pJaM
Daniel Alvear e, Jorge A. Capote e, Allan Jowsey a,b, Cecilia Abecassis-Empis a, Pedro Reszka a
Packer Engineering Inc., Chicago, USAd CTICM, Paris, Francee GIDAI, Universidad de Cantabria, Santander, Spain
iversity
science and engineering. Its current applications range from
reugh,
afetylags
models have been published recently [714] and many more are
ARTICLE IN PRESS
Contents lists availab
.el
Fire Safety
Fire Safety Journal 44 (2009) 590602simulations to experiments have found them in reasonable Corresponding author. Tel.: +441316507214.forensic investigation to risk assessments, life safety, smokemovement and detection, sprinkler performance, structural
expected. In general, these aim at determining the level ofaccuracy and the range of applicability of given re models bymeans of comparison to experiments.
The large majority of the studies that have compared
agreement. These studies show that current modelling providesgood predictions of the average thermal effects of a re (e.g. hot
E-mail address: [email protected] (G. Rein).1 Now working at CSTB France.0379-71
doi:10.1its rst applications to real re engineering problems in the late1980s [2,3] and now is widely used [4,5] in many aspects of re
experimental understanding by about 10 years [6]. Several papersand standards addressing the verication and validation of reFire modelling is frequently used in current re safetyengineering practice but discussions have been ongoing for manyyears now about the accuracy and reliability of the process.
Computer re modelling was rst developed as a research toolin the 1970s [1] after the surge of computer resources. It reached
hot layers and smoke movement during every stage of thedevelopment, from ignition and ame spread to, and throashover and extinction.
Modelling is among the fastest developing areas in re sscience. However, its ability to reproduce re phenomenaFire modelling
Blind
Forecast
CFD
FDS
Validation
fuel packages, ignition source and ventilation conditions. The aim of the exercise was to forecast the re
development as accurately as possible and compare the results. The aim was not to provide an
engineering analysis with conservative assumptions or safety factors. Comparison of the modelling
results shows a large scatter and considerable disparity among the predictions, and between predictions
and experimental measurements. The scatter of the simulations is much larger than the error and
variability expected in the experiments. The study emphasises on the inherent difculty of modelling
re dynamics in complex re scenarios like Dalmarnock, and shows that the accuracy to predict re
growth (i.e. evolution of the heat released rate) is, in general, poor.
& 2009 Elsevier Ltd. All rights reserved.
1. Introduction behaviour, and design of re safety. Modelling is being used tostudy re dynamics in enclosures and to simulate ames, plumes,f Department of Fire Protection Engineering, Ung Arup, San Francisco, USAh Efectis, Paris, France
a r t i c l e i n f o
Article history:
Received 7 March 2008
Received in revised form
10 December 2008
Accepted 12 December 2008Available online 17 March 2009
Keywords:12/$ - see front matter & 2009 Elsevier Ltd. A
016/j.resaf.2008.12.008dinburgh, Scotland, UK
of Maryland, USA
a b s t r a c t
An international study of re modelling was conducted prior to the Dalmarnock Fire Test One in order to
assess the state-of-the-art of re simulations using a round-robin approach. This test forms part of the
Dalmarnock Fire Tests, a series of experiments conducted in 2006 in a high-rise building. The
philosophy behind the tests was to provide measurements in a realistic re scenario involving multiple
fuel packages and non-trivial re growth, and with an instrumentation density suitable for comparison
with computational uid dynamics models. Each of the seven round-robin teams independently
simulated the test scenario a priori using a common detailed description of the compartment geometry,a BRE Centre for Fire Safety Engineering, University of Eb Arup, London, UKcRound-robin study of a priori modellingTest One
Guillermo Rein a,, Jose L. Torero a, Wolfram Jahn a,Sylvain Desanghere d,1, Mariano Lazaro e, Frederick
journal homepage: wwwll rights reserved.redictions of the Dalmarnock Fire
mie Stern-Gottfried b,a, Noah L. Ryder c,owrer f, Andrew Coles g, Daniel Joyeux h,
le at ScienceDirect
sevier.com/locate/firesaf
Journal
studies were conducted for very simple re scenarios, that did not
event and thus is providing a forecast. Most re model validations
ARTICLE IN PRESS
Jouinclude realistic features such as multiple fuel packages, amespread, window breakage, non-trivial geometry, and the resource (heat released rate or mass loss rate of the fuel) wasprovided. However, re modelling is used in practice to studyscenarios that include several of these features and where the resource is unknown. Of the studies conducted to date, only a fewcompare modelling results with experimental data from realisticre scenarios [13,14], however none of these were conducted apriori or as blind simulations.
The validation of re models is an essential task for theadvancement of re safety engineering. One of the issues thatremain to be further explored is the evaluation of the entireprocess of re modelling, in which the mathematical model isonly a component. The assumptions made by the user, thecollection of data for input and the selection of the parametervalues (some available in the literature and from experiments,some not) are crucial components leading to the creation of theinput le. Furthermore, the interpretation, claimed accuracy andimplementation of the model output are key components thatfurther highlight the important role of the user in the interactionwith the model.
It is reasonable to consider that the current state-of-the-art ofre modelling is reected not only in the mathematical modelscapabilities, but also on how re dynamics is implementedthroughout the different stages of modelling. Thus, in order toassess the strengths and limitations of the process as a whole, allthe stages of re modelling need to be investigated bothindependently and as a whole. This study looks at the overallprocess but does not underestimate the value of analysing theindividual components.
The objective of this study is to compare the modelling resultsproduced a priori by different teams of modellers of a realistic rescenario, the Dalmarnock Fire Test One. Test One is part of theDalmarnock Fire Tests series [15], conducted in 2006 in a realhigh-rise building. The study provides a range of forecastedbehaviours and a sense of the robustness, consistency andsensitivity of current re modelling including the predictions ofre growth. The results are compared to the experimentalmeasurements to allow evaluation of the accuracy and reliabilityof the a priori process as a whole.
