isentropic
surfaces, while the terrain-following sigma might be more appropriate during convective
conditions when the BL structure is more controlled by the fluxes of heat and momentum from
the surface.

Table 4. Model sensitivity during ANATEX to various vertical motion options (Eq. 2) for the
horizontal-puff, vertical-particle version of the model. Statistical results are the same as those
defined in Table 1.
_____________________________________________________________
GGW STC
Model Simulation R C/M R C/M
_____________________________________________________________
T/P Base - Table 1 0.72 1.25 0.69 1.87
T/P-5 Isentropic 0.78 1.13 0.72 1.65
T/P-6 Isobaric 0.74 0.92 0.71 1.57
T/P-7 Isosigma 0.66 0.84 0.42 2.26
T/P-8 Isopycnic 0.75 0.90 0.70 1.57
_____________________________________________________________

Deposition
There are three different removal mechanisms available: dry deposition, wet depletion, and
radioactive decay. Dry deposition can be either explicitly defined by a deposition velocity, a
gravitational settling velocity for particles, or it may be computed using the resistance method.

15
Wet removal for soluble gases is computed from its Henry's Law constant and particle wet
removal is computed from a scavenging ratio (R/R) within the cloud and by an explicit scavenging
coefficient (s-1) for pollutants below the cloud base. For computational simplicity, the total
deposition from both dry and wet removal processes is expressed in terms of inverse time
constants which can be added and hence the total deposition over a time step becomes

Dwet+dry = m {1-exp[-)t ($dry+$gas+$inc+$bel ) ] }, (19)

where m is the pollutant mass of either the particle or puff. The pollutant mass is then reduced by
the deposition amount. Each of the constants, for dry deposition ($dry), wet removal for gases
($gas), in-cloud wet removal of particles ($inc), and below-cloud wet removal of particles ($bel),
will be discussed in more detail in the following sections.

Dry removal

Dry removal is computed only when the bottom of the puff or the particle center position
is within the surface layer, defined as the dispersion model’s second meteorological data level
(usually around 75 m). Then the mass deposited by dry removal simply equals the product of the
deposition velocity (Vd ) and the air concentration, assumed to be uniform over the depth of the
pollutant layer ()Zp - in this case the surface layer). The deposition velocity is converted to a
removal constant,

$dry = Vd )Zp-1. (20)

The deposition velocity may be directly specified in the model’s input or calculated as the
gravitational settling velocity. Particle settling is computed following Van der Hoven (1968) from
the particle diameter, air density, and particle density. In addition to the resulting mass depletion,
the settling velocity is also applied to the pollutant's vertical position each time step to permit the
gradual sinking of particles and (non-gaseous) puffs.

The dry deposition velocity may also be computed through the resistance method. Rather
than explicitly defining a dry deposition velocity for a pollutant, the total deposition velocity can
alternately be computed from the sum of various resistances (Hicks, 1986) and the settling
velocity for particles such that the total deposition velocity

Vd = [ Ra + Rb + Rc + Ra Rb Vg ]-1 + Vg, (21)

where the subscripts for the resistances R, represent the atmospheric layer (a), the quasi-laminar
sub-layer (b), and the canopy layer (c) which represents the bulk resistance of various surfaces.
Gravitational settling, Vg, is zero for gases and Rc is zero for particles. The resistance
components depend upon meteorological conditions as well as the properties of the surface.
More detail on each of the resistance components can be found in Draxler and Hess (1997),
however, the method very closely follows that proposed by Wesely (1989) as incorporated into

16
the RADM model (Chang, 1990), and updated by Walmsley and Wesely (1996).

Wet removal

Wet deposition is divided into two processes (Hicks, 1986), those in which the polluted
air is continuously ingested into a cloud from a polluted boundary layer and those in which rain
falls through a polluted layer. The dominant process is difficult to determine during a model
simulation using only the data available from most meteorological archives. Therefore, for
particulate pollutants, the simplifying assumption of a scavenging ratio is assumed for pollutants
located within a cloud layer and the scavenging coefficient is used for pollutant removal in rain
below a cloud layer. The scavenging ratio, which is the ratio of the pollutant's concentration in
water to its concentration in air, is expressed as the removal constant,

$inc = Ft Fb Sr P )Zp-1, (22)

where Sr is the average scavenging ratio by volume, P is the precipitation rate, )Zp is the depth
of the pollutant layer, and Ft is the fraction of the pollutant layer that is below the cloud top, and
Fb is the fraction of the pollutant layer that is above the cloud bottom. The cloud bottom is
defined at the level when the RH first reaches 80% and the cloud top is reached when the RH
drops below 60%. Below-cloud removal is defined from

$bel = Sc (1.0-Fb). (23)

where Sc is the average scavenging coefficient.

The wet deposition of gases depends upon their solubility and for inert non-reactive gases
it is a function of the Henry's Law constant, the ratio of the pollutant's equilibrium concentration
in water to that in air and therefore

$gas = Ft H R T P )Zp-1, (24)

where R is the universal gas constant, and T is temperature. The wet removal of gases is applied
at all levels from the ground to the top of the cloud-layer.

Deposition example - Chernobyl

A good way to illustrate the model’s applicability to emergency response situations and
evaluate the deposition algorithms is to rerun the accident at Chernobyl. Special meteorological
data sets are available from ECMWF (1995) for the accident period. A single 120-h forecast data
file, output at 1.125o horizontal resolution every 6 h, was selected for the dispersion simulation,
primarily because it contained precipitation fields, not available in the longer time period (12 day)
analysis archive. The ECMWF fields have similar vertical resolution as the NGM, but contain

17
additional fields of winds and temperatures, at 10 m and 2 m, respectively. Flux fields were not
available and vertical mixing was derived from the wind and temperature profiles.

HYSPLIT was run assuming an emission rate of 1015 Bq h-1 of Cs-137 for the first 24 h,
distributed uniformly in the layer 750-1500 m AGL. Using the default model parameters, i. e., an
explicit dry deposition velocity of 0.1 cm s-1 (IAEA,1994 - equivalent to a mono-dispersed
spectrum of particles of diameter - 4 µm), a within-cloud scavenging ratio S r of 3.2 x 105
(Hicks,1986), and a below-cloud removal coefficient Sc of 5 x 10-5 s-1 (Hicks, 1986), the model
results for Cs-137 deposition are shown in Fig. 5 in units of Bq m-2. The published deposition
map (Fig. 6 - from the ATMES report, 1992) for this period shows an deposition peak (at the 50
KBq m-2 contour level) over central Sweden while the model predicted a peak of only18 KBq m-2
just to the east over the Baltic Sea. In general, considering that the model is using only forecast
precipitation fields, the model performance is quite good. Although the radioactive release
occurred for a 10-day period, the limited 5-day forecast data file only permitted a realistic
simulation of the first day’s release. The release rate on the second day dropped to 25% of the
rate on the first day, and due to changing meteorological conditions, 90% of the deposition over
the Scandinavian countries can be accounted for by the emissions that occurred on the first day.
Deposition peaks over central and south-eastern Europe (see Fig. 6) require the simulation to
span the duration of the accident. The sensitivity of some of the assumptions affecting deposition

Figure 5 Reproduction of the Cesium-137 Figure 6 The top-hat/particle model simulation of

deposition map from the ATMES (1992) report. Cs-137 deposition (Bq/m2) for the first 24 h of
Includes accumulations from 27 April through emissions of the 10 day Chernobyl accident using the
10 May 1986 in units of K Bq/m2. 120 h meteorological forecast from the ECMWF.
The concentration grid resolution was 0.5 degrees
and the deposition was accumulated from 27 April.
Latitude and longitudes are indicated at 10 degree
intervals.

18
can be evaluated by turning on or off each of the appropriate parameters, the results of which are
summarized in Table 5. The default simulation was made using the same model configuration as
in all the previous simulations: the horizontal top-hat puff and vertical particle model (T/P).
However the same simulation using the puff assumption in the vertical as well as horizontal (T/T)
had the greatest effect on the deposition amount, raising the peak to over 50 KBq m-2, an amount
almost identical to what was measured. Deposition amounts are higher using the T/T variation of
the model because the corresponding air concentrations are higher. Air concentrations are higher
using the vertical puff model because by the time the pollutant reached Scandinavia it had mixed
up to 3000 m in the particle model but only to 1500 m in the puff model. Measurements in that
region (Puhakka et al., 1990) indicated a wide spatial variation in the vertical mixing of the
radionuclides, but typical values ranged from 1000 to over 2000 m, indicating that perhaps the
vertical puff approximation provided better results in this case, perhaps due to problems in
defining an appropriate mixed layer during convective precipitation conditions. The puff
representation effectively restrains pollutants within the mixed layer while the particle model will
permit a certain fraction to mix through the interfacial layer, the strength of which is defined by
Eq. 10.

Table 5. Model computed deposition peaks over the Baltic Sea for the Chernobyl accident using different
dispersion and deposition options.
_____________________________________________________________
Simulation Description Deposition (K Bq m-2)
_____________________________________________________________
T/P Horizontal top-hat / vertical particle 17.7
T/T Horizontal top-hat / vertical top-hat 55.8
T/P-9 Wet only, no dry ($dry = 0) 18.2
T/P-10 Dry only, no wet ($inc = $bel = 0) 8.1
T/P-11 In-cloud wet only ($bel = $dry = 0) 22.7
T/P-12 Below-cloud wet only ($inc = $dry = 0) 3.6
T/P-13 Double in-cloud wet ($bel = $dry = 0) 27.0
T/P-14 Half in-cloud wet ($bel = $dry = 0) 17.5
______________________________________________________________

Other tests showed that dry removal accounted for about one third of the deposition (T/P-
10). However, when dry deposition was turned off and only wet only removal computed (T/P-9),
the maximum deposition was actually greater than in the base case, when both wet and dry
removal were turned on (T/P). This result is not inconsistent, in that dry deposition is a
continuous process, while wet removal is a function of the spatial and temporal rainfall pattern.
Turning off dry deposition permits more material to be transported to the rainfall region where it
is subsequently deposited.

One significant feature of the wet deposition computation is that this case is dominated by
the in-cloud removal processes (T/P-11 versus T/P-12), accounting for almost all of the wet
removal. However increasing (T/P-13) or decreasing (T/P-14) the in-cloud removal coefficient by
a factor of two only had a 20% effect on the deposition amount, suggesting that more
complicated schemes to determine the removal coefficients will have little effect on the final

19
results and that wet removal is dominated by the rainfall prediction and the vertical distribution of
the pollutant.

Volcanic eruptions
All the previous simulations were confined to the lower troposphere, primarily the BL.
However a frequently more common application of atmospheric dispersion models, especially as
related to emergency response situations deals with the forecasting of the movement of debris
from volcanic eruptions, which may penetrate well into the stratosphere. Recent developments in
aircraft engine design have made them more sensitive to the effects of volcanic ash. Volcanic
eruptions occur much more frequently than nuclear accidents and also provide an opportunity for
verification of model performance (Potts, 1993).

HYSPLIT required no special modification to handle these kind of events. In the volcanic
ash simulation, an emission point would be represented by several particles or puffs, each
corresponding to a different size range of ash. The puffs representing each range need to be
treated independently because they will have different atmospheric transport pathways due to
gravitational settling. The sizes range from diameters of 0.3 to 30 µm, but they all have the
same density (2.5 g cm-3), typical of a mix of pumice, basalt, and similar materials. This follows
the approach developed by Heffter and Stunder (1993) from examination of particle distributions
measured after several eruptions in the U.S.A. The release amount for each size range is adjusted
according to what fraction of the total mass is represented by that size in the particle distribution.
Most of the mass (99%) resides in the 3 to 30 µm range, but the larger particles fall quickly, and
only the smallest particles are subject to long-range transport.

The simulation is treated in a manner similar to any point source except that the source is
defined at two different heights. The model will then uniformly distribute the pollution between
these levels, effectively simulating a vertical line source from an explosive eruption. The Rabaul
eruption of 2000 UTC 18 September 1994 was selected for the example simulation. A
continuous release (1 unit h-1) between 1000 m to 22000 m AGL was used to represent the
eruption. The simulation, using the 381 km NOAA final analysis data, was run for only 5 hours
and the output, shown on Fig. 7, represents the instantaneous concentration at 0030 UTC 19
September. This is the time of the GMS VIS satellite photograph shown in Fig. 8. The bulk of
the ash is at the upper levels and the correspondence, in terms of the extent and shape of the
cloud, is quite good. Although the southern edge of the visible cloud is not bright, it extends to
nearly 7oS. On the corresponding lower resolution IR image (not shown) some material appears
west of 150oE and north of 4oS; these features are present on the VIS image but are diffuse.
There is some uncertaintly about the eruption height, as is the case in most of these events,
assumed to be constant from the start of the release to the time of the satellite photo. Stereo
conical scans using other satellites (Prata, 1996) showed considerable height variations over the

20
 Figure 8 GMS VIS satellite photo for 0032 UTC
Figure 7 Simulation of the Rabaul volcanic 19 September 1994, a time that corresponds to the
eruption of 2000 UTC 18 September 1994 showing
time of the simulation shown in Fig. 7.
the vertical column-averaged snapshot air
concentrations at 0030 UTC 19 September 1994.
Concentrations are only qualitative as the eruption
mass was unknown. The concentration grid
resolution was 0.1 degrees. Latitude and longitudes
are indicated at 2 degree intervals.

cloud domain. Although the satellite verification shown here is rather qualitative, it does
demonstrate that the same model can be effectively applied to a variety of different air quality
simulations.

Summary
A description, overview of the equations, and example calculations from HYSPLIT_4, a
Lagrangian model that can be used to calculate trajectories and air concentrations for analytical
studies and to support atmospheric emergency response applications, has been presented. The
configuration of the model was shown to be generic in that it could be set up to perform a variety
of different scenarios. In general, Lagrangian models are well suited for quick calculations from
pollutant point sources and such a modelling approach is ideal for situations where quick
turnaround is essential; i.e., about 1 to 3 min CPU time per simulation day. The model’s
performance has been evaluated in a qualitative sense by comparing the calculations for a variety
of different applications to real data observations, such as measured balloon trajectories, measured
air concentrations of inert tracers, measured radioactive deposition, and satellite photographs of
ash from volcanic eruptions. In general the model’s performance compared favourably to the
observations in all areas, although there is still room for improvement, with particular regard to
the vertical distribution of pollutants in different meteorological regimes. The model's
performance has been found to be sensitive to the low-level wind profile (for BL trajectories), the
vertical mixing coefficients (for air concentrations), and the interaction of the vertical pollutant

21
distribution and the rainfall field (for deposition). Clearly all aspects of model performance were
related to how well the vertical atmospheric structure is reproduced, either in the meteorological
fields of the input data driving the calculations, or in the parameterizations used to estimate
pollutant vertical mixing. Clearly this continues to be an active area of research and our
computational methods will be modified accordingly.

Acknowledgments
The authors would like to thank Steve Businger, University of Hawaii, for providing the
ACE balloon observations; Steve Siems, Monash University, Clayton, Victoria, for the ACE
trajectory calculations using the GASP data; Fred Prata, Division of Atmospheric Research,
CSIRO, Aspendale, Victoria, for providing rather unique insights to the details about the Rabaul
volcanic eruption; Gang Liang, BMRC, for development of the HYSPLIT graphical user
interface; Les Logan, BMRC, for developing some of the NCAR graphics display routines and an
interface for the BoM meteorological data; and David Pike, BMRC, for the preparation of
graphics and satellite photographs.

References
Abdella, K. and McFarlane, N.A. 1996. Parameterization of the surface-layer exchange
coefficients for atmospheric models. Bound. Lay. Met., 80, 233-248.

ATMES Report. 1992, Evaluation of long range atmospheric transport models using
environmental radioactivity data from the Chernobyl accident, W. Klug, G. Graziani, G.
Grippa, D. Pierce, and C. Tassone (Eds), Elsevier Science Publishers, Linton Rd., Essex,
England.

Bates, T.S., Huebert, B.J., Gras, J.L., Griffiths,F.B. and Durkee, P.A. 1998. International global
atmospheric chemistry (IGAC) project’s first aerosol characterization experiment (ACE
1): Overview. J. geophys. Res., 103, 16297-318.

Beljaars, A.C.M. and Betts, A.K. 1993. Validation of the boundary layer representation in the
ECMWF model. Proc. of a Seminar Held at ECMWF on Validation of Models over
Europe, Vol. II, 7-11 September, 1992, European Centre for Medium-Range Weather
Forecasts, Shinfield Park, UK, 159-195.

Beljaars, A.C.M. and Holtslag, A.A.M. 1991. Flux parameterizations over land surfaces for
atmospheric models, Jnl. appl. Met., 30, 327-341.

Betchov, R. and Yaglom, A.M. 1971. Comments on the theory of similarity as applied to

22
 turbulence in an unstably stratified fluid, Atmos. Oceanic Phys., 7, 829-832.

Bourke, W., Hart, T., Steinle, P., Seaman, R., Embery, G., Naughton, M. and Rikus, L. 1995.
Evolution of the Bureau of Meteorology global assimilation and prediction system. Part
2: Resolution enhancements and case studies, Aust. Met. Mag., 44, 19-40.

Chang, J.S. 1990. Appendix E, NAPAP Report 4, The regional acid deposition model and
engineering model, in Acid Deposition: State of Science and Technology, Vol I,
Emissions, Atmospheric Processes, and Deposition, U.S. Government Printing Office,
Washington D.C., 20402-9325.

Choudhury, B. J., Reginato, R. J. and Idso, S. B. 1986. An analysis of infrared temperature

observations over wheat and calculation of latent heat flux. Agric. Forest Meteor., 37, 75-
88.

Deardorff, J.W. 1973. Three-dimensional numerical modeling of the planetary boundary layer, p
271-311, Workshop on Micrometeorology, D.A. Haugen, Ed., American Meteorological
Society, 45 Beacon St., Boston, MA, 02108.

Draxler. R.R. 1990. The calculation of low-level winds from the archived data of a
regional primitive equation model, Jnl. appl. Met., 29, 240-248.

Draxler, R.R. 1991. The accuracy of trajectories during ANATEX calculated using
dynamic model analyses versus rawinsonde observations, Jnl. appl. Met., 30, 1446-1467.

Draxler, R.R. 1992. Hybrid single-particle Lagrangian integrated trajectories (HY-SPLIT):

Version 3.0 -- User’s guide and model description. NOAA Tech. Memo. ERL ARL-195,
26 pp. and Appendices. [Available from National Technical Information Service, 5285
Port Royal Road, Springfield, VA 22161.]

Draxler, R.R. 1996. Trajectory optimization for balloon flight planning, Weath. forecasting,
11, 111-114.

Draxler, R.R., Dietz, R., Lagomarsino, R.J. and Start, G. 1991, Across North America Tracer
Experiment (ANATEX): Sampling and Analysis, Atmos. Environ., 25A, 2815-2836.

Draxler, R.R. and Hess G.D. 1997. Description of the HSYPLIT_4 modelling system, NOAA
Technical Memorandum ERL ARL-224, December, 24 pp., [Available from National
Technical Information Service, 5285 Port Royal Road, Springfield, VA 22161.]

Draxler, R.R. and Stunder, B.J.B. 1988. Modeling the CAPTEX vertical tracer concentration
profiles, Jnl. appl. Met., 27, 617-625.

23
Draxler, R.R. and Taylor, A.D. 1982. Horizontal dispersion parameters for long-range transport
modeling, Jnl. appl. Met., 21, 367-372.

ECMWF 1995. The description of the ECMWF/WCRP level III-A global atmospheric data
archive, European Centre for Medium-Range Weather Forecasts, Reading, Berkshire,
England.

Fay, B., Glaab, H., Jacobsen, I. and Schrodin, R. 1995. Evaluation of Eulerian and Lagrangian
atmospheric transport models at the Deutscher Wetterdienst using ANATEX surface
tracer data, Atmos. Environ., 29, 2485-2497.

Heffter, J.L. and Stunder, B.J.B. 1993. Volcanic ash forecast transport and dispersion (VAFTAD)
Model, Weath. forecasting, 8, 533-541.

Hicks, B.B., 1986, Differences in wet and dry particle deposition parameters
between North America and Europe, in Aerosols: Research, Risk Assessment, and Control
Strategies, Lewis Publishers, Chelsea, MI, pp 973-982.

Holtslag, A.A.M. and Boville, B.A. 1993. Local versus nonlocal boundary-layer diffusion in a
global climate model, Jnl. climate, 6, 1825-1842.

Hurley, P. 1994. PARTPUFF - A Lagrangian particle-puff approach for plume dispersion

modeling applications, Jnl. appl. Met., 33, 285-294.

IAEA, 1994. Modelling the deposition of airborne radionuclides into the urban environment, First
report of the VAMP Urban Working Group, International Atomic Energy Agency,
Technical Document 760.

Kadar, B.A., and Perepelkin, V.G. 1989. Effect of the unstable stratification on wind and
temperature profiles in the surface layer, Atmos. Oceanic Phys., 25, 583-588.

Kreyszig, E. 1968, Advanced engineering mathematics. 2nd Ed. J. Wiley and Sons, New York,
898 p.

Launiainen, J. 1995. Derivation of the relationship between the Obukhov stability parameter and
the bulk Richardson number for flux-profile studies. Bound. Lay. Met., 76, 165-179.

Legg, B.J. and Raupach, M. 1982. Markov chain simulation of particle dispersion in
inhomogeneous flows: the mean drift velocity induced by a gradient in Eulerian velocity
variance, Bound. Lay. Met., 24, 3-13.

Peterson, R.A. and Stackpole, J.D. 1989. Overview of the NMC production suite, Weath.
forecasting, 4, 313-322.

24
Petterssen, S. 1940. Weather analysis and forecasting, McGraw-Hill Book Company, New York,
pp 221-223.

Prata, F. 1996, Private communication.

Potts, R.J. 1993, Satellite observations of Mt. Pinatubo ash clouds, Aust. Met. Mag., 42, 59-68.

Puhakka, T., Jylhä, K., Saarikivi, P., Koistinen, J., 1990. Meteorological factors influencing the
radioactive deposition in Finland after the Chernobyl accident, Jnl. appl. Met., 29, 813-
829.

Smagorinsky, J. 1963. General circulation experiments with the primitive equations: 1. The basic
experiment, Mon. Weath. Rev., 91, 99-164.

Taylor, A.D. 1997. Conformal map transformations for meteorological modelers, Computers and
Geosciences, 23, 63-75.

Troen, I., and Mahrt, L. 1986. A simple model of the atmospheric boundary layer: sensitivity to
surface evaporation, Bound. Lay. Met., 37, 129-148.

Van der Hoven, I. 1968. Deposition of particles and gases, in Meteorology and
Atomic Energy, D. Slade (Ed.), TID-24190, NTIS, Springfield, VA, 445p.

Walmsley, J. L. and Wesely, M. L. 1996. Modification of coded parametrizations of surface

resistances to gaseous dry deposition. Atmos. Environ., 30, 1181-1188.

Wesely, M.L. 1989. Parameterizations of surface resistances to gaseous dry deposition in

regional-scale numerical models, Atmos. Environ., 23, 1293-1304.

25