187
A publication of
CHEMICAL ENGINEERING TRANSACTIONS
The Italian Association
VOL. 31, 2013 of Chemical Engineering
Online at: www.aidic.it/cet
Guest Editors: Eddy De Rademaeker, Bruno Fabiano, Simberto Senni Buratti
Copyright © 2013, AIDIC Servizi S.r.l.,
ISBN 978-88-95608-22-8; ISSN 1974-9791
A Matter of Life and Death: Validating, Qualifying and
Documenting Models for Simulating Flow-Related Accident
Scenarios in the Process Industry
Trygve Skjold*, Helene H. Pedersen, Laurence Bernard, Prankul Middha,
Vagesh D. Narasimhamurthy, Thomas Landvik, Tommy Lea, Lars Pesch
GexCon AS, Fantoftvegen 38, P.O. Box 6015 Bergen Bedriftssenter, NO-5892 Bergen, Norway
trygve@gexcon.com
This paper describes an integrated approach for validating, qualifying and documenting numerical models
for simulating complex systems. Although the example used to illustrate the process entails simulations of
accident scenarios in the petroleum and process industries by means of computational fluid dynamics
(CFD), the methodology is not restricted to any particular model or system. CFD tools are applicable to
various aspects of societal safety, including transportation, storage and use of various energy carriers, as
well as malicious attacks involving toxic gas or condensed explosives. The approach adopted involves a
continuous process where relevant validation cases are classified according to the physical phenomena
involved, and prioritized based on parameters such as relevance for typical applications of the model
system, measurement quality and repeatability, availability of data, spatial scale, materials or substances
used, etc. A model evaluation protocol (MEP) provides guidelines for prioritizing the various validation
cases, and for evaluating the simulation results. Statistical methods and visualization techniques are
employed for describing the validation range and the associated uncertainties of the model system. Use of
the methodology is illustrated for a typical application of the commercial CFD tool FLACS: large-scale gas
explosions in congested geometries. The results highlight some of the inherent challenges associated with
the interpretation of results from large-scale experiments, and demonstrate how such challenges can be
addressed during the model evaluation process. The methodology can be extended to include sensitivity
studies and advanced optimization schemes for key model parameters.
1. Introduction
Major disasters continue to cause severe losses in the process industry and society in general. The
majority of the 100 largest property losses in the hydrocarbon industries from 1972 to 2011 involved fires
and explosions (Marsh, 2012). The Macondo disaster in 2010 demonstrated the devastating effects such
accidents can have on the environment (DHSG, 2011). Many organisations have adopted quantitative risk
analysis (QRA) as part of their approach for achieving satisfactory levels of safety (Vinnem, 2007).
However, there are significant uncertainties associated with most risk assessments, including the
completeness of the hazard identification processes, lack of relevant data for estimating the frequencies of
events such as loss of containment and ignition of flammable mixtures, and the topic of the present work:
how accurate are the models used for estimating the consequences of specific hazardous events?
Many accidents in the process industry involve complex fluid flow phenomena, with or without chemical
reactions (Mannan, 2012): release and dispersion of toxic, asphyxiating, radioactive or flammable material
in gaseous, liquid or solid form; gas, vapour, mist, dust or hybrid explosions; detonation of condensed
explosives and propagation of blast waves; jet and pool fires; etc. The type of models used for assessing
the consequences of such events range from the analytical expressions and empirical correlations or
nomographs in standards and guidelines, to phenomenological tools of varying complexity, and finally
sophisticated numerical model systems that solve conservation equations for fundamental parameters
such as mass, momentum and energy. Regardless of the complexity of the models, it is essential for the
188
quality of QRAs, and hence for safety and security, that risk analysts understand the underlying
assumptions and inherent limitations of the tools they use, as well as the level of accuracy they can expect
in the results. Both government bodies and industry show increasing awareness of the need to qualify
models for particular applications, for instance by requiring modellers to demonstrate the capabilities of
their models by reproducing results from specific sets of experiments (Ivings et al., 2007).
The validation and documentation process represents a fundamental challenge for developers of any
model system that aspire to describe a wider range of physical phenomena, or other initial and boundary
conditions, than the ones that can be mapped out by a finite number of experiments. Although the
governing equations for turbulent fluid flow are well established (Bradshaw, 1994), analytical solutions are
primarily of academic interest, and discrete solutions by direct numerical simulation (DNS) can still only be
realized for relatively simple systems. Models based on large eddy simulations (LES) have gained
increasing popularity in recent years. However, within the context of simulating industrial accident
scenarios, most commercial CFD tools still rely on turbulence models based on Reynolds-averaged
Navier-Stokes (RANS) equations (Launder and Spalding, 1974), often complemented with sub-grid models
to account for the influence of objects that cannot be resolved on the computational grid. For turbulent
reactive flows it is necessary to add models for chemical reactions, and to couple the resulting model
system (Hjertager, 1982). Several CFD codes for engineering applications have adopted the concept of
turbulent burning velocity ST for simulating premixed combustion. The speed of the propagating flame front
relative to the unburnt mixture is determined by an empirical expression on the form:
ST ∝ u′ A LB S LC ν D (1)
where u’ is the root-mean-square of the turbulent velocity fluctuations, L is a turbulent length scale, SL is
the laminar burning velocity, and ν is kinematic viscosity (or thermal diffusivity). Table 1 summarizes some
published values of the exponents in Eq. (1). It is evident that the values from literature span a
considerable range, and in Section 4 it will be shown how the validation system can be extended to
parameter optimization. Figure 1 shows the geometry model implemented in the commercial CFD code
FLACS for the test rig that will be used to illustrate the methodology presented in this paper.
Table 1: Examples of exponents in Eq. (1) from published combustion models; see references for details.
Publication A B C D
Bray (1990), used in FLACS v9.1 0.412 0.196 0.784 -0.196
Peters (1992,1999) 0.500 0.500 1.000 -0.500
Bradley et al. (1992) 0.550 0.150 0.600 -0.150
Zimont and Mesheriakov (1988) 0.750 0.250 0.500 -0.250
Kerstein (1988) 0.875 0.375 0.500 -0.375
Sensitivity range investigated in Figure 7 0.412-0.536 0.196 0.784 -0.196
Figure 1: Geometry model implemented in FLACS for the HSE test rig; dimensions 28 m x 12 m x 8 m.
189
2. Experiment and simulations
The methodology for validating, qualifying and documenting models for simulating flow-related accident
scenarios in the process industry will be illustrated for two repeated test series performed in a congested
offshore module (Evans et al., 1999): one series of five repeated experiments with central ignition (Alpha),
and one series of six repeated experiments with end ignition (Beta). The work was commissioned by the
UK Health & Safety Executive (HSE) and executed by BG Technology on the Spadeadam test site. There
were changes in the level of congestion between the two series, but all experiments were performed with
mixtures of natural gas and air for equivalence ratios in the range 1.05-1.14. The mixtures were ignited by
low-energy electrical discharges under initially quiescent conditions. The simulations have been performed
with the commercial CFD code FLACS v9.1 (GexCon, 2011; Pedersen and Middha, 2012).
3. Methodology and results
Figure 2 shows a schematic representation of the proposed methodology. Each potential validation case,
or instance, is classified according to the physical phenomena it represents. Relevant validation cases can
be experiments, accidents, detailed simulation results, or analytical solutions to idealized problems. The
schemes for classification and prioritization are illustrated in Figure 3. The categories defined for validating
modules in the CFD code FLACS include: wind (atmospheric flow); release and dispersion; fire; gas, mist,
dust and hybrid explosions; and blasts generated by condensed explosives or physical explosions. Each
category is further divided according to specific criteria, such as degree of congestion and confinement in
the case of gas explosions. The experiments in the HSE rig would typically belong to group 1B in Figure 3.
It is a challenge to define objective and unambiguous scales for categorizing validation cases based on
relevance, spatial scale, repeatability, etc. However, as long as a significant number of cases are
simulated the resulting uncertainty has limited influence on the outcome of the overall analysis.
Basic characteristics of each instance are registered in a database, together with relevant data for the
cases where the average score exceeds a certain threshold. Of particular concern with respect to
predicting the consequences of major accidents in industry is the lack of repeated large-scale experiments
of high quality. In this respect, the two test series from the HSE rig are quite unique. Figure 4 shows that
the repeatability is somewhat limited in both series, and Figure 5 shows measured and simulated
pressures as a function of distance from the ignition point. The spread in experimental results highlights
the inherent limitation with respect to the accuracy that can be achieved in CFD simulations.
Figure 2: Flow chart illustrating the work flow and main components in the integrated system for validating,
qualifying and documenting models for process safety applications.
190
Figure 3: Classification of validation cases (left), and categories for prioritization (right).
Figure 4: Spread in experimental results for repeated tests in the HSE rig; end ignition on the right.
Figure 5: Maximum measured and simulated pressures in the HSE rig as a function of distance from the
ignition point to the respective sensors; end ignition on the right. The experimental data were obtained
after smoothing the pressure traces with a 1.5 ms moving average. The simulation results were obtained
with FLACS v9.1 and cubical grid cells of size 0.8 m.
The validation cases that are registered in the database will be simulated according to their assigned
priorities, with particular focus on sensitivity analysis for variables such as critical model constants, spatial
and temporal resolution, initial and boundary conditions, etc. The model system includes tailor-made tools
for setting up, documenting and running simulations, quality assurance (QA), and utility programs for data
extraction and data reduction. Standard file formats for storing experimental data allows for visualization of
experimental and simulated results directly in the post-processor for the CFD tool. The performance of the
model system is determined based on criteria outlined in the model evaluation protocol (MEP), which for
most practical purposes follows the recommendations from MEGGE (1996). Figure 6 summarizes
simulation results for five grid resolutions: 0.4, 0.5, 0.8, 1.0 and 2.0 m cubical cells. The 2 m grid is clearly
too coarse for this problem, with only four cells across the flammable cloud, and it is not surprising that
these simulations severely under-predict the explosion pressures for both central and end ignition.
191
Figure 6: Geometric mean vs. geometric variance for the ratio of the predicted maximum overpressure to
the observed maximum overpressure for five grid resolutions; end ignition on the right.
For the remaining grid resolutions, the results from the CFD simulations are in fairly good agreement with
experimental values. The model under-predicts the mean explosion pressure, particularly for scenarios
with end ignition. However, Figure 5 shows that the main reason for this deviation can be traced to a
limited number of pressure sensors located in the far end of the module. The same measurements are the
primary cause of the spread in experimental results shown in Figure 4. It should be noted that FLACS has
been developed for simulating deflagrations, not detonations, and that the spatial and temporal resolution
in the simulations probably would need to be increased significantly in order to capture the most extreme
pressure peaks observed for scenarios with end ignition. The results highlight the need for developing
reliable criteria for predicting deflagration-to-detonation transition (DDT) in complex geometries. Previous
validation work has shown that FLACS performs significantly better for geometries with a higher degree of
confinement (Foisselon et al., 1998). As indicated in Figure 2, the validation framework is designed to
facilitate documentation of the software, including compilation of comprehensive validation reports. The
content of the validation database can be made available to users of the software through an online web
interface. Selected parts of the validation results should be included in user manuals and training material
for the CFD tool. The instances in the validation database will also be used for automated testing of the
software.
4. Parameter optimization
Once the system illustrated in Figure 2 is operational, it is straightforward to perform sensitivity studies and
parameter optimization. CFD codes for engineering applications rely on empirical constants, such as the
exponents in the ST correlations in Eq. (1). Figure 7 illustrates the potential for parameter optimization: by
increasing the value of the exponent A for u’ in the Bray correlation by 30 %, the model predictions for both
experimental series change from severe under-prediction to slight over-prediction. Table 1 shows that the
modified value, A = 0.536, is still well within the range of values that have been reported for this exponent
by other researchers. To modify the default values of model constants in a commercial CFD tool, such as
FLACS, would obviously require a thorough analysis of numerous experimental results. However, once the
validation database has been populated, the actual optimization process may proceed according to
methods known from chemical kinetics (Davis et al., 2004).
Figure 7: Geometric mean and geometric variance for the ratio of the maximum predicted overpressure
(0.8 m grid cells) to the maximum observed overpressure for different values of the exponent A in Table 1.
192
5. Conclusions
Developers of complex model systems for industrial applications can benefit significantly from adopting an
integrated approach to testing, validation, qualification and documentation. The proposed model evaluation
process entails a continuous process towards an extensive database of prioritized validation cases.
Standards for file formats, prioritization criteria, model evaluation, documentation, etc. facilitate efficient
validation, test driven development, QA, and preservation of corporate knowledge. It is straightforward to
extend the methodology to include sensitivity studies and optimization schemes for key model parameters.
Major disasters continue to cause severe losses in the process industry and society, and better validated
and more accurate models for consequence assessment may turn out to be a matter of life and death.
References
Bradley, D., Lau, A.K.C. and Lawes, M., 1992, Flame stretch rate as a determinant of turbulent burning
velocity, Philosophical Transactions: Physical Sciences and Engineering, 338, 359-387.
Bradshaw, P., 1994, Turbulence: the chief outstanding difficulty of our subject, Experiments in Fluids, 16,
203-216.
Bray, K.N.C., 1990, Studies of the turbulent burning velocity, Proceedings Royal Society A, 431, 315–335.
Davis, S.G., Mhadeshwar, A.B., Vlachos, D.G. and Wang, H., 2004, International Journal of Chemical
Kinetics, 36, 94–106.
DHSG, 2011, Final report on the investigation of the Macondo well blowout, Deepwater Horizon Study
Group (DHSG), <ccrm.berkeley.edu/pdfs_papers/bea_pdfs/dhsgfinalreport-march2011-tag.pdf>,
accessed 15.01.2013.
Evans, J.A., Exon, R. and Johnson, D.M., 1999, The repeatability of large scale explosion experiments,
Offshore Technology Report – OTO 1999 042, Health & Safety Executive, October 1999,
<www.hse.gov.uk/research/otopdf/1999/oto99042.pdf> accessed 15.01.2013.
Foisselon, P., Hansen, O.R. and van Wingerden, K., 1998, Detailed analysis of FLACS performance in
comparison to full-scale experiments, Christian Michelsen Research, Report CMR-98-F30058.
GexCon, 2011, FLACS v9.1 User’s Manual, <www.flacs.com>, accessed 31.12.2012.
Hjertager, B., 1982, Simulation of transient compressible turbulent reactive flows, Combustion Science &
Technology, 24, 159-170.
Ivings, M.J., Jagger, S.F., Lea, C.J. and Webber, D.M., 2007, Evaluating dispersion models for safety
analysis of LNG facilities, The Fire Protection Research Foundation, Quincy, MA, USA.
Kerstein, A.R., 1988, Pair-exchange model of turbulent premixed flame propagation, Proceedings of the
Combustion Institute, 21, 1281-1289.
Lauder, B.E. and Spalding, D.P., 1974, The numerical copmputation of turbulent flows, Computer Methods
in Applied Mechanics and Engineering, 3, 269-289.
Mannan, S., 2012. Lees’ Loss Prevention in the process industries: hazard identification, assessment and
th
control, 4 ed., Elsevier Butterworth Heinemann, Amsterdam, the Netherlands.
MEGGE, 1996, Gas explosion model evaluation protocol, Gas Explosion Model Evaluation Group,
European Communities, Directorate-General XII, Science Research and Development, The Steel
Construction Institute, Berkshire, UK.
Pedersen, H.H. and Middha, P., 2012, Modelling of vented gas explosions in the CFD tool FLACS,
Chemical Engineering Transactions, 26, 357-362.
Peters, N., 1992, A spectral closure for premixed turbulent combustion in flamelet regime, Journal of Fluid
Mechanics, 242, 611-629.
Peters, N., 1999, The turbulent burning velocity for large-scale and small-scale turbulence, Journal of Fluid
Mechanics, 384, 107-132.
Marsh, 2012, The 100 largest losses 1972-2009 – Large property damage losses in the hydrocarbon
nd
industry, 22 ed., Marsh Global Energy Risk Engineering, London, UK.
Vinnem, J.E., 2007, Offshore risk assessment: principles, modelling and applications of QRA studies,
Second edition, Springer, London, UK.
Zimont, V.L. and Mesheriakov, E.A., 1988, A model of combustion of partially premixed gases, in:
Structure of gas flames, Proceedings of international colloquium, Part I.I, Novosibirsk: ITPM SO AN
SSSR, 35-43 (in Russian).
