NIST Special Publication 1018-5

Fire Dynamics Simulator (Version 5) Technical Reference Guide

Volume 2: Verication

Randall McDermott Kevin McGrattan Simo Hostikka Jason Floyd

NIST Special Publication 1018-5

Fire Dynamics Simulator (Version 5) Technical Reference Guide

Volume 2: Verication

Randall McDermott Kevin McGrattan Fire Research Division Building and Fire Research Laboratory Simo Hostikka VTT Technical Research Centre of Finland Espoo, Finland Jason Floyd Hughes Associates, Inc. Baltimore, Maryland

April 8, 2009 FDS Version 5.3 SV NRepository Revision : 3685

AR

E TM

N T OF C O M

C ER

D EP

ST

ATES OF

AM

U.S. Department of Commerce Gary Locke, Secretary National Institute of Standards and Technology Patrick Gallagher, Acting Director

ER

ICA

UN
IT
E

Preface
This is Volume 2 of the FDS Technical Reference Guide. Volume 1 describes the mathematical model and numerical method. Volume 3 documents past and present experimental validation work. Instructions for using FDS are contained in a separate Users Guide [1]. The three volumes of the FDS Technical Reference Guide are based in part on the Standard Guide for Evaluating the Predictive Capability of Deterministic Fire Models, ASTM E 1355 [2]. ASTM E 1355 denes model evaluation as the process of quantifying the accuracy of chosen results from a model when applied for a specic use. The model evaluation process consists of two main components: verication and validation. Verication is a process to check the correctness of the solution of the governing equations. Verication does not imply that the governing equations are appropriate; only that the equations are being solved correctly. Validation is a process to determine the appropriateness of the governing equations as a mathematical model of the physical phenomena of interest. Typically, validation involves comparing model results with experimental measurement. Differences that cannot be explained in terms of numerical errors in the model or uncertainty in the measurements are attributed to the assumptions and simplications of the physical model. Evaluation is critical to establishing both the acceptable uses and limitations of a model. Throughout its development, FDS has undergone various forms of evaluation, both at NIST and beyond. This volume provides a survey of verication work conducted to date to evaluate FDS.

ii

About the Authors

Randall McDermott joined the research staff of the Building and Fire Research Lab in 2008. He received a B.S. degree from the University of Tulsa in Chemical Engineering in 1994 and a doctorate at the University of Utah in 2005. His research interests include domain decomposition, Lagrangian particle methods, and adaptive mesh renement. Kevin McGrattan is a mathematician in the Building and Fire Research Laboratory (BFRL) of NIST. He received a bachelors of science degree from the School of Engineering and Applied Science of Columbia University in 1987 and a doctorate at the Courant Institute of New York University in 1991. He joined the NIST staff in 1992 and has since worked on the development of re models, most notably the Fire Dynamics Simulator. Simo Hostikka is a Senior Research Scientist at VTT Technical Research Centre of Finland. He received a master of science (technology) degree in 1997 and a doctorate in 2008 from the Department of Engineering Physics and Mathematics of the Helsinki University of Technology. He is the principal developer of the radiation and solid phase sub-models within FDS. Jason Floyd is a Senior Engineer at Hughes Associates, Inc., in Baltimore, Maryland. He received a bachelors of science and Ph.D. in the Nuclear Engineering Program of the University of Maryland. After graduating, he won a National Research Council Post-Doctoral Fellowship at the Building and Fire Research Laboratory of NIST, where he developed the combustion algorithm within FDS. He is currently funded by NIST under grant 60NANB5D1205 from the Fire Research Grants Program (15 USC 278f). He is the principal developer of the multi-parameter mixture fraction combustion model and control logic within FDS.

iii

iv

Acknowledgments
Partial support for the preparation of the FDS Technical Reference Guides has been provided by the Ofce of Nuclear Regulatory Research of the US Nuclear Regulatory Commission (US NRC). Special thanks to Mark Salley and Jason Dreisbach for their efforts. Thanks to Chris Lautenburger and Carlos Fernandez-Pello for their assistance with the two-reaction test case.

vi

Contents
Preface About the Authors Acknowledgments 1 2 What is Verication? Survey of Past Verication Work 2.1 Analytical Tests . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Numerical Tests . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Grid Sensitivity . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Sensitivity of Large Eddy Simulation Parameters . . . . 2.3.3 Sensitivity of Radiation Parameters . . . . . . . . . . . 2.3.4 Sensitivity of Thermophysical Properties of Solid Fuels . 2.4 Code Checking . . . . . . . . . . . . . . . . . . . . . . . . . . The Basic Flow Solver 3.1 Single Mesh, Non-reacting Periodic Test Cases . 3.1.1 2D Analytical Solution to Navier-Stokes . 3.1.2 Decaying Isotropic Turbulence . . . . . . 3.1.3 The Dynamic Smagorinsky Model . . . . i iii v 1 3 3 4 5 5 7 7 8 9 11 11 11 15 18

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

Thermal Radiation 21 4.1 Radiation inside a box (radiation_in_a_box) . . . . . . . . . . . . . . . . . . . . . . . . . 22 4.2 Radiation from a plane layer (radiation_plane_layer) . . . . . . . . . . . . . . . . . . . . 23 4.3 Wall Internal Radiation (wall_internal_radiation) . . . . . . . . . . . . . . . . . . . . . . 24 Heat Conduction 25 5.1 Simple Heat Conduction Through a Solid Slab (heat_conduction) . . . . . . . . . . . . . . 26 5.2 Temperature-Dependent Thermal Properties (heat_conduction_kc) . . . . . . . . . . . . . 27 Pyrolysis 29 6.1 A Simple Two-Step Pyrolysis Example (two_step_solid_reaction) . . . . . . . . . . . . . . 30 6.2 Development of surface emissivity (emissivity) . . . . . . . . . . . . . . . . . . . . . . . . 31 6.3 Enthalpy of solid materials (enthalpy) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 vii

Droplets 33 7.1 Water Droplet Evaporation (water_evaporation) . . . . . . . . . . . . . . . . . . . . . . . 33 35

Bibliography

viii

Chapter 1

What is Verication?
The terms verication and validation are often used interchangeably to mean the process of checking the accuracy of a numerical model. For many, this entails comparing model predictions with experimental measurements. However, there is now a fairly broad-based consensus that comparing model and experiment is largely what is considered validation. So what is verication? ASTM E 1355 [2], Standard Guide for Evaluating the Predictive Capability of Deterministic Fire Models, denes verication as The process of determining that the implementation of a calculation method accurately represents the developers conceptual description of the calculation method and the solution to the calculation method. and it denes validation as The process of determining the degree to which a calculation method is an accurate representation of the real world from the perspective of the intended uses of the calculation method. Simply put, verication is a check of the math; validation is a check of the physics. If the model predictions closely match the results of experiments, using whatever metric is appropriate, it is assumed by most that the model suitably describes, via its mathematical equations, what is happening. It is also assumed that the solution of these equations must be correct. So why do we need to perform model verication? Why not just skip to validation and be done with it? The reason is that rarely do model and measurement agree so well in all applications that anyone would just accept its results unquestionably. Because there is inevitably differences between model and experiment, we need to know if these differences are due to limitations or errors in the numerical solution, or the physical sub-models, or both. Whereas model validation consists mainly of comparing predictions with measurements, as documented for FDS in Volume 3 of the Technical Reference Guide, model verication consists of a much broader range of activities, from checking the computer program itself to comparing calculations to analytical (exact) solutions to considering the sensitivity of the dozens of numerical parameters. The next chapter discusses these various activities, and the rest of the Guide is devoted mainly to comparisons of various sub-model calculations with analytical solutions.

Chapter 2

Survey of Past Verication Work

This chapter documents work of the past few decades at NIST, VTT and elsewhere to verify the algorithms within FDS.

2.1

Analytical Tests

Most complex combustion processes, including re, are turbulent and time-dependent. There are no closedform mathematical solutions for the fully-turbulent, time-dependent Navier-Stokes equations. CFD provides an approximate solution for the non-linear partial differential equations by replacing them with discretized algebraic equations that can be solved using a powerful computer. While there is no general analytical solution for fully-turbulent ows, certain sub-models address phenomenon that do have analytical solutions, for example, one-dimensional heat conduction through a solid. These analytical solutions can be used to test sub-models within a complex code such as FDS. The developers of FDS routinely use such practices to verify the correctness of the coding of the model [3, 4]. Such verication efforts are relatively simple and routine and the results may not always be published nor included in the documentation. Examples of routine analytical testing include: The radiation solver has been veried with scenarios where simple objects, like cubes or at plates, are positioned in simple, sealed compartments. All convective motion is turned off, the object is given a xed surface temperature and emissivity of one (making it a black body radiator). The heat ux to the cold surrounding walls is recorded and compared to analytical solutions. These studies help determine the appropriate number of solid angles to be set as the default. Solid objects are heated with a xed heat ux, and the interior and surface temperatures as a function of time are compared to analytical solutions of the one-dimensional heat transfer equation. These studies help determine the number of nodes to use in the solid phase heat transfer model. Similar studies are performed to check the pyrolysis models for thermoplastic and charring solids. Early in its development, the hydrodynamic solver that evolved to form the core of FDS was checked against analytical solutions of simplied uid ow phenomena. These studies were conducted at the National Bureau of Standards (NBS)1 by Rehm, Baum and co-workers [5, 6, 7, 8]. The emphasis of this early work was to test the stability and consistency of the basic hydrodynamic solver, especially the velocity-pressure coupling that is vitally important in low Mach number applications. Many numerical algorithms developed up to that point in time were intended for use in high-speed ow applications, like aerospace. Many of the techniques adopted by FDS were originally developed for
1 The

National Institute of Standards and Technology (NIST) was formerly known as the National Bureau of Standards.

meteorological models, and as such needed to be tested to assess whether they would be appropriate to describe relatively low-speed ow within enclosures. A fundamental decision made by Rehm and Baum early in the FDS development was to use a direct (rather than iterative) solver for the pressure. In the low Mach number formulation of the NavierStokes equations, an elliptic partial differential equation for the pressure emerges, often referred to as the Poisson equation. Most CFD methods use iterative techniques to solve the governing conservation equations to avoid the necessity of directly solving the Poisson equation. The reason for this is that the equation is time-consuming to solve numerically on anything but a rectilinear grid. Because FDS is designed specically for rectilinear grids, it can exploit fast, direct solvers of the Poisson equation, obtaining the pressure eld with one pass through the solver to machine accuracy. FDS employs double-precision (8 byte) arithmetic, meaning that the relative difference between the computed and the exact solution of the discretized Poisson equation is on the order of 1012 . The delity of the numerical solution of the entire system of equations is tied to the pressure/velocity coupling because often simulations can involve hundreds of thousands of time steps, with each time step consisting of two solutions of the Poisson equation to preserve second-order accuracy. Without the use of the direct Poisson solver, build-up of numerical error over the course of a simulation could produce spurious results. Indeed, an attempt to use single-precision (4 byte) arithmetic to conserve machine memory led to spurious results simply because the error per time step built up to an intolerable level.

2.2

Numerical Tests

Numerical techniques used to solve the governing equations within a model can be a source of error in the predicted results. The hydrodynamic model within FDS is second-order accurate in space and time. This means that the error terms associated with the approximation of the spatial partial derivatives by nite differences is of the order of the square of the grid cell size, and likewise the error in the approximation of the temporal derivatives is of the order of the square of the time step. As the numerical grid is rened, the discretization error decreases, and a more faithful rendering of the ow eld emerges. The issue of grid sensitivity is extremely important to the proper use of the model and will be taken up in the next chapter. A common technique of testing ow solvers is to systematically rene the numerical grid until the computed solution does not change, at which point the calculation is referred to as a Direct Numerical Solution (DNS) of the governing equations. For most practical re scenarios, DNS is not possible on conventional computers. However, FDS does have the option of running in DNS mode, where the NavierStokes equations are solved without the use of sub-grid scale turbulence models of any kind. Because the basic numerical method is the same for LES and DNS, DNS calculations are a very effective way to test the basic solver, especially in cases where the solution is steady-state. Throughout its development, FDS has been used in DNS mode for special applications. For example, FDS (or its core algorithms) have been used at a grid resolution of roughly 1 mm to look at ames spreading over paper in a microgravity environment [9, 10, 11, 12, 13, 14], as well as "g-jitter" effects aboard spacecraft [15]. Simulations have been compared to experiments performed aboard the US Space Shuttle. The ames are laminar and relatively simple in structure, and the comparisons are a qualitative assessment of the model solution. Similar studies have been performed comparing DNS simulations of a simple burner ame to laboratory experiments [16]. Another study compared FDS simulations of a counterow diffusion ames to experimental measurements and the results of a one-dimensional multi-step kinetics model [17]. Early work with the hydrodynamic solver compared two-dimensional simulations of gravity currents with salt-water experiments [18]. In these tests, the numerical grid was systematically rened until almost perfect agreement with experiment was obtained. Such convergence would not be possible if there were a fundamental aw in the hydrodynamic solver. 4

2.3

Sensitivity Analysis

A sensitivity analysis considers the extent to which uncertainty in model inputs inuences model output. Model parameters can be the physical properties of solids and gases, boundary conditions, initial conditions, etc. The parameters can also be purely numerical, like the size of the numerical grid. FDS typically requires the user to provide several dozen different types of input parameters that describe the geometry, materials, combustion phenomena, etc. By design, the user is not expected to provide numerical parameters besides the grid size, although the optional numerical parameters are described in both the Technical Reference Guide and the Users Guide. FDS does not limit the range of most of the input parameters because applications often push beyond the range for which the model has been validated. FDS is still used for research at NIST and elsewhere, and the developers do not presume to know in all cases what the acceptable range of any parameter is. Plus, FDS solves the fundamental conservation equations and is much less susceptible to errors resulting from input parameters that stray beyond the limits of simpler empirical models. However, the user is warned that he/she is responsible for the prescription of all parameters. The FDS manuals can only provide guidance. The grid size is the most important numerical parameter in the model, as it dictates the spatial and temporal accuracy of the discretized partial differential equations. The heat release rate is the most important physical parameter, as it is the source term in the energy equation. Property data, like the thermal conductivity, density, heat of vaporization, heat capacity, etc., ought to be assessed in terms of their inuence on the heat release rate. Validation studies have shown that FDS predicts well the transport of heat and smoke when the HRR is prescribed. In such cases, minor changes in the properties of bounding surfaces do not have a signicant impact on the results. However, when the HRR is not prescribed, but rather predicted by the model using the thermophysical properties of the fuels, the model output is sensitive to even minor changes in these properties. The sensitivity analyses described in this chapter are all performed in basically the same way. For a given scenario, best estimates of all the relevant physical and numerical parameters are made, and a baseline simulation is performed. Then, one by one, parameters are varied by a given percentage, and the changes in predicted results are recorded. This is the simplest form of sensitivity analysis. More sophisticated techniques that involve the simultaneous variation of several parameters are impractical with a CFD model because the computation time is too long and the number of parameters too large to perform the necessary number of calculations to generate decent statistics.

2.3.1

Grid Sensitivity

The most important decision made by a model user is the size of the numerical grid. In general, the ner the numerical grid, the better the numerical solution of the equations. FDS is second-order accurate in space and time, meaning that halving the grid cell size will decrease the discretization error in the governing equations by a factor of 4. Because of the non-linearity of the equations, the decrease in discretization error does not necessarily translate into a comparable decrease in the error of a given FDS output quantity. To nd out what effect a ner grid has on the solution, model users usually perform some form of grid sensitivity study in which the numerical grid is systematically rened until the output quantities do not change appreciably with each renement. Of course, with each halving of the grid cell size, the time required for the simulation increases by a factor of 24 = 16 (a factor of two for each spatial coordinate, plus time). In the end, a compromise is struck between model accuracy and computer capacity. Some grid sensitivity studies have been documented and published. Since FDS was rst publicly released in 2000, signicant changes in the combustion and radiation routines have been incorporated into the model. However, the basic transport algorithm is the same, as is the critical importance of grid sensitivity. In compiling sensitivity studies, only those that examined the sensitivity of routines no longer used have been 5

excluded. As part of a project to evaluate the use of FDS version 1 for large scale mechanically ventilated enclosures, Friday [19] performed a sensitivity analysis to nd the approximate calculation time based on varying grid sizes. A propylene re with a nominal heat release rate was modeled in FDS. There was no mechanical ventilation and the re was assumed to grow as a function of the time from ignition squared. The compartment was a 3 m by 3 m by 6.1 m space. Temperatures were sampled 12 cm below the ceiling. Four grid sizes were chosen for the analysis: 30 cm, 15 cm, 10 cm, 7.5 cm. Temperature estimates were not found to change dramatically with different grid dimensions. Using FDS version 1, Bounagui et al. [20] studied the effect of grid size on simulation results to determine the nominal grid size for future work. A propane burner 0.1 m by 0.1 m was modeled with a heat release rate of 1500 kW. A similar analysis was performed using Alperts ceiling jet correlation [21] that also showed better predictions with smaller grid sizes. In a related study, Bounagui et al. [22] used FDS to evaluate the emergency ventilation strategies in the Louis-Hippolyte-La Fontaine Tunnel in Montreal, Canada. Xin [23] used FDS to model a methane fueled square burner (1 m by 1 m) in the open. Engineering correlations for plume centerline temperature and velocity proles were compared with model predictions to assess the inuence of the numerical grid and the size of the computational domain. The results showed that FDS is sensitive to grid size effects, especially in the region near the fuel surface, and domain size effects when the domain width is less than twice the plume width. FDS uses a constant pressure assumption at open boundaries. This assumption will affect the plume behavior if the boundary of the computational domain is too close to the plume. Ierardi and Barnett [24] used FDS version 3 to model a 0.3 m square methane diffusion burner with heat release rate values in the range of 14.4 kW to 57.5 kW. The physical domain used was 0.6 m by 0.6 m with uniform grid spacings of 15, 10, 7.5, 5, 3, 1.5 cm for all three coordinate directions. For both re sizes, a grid spacing of 1.5 cm was found to provide the best agreement when compared to McCaffreys centerline plume temperature and velocity correlations [25]. Two similar scenarios that form the basis for Alperts ceiling jet correlation were also modeled with FDS. The rst scenario was a 1 m by 1 m, 670 kW ethanol re under a 7 m high unconned ceiling. The planar dimensions of the computational domain were 14 m by 14 m. Four uniform grid spacings of 50, 33.3, 25, and 20 cm were used in the modeling. The best agreement for maximum ceiling jet temperature was with the 33.3 cm grid spacing. The best agreement for maximum ceiling jet velocity was for the 50 cm grid spacing. The second scenario was a 0.6 m by 0.6 m 1000 kW ethanol re under a 7.2 m high unconned ceiling. The planar dimensions of the computational domain were 14.4 m by 14.4 m. Three uniform grid spacings of 60, 30, and 20 cm were used in the modeling. The results show that the 60 cm grid spacing exhibits the best agreement with the correlations for both maximum ceiling jet temperature and velocity on a qualitative basis. Petterson [26] also completed work assessing the optimal grid size for FDS version 2. The FDS model predictions of varying grid sizes were compared to two separate re experiments: The University of Canterbury McLeans Island Tests and the US Navy Hangar Tests in Hawaii. The rst set of tests utilized a room with approximate dimensions of 2.4 m by 3.6 m by 2.4 m and re sizes of 55 kW and 110 kW. The Navy Hangar tests were performed in a hangar measuring 98 m by 74 m by 15 m in height and had res in the range of 5.5 MW to 6.6 MW. The results of this study indicate that FDS simulations with grids of 0.15 m had temperature predictions as accurate as models with grids as small as 0.10 m. Each of these grid sizes produced results within 15 % of the University of Canterbury temperature measurements. The 0.30 m grid produced less accurate results. For the comparison of the Navy Hangar tests, grid sizes ranging from 0.60 m to 1.80 m yielded results of comparable accuracy. Musser et al. [27] investigated the use of FDS for course grid modeling of non-re and re scenarios. Determining the appropriate grid size was found to be especially important with respect to heat transfer at heated surfaces. The convective heat transfer from the heated surfaces was most accurate when the near 6

surface grid cells were smaller than the depth of the thermal boundary layer. However, a ner grid size produced better results at the expense of computational time. Accurate contaminant dispersal modeling required a signicantly ner grid. The results of her study indicate that non-re simulations can be completed more quickly than re simulations because the time step is not limited by the large ow speeds in a re plume.

2.3.2

Sensitivity of Large Eddy Simulation Parameters

FDS uses the Smagorinsky form of the Large Eddy Simulation (LES) technique. This means that instead of using the actual uid viscosity, the model uses a viscosity of the form LES = (Cs )2 |S| (2.1)

where Cs is an empirical constant, is a length on the order of the size of a grid cell, and the deformation term |S| is related to the Dissipation Function (see FDS Technical Reference Guide [28] for details). Related to the turbulent viscosity are comparable expressions for the thermal conductivity and material diffusivity: kLES = LES c p Prt ; (D)LES = LES Sct (2.2)

where Prt and Sct are the turbulent Prandtl and Schmidt numbers, respectively. Thus, Cs , Prt and Sct are a set of empirical constants. Most FDS users simply use the default values of (0.2,0.5,0.5), but some have explored their effect on the solution of the equations. In an effort to validate FDS with some simple room temperature data, Zhang et al. [29] tried different combinations of the Smagorinsky parameters, and suggested the current default values. Of the three parameters, the Smagorinsky constant Cs is the most sensitive. Smagorinsky [30] originally proposed a value of 0.23, but researchers over the past three decades have used values ranging from 0.1 to 0.23. There are also renements of the original Smagorinsky model [31, 32, 33] that do not require the user to prescribe the constants, but rather generate them automatically as part of the numerical scheme.

2.3.3

Sensitivity of Radiation Parameters

Radiative heat transfer is included in FDS via the solution of the radiation transport equation for a nonscattering gray gas, and in some limited cases using a wide band model. The equation is solved using a technique similar to nite volume methods for convective transport, thus the name given to it is the Finite Volume Method (FVM). There are several limitations of the model. First, the absorption coefcient for the smoke-laden gas is a complex function of its composition and temperature. Because of the simplied combustion model, the chemical composition of the smokey gases, especially the soot content, can effect both the absorption and emission of thermal radiation. Second, the radiation transport is discretized via approximately 100 solid angles. For targets far away from a localized source of radiation, like a growing re, the discretization can lead to a non-uniform distribution of the radiant energy. This can be seen in the visualization of surface temperatures, where hot spots show the effect of the nite number of solid angles. The problem can be lessened by the inclusion of more solid angles, but at a price of longer computing times. In most cases, the radiative ux to far-eld targets is not as important as those in the near-eld, where coverage by the default number of angles is much better. Hostikka et al. examined the sensitivity of the radiation solver to changes in the assumed soot production, number of spectral bands, number of control angles, and ame temperature. Some of the more interesting ndings were: Changing the soot yield from 1 % to 2 % increased the radiative ux from a simulated methane burner about 15 % 7

 Lowering the soot yield to zero decreased the radiative ux about 20 %. Increasing the number of control angles by a factor of 3 was necessary to ensure the accuracy of the model at the discrete measurement locations. Changing the number of spectral bands from 6 to 10 did not have a strong effect on the results. Errors of 100 % in heat ux were caused by errors of 20 % in absolute temperature. The sensitivity to ame temperature and soot composition are consistent with combustion theory, which states that the source term of the radiative transport equation is a function of the absorption coefcient multiplied by the absolute temperature raised to the fourth power. The number of control angles and spectral bands are user-controlled numerical parameters whose sensitivities ought to be checked for each new scenario. The default values in FDS are appropriate for most large scale re scenarios, but may need to be rened for more detailed simulations such as a low-sooting methane burner.

2.3.4

Sensitivity of Thermophysical Properties of Solid Fuels

An extensive amount of verication and validation work with FDS version 4 has been performed by Hietaniemi, Hostikka, and Vaari at VTT, Finland [34]. The case studies are comprised of re experiments ranging in scale from the cone calorimeter (ISO 5660-1) to full-scale re tests such as the room corner test (ISO 9705). Comparisons are also made between FDS results and data obtained in the SBI (Single Burning Item) Euro-classication test apparatus (EN 13823) as well as data obtained in two ad hoc experimental congurations: one is similar to the room corner test but has only partial linings and the other is a space to study res in building cavities. All of the case studies involve real materials whose properties must be prescribed so as to conform to the assumption in FDS that solids are of uniform composition backed by a material that is either cold or totally insulating. Sensitivity of the various physical properties and the boundary conditions were tested. Some of the ndings were: The measured burning rates of various materials often fell between two FDS predictions in which cold or insulated backings were assumed for the solid surfaces. FDS lacks a multi-layer solid model. The ignition time of upholstery is sensitive to the thermal properties of the fabric covering, but the steady burning rate is sensitive to the properties of the underlying foam. Moisture content of wooden fuels is very important and difcult to measure. Flame spread over complicated objects, like cables laid out in trays, can be modeled if the surface area of the simplied object is comparable to that of the real object. This suggests sensitivity not only to physical properties, but also geometry. It is difcult to quantify the extent of the geometrical sensitivity. There is little quantication of the observed sensitivities in the study. Fire growth curves can be linear to exponential in form, and small changes in fuel properties can lead to order of magnitude changes in heat release rate for unconned res. The subject is discussed in the FDS Validation Guide (Volume 3 of the Technical Reference Guide). where it is noted in many of the studies that predicting re growth is difcult. Recently, Lautenberger, Rein and Fernandez-Pello [35] developed a method to automate the process of estimating material properties to input into FDS. The methodology involves simulating a bench-scale test with the model and iterating via a "genetic" algorithm to obtain an optimal set of material properties for that particular item. Such techniques are necessary because most bench-scale apparatus do not provide a complete set of thermal properties. 8

2.4

Code Checking

An examination of the structure of the computer program can be used to detect potential errors in the numerical solution of the governing equations. The coding can be veried by a third party either manually or automatically with proling programs to detect irregularities and inconsistencies [2]. At NIST and elsewhere, FDS has been compiled and run on computers manufactured by IBM, HewlettPackard, Sun Microsystems, Digital Equipment Corporation, Apple, Silicon Graphics, Dell, Compaq, and various other personal computer vendors. The operating systems on these platforms include Unix, Linux, Microsoft Windows, and Mac OSX. Compilers used include Lahey Fortran, Digital Visual Fortran, Intel Fortran, IBM XL Fortran, HPUX Fortran, Forte Fortran for SunOS, the Portland Group Fortran, and several others. Each combination of hardware, operating system and compiler involves a slightly different set of compiler and run-time options and a rigorous evaluation of the source code to test its compliance with the Fortran 90 ISO/ANSI standard [36]. Through this process, out-dated and potentially harmful code is updated or eliminated, and often the code is streamlined to improve its optimization on the various machines. However, simply because the FDS source code can be compiled and run on a wide variety of platforms does not guarantee that the numerics are correct. It is only the starting point in the process because it at least rules out the possibility that erratic or spurious results are due to the platform on which the code is running. Beyond hardware issues, there are several useful techniques for checking the FDS source code that have been developed over the years. One of the best ways is to exploit symmetry. FDS is lled with thousands of lines of code in which the partial derivatives in the conservation equations are approximated as nite differences. It is very easy in this process to make a mistake. Consider, for example, the nite difference approximation of the thermal diffusion term in the i jkth cell of the three-dimensional grid: ( kT )i jk Ti+1, jk Ti jk Ti jk Ti1, jk 1 ki+ 1 , jk ki 1 , jk + 2 2 x x x Ti, j+1,k Ti jk Ti jk Ti, j1,k 1 ki, j+ 1 ,k ki, j 1 ,k + 2 2 y y y Ti j,k+1 Ti jk Ti jk Ti j,k1 1 k ki j,k 1 1 2 z i j,k+ 2 z z

which is written as follows in the Fortran source code: DTDX = (TMP(I+1,J,K)-TMP(I,J,K))*RDXN(I) KDTDX(I,J,K) = .5*(KP(I+1,J,K)+KP(I,J,K))*DTDX DTDY = (TMP(I,J+1,K)-TMP(I,J,K))*RDYN(J) KDTDY(I,J,K) = .5*(KP(I,J+1,K)+KP(I,J,K))*DTDY DTDZ = (TMP(I,J,K+1)-TMP(I,J,K))*RDZN(K) KDTDZ(I,J,K) = .5*(KP(I,J,K+1)+KP(I,J,K))*DTDZ DELKDELT = (KDTDX(I,J,K)-KDTDX(I-1,J,K))*RDX(I) + . (KDTDY(I,J,K)-KDTDY(I,J-1,K))*RDY(J) + . (KDTDZ(I,J,K)-KDTDZ(I,J,K-1))*RDZ(K) This is one of the simpler constructs because the pattern that emerges within the lines of code make it fairly easy to check. However, a mis-typing of an I or a J, a plus or a minus sign, or any of a hundred different mistakes can cause the code to fail, or worse produce the wrong answer. A simple way to eliminate many of these mistakes is to run simple scenarios that have perfectly symmetric initial and boundary conditions. For example, put a hot cube in the exact center of a larger cold compartment, turn off gravity, and watch the heat diffuse from the hot cube into the cold gas. Any simple error in the coding of the energy equation will show 9

up almost immediately. Then, turn on gravity, and in the absence of any coding error, a perfectly symmetric plume will rise from the hot cube. This checks both the coding of the energy and the momentum equations. Similar checks can be made for all of the three dimensional nite difference routines. So extensive are these types of checks that the release version of FDS has a routine that generates a tiny amount of random noise in the initial ow eld so as to eliminate any false symmetries that might arise in the numerical solution. The process of adding new routines to FDS is as follows: typically the routine is written by one person (not necessarily a NIST staffer) who takes the latest version of the source code, adds the new routine, and writes a theoretical and numerical description for the FDS Technical Reference Guide, plus a description of the input parameters for the FDS Users Guide. The new version of FDS is then tested at NIST with a number of benchmark scenarios that exercise the range of the new parameters. Provisional acceptance of the new routine is based on several factors: (1) it produces more accurate results when compared to experimental measurement, (2) the theoretical description is sound, and (3) any empirical parameters are obtainable from the open literature or standard bench-scale apparatus. If the new routine is accepted, it is added to a test version of the software and evaluated by external users and/or NIST grantees whose research is related to the subject. Assuming that there are no intractable issues that arise during the testing period, the new routine eventually becomes part of the release version of FDS. Even with all the code checking performed at NIST, it is still possible for errors to go unnoticed. One remedy is the fact that the source code for FDS is publicly released. Although it consists of on the order of 30,000 lines of Fortran statements, various researchers outside of NIST have been able to work with it, add enhancements needed for very specic applications or for research purposes, and report back to the developers bugs that have been detected. The source code is organized into 27 separate les, each containing subroutines related to a particular feature of the model, like the mass, momentum, and energy conservation equations, sprinkler activation and sprays, the pressure solver, etc. The lengthiest routines are devoted to input, output and initialization. Most of those working with the source code do not concern themselves with these lengthy routines but rather focus on the nite-difference algorithm contained in a few of the more important les. Most serious errors are found in these les, for they contain the core of the algorithm. The external researchers provide feedback on the organization of the code and its internal documentation, that is, comments within the source code itself. Plus, they must compile the code on their own computers, adding to its portability.

10

Chapter 3

The Basic Flow Solver

3.1 Single Mesh, Non-reacting Periodic Test Cases

In this section we present test cases aimed at exercising the advective, pressure and viscous terms and the time integration for non-reacting ows with periodic boundary conditions in all directions (this eliminates the possibility of gravitational stratication). In Section 3.1.1 we present an analytical solution that is useful for conrming the convergence rates of the truncation errors in the discretization schemes. In Section 3.1.2 we present a canonical ow for LES which tests that the subgrid stress model has been coded properly.

3.1.1

2D Analytical Solution to Navier-Stokes

Consider the 2D incompressible Navier-Stokes equations u + u u = p + 2 u , t (3.1)

where the velocity is given by u = [u, v]T , and the kinematic viscosity and pressure are denoted and p, respectively. An analytical solution of these equations is given by [37] u(x, y,t) = 1 A cos(x t) sin(y t) e2t , v(x, y,t) = 1 + A sin(x t) cos(y t) e p(x, y,t) =
2t

(3.2) (3.3)

A2 [cos(2(x t)) + cos(2(y t))] e4t . (3.4) 4 Here, A represents an arbitrary amplitude and is assumed to take a value of 2 in this example. Note that this solution satises continuity for all time, u = 0, (3.5) is spatially periodic on an interval 2 in each direction, and is temporally periodic on 2 if = 0; otherwise, the solution decays exponentially. Below we present two series of tests which demonstrate the second-order accuracy of the FDS numerical scheme and thus provide a strong form of code verication for the advective and viscous terms which are exercised. The physical domain of the problem is a square of side L = 2. The grid spacing is uniform x = y = L/N in each direction with N = {8, 16, 32, 64} for each test series. The staggered grid locations are denoted xi = i x and y j = j y, and the cell centers are marked by an overbar, xi = xi x/2 and y j = y j y/2. First, we present qualitative results for the case in which = 0. Thus, only the advective discretization and the time integration are being tested. Figure 3.1 shows the initial and nal (t = 2) numerical solution 11

Figure 3.1: Initial and nal states of the u-component of velocity.

for the case N = 64. As mentioned, with = 0 the solution is periodic in time and this gure demonstrates that, as should be the case, the FDS numerical solution is unaltered after one ow-through time. Next, in Figure 3.2, we show time histories of the u-component of velocity at the center of the domain for the case in which = 0.1. It is clearly seen that the FDS solution (thin line) converges to the analytical solution (thick line). Note that the analytical solution is evaluated at the same location as the FDS staggered grid location for the u-component of velocity, (xN/2 , yN/2 ), which is different in each case, N = {8, 16, 32, 64}. Figure 3.3 is the key quantitative result of this verication test. In this gure we plot the rms error, rms , in the u-component of velocity against the grid spacing. The error is dened by rms
2 1 M Uikj u(xi , y j ,tk ) , M k=1

(3.6)

where M is the number of time steps and k is the time step index. The spatial indices are (i = N/2, j = N/2) and Uikj represents the FDS value for the u-component of velocity at the staggered storage location for cell (i, j) at time step k; u(xi , y j ,tk ) is the analytical solution for the u-component at the corresponding location in space and time. The gure conrms that the advective terms, the viscous terms, and the time integration in the FDS code are convergent and second-order accurate.

12

Analytical (u-vel) FDS (UVEL)

Analytical (u-vel) FDS (UVEL)

2.0 Velocity (m/s) 1.5 1.0 0.5 0 1 2

U-Velocity (ns2d_8_nupt1) Velocity (m/s)

2.0 1.5 1.0 0.5

U-Velocity (ns2d_16_nupt1)

3 4 Time (s)

3 4 Time (s)

Analytical (u-vel) FDS (UVEL)

Analytical (u-vel) FDS (UVEL)

2.0 Velocity (m/s) 1.5 1.0 0.5 0 1 2

U-Velocity (ns2d_32_nupt1) Velocity (m/s)

2.0 1.5 1.0 0.5

U-Velocity (ns2d_64_nupt1)

3 4 Time (s)

3 4 Time (s)

Figure 3.2: Time history of the u-component of velocity half a grid cell below the center of the domain for a range of grid resolutions. The domain is a square of side L = 2 m. The N N grid is uniform. Progressing from left to right and top to bottom we have resolutions N = {8, 16, 32, 64} clearly showing convergence of the FDS numerical solution (open circles) to the analytical solution (solid line). The case is run with constant properties, = 1 kg/m3 and = 0.1 kg/m/s, and a CFL of 0.25.

13

100

100

RMS Error (m/s)

RMS Error (m/s)

O (x) O (x2

10-1

10-1

10-2

) FDS (rms error)

10-2 10-1

O (x) O (x2

) FDS (rms error)

10-1

Grid Spacing, x (m)

Grid Spacing, x (m)

Figure 3.3: (Left) Convergence rate for the u-component of velocity with = 0 showing that the advective terms in the FDS code are second-order accurate. The triangles represent the rms error in the u-component for grid spacings of x = L/N where L = 2 m and N = {8, 16, 32, 64}. The solid line represents rst-order accuracy and the dashed line represents second-order accuracy. The simulation is run to a time of t = 2 s with a CFL of 0.25. The u-component at the center of the domain is compared with the analytical solution at the same location. (Right) Same case, except = 0.1, showing that the viscous terms in the FDS code are second-order accurate.

14

3.1.2

Decaying Isotropic Turbulence

In some cases the difference between verication and validation is not so clear. Once a model is wellestablished and validated it may actually be used as a form of verication. Granted, such a test is not as strong a verication as the convergence study shown in Section 3.1.1. Nevertheless, these tests are often quite useful in discovering problems within the code. The case we examine in this section, decaying isotropic turbulence, is highly sensitive to errors in the advective and diffusive terms because the underlying physics is inherently three-dimensional and getting the problem right depends strongly on a delicate balance between vorticity dynamics and dissipation. An even more subtle yet extremely powerful verication test is also presented in this section when we set both the molecular and turbulent viscosities to zero and conrm that the integrated kinetic energy within the domain remains constant. In the absence of any form of viscosity, experience has shown that the slightest error in the advective terms or the pressure projection will cause the code to go unstable. This verication is therefore stronger than one might initially expect. In this section we test the FDS model against the low Reynolds number (Re) data of Comte-Bellot and Corrsin (CBC) [38]. Viscous effects are important in this data set for a well-resolved LES, testing the models Re dependence. Following [39], we use a periodic box of side L = 9 2 centimeters ( 0.566 m) and = 1.5 105 m2 /s for the kinematic viscosity. The non-dimensional times for this data set are: x/M = 42 (initial condition), 98, and 171, where M is the characteristic mesh spacing of the CBC wind tunnel and x is the downstream location of the data station. Considering the mean velocity in the CBC wind tunnel experiment, these correspond to dimensional times of t = 0.00, 0.28, and 0.66 seconds in our simulations. The initial condition for the FDS simulation is generated by superimposing Fourier modes with random phases such that the spectrum matches that of the initial CBC data. An iterative procedure is employed where the eld is allowed to decay for small time increments subject to Navier-Stokes physics, each wavenumber is then injected with energy to again match the initial ltered CBC spectrum. The specic lter used here is discussed in [40]. To provide the reader with a qualitative sense of the ow, Figure 3.4 shows the initial and nal states of the velocity eld in the 3D periodic domain. The ow is unforced and so if viscosity is present the total energy decays with time due to viscous dissipation. Because the viscous scales are unresolved, a subgrid stress model is required. Here the stress is closed using the gradient diffusion hypothesis and the eddy viscosity is modeled by the constant coefcient Smagorinsky model with the coefcient taken to be Cs = 0.2 (see the Technical Reference Guide for further details). The decay curves for two grid resolutions are shown plotted on the left in Figure 3.5. For an LES code such as FDS which uses a physically-based subgrid model, an important verication test is to run this periodic isotropic turbulence simulation in the absence of both molecular and turbulent viscosity. For so-called energy-conserving explicit numerics the integrated energy will remain nearly constant in time. This is demonstrated by the blue line in the top-left plot in Figure 3.5. The deviations from identical energy conservation (to machine precision) are due solely to the time discretization (the spatial terms are conservative as discussed in [41]) and converge to zero as the time step goes to the zero. Note that strict energy conservation requires implicit time integration [42, 43] and, as shown by the red curve on the same plot where only molecular viscosity is present in the simulation, this cost is unwarranted given that the molecular dissipation rate clearly overshadows the relatively insignicant amount of numerical dissipation caused by the explicit method. The FDS result using the Smagorinsky eddy viscosity (the black solid line) matches the CBC data (red open circles) well for the 323 case (top-left). However, the FDS results are slightly too dissipative in the 643 case (bottom-left). This is due to a well-known limitation of the constant coefcient Smagorinsky model: namely, that the eddy viscosity does not converge to zero at the appropriate rate as the lter width (here equivalent to the grid spacing) is decreased. 15

Figure 3.4: Initial and nal states of the isotropic turbulence eld.

To the right of each decay curve plot in Figure 3.5 is the corresponding spectral data comparison. The three black solid lines are the CBC spectral data for the points in time corresponding to dimensional times of t = 0.00, 0.28, and 0.66 seconds in our simulations. As described above, the initial FDS velocity eld (represented by the black dots) is specied to match the CBC data up to the grid Nyquist limit. From there the spectral energy decays rapidly as discussed in [40]. For each of the spectral plots on the right, the results of interest are the values of the red and blue dots and how well these match up with the corresponding CBC data. For the 323 case (top-right) the results are remarkably good. Interestingly, the results for the more highly resolved 643 case are not as good. This is because the viscous scales are rather well-resolved at the later times in the experiment and, as mentioned, it is well-known that the constant coefcient Smagorinsky model is too dissipative under such conditions. Overall, the agreement between the FDS simulations and the CBC data is satisfactory and any discrepancies can be explained by limitations of the model. Therefore, as a verication the results here are positive in that nothing points to coding errors.

16

0.05 0.045 0.04 kinetic energy (m2/s2) 0.035 0.03 0.025 0.02 0.015 FDS zero visc FDS mol visc FDS Smag filtered CBC data E(k), m3/s2

10

Increasing time

10

10

LES Nyquist limit 0.01 0.005 10

6

0.1

0.2

0.3 0.4 time (s)

0.5

0.6

0.7

10

10 k, 1/m

10

0.07 FDS Smag filtered CBC data 0.06

10

Increasing time

0.05 kinetic energy (m2/s2) E(k), m3/s2

10 0.04

0.03

10 0.02

0.01
6

LES Nyquist limit

0.1

0.2

0.3 0.4 time (s)

0.5

0.6

0.7

10

10

10 k, 1/m

10

Figure 3.5: (Left) Time histories of integrated kinetic energy corresponding to the grid resolutions on the right side
of the gure. In the 323 case (top), the CBC data (open circles) are obtained by applying a lter to the CBC energy spectra at the Nyquist limit for an N = 32 grid. Similarly, for the 643 case (bottom), the CBC data are obtained from ltered spectra for an N = 64 grid. Notice that the integrated FDS results for the 323 case compare better with the ltered CBC data than the 643 results. This is a well-known limitation of the constant coefcient Smagorinsky model: namely, that the eddy viscosity does not converge to zero at the appropriate rate as the lter width (here equivalent to the grid spacing) is decreased. (Right) Energy spectra for the 323 case (top) and the 643 case (bottom). The solid black lines are the spectral data of Comte-Bellot and Corrsin at three different points in time corresponding to downstream positions in the turbulent wind tunnel. The initial condition for the velocity eld (spectra shown as black dots) in the FDS simulation is prescribed such that the energy spectrum matches the initial CBC data. The FDS energy spectra corresponding to the subsequent CBC data are shown by the red and blue dots.

17

Figure 3.6: Smagorinsky coefcient for a 643 simulation of the CBC experiment.

3.1.3

The Dynamic Smagorinsky Model

In the previous section, all calculations were performed with a constant and uniform Smagorinsky coefcient, Cs = 0.2. For the canonical case of homogeneous decaying isotropic turbulence at sufciently high Reynolds number this model is sufcient. However, we noticed that, even for the isotropic turbulence problem, when the grid Reynolds number is low (i.e., the ow is well-resolved) the constant coefcient model tends to over predict the dissipation of kinetic energy (see Figure 3.5). This is because the eddy viscosity does not converge to zero at the proper rate; so long as strain is present in the ow (the magnitude of the stain rate tensor is nonzero), the eddy viscosity will be nonzero. This violates a guiding principle in LES development: that the method should converge to a DNS if the ow eld is sufciently resolved [44]. The dynamic procedure for calculating the model coefcient (invoked by setting DYNSMAG=.TRUE. on the MISC line) alleviates this problem. The basis of the model is that the coefcient should be the same for two different lter scales within the inertial subrange. Details of the procedure are explained in the following references [45, 46, 47, 48, 49]. Here we present results for the implementation of the dynamic model in FDS. In Figure 3.6 we show contours of the Smagorinsky coefcient Cs (x,t) at a time midway through a 643 simulation of the CBC experiment. Notice that the coefcient ranges from 0.00 to roughly 0.30 within the domain with the average value falling around 0.17. Next, in Figure 3.7, we show results for the dynamic model analogous to Figure 3.5. For the 323 case the result is not dramatically different than the constant coefcient model. In fact, one might argue that the 323 constant coefcient results are slightly better. But there are several reasons why we should not stop here and conclude that the constant coefcient model is superior. First, as pointed out in Pope Exercise 13.34 [50], 383 is required to resolve 80% of the total kinetic energy (for this ow) and thus put the cutoff wavenumber within the inertial subrange of turbulent length scales. Pope recommends that simulations which are underresolved by this criterion should be termed very large-eddy simulations weather forecasting is a typical example. For a 323 LES, the test lter width in the dynamic model falls at a resolution of 163 , clearly outside the inertial range. A tacit assumption underlying the original interpretation of the dynamic model is that both the grid lter scale and the test lter scale should fall within the inertial range, since this is the 18

0.05 0.045 0.04 kinetic energy (m2/s2) 0.035 0.03 0.025 0.02 0.015 0.01 0.005 E(k), m3/s2 FDS Dyn Smag filtered CBC data

10

Increasing time

10

10

LES Nyquist limit

0.1

0.2

0.3 0.4 time (s)

0.5

0.6

0.7

10

10

10 k, 1/m

10

0.07 FDS Dyn Smag filtered CBC data 0.06

10

Increasing time

0.05 kinetic energy (m2/s2) E(k), m3/s2

10 0.04

0.03

10 0.02

LES Nyquist limit 0.01

6

0.1

0.2

0.3 0.4 time (s)

0.5

0.6

0.7

10

10

10 k, 1/m

10

Figure 3.7: Dynamic Smagorinsky model results (analogous to Figure 3.5) for integrated kinetic energy (left) and
spectra (right).

range in which the scales of turbulent motion (in theory) exhibit fractal-like, scale similar behavior (recently the procedure has been derived from other arguments [51]). With this in mind, it is perhaps not surprising that the dynamic model does not perform optimally for the low resolution case. In the higher resolution 643 case, however, the dynamic model does perform better than the constant coefcient model and this is the desired result: we want better performance at higher resolution. As can be seen from the energy spectra (lower right), the energy near the grid Nyquist limit is more accurately retained by the dynamic model. This equates to better ow structure with fewer grid cells. Thus, for practical calculations of engineering interest the small computational overhead of computing the coefcient may recuperated by a reduction is cell count.

19

20

Chapter 4

Thermal Radiation
The Radiative Transport Equation (RTE) for an absorbing/emitting and scattering medium is s I (x, s) = (x, ) + s (x, ) I (x, s) + B(x, ) + s (x, ) 4 (s, s ) I (x, s ) ds (4.1)

where I (x, s) is the radiation intensity at wavelength , s is the direction vector of the intensity, (x, ) and s (x, ) are the local absorption and scattering coefcients, respectively, and B(x, ) is the emission source term. The integral on the right hand side describes the in-scattering from other directions. In the case of a non-scattering gas the RTE becomes s I (x, s) = (x, ) Ib (x) I (x, s) (4.2)

where Ib (x) is the source term given by the Planck function (see below). In practical simulations the spectral () dependence cannot be solved accurately. Instead, the radiation spectrum is divided into a relatively small number of bands and a separate RTE is derived for each band. The band specic RTE is s In (x, s) = n (x) [Ib,n (x) In (x, s)] , n = 1...N (4.3)

where In is the intensity integrated over the band n, and n is the appropriate mean absorption coefcient inside the band. The source term can be written as a fraction of the blackbody radiation Ib,n = Fn (min , max ) T 4 / (4.4)

where is the Stefan-Boltzmann constant. The calculation of factors Fn is explained in Ref. [52]. When the intensities corresponding to the bands are known, the total intensity is calculated by summing over all the bands
N

I(x, s) =

n=1

In (x, s)

(4.5)

There are numerous examples in the heat transfer literature of exact solutions, for simple congurations of hot and cold objects, of the radiation transport equation.

21

4.1

Radiation inside a box (radiation_in_a_box)

This verication case tests the computation of three-dimensional conguration factor inside a cube box with one hot wall and ve cold (0 K) walls. An overview of the test geometry is shown here:

H4 H3 c a b (y,z) H2 H1 dA

The conguration factors are calculated at the diagonal of the cold wall opposite to the hot wall. The exact values of the conguration factor from plane element dA to parallel rectangle H are calculated using the analytical solution [52] (y,z) 0.025 0.075 0.125 0.175 0.225 HdA 0.1457 0.1603 0.1748 0.1888 0.2018 (y,z) 0.275 0.325 0.375 0.425 0.475 HdA 0.2135 0.2233 0.2311 0.2364 0.2391

Different variations of the case include the mesh resolution (203 and 1003 cells) and the number of radiation angles (50, 100, 300, 1000, 2000). The exact and FDS results are shown here:

Incident Heat Flux (radiation_box) 0.30 Heat Flux (kW/m2 ) 0.25 0.20 0.15 0.10 0.05 0.00 0.0 0.2
Analytical (Phi_HdA) FDS (Flux_20_50) FDS (Flux_20_100) FDS (Flux_20_300) FDS (Flux_20_1000) FDS (Flux_20_2000)

Incident Heat Flux (radiation_box) 0.30 Heat Flux (kW/m2 ) 0.25 0.20 0.15 0.10 0.05 0.8 1.0 0.00 0.0 0.2
Analytical (Phi_HdA) FDS (Flux_100_50) FDS (Flux_100_100) FDS (Flux_100_300) FDS (Flux_100_1000) FDS (Flux_100_2000)

0.4 0.6 Position (m)

0.4 0.6 Position (m)

0.8

1.0

22

4.2

Radiation from a plane layer (radiation_plane_layer)

This case tests the computation of three-dimensional radiation from a homogenous, innitely wide layer of hot gases. The temperature of the layer is 1273.15 K and the absorption coefcient, , is varied. The thickness of the layer is xed at 1 m, and the optical depth is = (1 ) m1 . Wall temperatures are set to 0 K. The results are compared against the exact solution S() presented in [53] S() = Sb [1 2E3 ()] (4.6)

where Sb = T 4 is the black-body heat ux from the radiating plane and E3 () is the exponential integral function (order 3) of the optical depth . The FDS results are computed at two mesh resolutions in the x-direction (I=20 and I=150). For I=20, both one-band and six-band versions are included to test the correct integration of heat uxes over multiple bands. For I=20, 2-D versions are also computed (J=1). The limiting case, = , using a solid wall of temperature 1273.15 K, is computed to test the wall heat ux computation. The exact values and FDS predictions of the wall heat uxes are given in the table below. (m1 ) 0.01 0.1 0.5 1.0 10 S() (kW/m2 ) 2.8970 24.9403 82.9457 116.2891 148.9698 148.9709 FDS (I=20,J=20) 1 band 6 bands 2.9214 2.9104 25.5668 25.4705 83.1353 82.8224 115.4055 114.9711 148.9619 148.4011 147.7533 147.1970 FDS (I=20,J=1) 1 band 6 bands 2.8364 2.8257 25.1078 25.0133 84.3719 84.0542 117.8011 117.3576 148.9677 148.4069 147.9426 147.3856 FDS (I=150) 1 band 2.9285 25.7191 84.0311 116.7755 148.9695 147.9419

23

4.3

Wall Internal Radiation (wall_internal_radiation)

In-depth absorption of thermal radiation in a solid is computed using a two-ux model. In this example, the accuracy of the two-ux model is tested in the computation of the emissive ux from a 0.10 m thick, homogenous layer of material at a temperature of 1273.15 K surrounded by an ambient temperature of 10 K. The absorption coefcient is varied to cover a range [0.01, 10] of optical depth, . The exact solutions for radiative ux are the analytical solutions of plane layer emission [53] S() = Sb [1 2E3 ()] (4.7)

where Sb = T 4 is the black-body heat ux from the radiating plane and E3 () is the exponential integral function (order 3) of optical depth, . The exact solutions and FDS results are shown in the table below. (m1 ) 0.01 0.1 0.5 1.0 10. S() (kW/m2 ) 2.897 24.94 82.95 116.3 149.0 FDS (kW/m2 ) 2.950 26.98 93.90 128.4 149.0

24

Chapter 5

Heat Conduction
This chapter contains examples that test the one-dimensional heat conduction solver in FDS. A one-dimensional heat conduction equation for the solid phase temperature Ts (x,t) is applied in the direction x pointing into the solid (the point x = 0 represents the surface) s cs Ts Ts = ks + qs t x x (5.1)

In cylindrical and spherical coordinates, the heat conduction equation is written s cs Ts Ts 1 = rks t r r r + qs ; s cs Ts 1 Ts = 2 r 2 ks t r r r + qs (5.2)

FDS offers the user these options, with the assumption that the obstruction is not actually recti-linear, but rather cylindrical or spherical in shape. This option is useful in describing the behavior of small, complicated targets like cables or heat detection devices.

25

5.1

Simple Heat Conduction Through a Solid Slab (heat_conduction)

Analytical solutions of transient, one-dimensional heat conduction through a slab can be found in Refs. [54] and [55]. Four cases are examined here. In each, a slab of thickness L = 0.1 m is exposed on one face to an air temperature of Tg = 120 C. The other face is insulated (adiabatic). The convective heat transfer from the gas to the slab is qc = h (Tg Ts ), where h is constant, and Ts is the slab face temperature. No thermal radiation is included. Case A B C D k (W/m/K) 0.1 0.1 1.0 10.0 (kg/m3 ) 100 100 1000 10000 c (kJ/kg/K) 1 1 1 1 h (W/m2 /K) 100 10 10 10 Bi hL/k 100 10 1 0.1

140 Surface Temperature (heat_conduction_a) 120 Temperature ( C) 100 80 60 40 20 0 0 5 10 15 20 Time (min) 25 30 Temperature ( C)
Analytical (Front) Analytical (4 cm) Analytical (Back) FDS (Front) FDS (4 cm) FDS (Back)

140 Surface Temperature (heat_conduction_b) 120 100 80 60 40 20 0 0 50 100 150 200 Time (min)
Analytical (Front) Analytical (4 cm) Analytical (Back) FDS (Front) FDS (4 cm) FDS (Back)

250

300

140 Surface Temperature (heat_conduction_c) 120 Temperature ( C) 100 80 60 40 20 0 0 100 200 300 400 Time (min)
Analytical (Front) Analytical (4 cm) Analytical (Back) FDS (Front) FDS (4 cm) FDS (Back)

140 Surface Temperature (heat_conduction_d) 120 Temperature ( C) 100 80 60 40 20 0 0 100 200 300 400 500 600 700 800 Time (min)
Analytical (Front) Analytical (4 cm) Analytical (Back) FDS (Front) FDS (4 cm) FDS (Back)

500

600

26

5.2

Temperature-Dependent Thermal Properties (heat_conduction_kc)

This example demonstrates the 1-D heat conduction in cartesian, cylindrical and spherical geometries with temperature-dependent thermal properties. The cartesian solution was computed using HEATING (version 7.3), a multi-dimensional, nite-difference, general purpose heat transfer model [56]. The cylindrical and spherical solutions were computed using a commercial nite-element solver, ABAQUS. The sample of homogenous material is initially at 0 C and at t > 0 exposed to a gas at 700 C. A xed heat transfer coefcient of 10 W/m2 /K is assumed. The density of the material is 10000 kg/m3 . The conductivity and specic heat are functions of temperature with the following values: k(0) = 0.10 W/m/K, k(200) = 0.20 W/m/K, c(0) = 1.0 kJ/kg/K, c(100) = 1.2 kJ/kg/K, c(200) = 1.0 kJ/kg/K. The thickness (radius) of the sample is 0.01 m. In the cartesian case, the back surface of the material is exposed to a gas at 0 C. In the gure below, the light colored solid lines are FDS results and the dark lines are the HEATING results. An example input with cylindrical geometry looks like:
&MATL ID='MAT_1' EMISSIVITY = 0.0 CONDUCTIVITY_RAMP='K_RAMP' SPECIFIC_HEAT_RAMP = 'C_RAMP' DENSITY=10000. / &RAMP &RAMP &RAMP &RAMP &RAMP &RAMP ID ID ID ID ID ID = = = = = = 'K_RAMP' 'K_RAMP' 'K_RAMP' 'C_RAMP' 'C_RAMP' 'C_RAMP' T=0, T=100, T=200, T=0, T=100, T=200, F= F= F= F= F= F= 0.10 0.15 0.20 1.00 1.20 1.00 / / / / / /

&SURF ID='SLAB' STRETCH_FACTOR = 1.0 GEOMETRY = 'CYLINDRICAL' MATL_ID='MAT_1' THICKNESS=0.01 /

250 200 Temperature ( C)

Surface Temperature (heat_conduction_kc) Temperature ( C)

50 40 30 20 10 16 0 0

Back/Inner Temperature (heat_conduction_kc)

150 100 50 0 0 2 4 6 8 10 Time (min)

HEATING (cart_surf) ABAQUS (cyl_surf) ABAQUS (sph_surf) FDS (cart_front) FDS (cyl_front) FDS (sph_front)

HEATING (cart_back) ABAQUS (cyl_back) ABAQUS (sph_back) FDS (cart_back) FDS (cyl_back) FDS (sph_back)

12

14

8 10 Time (min)

12

14

16

27

28

Chapter 6

Pyrolysis
Solids can undergo simultaneous reactions under the following assumptions: instantaneous release of volatiles from solid to the gas phase, local thermal equilibrium between the solid and the volatiles, no condensation of gaseous products, and no porosity effects Each material component may undergo several competing reactions, and each of these reactions may produce some other solid component (residue), gaseous fuel, and/or water vapor according to the yield coefcients s , f and w , respectively. These coefcients should usually satisfy s + f + w = 1, but smaller yields may also be used to take into account the gaseous products that are not explicitly included in the simulation.

29

6.1

A Simple Two-Step Pyrolysis Example (two_step_solid_reaction)

Consider the set of ordinary differential equations describing the mass fraction of three components of a solid material undergoing thermal degradation: dYa dt dYb dt dYc dt = KabYa = KabYa KbcYb = KbcYa (6.1)

where the mass fraction of component a is 1 initially. The analytical solution is: Ya (t) = exp(Kabt) Kab Yb (t) = exp(Kabt) exp(Kbct) Kbc Kab Yc (t) = [Kab (1 exp(Kbct)) + Kbc (exp(Kabt) 1)] /(Kab Kbc )

(6.2) (6.3)

The analytical and numerical solution for the parameters Kab = 0.389 and Kbc = 0.262 are shown here:

1.0 0.8 Temperature ( C) 0.6 0.4 0.2 0.0 0

Solid Density (two_step_solid_reaction)

Analytical (Ya) Analytical (Yb) Analytical (Yc) FDS (Ya) FDS (Yb) FDS (Yc)

10

15 20 Time (s)

25

30

30

6.2

Development of surface emissivity (emissivity)

For thermally thick surfaces, the surface emissivity is computed as a mass-weighted sum of the individual material emissivities in the rst condensed phase grid cell. In the verication test, the initial material, having emissivity of 1.0, is converted to another material, having emissivity of 0.0, at a constant rate of 0.1 s1 . As a result, the surface emissivity should change linearly from 1.0 to 0.0 in 10 s.

1.0 0.8 Emissivity 0.6 0.4 0.2

Analytical (Emissivity) FDS (EMISSIVITY)

Surface Emissivity (emissivity)

0.0

Time (s)

10

31

6.3

Enthalpy of solid materials (enthalpy)

Consider a thin plate of conductive material that is exposed on one side to an elevated temperature heat source and exposed on the other to an ambient temperature void. In the thermally-thin limit, the temperature of the slab is governed by the following equation dTs qfront + qback = dt cs s (6.4)

In this example, the initial exposure to the front side of the slab is 3 kW/m2 . The original material (call it A) undergoes a reaction to form material B. The reaction rate is constant, 0.2 s1 , which in this case means that material A disappears in exactly 5 s. This is achieved by setting ns and E to 0 and A to 0.2 in the reaction rate term: s,A ns E r= (6.5) A exp s0 RTs The density and conductivity of both materials are 30 kg/m3 and 10 W/m/K, respectively. The emissivity of front and back is 1. The specic heat of material A changes from 1.0 kJ/kg/K to 0.1 kJ/kg/K above 80 C, while the specic heat of material B is constant at 1.0 kJ/kg/K. The slab is 1 mm thick.

200 150 100 50

Surface Temperature (enthalpy)

Temperature ( C)

Analytical (T_slab) FDS (T_slab)

0 0.0

0.5

1.0

1.5 2.0 2.5 Time (s)

3.0

3.5

4.0

Note that the analytical solution is actually a simple numerical integration of the equations above with a small time step to ensure accuracy. This example tests a number of features, including the reaction rate, mass weighted specic heats, and radiation boundary conditions. Note that the convective heat transfer has been turned off, and the correct steady-state temperature is calculated by FDS.

32

Chapter 7

Droplets
7.1 Water Droplet Evaporation (water_evaporation)

The test case called water_evaporation involves stationary water droplets in a box with dimensions of 1 m on a side. The walls of the box are assumed adiabatic. The air within the box is stirred to maintain uniform conditions, and there are no leaks or heat losses. Initially, the air temperature is 20 C, the median volumetric diameter of the droplets is 100 m, the water temperature is 90 C, and the total mass of water droplets is 0.2 kg. It is expected that a steady-state will be achieved after about 10 s. The initial energy content, the sum of the enthalpy of the water and air, is 75.3 kJ + 540.1 kJ = 615.4 kJ. After a short period of time, 0.0246 kg of water evaporates, the mass fraction of water vapor rises to 0.0205 kg/kg, and the air temperature increases to 22.3 C. At this point the energy content of the water droplets and humid air is 16.2 kJ + 604.9 kJ = 621.1 kJ. At 22.3 C, the expected mass fraction of water vapor is 0.0201 kg/kg, but the predicted value is 0.0205 kg/kg, an over-prediction of about 2 %. The energy gained by the air is over-predicted by about 8 %. 120 600 Enthalpy (water_evaporation) Relative Humidity (water_evaporation) 100 500
Humidity (%)
Analytical (h_gas) Analytical (h_water) FDS (h_gas) FDS (h_water)

Enthalpy (kJ)

400 300 200 100 0 0 2 4 Time (s) 6

80 60 40 20
Analytical (Rel. Hum) FDS (humid)

10

Time (s)

10

33

34

Bibliography
[1] K.B. McGrattan, B.W. Klein, S. Hostikka, and J.E. Floyd. Fire Dynamics Simulator (Version 5), Users Guide. NIST Special Publication 1019-5, National Institute of Standards and Technology, Gaithersburg, Maryland, October 2007. i [2] American Society for Testing and Materials, West Conshohocken, Pennsylvania. ASTM E 1355-04, Standard Guide for Evaluating the Predictive Capabilities of Deterministic Fire Models, 2004. i, 1, 9 [3] W. Mell, K.B. McGrattan, and H. Baum. Numerical Simulation of Combustion in Fire Plumes. In Twenty-Sixth Symposium (International) on Combustion, pages 15231530. Combustion Institute, Pittsburgh, Pennsylvania, 1996. 3 [4] K.B. McGrattan, H.R. Baum, and R.G. Rehm. Large Eddy Simulations of Smoke Movement. Fire Safety Journal, 30:161178, 1998. 3 [5] H.R. Baum, R.G. Rehm, P.D. Barnett, and D.M. Corley. Finite Difference Calculations of Buoyant Convection in an Enclosure, Part I: The Basic Algorithm. SIAM Journal of Scientic and Statistical Computing, 4(1):117135, March 1983. 3 [6] H.R. Baum and R.G. Rehm. Finite Difference Solutions for Internal Waves in Enclosures. SIAM Journal of Scientic and Statistical Computing, 5(4):958977, December 1984. 3 [7] H.R. Baum and R.G. Rehm. Calculations of Three Dimensional Buoyant Plumes in Enclosures. Combustion Science and Technology, 40:5577, 1984. 3 [8] R.G. Rehm, P.D. Barnett, H.R. Baum, and D.M. Corley. Finite Difference Calculations of Buoyant Convection in an Enclosure: Verication of the Nonlinear Algorithm. Applied Numerical Mathematics, 1:515529, 1985. 3 [9] K.B. McGrattan, T. Kashiwagi, H.R. Baum, and S.L. Olson. Effects of Ignition and Wind on the Transition to Flame Spread in a Microgravity Environment. Combustion and Flame, 106:377391, 1996. 4 [10] T. Kashiwagi, K.B. McGrattan, S.L. Olson, O. Fujita, M. Kikuchi, and K. Ito. Effects of Slow Wind on Localized Radiative Ignition and Transition to Flame Spread in Microgravity. In Twenty-Sixth Symposium (International) on Combustion, pages 13451352. Combustion Institute, Pittsburgh, Pennsylvania, 1996. 4 [11] W. Mell and T. Kashiwagi. Dimensional Effects on the Transition from Ignition to Flame Spread in Microgravity. In Twenty-Seventh Symposium (International) on Combustion, pages 26352641. Combustion Institute, Pittsburgh, Pennsylvania, 1998. 4 35

[12] W. Mell, S.L. Olson, and T. Kashiwagi. Flame Spread Along Free Edges of Thermally-Thin Samples in Microgravity. In Twenty-Eighth Symposium (International) on Combustion, pages 28432849. Combustion Institute, Pittsburgh, Pennsylvania, 2000. 4 [13] K. Prasad, Y. Nakamura, S.L. Olson, O. Fujita, K. Nishizawa, K. Ito, and T. Kashiwagi. Effect of Wind Velocity on Flame Spread in Microgravity. In Twenty-Ninth Symposium (International) on Combustion, pages 25532560. Combustion Institute, Pittsburgh, Pennsylvania, 2002. 4 [14] Y. Nakamura, T. Kashiwagi, K.B. McGrattan, and H.R. Baum. Enclosure Effects on Flame Spread over Solid Fuels in Microgravity. Combustion and Flame, 130:307321, 2002. 4 [15] W.E. Mell, K.B. McGrattan, and H.R. Baum. g-Jitter Effects on Spherical Diffusion Flames. Microgravity Science and Technology, 15(4):1230, 2004. 4 [16] A. Mukhopadhyay and I.K. Puri. An Assessment of Stretch Effects on Flame Tip Using the Thin Flame and Thick Formulations. Combustion and Flame, 133:499502, 2003. 4 [17] A. Hamins, M. Bundy, I.K. Puri, K.B. McGrattan, and W.C. Park. Suppression of Low Strain Rate Non-Premixed Flames by an Agent. In Proceedings of the 6th International Microgravity Combustion Workshop, NASA/CP-2001-210826, pages 101104. National Aeronautics and Space Administration, Lewis Research Center, Cleveland, Ohio, May 2001. 4 [18] K.B. McGrattan, R.G. Rehm, and H.R. Baum. Fire-Driven Flows in Enclosures. Journal of Computational Physics, 110(2):285291, 1994. 4 [19] P. Friday and F. W. Mowrer. