You are on page 1of 13

Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

Contents lists available at ScienceDirect

Journal of Loss Prevention in the Process Industries

journal homepage:

Sensitivity analysis of Phasts atmospheric dispersion model for three toxic

materials (nitric oxide, ammonia, chlorine)
Nishant Pandya a, b, c, Nadine Gabas a, b, *, Eric Marsden c

Universit de Toulouse, INPT, UPS, Laboratoire de Gnie Chimique, 4, Alle Emile Monso, F-31432 Toulouse, France
CNRS, Laboratoire de Gnie Chimique, 4, Alle Emile Monso, BP 84234, F-31432 Toulouse, France
Institut pour une Culture de Scurit Industrielle, 6 Alle Emile Monso, BP 34038, 31029 Toulouse, France

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 9 March 2011
Received in revised form
10 June 2011
Accepted 10 June 2011

We present the results of a parametric sensitivity analysis of a widely used model for atmospheric
dispersion of toxic gases, in order better to understand the inuence of user-adjustable parameters on
model outputs. We have studied 60 min continuous release scenarios for three different products (nitric
oxide, ammonia and chlorine), chosen to cover a range of physical characteristics and storage conditions.
For each product, we have broken down base-case scenarios into a number of sub-scenarios corresponding to different release conditions which determine physical phenomena (ow rate, release angle,
release elevation and atmospheric stability class). The use of statistical tools to analyze the results of
a large number of model executions allows us to rank model parameters according to their inuence on
the variability of a number of model outputs (distances and concentrations), on a per-scenario and perproduct basis. Analysis of the results allows us to verify our understanding of the modeling of cloud
2011 Elsevier Ltd. All rights reserved.

Atmospheric dispersion
Consequence modeling
Toxic materials
Sensitivity analysis
Phast software

1. Introduction
The prevention of technological risks requires industrial sites to
estimate the consequences of different accident scenarios based on
a probabilistic risk assessment. An important contribution to the
calculation of the consequences comes from the modeling of atmospheric dispersion, particularly of the accidental release of toxic
products. Given the implications in terms of land-use planning, it is
important that the calculations be based on the best scientic
knowledge available. This subject has motivated numerous studies
since the early 1980s through the development of numerical models
which are currently used for loss prevention purposes in chemical
process engineering. In order to increase condence in these
programs, a subsequent effort has been spent to validate dispersion
models by comparing measured and computed data (Calay & Holdo,
2008; Hanna, Hansen, Ichard, & Strimaitis, 2009; Kisa & Jelemensky,
2009; Luketa-Hanlin, Koopman, & Ermak, 2007; Middha, Hansen, &

* Corresponding author. Laboratoire de Gnie Chimique, Campus INP-ENSIACET,

4 Alle Emile Monso, BP 84234, F-31432 Toulouse Cedex 4, France. Tel.: 33 5 34 32
36 76; fax: 33 5 34 32 33 99.
E-mail addresses: (N. Pandya), nadine.gabas@ (N. Gabas), (E. Marsden).
0950-4230/$ e see front matter 2011 Elsevier Ltd. All rights reserved.

Storvik, 2009) according to recommendations edited by ASTM (2005)

for example.
Dispersion models can be classied into three categories, which
are, from the less to the more complex, Gaussian models, integraltype models and 3D or computational uid dynamics (CFD) models.
Gaussian models are derived from the diffusion equation and from
observations made by experimental work, i.e. the pollutant concentration follows a Gaussian distribution, whose standard deviations are
dependent on the atmospheric turbulence and the distance from the
source or the duration since the beginning of the release. These
models are appropriate for passive clouds and therefore for the last
stage of heavy-gas dispersion (passive dispersion).
Integral-type models are simplications of the conservation
equations for mass, momentum and energy. They model the transitions between different stages of heavy-gas dispersion: slumping
and creeping, transition phase and nally passive dispersion. These
box models provide relatively easy and fast dispersion estimations.
Some of them, like ALOHA, DEGADIS, HEGADAS and Phasts UDM,
are among the most popular and widely used in safety engineering
applications. Despite the convenience they offer in their application,
they appear to have some drawbacks: some physical phenomena
use semi-empirical relationships whose parameters have been
tuned on eld test data; since the trials usually do not include
obstacles, they can provide reliable results only in open eld

N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

In order to analyze the effect of terrain and of large obstacles on

gas dispersion, CFD codes (such as CFX, FLUENT, FLACS and FDS)
have been developed. This approach (simultaneous resolution of
balance equations of mass, momentum and energy) allows a full
three-dimensional analysis to be performed. The wind velocity is
completely resolved, unlike simpler models where velocity is
a single value or a function of height. They can deal with heavy,
neutral or light gas dispersion. While providing more detailed
results, they require more computational resources and analyst skill.
They are starting to play an important role in risk assessments for
the process industry because they have the potential better to assess
the impact of certain barrier systems and of terrain effects. In this
context, a number of recent papers have compared simulation
results obtained by CFD and integral models (Fiorucci et al., 2008;
McBride, Reeves, Vanderheyden, Lea, & Zhou, 2001; Mouilleau &
Champassith, 2009; Pontiggia, Derudi, Alba, Scaioni, & Rota, 2010;
Riddle, Carruthers, Sharpe, McHugh, & Stocker, 2004) or aimed to
improve turbulence modeling (Pontiggia, Derudi, Busini, & Rota,
2009; Sklavounos & Rigas, 2004), often the main cause of gaps
between observations and numerical simulations.
The modeling of the impact generated by an accidental release
of hazardous chemical depends on a number of parameters related
to the release type, to the product and to the software itself:
conditions under which the dispersion occurs (meteorological and
environmental), physical properties and toxicity of the chemical
and internal parameters of the modeling tool. Simulation results
may depend quite strongly on the values chosen for some of these
parameters. While exibility in the parameter choice is useful, it
can lead to effect distances that vary considerably when different
experts study the same scenario (CCPS, 1996; MEEDDAT, 2008).
An important technique for developing condence in ones
understanding of a model is sensitivity analysis, which evaluates
how variations in a models outputs can be apportioned to variations
in the inputs. The most basic sensitivity analysis methods consist of
varying input parameters one at a time (OAT) while holding other
parameters at central values, so the sensitivity indices derived are
dependent on these central values. More sophisticated sensitivity
analysis techniques examine the global response (averaged over the
variation of all parameters) of model outputs by exploring the entire
input space: these are global sensitivity analysis methods (Saltelli,
Chan, & Scott, 2004).
Previous work on sensitivity analysis of atmospheric dispersion
models has been limited to OAT methods. In 1989, Kakko published
a quantitative sensitivity analysis of the RISKIT program by varying
source term parameters, surface roughness and local weather
characteristics. Kok, Eleveld, and Twenhfel (2004) have carried out
a sensitivity analysis of NPK-PUFF, a Lagrangian code used to model
release scenarios of radioactive contaminants. Ferenczi (2005)
undertook an OAT sensitivity analysis of a RIMPUFF code which
models radioactive pollutant dispersion. Bubbico and Mazzarotta
(2008) have applied an OAT method to 15-min accidental toxic
release scenarios of hydrogen chloride, ammonia, trimethylamine
and bromine using ALOHA and Trace 9.0 software tools. More
recently, Cormier, Qi, Yun, Zhang, and Mannan (2009) have carried
out a sensitivity analysis of the CFX CFD code concerning a limited
number of parameters (turbulence models, source term and meteorological conditions) to assess the effects on the distance to Lower
Flammability Limit and the concentration levels of LNG releases.
This paper presents our work on a global parametric sensitivity
analysis of the Unied Dispersion Model (UDM) of Phast, one of
the most comprehensive computer programs for the modeling
of accidental releases, used by companies and the competent
authorities. We present results concerning three materials which
are relevant for safety reports: nitric oxide (NO), ammonia (NH3)
and chlorine (Cl2).


2. Theoretical background
2.1. Dispersion modeling
2.1.1. Phast software tool
Phast (Process Hazard Analysis Software Tool) is a commercial
code for consequence assessment developed by DNV Software. It
simulates the evolution of an accidental release from the initial
point to far-eld dispersion, including modeling of rainout, pool
vaporization and evaporation, while accounting for ammable and
toxic effects. Phast is able to simulate various source terms such as
leaks, line ruptures, long pipeline releases and tank roof collapses in
both pressurized and unpressurized vessels and pipes, which are
combined with Phasts Unied Dispersion Model (UDM) to obtain
desired consequence results: for example, i) concentration at
a given distance, ii) distance to hazardous concentration of interest,
iii) transition through various stages such as jet phase, heavy phase,
transition phase and passive dispersion phase, and iv) footprint of
the cloud at a given time. Phast version 6.54 has been used in this
2.1.2. UDM
UDM models jet, dense, buoyant and passive dispersion
including droplet rainout and re-evaporation (Cook & Woodward,
1995; Witlox, 2006). In this model, the simulation of the progression of the cloud resulting from a release follows several phases
without discontinuous transitions between them. A similar formulation is used for both continuous and instantaneous releases, with
a transition from the former to the latter for short duration releases.
The model has been validated against eld scale experimental data.
In the case of a continuous release of heavy gas, Fig. 1 shows the
different phases of the cloud along downwind distance: jet
(elevated, touching down then ground level), heavy, transition and
passive (Gaussian concentration prole). The jet phases are dominant initially, followed by the transition phase and the fully passive
one. The transition phase is implemented to avoid discontinuities in
the cloud behavior between near-eld and far-eld dispersion.
A set of differential equations is integrated to give the key
variables as a function of distance (for a continuous release) or time
(for an instantaneous release). These are solved to obtain other
variables describing the dispersing cloud. The same type of differential equations is resolved throughout all phases of dispersion (jet,
dense and passive), although the terms vary as the cloud passes
from one phase to the next. The model equations for the overall
behavior of the dispersing cloud implement conservation of total
cloud mass, conservation of heat transfer, conservation of
momentum and cross-wind cloud spreading in heavy phase.
Total air entrainment (Etot) into a plume is modeled in Phast as
the sum of several mechanisms (the meaning of symbols is presented in Nomenclature section):
 jet entrainment (Ejet) caused by turbulence resulting from the
difference between plume speed and ambient wind speed; it is

Fig. 1. Dispersion phases according to the UDM (case of heavy-gas behavior).

(: centerline of the cloud, wind direction according to x).


N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

present both for a jet and a plume which moves less fast than
the wind. It is proportional to the a1 parameter according to the
following equation:


0:5$p0:5 $a1 $Pabove $ra $jucld  ua cos qj


 cross-wind entrainment (Ecross) associated with the formation,

in the wake of a rising or falling plume, of trailing vortices in
response to the deection by the release plume of ambient air.
It is proportional to the a2 parameter according to the following

Ecross a2 $ra $Pabove $jua sin qj


Ecross is zero for a horizontal release, and maximal for a vertical

 passive entrainment caused by ambient turbulence. It is
present both in the near eld (Enf
pas) and the far eld (Epas). In
the near eld, it is proportional to the epas parameter according
to the following equation:


$p$ra $epas $31=3 $ l4=3


 heavy-gas entrainment (Ehvy) due to the difference between

plume density and ambient air density.
During the heavy phase, cloud spreading along cross-wind
direction is proportional to the dense cloud spreading parameter
(CE) according to the following equation:


ux $Cm

g$max0; rcld  ra z zcld $Heff 1 hd



 cloud density is sufciently close to the ambient density

(depending on parameter rpas
ro )
 cloud speed is sufciently close to the ambient speed
(depending on parameter rpas
u )
 passive entrainment is sufciently close to total entrainment
(depending on parameter rpas
E )
 after touchdown, the Richardson number is sufciently small
(depending on parameter Ripas
* ).
Parameter CDa intervenes in the calculation of air-borne drag
force, Fair
drag, which is perpendicular to the cloud propagation axis.
During the jet phase, Fair
drag between the cloud and the surrounding
air is proportional to the drag coefcient CDa per the following

 class A: source term and weather condition variables (Pst, Tst,

DO, Lh, Ta, Pa, Ha, Dur, ZR, q, Z0, ua, Cpa)
 class B: parameters permitting to select sub-models of Phast
(see Table 1 for examples)
 class C: internal parameters of physical models (see Table 1 for
 class D: numerical resolution parameters (see Table 1 for
In the following, we have studied the impact of varying all of
these parameters simultaneously.
2.2. Sensitivity analysis
A key stage of the study is to select a suitable sensitivity analysis
technique and relevant software for its implementation.
2.2.1. Methods
There are three types of sensitivity analysis techniques:
screening, local and global methods.
The aim of screening methods is to identify the most important
parameters from amongst a large number that affect model outputs.
They are useful for models which are computationally expensive to
evaluate and/or have a large number of input parameters. Various
strategies and methods have been discussed in several articles
with illustrative examples (Campolongo, Cariboni, & Saltelli, 2007;
Morris, 2006).
Local methods calculate the slope of the model output at a given
point in the input space by calculating the partial derivative at that
Table 1
Studied parameters and values.

The transition to passive dispersion occurs if the following four

conditions are satised:

drag CDa $Pabove $ra $ua $sin q

2.1.3. Classication of Phast parameters

We have classied the Phast input variables and parameters as


[default value in Phast]

Meaning of the discrete parameters

Wp [2]
TPp [3]

1: Constant, 2: log-Power law

1: Both constant, 2:both linear,
3: logarithmic temperature and linear
1: None, 2: heat only
1: Yes, 2: no
1: Wet soil, 2: dry soil, 3: concrete,
4: insulating concrete
1: Rainout, equilibrium, 2: rainout,
2: Synchronized, 6: rigorous

HT [2]
Mixing [1]
Spool [3]
Mthrm [2]
Mdrop [2]

[default value in Phast]

of the range

of the range

a1 [0.17]
a2 [0.35]
CDa [0]
g [0]
CE [1.15]
epas [1]
u [0.1]
ro [0.015]
rtr [2]
Ri [20]
rquasi [0.8]
Ripool [0.015]
Entpool [1.5]



Hmineff [0.02]



The default value of the CDa parameter in the version of Phast

which we have tested is 0, which cancels the contribution of this
Finally, Wp is a discrete parameter which may inuence
dispersion cloud phenomena. It determines the vertical wind
prole, which can be either a constant or a power law increasing
with distance from ground level.

N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

point (Turnyi & Rabitz, 2004). While easy to implement, local

sensitivity analysis is an OAT method, and the sensitivity of
a specic parameter depends on the central values chosen for the
other parameters.
Global techniques are more complex, but incorporate the whole
range of variation and the probability density function of the input
parameters to calculate their inuence on the output. An important
class is variance-based methods such as Fourier Amplitude Sensitivity Test (FAST), one of the most popular, chosen in this study. The
core feature of FAST is that it explores the multidimensional space
of the inputs using a search curve that scans the entire input space.
The method provides quantitative measures of sensitivity in the
form of sensitivity indices. The rst-order sensitivity index, Si, of an
input parameter Xi, is a measure of the intrinsic effect of Xi on the
output variance Var(Y):




Second-order sensitivity indices Sij (where i s j), measure the

interaction effect of Xi and Xj on the output variance:



Var E Y=Xi ; Xj  EY=Xi  E Y=Xj


Other higher-order indices (such as S123) are dened in the same

manner. Note that according to these equations all the sensitivity
indices are comprised between 0 and 1. In 1999, Saltelli et al.
(Saltelli, Tarantola, & Chan, 1999) introduced an Extended version
of FAST (EFAST) to compute the so-called total sensitivity indices
especially suited to apportion the model output variation to the
input factors in a comprehensive manner. The total sensitivity
index is dened as the sum of all the sensitivity indices involving
the factor in question. For example, supposing that we have three
factors in our model, the total effect of factor X1 (ST1) on the output
variance Var(Y) is given by:

ST1 S1 S12 S13 S123


The EFAST method permits an evaluation of model sensitivity

through calculation of rst-order and total sensitivity indices to
 the input parameters that are very inuent alone (direct
inuence) on an output variance: those having highest rstorder sensitivity indices,
 the input parameters that have little direct impact on the
considered output: those having low rst-order sensitivity
indices (near 0),
 the input parameters that show signicant total inuence
(directly and/or in interaction with other parameters) on an
output variance: those having highest total sensitivity indices.


Table 2
Bifurcation parameters and corresponding release conditions.








Bifurcation parameters
q (degree)
ZR (m)
SCa/ua (m s1)
Q (kg s1)






Release conditions
Vf (m s1)
Tf (K)
Lf ()







SC means Pasquill stability class.

D corresponds to neutral atmospheric conditions and F to stable conditions.

package developed by the European Commissions Joint Research

Centre, which is freely available and implements the EFAST
method. An advantage of SimLab is that the user can link it to an
external model via executable les.
3. Selected hazardous materials
Three toxic materials stored under various conditions were
chosen: nitric oxide (NO), ammonia (NH3) and chlorine (Cl2).
Considering their physico-chemical properties, they are likely to
exhibit different behavior during dispersion. They also cover a range
of common scenarios in safety studies.
3.1. NO
NO is a neutral gas (molar mass of 30.01 g mol1) whose
behavior is similar to that of ambient air. It is usually stored in
a pressurized tank. Only three phases are likely to occur during the
dispersion of its cloud: jet phase, transition phase and passive
phase. It means that the jet phase at ground level, shown in Fig. 1,
may not exist so the plume centerline remains above the ground
level even when the cloud edge touches the ground.
3.2. NH3
Ammonia is widely used in industry. It is usually stored in the
liquid phase in pressurized vessel. After its emission, a two-phase
ow occurs forming an ammonia cloud composed of vapor and
very ne droplets that do not fall to the ground. The droplets
evaporate quickly, cooling the air. It results a cold mixture of air and
ammonia, denser than the ambient air, even though pure gaseous
ammonia is lighter than air at ambient temperature (vapor density
of 0.73 at 288.15 K). For NH3 releases, all the dispersion phases
normally occur. The dispersion of ammonia releases has been widely

2.2.2. SimLab software tool

There are several specialized tools for performing sensitivity
analysis (Chan, Scott, & Andres, 2004). We have chosen the SimLab



Sensitivity indices
Fig. 2. Tree of scenarios.

Fig. 3. Methodology for sensitivity analysis.



N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

Table 3
Model outputs.



Concentration at a given downwind distance, x

C200, ppm
C1k, ppm
C10k, ppm

x 200 m
x 1000 m
x 10 000 m

Downwind distance corresponding to SEIa

Downwind distance corresponding to SEL1b

X(SEI), m
X(SEL1), m

SEI 80 ppm
SEL1 600 ppm




SEI 354 ppm

SEL1 3400 ppm

SEI 19 ppm
SEL1 110 ppm

SEI: threshold for irreversible effects.

SEL1: threshold for 1% lethality.

studied (Deaves, 1992; Hanna, Strimaitis, & Chang, 1991; Khan &
Abbasi, 1999; Kisa & Jelemensky, 2009; Pontiggia et al., 2010).

scenarios shown in Fig. 2. As this gure shows, the release rate (Q),

q, Pasquill stability class (SC)/ua and ZR have been chosen as

Chlorine is a highly toxic gas produced and used in large

quantities all over the world. It has a wide range of uses that need to
be transported and stored as a liqueed compressed gas. Despite
the efforts conducted to achieve a high degree of safety in the
production, transportation and handling of chlorine, accidental
releases still occur with quantities of chlorine which may vary from
hundreds of kilograms to tons (BARPI, 2002). The high vapor
density (3.04 at 288.15 K) of chlorine leads to the development of
denser than air clouds, whose dispersion and dilution are lower
than that of passive ones. Such dense clouds may stay and persist at
ground level, which corresponds to human breathing level. Similarly to NH3, Cl2 discharges ash on expansion to atmospheric
conditions, which can lead to aerosol formation, rainout and pool
vaporization. Release scenarios of chlorine have been widely
studied, even recently (Deaves, 1992; Deaves & Hall, 1990; Hanna
et al., 2009; Janour, Jurcakova, Brych, Dittrt, & Dittrich, 2010;
Kakko, 1989; Khan & Abbasi, 1999; Lines, Daycock, & Deaves, 2001).

bifurcation parameters, because they may lead to different

physical phenomena during dispersion. We therefore study two
different values of these bifurcation parameters in separate subscenarios, leading to a total of 16 sub-scenarios for each product.
Q is a pseudo-parameter resulting from the combination of DO, Tst,
Pst input parameters. Note that the typical values of Q e Q1 (low
release rate) and Q2 (high release rate) e were obtained from
standard industrial storage conditions of the products (Tst and Pst):
Q1 corresponds to DO equal to 0.025 m (NO and Cl2) or 0.05 (NH3)
and Q2 to DO equal to 0.16 m (NH3 and Cl2) or 0.5 m (NO). The
release conditions studied are reported in Table 2.
For each sub-scenario, sensitivity analyses have been undertaken by simultaneously varying parameters listed in Table 1
(parameters in italic are not relevant for NO scenarios whose
release is only in gas phase). The parameter ranges used for the
sensitivity analysis were chosen with the help of expert users who
collaborated on the project or from the lower and upper bounds
allowed by the software. In the following, a scenario is identied by
the product formula associated with its number in the tree of
scenarios: for example NO-1 for the scenario n 1 of NO.

4. Methodological approach

4.2. Methodology

In this paragraph, we describe the strategy adopted to select the

release scenarios for each product and the methodology developed
to execute sensitivity analysis of Phast.

A methodology to carry out sensitivity studies has been developed by linking Phast and SimLab and running Monte-Carlo
experiments. It comprises six steps, as shown in Fig. 3:

3.3. Cl2

4.1. Analysis strategy

For each material, we have carried out an initial screening test
by varying all the parameters belonging to the four classes (A, B, C
and D). The minimum and maximum bounds for each input
parameter were selected with the help of experienced Phast users
to be representative of industrial use of each material. The details of
these simulations are presented elsewhere (Pandya, 2009). We
conrmed that class A variables (controlling source term and
weather conditions, such as DO, ZR, q and ua), have a greater
inuence than parameters from classes B, C and D (a contrary
nding would not have been reassuring for the quality of the
model). Consequently, the effects of parameters in classes B, C and
D, which are less well known by users of the tool, are potentially
masked. In order better to understand the inuence of these
parameters, we have undertaken detailed sensitivity analyses of
scenarios where the class A parameters do not vary.
For each product, we have dened a base-case scenario
(Ta 288.15 K, Pa 1.013  105 Pa, Ha 0.7, Z0 0.951 m1, Tgnd Ta)
divided into representative sub-scenarios according to the tree of

Corresponding to Regular large obstacle coverage (suburb, forest) in Phast.

1. The description of each input parameter (minimum and

maximum values, distribution type) is specied by the user,
saved in the controller and sent to SimLab. In the results presented, we have always used uniform distributions.
2. The set of sample points created by SimLab is saved in the
controller. We have used sample sizes of 20 000 points, as
a compromise between uncertainty in sensitivity index estimation and experiment running time.
3. For each sample point, Phast calculates the selected outputs.
4. The outputs are sent to the controller. Steps 3 and 4 take around
24 h, distributed over several computers.
5. The set of Phast outputs is transferred to SimLab.
6. SimLab calculates various sensitivity indices (rst order and
This experimental work led us to execute Phast several million
times over a two-year period.
5. Results and discussion
The results are relative to continuous 1 h discharges from a hole
in the storage tank (Phasts leak module). The cloud is assumed to
progress in an open eld (no impingement). The toxic averaging

N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32


Fig. 4. Plume of NO corresponding to scenario 4 simulated by Phast software.

time is taken equal to the release duration, as recommended in the

softwares User Manual. A number of relevant outputs related to
toxic threshold values have been selected as shown in Table 3. All
outputs are calculated at the reference height of 1.5 m, generally
used in safety case studies.
After describing details of the release dispersion for one of the
scenarios, we analyze the results per scenario and per parameter.

5.1. Description of cloud dispersion for scenario 4

As an example, we discuss scenario 4, which corresponds to
a horizontal release 20 m above the ground with a low release rate
and F/3 atmospheric conditions (stable, 3 m s1 wind speed).

Simulations with Phast give the evolution of the cloud as shown in

Fig. 4 for NO, in Fig. 5 for NH3 and in Fig. 6 for Cl2.
In Fig. 4, we can notice that the NO cloud centerline tends to
remain at the same height (about 17 m at 10 km downwind
distance). Density of the cloud passes from 1.59 (Tf 230.1 K) at
0 m downwind distance to 1.23 (Tf 288.1 K) at 10 km downwind
distance. The cloud edge touches down at 184 m distance. The
dispersion is passive from 123 m downwind distance.
For NH3, according to Phast simulation, the initial liquid fraction
is near 90% with total droplet evaporation occurring at 80 m from
the source (rainout equal to 0.06%). In Fig. 5, we can see that after
the jet phase the cloud falls rapidly and the cloud centerline
touches the ground at 224 m. At the source, cloud density is relatively high (8.14, Tf 239.6 K); when the dispersion becomes

Fig. 5. Plume of NH3 corresponding to scenario 4 simulated by Phast software.


N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

Fig. 6. Plume of Cl2 corresponding to scenario 4 simulated by Phast software.

passive (at about 300 m) the cloud density is equal to 1.22

(Tf 285.2 K).
For Cl2 the initial liquid fraction is near 85% with total droplet
evaporation occurring at 82 m from the source (no rainout is
observed). The edge of the cloud touches the ground around 80 m
(Fig. 6). Cl2 centerline touches the ground at 191 m downwind
distance. Density of the cloud passes from 22.7 at 0 m (Tf 238.8 K)
to 8.0 at 1 m (Tf 232.3 K) and to 1.23 at 60 m (Tf 285.6 K).
5.2. Per-scenario analysis
We represent the sensitivity indices on a color gradation scale,
where red means sensitivity index equal to 1 (output variability can
entirely be attributed to the variation of this parameter) and light
gray means the index is equal to 0 (parameter has no impact on the
output). An upward arrow in a matrix position indicates that the
Pearson correlation coefcient is greater than 0.2, meaning that
when the parameter increases, the output also tends to increase.
A downward arrow indicates that the correlation coefcient is less
than -0.2, meaning that the output decreases when the parameter
increases. The absence of an arrow indicates that there is no
signicant linear correlation.
We choose to present results relative to 3 selected scenarios:
number 2 (horizontal elevated release under D/5 weather conditions), number 4 (horizontal elevated release under F/3 weather

Table 4
First-order sensitivity indices (Si) for scenario NO-2.

conditions) and number 5 (vertical release near the ground under

D/5 weather conditions). Given that toxic threshold values of the
products are different (Table 3), we only compare concentration
outputs in the near, intermediate and far elds. Therefore, distance
outputs are discussed separately in Section 5.4. We list below the
most inuential parameters for each scenario. The analysis of the
inuence of these parameters is given in Section 5.3.
5.2.1. Scenario 2
Tables 4, 6 and 8 represent the rst-order indices and Tables 5, 7
and 9 the total order indices for scenarios NO-2, NH3-2 and Cl2-2
According to Table 4 relative to NO-2, the parameters which have
the greatest impact on concentration outputs are, in descending order:
 C200: epas (Si 0.22), Wp (Si 0.21), rpas
E (Si 0.045)
 C1k: epas (Si 0.32), Wp (Si 0.08), a1 (Si 0.057)
 C10k: Wp (Si 0.7).
For C200, this means that around 22% of the variability in the
near-eld concentration is caused by variations in the near-eld
passive entrainment parameter (epas), 21% by the vertical wind
prole parameter (Wp) and the remainder mostly to interactions
between parameters (non-additive contributions to the output

N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32


Table 5
Total order sensitivity indices (STi) for scenario NO-2.

Total order indices for scenario NO-2 represented in Table 5,

reveal that several parameters have an interaction effect on the
outputs. For example, for C200, in addition to the direct effect of
parameters listed above, jet entrainment coefcient (a1) inuences
the output in combination with other parameters. Besides, few
parameters have an interaction effect on C10k. This is due to the
high inuence of Wp (Si w 0.70) alone on this output. This also
indicates that the cloud dispersion is governed by several mechanisms in the near to medium eld and when the cloud becomes
passive, it is mainly affected by the wind prole (variation of wind
speed according to height).
According to Table 6 relative to NH3-2, the parameters affecting
concentration outputs are in descending order:
 C200: a1 (Si 0.26), a2 (Si 0.23), epas (Si 0.039)
 C1k: Wp (Si 0.18), a1(Si 0.12), a2 (Si 0.11), CE (Si 0.1),
g(Si 0.037)
 C10k: Wp (Si 0.8)
Concerning total order indices for scenario NH3-2 (Table 7), we
can note that some parameters which have no rst-order inuence
on C200 have strong higher-order inuence (for example solar
radiation ux (Sux), Wp and cross-wind spreading parameter (CE)).
Note that higher-order inuences, representing the impact of joint
variation of parameters, would not be detected using a simpler
OAT-type sensitivity analysis. They cannot be decomposed in detail
with the sensitivity analysis method that we have used (we are only
able to measure the sum of the higher-order sensitivity indices).
Regarding C10k and Wp, the comment made above for NO-2 is also
true for NH3-2.
According to Table 8 relative to Cl2-2, the parameters with the
greatest impact on concentration outputs are, in descending order:
 C200: epas (Si 0.17), a1 (Si 0.12), a2 (Si 0.08)
 C1k: Wp (Si 0.13), epas (Si 0.1), Mdrop (Si 0.04)
 C10k: Wp (Si 0.45)
Total order indices for scenario Cl2-2, represented in Table 9,
show a generalized strong interaction effect in the three elds.
Fig. 7 shows a comparison of scenario 2 rst-order sensitivity
indices for NH3, NO and Cl2 (C200 output). Unlike scenarios NH3-2
Table 6
First-order sensitivity indices (Si) for scenario NH3-2.

and Cl2-2, the ground-level concentration in the near eld of NO-2

is only weakly affected by the jet (a1) and cross entrainment (a2)
parameters (cf. explanations in Section 5.3). Moreover, when
comparing scenarios NH3-2 and Cl2-2, we found that, for Cl2-2, epas
has a greater inuence on C200 than in the case of NH3-2. This
could be due to the fact that the NH3 cloud touches down more
quickly than Cl2 and near-eld passive entrainment parameter
(epas) has no inuence at ground level.
5.2.2. Scenario 4
Fig. 8 shows a comparison of scenario 4 rst-order sensitivity
indices for NH3, NO and Cl2 (C200 output). Note that only vertical
wind prole (Wp) has an inuence on NO-4, which is a gas phase
release where the cloud does not touchdown (Fig. 4). Furthermore,
for the three products, note that jet entrainment coefcient (a1) is
relatively inuential since the release is a jet with high initial speed.
The cross-wind entrainment coefcient (a2) is inuential only for
NH3 and Cl2, since in these two cases the cloud descends quickly in
the near eld (Figs. 5 and 6). For the chlorine release, maximum
non-passive entrainment fraction parameter (rpas
E ) is relatively
inuential since it affects the start of the transition phase: since this
gas is heavier than NH3 and NO, non-passive entrainment is mostly
due to heavy-gas entrainment.
5.2.3. Comparison of scenarios 2 and 4
Comparing scenarios 2 and 4 allows us to analyze the effect of
weather conditions on the dominant parameters. Scenario 2 is
associated with D/5 conditions (neutral Pasquill stability class
with 5 m s1 wind) and scenario 4 with F/3 (stable with 3 m s1
wind). This comparison is limited to the near eld (concentration
For the NO release, the median value of C200 is 115 ppm for NO2 and 228 ppm for NO-4. This is compatible with the fact that more
stable atmospheric conditions lead to less dispersion. The main
difference between these scenarios concerns the a1 parameter. It is
dominant in scenario 4 but has negligible effect in scenario 2. The
jet phase is signicant in the former case (jet entrainment rate,
Ejet [ near-eld passive dispersion entrainment rate, Enf
whereas the jet disperses rapidly in the latter case (Enf
pas [ Ejet).
Concerning NH3, we do not observe a signicant difference
between F/3 and D/5 conditions in the near eld. As expected, the


N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

Table 7
Total order sensitivity indices (STi) for scenario NH3-2.

median value of C200 is smaller for NH3-2 (2400 ppm) than for
NH3-4 (3600 ppm).
Concerning Cl2, a very heavy gas, note that the most signicant
parameter for Cl2-2 is near-eld passive entrainment parameter
(epas) and maximum non-passive entrainment fraction parameter
E ) for Cl2-4. The median values for C200 are 300 ppm for Cl2-2
and 650 ppm for Cl2-4. We know (cf. Section 2.1.2) that epas mainly
impacts Enf
pas and rE impacts heavy-gas entrainment rate (Ehvy). In
stable (F/3) conditions, the release touches down more quickly,
thus Ehvy has more impact than Enf
5.2.4. Scenario 5
Scenario 5 represents a vertical release at height of 1 m above
the ground with low release rate and atmospheric conditions D/5.
The cloud of NH3 or Cl2 ascends due to the vertical jet (with high
release velocity) and then falls to the ground because of the high
initial density of the cloud.
Fig. 9 shows the main inuential parameters for NO-5, NH3-5
and Cl2-5 in the near eld. For NO-5, the two most inuential
parameters are maximum non-passive entrainment fraction
parameter (rpas
E ) (Si 0.15) and drag coefcient of plume in air (CDa)
(Si 0.13) (cf. explanations in Section 5.3). The cloud of NO initially
forms a jet which rises in the atmosphere and remains at an
elevated height. For elevated clouds, the passive transition distance
depends on rpas
E , which determines the contribution of Epas to cloud
dilution and thus the near-eld concentration at ground level.
For NH3-5, the most inuential parameter is a2 (Si 0.18) which
appears in the calculation of cross-wind entrainment rate (Ecross)
(eq. (2)). Indeed, cross-wind entrainment is predominant because
the release is vertical (cf. more explanations in Section 5.3.2)
For Cl2-5, only epas has a direct (rst order) impact. Note
however that numerous parameters (a2, rpas
E , epas, g, CE, .) have an
interaction effect in the near eld, shown by their total sensitivity
indices (STi).
5.3. Per-parameter analysis
In the following analysis, we discuss the rst-order inuence of
parameters on all scenarios.
5.3.1. Effect of jet entrainment coefcient a1
Parameter a1 has a strong inuence in the near eld (output
C200) primarily for the horizontal release for all three products
Table 8
First-order sensitivity indices (Si) for scenario Cl2-2.

(Fig. 10). Indeed, in the near eld, the release velocity is much
higher than the wind speed, and thus jet entrainment dominates.
The importance of jet entrainment decreases as the dispersion
progresses and the clouds velocity decreases, explaining the lower
effect on C1k and lack of effect on C10k except for NO-11 and NO12. Fig. 10 shows that a1 is more inuential for the scenarios corresponding to a high release rate (scenarios 9e12), except NO-3.
Indeed, NO-3 represents a case where the difference between the
cloud speed and the wind speed (ucld  ua) is high (eq. (1)), because
the initial velocity is much greater than for Cl2 and NH3. Scenarios 1
and 3 (horizontal release near ground, for all products) also appear
in Fig. 10. This can be explained by the fact that the other parameters which potentially affect near-eld entrainment (a2, CDa, epas)
have little effect in this type of scenario (cf. explanations below).
5.3.2. Effect of cross-wind entrainment coefcient a2
Parameter a2 has a signicant inuence on output C200 for
almost all vertical release scenarios where release speed is high. A
vertical release produces a jet which rises then (for heavy gases)
falls to the ground. In certain horizontal and elevated (ZR 20 m)
releases, the a2 parameter shows the same inuence as for heavy
gases (for example, scenarios NH3-2 and NH3-4).
The impact of a2 is very low (Si < 0.02) for the NO releases,
because of the passive behavior of the cloud. For the NH3 and Cl2
releases, a2 has an impact until the intermediate eld, in particular
for vertical releases with a high release rate and for horizontal
releases at 20 m above ground level. The inuence of this parameter disappears (Si < 0.075) in the far eld for all scenarios.
5.3.3. Effect of near-eld passive entrainment parameter epas
Parameter epas (near-eld passive entrainment) has no inuence
for ammonia releases, because the jet effect has a greater impact
than ambient turbulence. For chlorine releases, storage conditions
lead to a lower release speed than for ammonia, so ambient
turbulence has a greater inuence. For NO releases, the inuence is
limited to the near and intermediate eld and to elevated releases.
Indeed, passive entrainment is greater for an elevated release than
for a release at ground level. The release velocity for these scenarios
is very high (500 m s1).
Parameter epas intervenes in the near-eld passive entrainment
calculation. This entrainment disappears when the cloud centerline
touches the ground. Hence, if a cloud remains above the ground for
a longer time, this parameter plays an important role. For scenario 2

N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32


Table 9
Total order sensitivity indices (STi) for scenario Cl2-2.

with an elevated release, this parameter appears to be inuential

for all three products, NO, NH3 and Cl2, despite their distinguished
physical properties. The inuence of epas is above all important for
NO because in most cases the cloud center does not touch the

5.4. Summary of results for concentration outputs

5.3.5. Effect of vertical wind prole Wp

Wp is a discrete parameter which controls the choice of vertical
wind speed prole, either constant or a power law increasing with z
(the second choice corresponding better to physical reality). It is the
only discrete parameter which has a signicant inuence in these
experiments. It leads to output histograms having a bimodal
distribution which corresponds to the two wind speed proles. This
parameters impact is mostly in the far eld, and a little in the
intermediate eld. Indeed, as distance from the source increases,
the cloud is more likely to have transitioned to passive dispersion,
when it will be more sensitive to wind speed.

Tables 10 and 11 summarize the main results of this work

relative to the three different toxic materials. They show the most
inuential Phast parameters, in decreasing order of importance,
corresponding to rst-order sensitivity index greater than 0.1. Note
that in these tables, vertical wind prole (Wp) is the only class B
parameter, the others belonging to class C.
Table 10 shows the near-eld concentration for scenarios 2, 4
and 5. Note that, for all scenarios, the kind of predominant
parameters varies as a function of material type, which conrms
that they exhibit different behavior during dispersion. For NO,
a typical neutral gas, the inuential parameters depend on the
release scenario. In the near eld, NH3 is in the jet phase for all
scenarios (2, 4 or 5) so the inuential parameters are a1 and/or a2.
For Cl2, the release speed being lower than for NO and NH3, the
near-eld passive entrainment parameter (epas) is inuential for
scenarios 2 and 5 (both D5 weather conditions). In scenario 4, due
to stable weather conditions (F3), jet and transition phase parameters are both predominant.
Table 11, specic to scenario 2, gives the inuential parameters
in the whole dispersion eld (near, intermediate and far). In the far
eld, Wp is the only dominant parameter because all the releases
are the passive phase. In the near and intermediate elds, the
inuential parameters vary according to the type of dispersion

Fig. 7. Comparative rst-order sensitivity indices (Si) for scenario 2 of NH3, NO and Cl2
(C200 output).

Fig. 8. Comparative rst-order sensitivity indices (Si) for scenario 4 for NH3, NO and
Cl2 (C200 output).

5.3.4. Effect of drag coefcient of plume in air CDa

This parameter has an effect in the near and intermediate eld
for NO releases and for vertical releases. This inuence essentially
disappears in the far eld, because q approaches zero (the clouds
axis is horizontal). This explains the low or zero inuence of this
parameter for horizontal releases (for all materials) and for twophase releases where the cloud is heavier than air (NH3, Cl2) and
thus touches down very quickly.


N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32
Table 10
Phast inuential parameters in the near eld for three selected scenarios.
Phast inuent parameters
(Si > 0.1)

(neutral gas)

light gas)

heavy gas)

Sc n 2: elevated, horizontal,
low release rate neutral
weather (D5)
Sc n 4: elevated, horizontal,
low release rate stable
weather (F3)
Sc n 5: near-ground vertical
low release rate neutral
weather (D5)

epas, Wp


epas, a1

Wp, a1


a2, rEpas, a1

rEpas, CDa



Table 11
Phast inuential parameters for scenario 2 (elevated, horizontal, low release rate,

Fig. 9. Comparative rst-order sensitivity indices (Si) for scenario 5 for NH3, NO and
Cl2 (C200 output).

Phast inuent parameters

(Si > 0.1)

(neutral gas)

light gas)

heavy gas)

Near eld (200 m)

Intermediate eld (1000 m)
Far eld (10 000 m)

epas, Wp

Wp, a1,a2

epas, a1
Wp, epas


5.5. Discussion concerning the distance outputs

5.6. Other applications of the sensitivity analysis
The distance outputs X(SEI) and X(SEL1) represent the
maximum downwind distance corresponding to the threshold for
irreversible effects and threshold for 1% lethality, respectively.
These effect distances are used in risk assessments to determine the
safety perimeters around industrial sites. Sensitivity analysis
results of these outputs for the scenarios 2 are presented in
Tables 4e9. In Table 12, the median values of X(SEI) and X(SEL1)
indicate whether the corresponding output is located in the near,
intermediate or far eld.
The NO horizontal elevated release (scenarios 2 and 4) does not
have an output for X(SEL1), because the threshold concentration
(measured at a height of 1.5 m above ground level) is not reached.
Note that the parameter sensitivities for output X(SEI) for the
ammonia release behaves identically to output C1k (inuence of
parameters Wp and a1). This is due to the fact that the mean value
of this output (Table 12) is around 1000 m. The same observation
also concerns the other outputs, which behave similarly to one of
C200, C1k or C10k. For this reason, we do not discuss parameter
sensitivity further for these outputs.

Our sensitivity analysis work provides information which is

useful to both users and developers of the modeling tool. Users
have little intuition on the effect of certain parameters in given
release conditions and for specic product characteristics, and the
information available in the software documentation is not sufcient to fully understand its impact. They may wish to adjust
certain parameters better to correspond to specic characteristics
of the release being studied, but will hesitate if they are unsure of
the parameters effects. Using a tool that we have developed to
quickly access sensitivity results for a specic product and release
condition, users are able better to understand the behavior of their
modeling tool.
Concerning the developers of the modeling tool, sensitivity
analysis techniques may help identify modeling inadequacies or
programming errors, if an internal parameter has a larger or lower
inuence than expected by modelers. It is also a useful software
testing tool, since sensitivity analysis runs involve executing the
tool with unusual combinations of inputs, which are rarely tested
by developers. We have encountered a number of situations where
the modeling tool raised internal errors for unusual (and often
physically unrealistic) combinations of inputs.

Table 12
Median values of distance outputs for scenarios 2, 4 and 5.

Fig. 10. Sensitivity ranking for parameter a1 and C200 output.

X(SEI) (m)

X(SEL1) (m)

Sc n 2



0 (No output)

Sc n 4



0 (No output)

Sc n 5




N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

6. Conclusion
In this paper, we have carried out a parametric sensitivity
analysis of Phasts Unied Dispersion Model in order better to
understand the inuence of user-adjustable input parameters on
model outputs. The EFAST variance-based global method used has
provided rst-order sensitivity indices, to characterize the intrinsic
inuence of each parameter, and total sensitivity indices which
represent a parameters inuence including its joint interaction
with all other input parameters. We have examined 1 h continuous
release scenarios relative to three toxic materials important for
safety studies: nitric oxide (passive gas stored as a pressurized gas),
ammonia (lighter than air stored as liqueed gas) and chlorine
(denser than air stored as liqueed gas).
The numerous input variables and parameters of Phast have been
distributed into four classes: (A) source term and weather conditions,
(B) model selection parameters, (C) internal model parameters and
(D) numerical resolution parameters. For each material, the base-case
scenario is divided into representative sub-scenarios according to the
release conditions which determine physical phenomena (ow rate,
release angle, release elevation and atmospheric stability class). A
number of relevant outputs related to toxic threshold values have
been selected (distances and concentrations). For each sub-scenario,
sensitivity analyses have been undertaken by simultaneously
varying the parameters of class B, C and D over wide input ranges.
The results have allowed ranking model parameters according
to their direct inuence on the variability of selected model
outputs, per-scenario and per-product. Moreover, interpretations of
the results have been proposed by considering the specic physicochemical properties of the selected materials which exhibit
different behavior during dispersion. Parameters impact is studied
separately for the near, intermediate and far eld, allowing us to
verify our understanding of how the cloud disperses.
This work provides users of Phast a better understanding of the
inuence of a number of parameters on the simulation of accidental
dispersion scenarios, which is particularly interesting considering its
implication in the calculation of safety perimeters and consequently
in land-use planning. Note that in ordinary use of Phast, only variables
in class A would be set by the user, with other parameters left at their
default setting. Parameters in classes B, C and D are occasionally
changed by users better to match specic features of the release being
modeled, or to overcome convergence problems.
From the point of view of Phast developers, it helps to distinguish
between adjustable parameters whose inuence is signicant for
certain products and release conditions (and which may deserve
further investigation), and parameters whose inuence is negligible.
In the continuity of this study, we will analyze the sensitivity of
Phast parameters in a narrow range around their default values,
which better corresponds to ordinary use of the tool.
The authors would like to acknowledge valuable technical
assistance from experienced Phast users working for the industrial
sponsors of this work (Air Liquide, Arkema, IRSN, Rhodia, SanoAventis, Technip, Total, URS). We also thank DNV Software for
useful feedback on our analysis.
ASTM (American Society for Testing and Materials). (2005). Standard guide for
statistical evaluation of atmospheric dispersion model performance. Designation
BARPI (Bureau dAnalyse des Risques et des Pollutions Industrielles). (2002). Accidents franais impliquant du chlore gazeux/liquide sous pression e Etat au 23/
10/2002. In ED6289.


Bubbico, R., & Mazzarotta, B. (2008). Accidental release of toxic chemicals: inuence
of the main input parameters on consequence calculation. Journal of Hazardous
Materials, 151, 394e406.
Calay, R. K., & Holdo, A. E. (2008). Modelling the dispersion of ashing jets using
CFD. Journal of Hazardous Materials, 154, 1198e1209.
Campolongo, F., Cariboni, J., & Saltelli, A. (2007). An effective screening design for
sensitivity analysis of large models. Environmental Modelling & Software, 22,
CCPS (Center for Chemical Process Safety Center). (1996). Guidelines for use of vapor
cloud dispersion models (2nd ed.). American Institute of Chemical Engineers.
Chan, K., Scott, E. M., & Andres, T. (2004). Software for sensitivity analysis e a brief
review. In A. Saltelli, K. Chan, & E. M. Scott (Eds.), Sensitivity analysis (pp. 451e462).
Chichester: John Wiley & Sons Publication.
Cook, J., & Woodward, J. L. (1995). Further development of the unied dispersion
model. In J. J. Mewis, H. J. Pasman, & E. E. De Rademaeker (Eds.), Loss prevention
and safety promotion in the process industries, Vol. II (pp. 93e104). Amsterdam:
Elsevier Science BV.
Cormier, B. R., Qi, R., Yun, G., Zhang, Y., & Mannan, M. S. (2009). Application of
computational uid dynamics for LNG vapor dispersion modeling: a study of
key parameters. Journal of Loss Prevention in the Process Industries, 22, 332e352.
Deaves, D. M. (1992). Dense gas dispersion modeling. Journal of Loss Prevention in
the Process Industries, 5, 219e227.
Deaves, D. M., & Hall, R. C. (1990). The effects of sloping terrain on dense gas
dispersion. Journal of Loss Prevention in the Process Industries, 3, 142e145.
Ferenczi, Z. (2005). Sensitivity analysis of the mesoscale puff model e RIMPUFF. In
Proceedings of 10th international conference on harmonisation within atmospheric
dispersion modelling for regulatory purposes, Sissi, Crete (pp. 599e603).
Fiorucci, A., Pontiggia, M., Derudi, M., Alba, M., Scaioni, M., Pendino, R., et al. (2008).
Risk assessment of dangerous products release and dispersion: a comparison
between CFD and integral models. Chemical Engineering Transactions, 13, 195e202.
Hanna, S. R., Hansen, O. R., Ichard, M., & Strimaitis, D. (2009). CFD model simulation
of dispersion from chlorine railcar releases in industrial and urban area.
Atmospheric Environment, 43, 262e270.
Hanna, S. R., Strimaitis, D. G., & Chang, J. C. (1991). Evaluation of fourteen hazardous
gas models with ammonia and hydrogen uoride eld data. Journal of
Hazardous Materials, 26, 127e158.
Janour, Z., Jurcakova, K., Brych, K., Dittrt, F., & Dittrich, F. (2010). Potential risks at an
industrial site: a wind tunnel study. Process Safety and Environmental Protection
Process, 88, 185e190.
Kakko, R. (1989). Vapour cloud modelling in the assessment of major toxic hazards.
Journal of Loss Prevention in the Process Industries, 2, 102e107.
Khan, F. I., & Abbasi, S. A. (1999). Modelling and control of the dispersion of hazardous
heavy gases. Journal of Loss Prevention in the Process Industries, 12, 235e244.
Kisa, M., & Jelemensky, L. (2009). CFD dispersion modelling for emergency
preparedness. Journal of Loss Prevention in the Process Industries, 22, 97e104.
Kok, Y. S., Eleveld, H., & Twenhfel, C. J. W. (2004). Sensitivity and uncertainty
analyses of the atmospheric dispersion model NPK-PUFF. In Proceedings of 9th
international conference on harmonisation within atmospheric dispersion modelling for regulatory purposes, Garmisch-Partenkirchen, Germany (pp. 79e83).
Lines, I. G., Daycock, J. H., & Deaves, D. M. (2001). Guidelines for the inclusion of low
wind speed conditions into risk assessments. Journal of Hazardous Materials, 83,
Luketa-Hanlin, A., Koopman, R. P., & Ermak, D. L. (2007). On the application of
computational uid dynamics codes for liqueed natural gas dispersion. Journal
of Hazardous Materials, 140, 504e517.
McBride, M. A., Reeves, A. B., Vanderheyden, M. D., Lea, C. J., & Zhou, X. X. (2001).
Use of advanced techniques to model the dispersion of chlorine in complex
terrain. Transactions of IChemE, 79(Part B), 89e102.
MEEDDAT (Ministre de lEcologie, de lEnergie, du Dveloppement Durable et de
lAmnagement du Territoire). (2008). Recueil de bonnes pratiques lies
lutilisation de logiciels de modlisation de la dispersion dammoniac (groupe
de travail national ammoniac).
Middha, P., Hansen, O. R., & Storvik, I. E. (2009). Validation of CFD-model for
hydrogen dispersion. Journal of Loss Prevention in the Process Industries, 22,
Morris, M. D. (2006). Input screening: nding the important model inputs on
a budget. Reliability Engineering & System Safety, 91, 1252e1256.
Mouilleau, Y., & Champassith, A. (2009). CFD simulations of atmospheric gas
dispersion using the re dynamics simulator (FDS). Journal of Loss Prevention in
the Process Industries, 22, 316e323.
Pandya, N. (2009). Analyse de sensibilit paramtrique dun outil de modlisation des
consquences de scnarios daccidents. Application la dispersion atmosphrique
de rejets avec le logiciel Phast. PhD thesis, Institut National Polytechnique de
Toulouse, France.
Pontiggia, M., Derudi, M., Alba, M., Scaioni, M., & Rota, R. (2010). Hazardous gas
releases in urban areas: assessment of consequences through CFD modeling.
Journal of Hazardous Materials, 176, 589e596.
Pontiggia, M., Derudi, M., Busini, V., & Rota, R. (2009). Hazardous gas dispersion:
a CFD model accounting for atmospheric stability classes. Journal of Hazardous
Materials, 171, 739e747.
Riddle, A., Carruthers, D., Sharpe, A., McHugh, C., & Stocker, J. (2004). Comparisons
between FLUENT and ADMS for atmospheric dispersion modeling. Atmospheric
Environment, 38, 1029e1038.
Saltelli, A., Chan, K., & Scott, E. M. (2004). Sensitivity analysis. Chichester: John Wiley
& Sons Publication.


N. Pandya et al. / Journal of Loss Prevention in the Process Industries 25 (2012) 20e32

Saltelli, A., Tarantola, S., & Chan, K. (1999). A quantitative, model independent
method for global sensitivity analysis of model output. Technometrics, 41, 39e56.
Sklavounos, S., & Rigas, F. (2004). Validation of turbulence models in heavy gas
dispersion over obstacles. Journal of Hazardous Materials, 108, 9e20.
Turnyi, T., & Rabitz, H. (2004). Local methods. In A. Saltelli, K. Chan, & E. M. Scott
(Eds.), Sensitivity analysis (pp. 81e99). Chichester: John Wiley & Sons Publication.
Witlox, H. W. M. (2006). UDM theory document. London: DNV Software.

CDa: drag coefcient of plume in air (e)
CE: cross-wind spreading parameter (e)
Cm: conversion factor between cloud half-widths
Cpa: atmospheric specic heat (J kg1 K1)
DO: orice diameter (m)
Dur: release duration (s)
E(Y/X): conditional expectation
Ecross: cross-wind entrainment rate (kg m1 s1)
Ehvy: heavy-gas entrainment rate (kg m1 s1)
Ejet: jet entrainment rate (kg m1 s1)
Effpas: far-eld passive dispersion entrainment rate (kg m1 s1)
1 1
s )
pas: near-eld passive dispersion entrainment rate (kg m
Entpool: pool vaporization entrainment parameter (kg m1 s1)
epas: near-eld passive entrainment parameter (e)
Etot: total dispersion entrainment rate (kg m1 s1)
drag: air-borne drag force (N m )
g: gravitational acceleration (m s2)
Ha: relative atmospheric humidity (e)
Heff: effective height of cloud after full touchdown (m)
hd: fraction of bottom half of cloud which is above ground (e)
Hmineff: minimum effective height of cloud (m)
HT: ag for heat/water vapor transfer (e)
Lf: release liquid fraction (e)
Lh: liquid height in storage tank above the release point (m)
ly: horizontal turbulent eddy length scale (m)
lz: vertical turbulent eddy length scale (m)
Mdrop: droplet equation solution method (e)
Mixing: mixing height constrain ag (e)
Mthrm: droplet evaporation thermodynamic model (e)
Pa: atmospheric pressure (Pa)
Pabove: cloud perimeter (m)
Pst: storage pressure (Pa)

Q: release rate (kg s1)

u : maximum cloud/ambient velocity difference parameter (e)
ro : maximum cloud/ambient density difference parameter (e)
E : maximum non-passive entrainment fraction parameter (e)
* : maximum Richardson number (e)
tr : distance multiple for phasing in full passive entrainment (e)
Ry: cloud cross-wind radius (m)
Ri: Richardson number for lift-off criterion (e)
rquasi: quasi-instantaneous transition parameter (e)
Ripool: Richardson number for passive transition above pool (e)
Sux: solar radiation ux (W m2)
Si: rst-order sensitivity index (e)
Spool: pool evaporation surface type ag (e)
STi: total sensitivity index (e)
Ta: atmospheric temperature (K)
Tf: release temperature (K)
Tgnd: ground temperature (K)
Tst: storage temperature (K)
TPp: vertical atmospheric temperature and pressure prole (e)
ua: wind speed (m s1)
ucld: cloud speed (m s1)
ux: wind speed at downwind distance x (m s1)
Var: variance
Var(E(Y/X)): variance of the conditional expectation
Vf: release velocity (m s1)
Wgnd: footprint half-width of the plume (m)
Wp: vertical wind prole (e)
x: horizontal downwind distance (m)
Xi: input parameter
Yi: output parameter
z: direction perpendicular and vertical to the downwind distance
zcld: height above ground of cloud centerline (m)
Z0: surface roughness length (m)
ZR: release height above ground (m)
Greek letters
a1: jet entrainment coefcient (e)
a2: cross-wind entrainment coefcient (e)
q: release angle to horizontal ( )
g: dense cloud side entrainment parameter (e)
ra: density of ambient air
rcld: density of cloud
3: dissipation rate of kinetic energy (m2 s3)