2. Round-robin studies in re science
A round-robin study involves the analysis of a commonscenario by several independent teams and then draws conclu-sions from the comparison of all results [16]. In re safety sciencethe most renowned round-robin study was that conducted byEmmons around 1968 [17] after his trip around the world visiting40 re laboratories in order to compare the different ammabilityratings of the same set of commonmaterials. The results showed alarge discrepancy and made a case for the critical need oflayer temperature) but do not address the accuracy of thepredictions of the re development (e.g. re growth and/or theheat released rate) or spatial resolution. In the great majority ofcases, the simulations have been conducted after the tests andwith good access to the recorded experimental data (this is knownas a posteriorimodelling). Thus, the comparisons are not blind andthe modelling may include some bias due to prior knowledge ofthe event being modelled. This bias may or may not be explicitlyreported along the modelling results. Only a few a prioriprediction studies are available in the literature [1113]. These
G. Rein et al. / Fire Safetyfundamental understanding of ammability tests. The publicationof his results prompted the international standardisation ofmaterial re testing. A more recent round-robin study [18] foundare open simulations, also called a posteriori, where the modelleris also provided with the results from the experiment. Only apriori simulations are free of the bias that could be introduced byprior knowledge of the development of the event. In speciedsimulations the modeller is directly given the input le to be runin the model.
In re modelling, only a few round-robins can be found in theliterature [16,19]. CIB organised and conducted a large andinternational blind round-robin for re models [11] but theresults were not made publicly available. Only one team publishedtheir results independently [12]. However, the essential objectiveof a round-robin is to present all the results together for anunbiased comparison. Presenting only a few selected cases canprovide a distorted view of the results. The validation exercisepublished in NUREG [8] could be viewed as a round-robin study;however it was conducted as an open exercise with access to theexperimental results. Very recently, a new blind modelling round-robin was conducted to simulate a re in a small hotel room [20].Nevertheless the lack of ample round-robin studies in remodelling is a pending issue of the discipline.
The present round-robin study involves a pool of participantscomposed of independent international teams, all working in theeld of re engineering and using re modelling as part of theirprofessional practice. There are representatives from mostbranches of re modelling, from fundamental and appliedresearch to nal engineering design. Due to the wide range ofparticipants, the results pertain to a wide range of end-users andallow certain conclusions to be made that reect on the state-of-the-art of re modelling. The participants worked independentlyand had access to a large, common pool of data obtained from theinitial conditions prior to the ignition of the re. Each teamsubmitted one or two simulations that, in their view, representedthe best prediction of the process based on their a prioriknowledge.
3. The Dalmarnock Fire Tests
The large-scale Dalmarnock Fire Tests [15] consist of three testsconducted in a 23-storey reinforced concrete building in Glasgow(UK), July 2006. The two tests of main interest here (henceforthreferred to as Test One and Test Two) were those conducted in twoidentical ats, whereas the third test involved only smokemanagement in a stairwell and will not be further discussed.The Dalmarnock Tests were set up to recreate a realistic rescenario involving multiple fuel packages arranged in an ordinaryfashion, consistent with real dwellings. Arrangements of this typeinvariably result in re growth that is not readily obvious and thusprediction of re development can be a challenge. Nevertheless,good agreement between the calibrations of heat ux gaugesperformed by different re laboratories, thus reinforcing thereliability of these calibrations.
A round-robin of modelling results typically involves theproduction of independent predictions of a common event.Conclusions are then drawn from the comparison of the differentpredictions and the real behaviour. ASTM E 1355 [7] denes threetypes of simulation: blind, open and specied. In blind simula-tions, also called a priori, the modeller is provided only thedescription of the initial scenario and is responsible for develop-ing the appropriate input from this description, including detailsof the geometry, material properties, re development, etc. Themodeller has no access to the experimental measurements of the
rnal 44 (2009) 590602 591the Dalmarnock compartment test was designed to maximise itsrepeatability. Ignition procedures and fuel distribution weredened in a manner such that potential variations could be kept
to a minimum. Furthermore, the comparison of the results fromDalmarnock Fire Test One and Test Two conrms that therepeatability of the tests was high [15, Chapter 4]. This studyconsiders only Test One for comparison against re modelling,however Test Two is briey discussed.
3.1. Description of Test One
Test One was held in a two-bedroom single family at, with theliving room set up as the main experimental compartment. TestTwo was conducted in an identical at but two oors below TestOne. An identical fuel arrangement was used in both tests. Bothres grew to ashover conditions but only Test One was allowedto continue burning during post-ashover. A detailed descriptionof the compartments, the fuel layout and the measurements hasbeen given by Abecassis-Empis et al. [21] and Rein et al. [15,Chapter 2], but an overview is included here for quick reference.
The at comprised a main corridor off which were twobedrooms, a bathroom and a living room, with a small kitchenoff to the side of the living room, as shown in Fig. 1. The mainexperimental compartment was the 3.50m4.75m, 2.45m highliving room, with a 2.35m1.18m set of windows (two panes) on
the west-facing wall, 1.11m above the oor (see Fig. 1). It wasfurnished as a regular living room/ofce. The general layout wassuch that most of the fuel was concentrated towards the backcorner (NE) of the compartment, away from the window and thedoors, with a fairly even fuel loading throughout the rest ofthe compartment (see Fig. 2) and no further loading elsewhere inthe at.
While the main source of fuel was a two-seat sofa stuffed withretardant exible polyurethane foam, the compartments alsocontained two ofce desks with a computer and a padded chaireach, as well as three tall wooden bookcases, a short plasticcabinet, three small wooden coffee tables, a range of paper itemsand two tall, plastic lamps. The bookcases were fully-laden withbooks, video tapes, paper-lled cardboard les and several otherplastic items, as was the small cabinet. Other minor living room/ofce items were arranged as they would be in real dwellings.Fig. 3a shows a photograph, taken before the test, of thecompartment corner where the ignition source, the sofa andnearby items are located. The fuel load density in the maincompartment was estimated to be 32kg/m2 of wood equivalent,whereas a typical value for ofce buildings is around 25kg/m2
[22]. The ignition source was a plastic wastepaper basket lledwith crumpled newspaper and approximately 500ml of heptane.It was placed in-between the sofa and a bookcase, directly below ablanket that was draped over the sofa arm.
ARTICLE IN PRESS
Corridor
CupboardBathroom
Hallway
Bedroom-2Bedroom-1
1 mN
Kitchen
MainCompartment
Fig. 1. Plan view of the at layout showing basic dimensions (to scale), rooms andwindows [21,15].
G. Rein et al. / Fire Safety Journal 44 (2009) 590602592Fig. 2. Furniture layout in the mFig. 3. View of the ignition source, the sofa and nearby items in the maincompartment: (a) before the re and (b) after the re [21,15].ain compartment [21,15].
ARTICLE IN PRESS
o [
st Tw
Journal 44 (2009) 590602 593During Test One all the doors in the at were left open, exceptthe bathroom and cupboard doors. The front entrance doorcommunicating to the oor access corridor was also left open.Windows of all compartments, excluding the kitchen and bed-room-2, were left closed. The kitchen window was left partiallyopen, creating one vent at the top and another vent at the bottomdue to the pivoting mechanism of the pane.
A large number of sensors were installed throughout the at inorder to obtain detailed measurements of the re development.The instrumentation layout was designed to provide a highdensity of sensors. More than 270 thermocouples were distrib-uted throughout the main compartment, at different locations andat several heights, to provide gas temperature data at high spatialresolution suitable for comparison with computational uiddynamics (CFD) simulations. Smoke obscuration was measuredusing 8 pairs of laser-receiver sensors. Gas velocity was measuredat the ventilation openings of the main compartment using 14 bi-
Fig. 4. Evolution of the average compartment temperature for Test One and Test TwOne was allowed to continue burning during the post-ashover stages whereas Te
G. Rein et al. / Fire Safetydirectional velocity probes in total. Additionally, more than 15video cameras, spread throughout the at, provided visualrecordings which allowed monitoring of the re development.Other measurements include temperatures within and heat uxesto the east wall (fully instrumented light-steel framing wall) andthe ceiling of the main compartment, a dozen smoke detectors indifferent rooms, and strain and deection gauges used to monitorthe deformation of key structural elements.
The test was initiated by igniting the contents of thewastepaper basket. The re rapidly spread to the blanket and afew seconds later to the cushions on the sofa. After about 4.5minof localised re on the sofa, the rst bookcase ignited, quicklyfollowed by the onset of ashover (5min after ignition), withsudden reduced visibility and the ignition of most exposed paperitems throughout the compartment. The re continued to burnand broke the kitchen window (glass fell out) 12min into the test.The north pane of the compartment window was externallyforced to break (glass fell out) at 13min. The other pane fell outdue to the re shortly before 19min into the re. The re was thenput out by the re brigade after 19min of burning. Most of the fuelload was burnt and the majority of the remains were unrecogni-sable. Fig. 3b shows a photograph, taken after the test, of theremains at the compartment corner where the ignition source, thesofa and nearby items were located. Fig. 4 shows the evolution ofthe average temperature in the main compartment during TestOne, where the shaded area indicates standard deviation. A moredetailed timeline of the test is described by Abecassis-Empis et al.[21] and Rein et al. [15, Chapter 3].
Test Two had an identical at geometry and fuel conguration.The only two signicant variations were a smaller amount ofheptane used for the ignition protocol and a drastically alteredventilation conditions. The door to the kitchen remained closedfor approximately 2min after ignition, the door to the corridor for3min, and the main door for 4min. This opening sequence wasmanaged through remote control by an operator and the decisionswere taken based on the remote real-time monitoring of the redevelopment [15, Chapter 4]. The main compartment windowswere also opened by remote control, approximately 45 s afterignition. All other windows of the at remained fully openedthroughout the test. In addition, ventilation conditions in Test Twoincluded a 1.2m high1.4m wide hole was previously made inthe south wall of the main compartment linking it with bedroom-1 (see Fig. 1). The large hole served a double purpose of providing
21,15]. The shaded areas indicate standard deviation within the compartment. Test
o was extinguished immediately after ashover.extra ventilation and allowing the test crew2 an alternative accessand evacuation path from the compartment. Despite these drasticdifferences, the Test Two re was seen to spread following thesame pattern as Test One. Fig. 4 juxtaposes the results of Tests Oneand Two. It can be noted that both tests show an almost identicaltime to ashover (only 10 s difference). The difference in theignition and ventilation conditions between both tests led to areduction in the average pre-ashover temperature by approxi-mately 50 1C and to signicantly lower spatial scatter. Post-ashover temperatures cannot be compared because re suppres-sion by re ghters immediately followed the onset of ashover inTest Two. Although both tests did not produce identical res thedifferences are relatively small, and it shows that the Dalmarnockcompartment test conguration provided a robust and reasonablyrepeatable re scenario for benchmarking.
3.2. The philosophy of the tests
It is a common conception that re experiments couldsignicantly differ from each other when run under the sameconditions, and therefore many repeats are necessary to achievevalidation data. Given the cost and complexity of large-scale re
2 Two members of the test crew were present in the room during the re
accompanied by Fire Service personnel.
individual items before ashover and the windows breaking and
instrumentation; and individual descriptions, material, dimen-sionand
men
panes was externally forced to break at 800 s after ignition, and
In total, ten simulations were submitted: eight CFD simula-tions using FDS4 [24], and two zone model simulations using
ARTICLE IN PRESS
G. Rein et al. / Fire Safety Jou594falling out. The ignition issue was addressed by dening anignition protocol for the rst item and by placing it in closeproximity to a large quantity of fuel arranged vertically, in theform of a bookshelf. This provided an ISO room corner test type ofconguration where it is guaranteed that ashover will beattained soon after the ignition of the secondary item. Thepotential bounds of variability of the data were thus establishedby using very different ventilation and ignition conditions for bothtests. For Test One, only the doors of the main compartment wereleft open, while for Test Two, all doors and windows were open,ensuring, together with the large opening in the south wall, freeow of air and smoke in and out of the compartment.
Comparison of the data from both experiments establishes thatDalmarnock Test One is a robust and reasonably repeatable rescenario. The test can therefore be used to assess different aspectsof re safety engineering practicein this case, re modelling.
4. Dalmarnock round-robin study
The aim of the study was to forecast the re dynamics of theset scenario. The teams were asked to forecast the test results asaccurately as possible, and to avoid an engineering analysis withconservative assumptions or safety factors, as is common for usein re safety design. All teams were given access to a commonpool of information about the test experimental set-up and initialconditions.3
4.1. Common description of the scenario
Each team was free to use the information provided as theysaw t according to their own criteria. There were no limitationsimposed on the teams for consulting the literature and to searchfor additional data regarding other re experiments or similartests. Any missing information, unclear information or additionaldetails were intended to be complemented by the teamsassumptions, research and external sources, as is common in remodelling work, frequently conducted in each teams practice.tests, it is evident that the number of repeats that will establishstatistical validity for the data will never be achieved. This featuredistinctive to re is associated to its inherent complexity and hasalways cast a shadow of uncertainty upon any modellingvalidation exercise. Experimental data used for comparison withmodelling results can never be considered as robust as it could bein other elds, such as small ames. For example, the SandiaFlames by Barlow et al. [23] have an excellent reproducibility andcount on diagnostics that provide detailed information ontemperature, velocities, species and turbulence statistics. Theseames have been extensively used to benchmark turbulentcombustion models. Such a high level of reproducibility anddetailed diagnostics cannot be achieved in large res, therefore, adifferent approach was taken in this study.
The philosophy behind the Dalmarnock Fire Tests was toprovide instrumentation density suitable for comparison withgrids used in CFD simulations and to arrange ignition, fuel andventilation such as to guarantee a robust test scenario. In order totest this robustness, the fuel layout was dened to minimisevariations in the re spread pattern and identical compartmentswere used in both tests. The two main processes that can lead todrastic differences in re development are the ignition of3 In order to avoid bias in the predictions from The University of Edinburgh
team, the modellers were kept apart from the experimentalists and submitted
their input le before the actual test was conducted.CFAST [25]. No limitations or suggestions were given regardingthe re model to be used. Each teamwas free to choose the modeldeemed most suitable or preferred for the task. The organisersendeavoured to include as many different models as possible inthe round-robin study, but users of other codes declined theinvitation to participate or withdrew.
Table 1 summarises the most important assumptions made ineach simulation. Each simulation domain is shown in Fig. 5. Moreinformation and detailed descriptions of each input le areprovided by Rein et al. [15, Chapter 10].
The teams were asked to provide results in three ascendinglevels of complexity:
(1) General re behaviour and time to major events (e.g. ignitionof nearby objects, ashover, window breakage, burn-out).
(2) Transient re behaviour by zones (e.g. average temperature ofdifferent layers and rooms, growth of smoke layer, ignition ofother items).
(3) Transient re behaviour by elds, both in space and time(e.g. temperature, ow and species concentration elds). Thislevel suits CFD models only.
The process of converting the data from CFDmodel-type to zonemodel-type information and the assumptions inherent to theprocess were the responsibility of each team and considered aspart of the round-robin study. The conversion of the experimentaldata point measurements to zone-type data was done assumingthis information was also provided to the teams. Meteorologicaldata from two nearby weather stations was also available.Media coverage was inevitable and thus all teams were providedwith copies of selected news articles which included photographsand journalist descriptions of the event as seen from the outside.A 5-min video recorded with a hand-held camera, summarisingthe event, the compartment before and after the re, and the redevelopment as seen from outside the building was also providedto the teams as part of the round-robin set of data. It is importantto note that the extent of this information by far exceeds thetypical set of data available for a user when attempting tosimulate a re of this nature.
4.2. Input and output les for the simulationstheof ththaA spromeasurement was also provided to the teams. Information onventilation conditions included size, photographs and statuse windows and doors. One of the main compartment windowconratesured in the furniture calorimeter. This calorimeter experi-t was allowed to burn until one third of the sofa mass wassumed and then the re was extinguished. This heat releasetionmeas and photographs of each furniture item. A replica of the sofathe wastepaper basket were tested under laboratory condi-s, and the initial heat release rate of the ensemble wasThe teams were given all the details available up to ignition ofthe re and a general overview related to the aftermath. Thisincluded: the geometry and dimensions of the at; a detailed andmeasured layout of the room furniture; 50 photographs of thewhole compartment nal set-up, windows, fuel packages and
rnal 44 (2009) 590602t the smoke layer interface is located near the 100 1C isotherm.ensitivity study for this criterion was conducted and resultsvided include isotherms in the range from 90 to 250 1C.
ARTICLE IN PRESS
n
lly p
item
t. HR
n of
JouTable 1Summarised information of each simulations input le.
Sim# Fire
model
ERTa
(h)
Grid (mm) General description of input to the simulatio
A1 CFAST 0.01 Domain includes the whole at. HRR is partia
NIST sofa experiment. Ignition of secondary
A2 FDS 4 153 50 Domain includes only the main compartmen
the HRR from a NIST experiment sofa. Ignitio
prescribed surface burning rate.
G. Rein et al. / Fire Safety5. Comparison and analysis of the results
It is important to keep in mind while analysing the resultsthat all simulations were forecasts conducted a priori. Inother words, quoting the words attributed to Sir WinstonChurchill (circa 1945): I always avoid prophesying beforehandbecause it is much better to prophesy after the event has already
taken place.The large pool of data submitted by the round-robin partici-
pants greatly exceeds what can be presented in this paper. Thus,only a selection of the most important results is presented anddepicted in Figs. 614.
B FDS 4 23 From 5 to
500
Domain includes the whole at. HRR is partially p
HRR from sofa replica experiment plus the rema
predicted by ignition temperature, material prop
C CFAST 0.01 Domain includes the whole at and the main o
source prescribed using the measured HRR from
temperature and material properties.
D1 FDS 4 19 100 Domain includes the whole at. HRR is fully presc
on ventilation conditions.
D2 FDS 4 128 From 50 to
100
Domain includes the whole at. HRR is fully pre
model for ame spread.
E1 FDS 4 55 100 Domain includes the whole at. HRR is fully pred
items predicted by ignition temperature and ma
E2 FDS 4 33 100 Domain includes the whole at. HRR is fully presc
on ventilation conditions.
F1 and
F2
FDS 4 170 90 Domain includes the main compartment, kitchen
re source prescribed using the measured HRR fr
2/3 of sofa mass. The peak HRR is raised by 20% in
material properties and prescribed surface burni
More information and detailed descriptions of each of the input le are provided by Ra Estimated running time on the respective computers used.
Fig. 5. Computational domain frescribed and partially predicted. Initial re source prescribed using the HRR from a
s predicted by ignition temperature and material properties.
R is partially prescribed and partially predicted. Initial re source prescribed using
secondary items predicted by ignition temperature, material properties and
rnal 44 (2009) 590602 5955.1. General re behaviour
The predicted results for the maximum average temperature inthe compartment together with the time to reach ashover(and experimental values for comparison) are summarised inTable 2. The predicted times to ashover varied approximatelybetween a 180% over-prediction down to a 100% under-predictionand fell into two main groups, those predicted at 800780 s (13min) very close to the time for forced windowbreakage at 800 s and those that predicted ashoverbefore 180780 s (3min). One simulation predicted that noashover would occur. Similarly, the predicted maximum
rescribed and partially predicted. Initial re source prescribed using the measured
ining 2/3 sofa mass that way allowed to burn further. Ignition of secondary items
erties and prescribed heat of vaporization
or access corridor. HRR is partially prescribed and partially predicted. Initial re
sofa replica experiment as given. Ignition of secondary items predicted by ignition
ribed using initially a uniform t-square re over the sofa area and then values based
dicted. Ignition of secondary items predicted by material properties and pyrolysis
icted except ignition that is a small wastepaper basket re. Ignition of secondary
terial properties.
ribed using initially a uniform t-square re over the sofa area and then values based
, bedroom-1 and hallway. HRR is partially prescribed and partially predicted. Initial
om sofa replica experiment but extrapolated with a t-square re for the remaining
F1 and by 40% in F2. Ignition of secondary items predicted by ignition temperature,
ng rate.
ein et al. [15, Chapter 10].
or each of the simulations.
ARTIC
LEIN
PRESS
0 200 400 600 800 1000 12000
2000
4000
6000
8000
10000
time [s]
H
R
R
[
k
W
]
A2
D1
E2
B
F1
A1
E1
F2
D2
C
EXP
D2E1
E2
D1
B
A2
CA1
A2
F1
F2
D1
Fig. 6. Evolution of the global heat release rate within the compartment. Legend for the different curves: continuous line for CFD simulations; dashed line for zone model simulations; and dotted for the experimental data witherror bars.
G.Rein
etal./Fire
Safety
Journal44(2009)590602
596
ARTICLE IN PRESS
0 200 400 600 800 1000 1200time [s]
D2
E1
A1
F1B F2 A2
C
E2
D1D1
0
200
400
600
800
1000
1200
T [
C]
0 200 400 600 800 1000 12000
0.5
1
1.5
2
2.5
time [s]
Hei
ght [
m]
F1F2 E1
B
A1A2
E2
D2D1
Fig. 7. Evolution of the hot layer in the compartment: (a) average temperature and(b) height to the interface from compartment oor. Experimental values derived
assuming the smoke layer interface at the 100 1C isotherm. The legend in Fig. 6applies.
0 200 400 600 800 1000 12000
10
20
30
40
time [s]
[1/m
]
B E2 A1
A2E1
C
D2
D1
Fig. 8. Evolution of the average extinction coefcient in the hot layer and errorbars. The legend in Fig. 6 applies.
D2B]
G. Rein et al. / Fire Safety JouE1
F1A2
F2Hei
ght [
mtime 200 [s]
D1E2
rnal 44 (2009) 590602 597average temperatures in the compartment varied approxi-mately between a 50% over-prediction down to a 70% under-prediction.
5.2. The global heat release rate
The global heat release rate (HRR) is given in Fig. 6. The samelegend is used for the results in all the subsequent gures(continuous line for CFD models, dashed line for zone models anddotted for the experimental data). Three distinct stages areobserved: initial growth, rst post-ashover stage until compart-ment window breakage and subsequent second post-ashover
0 200 400 600 800 1000
T [C]
0 200 400 600 800 1000
time 700 [s]
D2
E1
E2
A2D1
B
F1
F2
0
0.5
1
1.5
2
2.5
Hei
ght [
m]
T [C]
0
0.5
1
1.5
2
2.5
0 200 400 600 800 1000
time 1100 [s]
Hei
ght [
m]
T [C]
E1
F2
F1
E2
B
D1A2
D2
Fig. 9. Gas-phase temperature vs. height from the oor in the main compartmentat three different times, in the northeast corner near the sofa (see Fig. 2). Only CFD
simulations and experimental data. The legend in Fig. 6 applies.
ARTICLE IN PRESS
0.5
Joutime 200 [s]
D2
E2
D1
E1
B
F1
A2
F2
0.5
1
1.5
2
2.5
Hei
ght [
m]
G. Rein et al. / Fire Safety598stage. The heat release rate inside the main compartment wascalculated during the test using the principle of oxygen depletion.The inow of air and outow of combustion products at theopenings of the compartment were established by means oftemperature and gas velocity probe measurements. It wasassumed that nearly all the oxygen was depleted inside thecompartment. The estimated experimental error associated withthis calculation is presented in Fig. 6, based on a conservativeoverestimation of the two main sources of error. The total airowinto the compartment was measured using only three probes pervent and results in an upper bound of 53% difference between theinow and outow values during the rst stage post-ashover anda 13% difference during the second stage. Relaxation of thecomplete oxygen depletion assumption provides a lower boundestimate of the HRR error of 18% if the oxygen concentration in the
0 200 400 600 800 1000T [C]
0
0 200 400 600 800 1000
time 700 [s]
D2
E2
D1E1
B
A2
F1
F2
0
0.5
1
1.5
2
2.5
Hei
ght [
m]
T [C]
0 200 400 600 800 10000
0.5
1
1.5
2
2.5time 1100 [s]
E2
D1
E1F2
F1A2
B
D2
Hei
ght [
m]
T [C]
Fig. 10. Gas-phase temperature vs. height from the oor in the main compartmentat three different times, in the southwest corner near the window (see Fig. 2). Only
CFD simulations and experimental data. The legend in Fig. 6 applies.time 200 s
D2
E1D1
E2
A2
F11
1.5
2
2.5
Hei
ght [
m]
rnal 44 (2009) 590602compartment were 4% in volume. The HRR measurements conveyan approximately steady 3MW re between the onset of ashover(at 300 s) and the compartment window breakage at 800 s.Thereafter the HRR is around 5MW until forced extinction, asshown in Fig. 6. These HRR measurements are in good agreementwith simple hand-calculations using ventilation factors. Thesecalculations and details of the HRR measurement technique canbe found in Refs. [21,15, Chapter 3].
The simulations show a wide scatter of predicted rebehaviours. One simulation (D2) over-predicts the HRR by 100%,another (E1) provides a reasonably good prediction and all othersimulations under-predicted the HRR in the range of 3090%. Notethat two simulations (F1 and F2) compare poorly to the
0 50 100 150Heat Flux [kW/m2]
0
50 0 50 100 150
time 700 s
D2
D1
A2
E2F1, F2
E1
0
0.5
1
1.5
2
2.5
Hei
ght [
m]
Heat Flux [kW/m2]
0 50 100 150
time 1100 s
F2D1
A2
D2
F1
E1, E2
0
0.5
1
1.5
2
2.5
Hei
ght [
m]
Heat Flux [kW/m2]
Fig. 11. Local total incident heat-ux vs. height on the east wall of the maincompartment at three different times. Only CFD simulations and experimental
data. The legend in Fig. 6 applies.
ARTICLE IN PRESS
T [C]
time 700 s
Jou50
100
150
200
250
[kW
/m2 ]
D2
F2
F1
E2A2
D1
G. Rein et al. / Fire Safetymeasurements but only in time, since the HRR values areadequately predicted for the post-ashover stages. Most teamsattempted to partially predict rather than fully prescribe theheat release rate. Only two models prescribed it completely(D1 and E2). For the two simulations that fully prescribed the HRR(D1 and E2), the prescribed values are not reached in the modeldue to unburnt fuel leaving the domain via the vents. It is worthnoting that all except one (D2) of the simulations predictingashover before 3min did not use the measured HRR for thereplica sofa. Other teams deemed the measured sofa HRR to bedecient or too slow for the re growth stage. The best averageresults and lowest scatter are obtained after the forced windowbreakage (at 800 s), as the teams were informed of the timing ofthis event.
The HRR curve is the single most important and comprehen-sive characteristic of re development, resulting from the timeevolution and coupling of many important re mechanisms. Thewide range of simulated HRR curves demonstrates the difculty inaccurately predicting the re growth in the case of a non-trivial,realistic re scenario.
0 200 400 600 800 1000 12000
time [s]
0 200 400 600 800 1000 12000
50
100
150
200
250
time [s]
[kW
/m2 ]
D2
E1
D1 E2
F1
F2
Fig. 12. Evolution of the local total incident heat-ux on the east wall of the maincompartment, at heights of: (a) 250 cm and (b) 50 cm from the oor. Only CFD
simulations and experimental data. The legend in Fig. 6 applies.0 200 400 600 800 1000
time 200 s
D2
E1
D1
E2
F1, F2
0.5
1
1.5
2
2.5
Hei
ght [
m]
rnal 44 (2009) 590602 5995.3. Zone results
The results for the hot layer are shown in Figs. 7 and 8. Fig. 7shows the evolution of the average hot layer temperature (Fig. 7a)and the hot layer height (Fig. 7b). The experimental values areaveraged over the entire layer assuming that during the growthphase the interface is at the 100 1C isotherm or at the height of thelargest temperature gradient when below 100 1C at the verybeginning of the test. The sensitivity of the height to the smokelayer to variations of the assumed isotherm value is also presentedin Fig. 7 for the range from 90 to 250 1C. The experimentallycalculated smoke layer temperatures were insensitive (less than3% change) to variations in smoke layer height criteria within thatrange. The smoke height calculations agree with visual esti-mates of the smoke layer height, during the re growth stage,obtained from the camera footage [21,15, Chapter 3]. During the
0 200 400 600 800 1000
D2
A2
E1D1
E2
F1, F2
0.5
1
1.5
2
2.5
Hei
ght [
m]
T [C]
0 200 400 600 800 10000.5
1
1.5
2
2.5time 1100 s
F1
E2
A2
F2
E1
D1
D2
Hei
ght [
m]
T [C]
Fig. 13. Local surface temperature vs. height on the east wall of the maincompartment for three different times. Only CFD simulations and experimental
data. The legend in Fig. 6 applies.
ARTICLE IN PRESS
Journal 44 (2009) 5906020
200
400
600
800
1000
1200
1400
T [
C]
D2
E1
E2 A2
D1 F1 F2
G. Rein et al. / Fire Safety600post-ashover stages, the video cameras conrm that the smokelayer engulfed the whole compartment and thus the height to thesmoke layer is zero.
There is a wide scatter of modelling results shown in bothgures. Most simulations under-predicted the hot layer tempera-ture. Four simulations fell within a 1040% under-predictionrange and the others were above the 50% range of over- andunder-predictions. The very wide range of behaviours predictedreects the inuence of the users assumptions when convertingeld results to zone results, as well as the difference in overallassumptions used for the input. It is worth highlighting that thesimulation that performed the best at predicting the HRR (E1),predicting it within a 10% range, predicted the average hot layertemperature within a 30% range but differs greatly from theexperimental measurements of smoke layer height.
Fig. 8 shows the extinction coefcient of the smoke layer. Theexperimental error for these measurements is estimated at 10% onaverage, based on the comparison of laser measurements of theextinction coefcient and hand-calculations using the visualrecordings from cameras [21,15, Chapter 3]. There is a widescatter of predicted values and in this case, the experimental
respect to the experimental measurements) with a bias towards
0 200 400 600 800 1000 1200time [s]
0 200 400 600 800 1000 12000
200
400
600
800
1000
1200
1400
time [s]
T [
C]
E1
D2 A2
E2 F1
F2
D1
Fig. 14. Evolution of the surface local temperature on the east wall of the maincompartment, at heights of: (a) 250 cm and (b) 50 cm from the oor. Only CFD
simulations and experimental data. The legend in Fig. 6 applies.under-prediction. A relatively lower scatter is observed during thesecond stage of the post-ashover period and also near the window,measurements are mid-range between the predictions, and thereis no global bias towards under- or over-prediction.
5.4. Field results
Localised results of eld variables are shown in Figs. 914. Figs. 9and 10 show the local gas-phase temperatures, as a function ofheight, at different times and locations throughout the compart-ment. Three times are chosen as representative of the three distinctstages: during re growth (200 s), rst stage post-ashover (700s),and the second stage post-ashover (1100 s); and two locations arechosen: northeast, near the sofa, and southwest, near the window(see locations marked in Fig. 2). The experimental error is plottedwith the temperature measurements and calculated to be amaximum of 20 1C due to small errors in the spatial locations andthe radiation correction for the thermocouple readings [21,15,Chapter 3]. This maximum error is conservatively applied to allthermocouple readings reported in the paper. In general, there is awide range of predicted temperature results (roughly 780% in
Table 2Comparison of simulated time to ashover and maximum average temperature in
the smoke layer of the main compartment with experimental data.
Simulation Time to ashover
(s)
Maximum average smoke layer
temperature (1C)
A1 850 790
A2 290 500
B 840 690
C No ashover 200
D1 200 720
D2 80 1150
E1 180 900
E2 180 610
F1 720 590
F2 850 720
Experimental 300 750away from the larger fuel load at the northeast corner. Thesimulation that predicts the HRR within 10% (E1) of the experi-mental measurements, over-predicts local temperatures up to 200%during the growth phase but post-ashover, the disparity is reducedto an average 25% difference.
Local results for the instrumented wall, east of the maincompartment (see Fig. 2), are shown in Figs. 1114. Figs. 11 and 12show the total incident heat ux4 close to the centre of the wall,both at different times and locations. The experimental error fromthe thin-skin calorimeter gauges is estimated by comparison ofdifferent smoothing techniques to treat the derivative of thereadings and adding the calibration and radiation errors [26]. Thisway, it is calculated that before ashover the error in heat ux isabout 13% and thereafter about 5%. In general, the scatter is large,particularly during the growth phase and at lower heights.Predictions are better higher up the wall and during the rstpost-ashover stage.
Similarly, Figs. 13 and 14 show the wall surface temperature atdifferent times and locations approximately at the centrelinebetween the north wall and the kitchen door. The experimental
4 Thin-skin calorimeter gauges measure total incident heat ux including
convective and radiative contributions.
used in subsequent studies [15, Chapter 11, 27,28] to show that itis possible to conduct a posteriori FDS simulations that reproduce
Nevertheless, the general behaviour captured by severalsimulations provides re features that may be good enough to
ARTICLE IN PRESS
Jouthe general re behaviour to a satisfactory level. This wasachieved due to the availability of sufcient experimental dataof the real behaviour for reference, allowing for an adequate set-up of the input le. Nevertheless, achieving simulation resultscomparable to the experimental data was a very laborious task.Several simulations were required, with repeated comparison toerror is shown with the temperature measurements. As with theheat ux, the general scatter is large, especially during the growthphase and at lower heights. However, in general terms, the scatteris higher for the wall temperatures than that of the wall heatuxes.
6. Discussion of results
The results indicate large scatter and considerable disparity,both amongst the predicted res and between the predicted resand the experimental data. The scatter of the simulations is muchlarger than the estimated experimental error. Moreover, compar-ison between the range of predicted hot layer temperatures (seenin Fig. 7a) and the differences between Test One and Test Two(seen in Fig. 4) further demonstrates that the scatter is also muchlarger than the expected experimental variability.
Although not the intent of the paper, the results show that outof the ten simulations; one provided good results; four providedacceptable results in some sense; and ve did poorly. It isimportant to emphasise that none of the predictions accuratelypredicted the time to ashover. One simulation predicted HRRdevelopment and wall heat uxes adequately, but diverged fromexperimental data on other local quantities. Thus, one conclusionis that in complex modelling scenarios, such as Dalmarnock, goodresults at the global compartment scale do not necessarilycorrelate to good results at the local scale. For example, it isobserved that while a simulation could provide good predictionsof the global compartment HRR, it may be simulating wrongly thelocation and height of the ames or the surface area of the re,hence failing to predicted local quantities.
The greatest source of scatter originates in the prediction of there growthi.e. the heat release rate. This is due to the inherentcomplexity in re growth modelling, particularly for ame spreadand ignition of secondary fuel items. Since most participants usedthe same re model, FDS4, it is reasonable to think that the widerange of predicted behaviour is mostly the result of theuncertainty associated with the denition of valid input dataunder the constraints of the model (assumptions and parametervalues). The large number of degrees of freedom (i.e. apparentpossible assumptions and uncertainty in the parameter values,among others) and the broad variability of the material propertiesavailable in the literature lead to large variability in the results.This variability needs to be considered when re modelling isused to predict re growth in complex scenarios. These conclu-sions are expected to be applicable to the full suite of re modelscurrently available, and not uniquely to the two models used inthis study. This is further corroborated in the recent round-robinstudy [20] conducted using ten different re models. These resultsalso indicate a signicant scatter around the experimentalmeasurements, even when the measured HRR of the re isprovided for direct input.
With the purpose of illustrating the difference between a prioriand a posteriori modelling, Dalmarnock Fire Test One has been
G. Rein et al. / Fire Safetythe experimental results and subsequent ne-tuning of inputparameters until convergence to an acceptable level of agreementwas achieved.be applied towards some engineering problems if a robust andconservative methodology is dened. A prerequisite for thismethodology is that it can use predictions with crude levels ofaccuracy and that it applies appropriate safety factors.
Acknowledgements
Our gratitude to the Building and Fire Research Laboratory atNIST for developing and the computer re models used in thisstudy and for making them freely available. We regret that usersof different re models withdrew from the study or declined ourinvitation to participate.
Thanks to Chris Lautenberger who provided very constructivereviews of this paper and helped to improve it.
The experiments and the round-robin study would not havebeen successful without the help of many colleagues at the BRECentre of Fire Safety Engineering at The University of Edinburgh.Specially, thanks must go to Thomas Steinhaus, Adam Cowlard,Hubert Biteau, Aitor Amundarain, Martin Gillie, Ricky Carvel, ChrisSchemel, Dougal Drysdale, Stephen Welch, Ruth Thompson andmany others.
The authors are grateful for the funding and the help of the manyorganisations involved in the Dalmarnock Fire Tests: BBC Horizonprogramme, Engineering and Physical Sciences Research Council,The University of Edinburgh, Glasgow Housing Authority, Strath-clyde Fire & Rescue Service, Glasgow Caledonian University, Lion TV,BRE Trust, Xtralis, Powerwall Systems and the EU Programme Alban.This work has formed part of FireGrid-www.regrid.org-and wasco-funded by the UK Government Strategy Boards CollaborativeResearch and Development programme.
References7. Concluding remarks
A realistic and repeatable re test was conducted underconditions that are particularly relevant to CFD modellingvalidation. The study is an assessment of the state-of-the-art ofre modelling in a non-trivial, realistic scenario and evaluates theprocess of re modelling as a whole, including the prediction ofthe heat release rate and the effect of different assumptions, inputparameter values, computational approaches and user interac-tions with the model.
The aim of the round-robin exercise was to forecast the testresults as accurately as possible, and not to provide an engineer-ing analysis with conservative assumptions or safety factors.Design for re safety was not the objective of the exercise. Theissue of how to use reliably re modelling for safety andengineering design is a very important issue currently underresearch in many institutions and rms.
The Dalmarnock Test involved multiple fuel packages and non-trivial re growth. The results presented here show that currentmodelling cannot provide good predictions of HRR evolution(i.e. re growth) in realistic complex scenarios like Dalmamock.Fire modelling is not yet able to predict the HRR and moreresearch efforts need to be tailored towards this issue. However,re environments away from the ames could be calculated if anaccurate HRR is part of the input data to the modelling process.This is because current modelling tools provide good predictionsof the effects of a re in the far eld once the re growth is known.
rnal 44 (2009) 590602 601[1] H.W. Emmons, The prediction of res in buildings, Proceedings of theCombustion Institute 17 (1978).
[2] G. Cox, R. Chitty, S. Kumar, Fire modelling and the kings cross reinvestigation, Fire Safety Journal 15 (1) (1989) 103106.
[3] G. Cox, Turbulent closure and the modelling of re by using computationaluid dynamics, Philosophical Transactions of the Royal Society A 356 (1748)(1998) 28352854.
[4] V. Novozhilov, Computational uid dynamics modeling of compartment res,Progress in Energy and Combustion Science 27 (6) (2001) 611666.
[5] G. Cox, S. Kumar, Modeling enclosure res using CFD, in: SFPE Handbook ofFire Protection Engineering, third ed. (Chapter 38), 2002.
[6] K.B. McGrattan, Fire Modeling: Where Are We? Where Are We Going?, FireSafety Science 8 (2005) 5368.
[7] ASTM E1355, Standard Guide for Evaluating Predictive Capability ofDeterministic Fire Models, American Society for Testing and Materials, WestConshohoken, PA, 2005.
[8] NUREG-1824 and EPRI 1011999, Verication and Validation of Selected FireModels for Nuclear Power Plant Applications, vols. 17, US Nuclear RegulatoryCommission, Washington, DC and Electric Power Research Institute, Palo Alto,CA, 2007.
[9] M.H. Salley, J. Dreisbach, K. Hill, R. Kassawara, B. Naja, F. Joglar, A. Hamins,K.B. McGrattan, R. Peacock, B. Gautier, Verication and validationhow todetermine the accuracy of re models, Fire Protection Engineering Magazine,Spring, 2007.
[10] J. Wen, K. Kang, T. Donchev, J. Karwatzki, Validation of FDS for the predictionof medium-scale pool res, Fire Safety Journal 42 (2) (2007) 127138.
[11] T. Hakkarainen, O. Keski-Rahkonen, L. Lindberg, CIBW14 Round Robin of CodeAssessment: Design Report of Scenario C, VTT Technical Research Centre ofFinland, 1999.
[12] S.D. Miles, S. Kumar, G. Cox, Comparisons of blind predictions of a CFD modelwith experimental data, Fire Safety Science 6 (2000) 543554.
[13] P.A. Reneke, M.J. Peatross, W.W. Jones, C.L. Beyler, R. Richards, A comparison ofCFAST predictions to USCG real-scale re tests, Journal of Fire ProtectionEngineering 11 (1) (2001) 4368.
[14] N. Pope, C. Bailey, Quantitative comparison of FDS and parametric re curveswith post-ashover compartment re test data, Fire Safety Journal 41 (2)(2006) 99110.
[15] Rein, G., Abecassis-Empis, C., Carvel, R. (Eds.), The Dalmarnock Fire Tests:Experiments and Modelling, The University of Edinburgh, 2007.
[16] A.N. Beard, On a priori, blind and open comparisons between theory andexperiment, Fire Safety Journal 35 (2000) 6366.
[17] H.W. Emmons, Fire research abroad, Fire Research Abstracts and Reviews 10(2) (1968) 133143.
[18] W.M. Pitts, A.V. Murthy, J.L. deRis, J.R. Filtz, K. Nygard, D. Smith, I. Wetterlund,Round robin study of total heat ux gauge calibration at re laboratories, FireSafety Journal 41 (6) (2006) 459475.
[19] A.N. Beard, Problems with using models for re safety, in: A.N. Beard,R. Carvel (Eds.), The Handbook of Tunnel Fire Safety, Thomas Telford, London,2005 (Chapter 14).
[20] A. Copalle, P. Van Hulle, D. Joyeux, L. Bustamante, E. Guillaume, D., Marquis,A. Thiry, M. Suzanne, L. Gay, P. Lamuth, C. Rome, A. Muller, P. Fromy,F. Demouge, P. Breton, S. Suard, S. Melis, A simulations exercise for CFD andzone models in the case of a bedroom re, in: Ninth International Symposiumon Fire Safety Science, Work-in-Progress Poster, Karlsruhe, September 2008.
[21] C. Abecassis-Empis, P. Reszka, T. Steinhaus, A. Cowlard, H. Biteau, S. Welch,G. Rein, J.L. Torero, Characterisation of Dalmarnock Fire Test One, Experi-mental Thermal and Fluid Science 32 (7) (2008) 13341343.
[22] J. Milke, V. Kodur, C. Marrioon, A overview of re protection inbuildings, FEMA 403-World Trade Center Building Performance Study,2002. Appendix A.
[23] R.S. Barlow, J.H. Frank, A.N. Karpetis, J.Y. Chen, Piloted methane/air jet ames:scalar structure and transport effects, Combustion and Flame 143 (2005)433449.
[24] K.B. McGrattan, G. Forney, Fire Dynamics Simulator (Version 4) Users Guide,NIST Special Publication 1019, National Institute of Standards and Technology,Gaithersburg, MD, USA, 2006.
[25] R.D. Peacock, P.A. Reneke, W.W. Jones, R.W. Bukowski, G. Forney, A Users Guidefor FAST: Engineering Tools for Estimating Fire Growth and Smoke Transport,National Institute of Standards and Technology, Special Publication 921, 2000.
[26] A. Amundarain, Assessment of the thermal efciency, structure andre resistance of lightweight building systems for optimized design,Ph.D. Thesis, University of Edinburgh, 2007. /http://hdl.handle.net/1842/2128S.
[27] W. Jahn, G. Rein, J.L. Torero, The effect of model parameters on the simulationof re dynamics, Fire Safety Science 9 (2008) 13411352.
[28] M. Lazaro, H. Boehmer, D. Alvear, J.A. Capote, A. Trouve, Numerical simulationof re growth, transition to ashover, and post-ashover dynamics in theDalmarnock Fire Test, Fire Safety Science 9 (2008) 13771388.
ARTICLE IN PRESS
G. Rein et al. / Fire Safety Journal 44 (2009) 590602602ISBN:978-0-9557497-0-4, November. Accessible at /http://www.era.lib.ed.ac.uk/handle/1842/2037S.
Round-robin study of a priori modelling predictions of the Dalmarnock Fire Test OneIntroductionRound-robin studies in fire scienceThe Dalmarnock Fire TestsDescription of Test OneThe philosophy of the tests
Dalmarnock round-robin studyCommon description of the scenarioInput and output files for the simulations
Comparison and analysis of the resultsGeneral fire behaviourThe global heat release rateZone resultsField results
Discussion of resultsConcluding remarksAcknowledgementsReferences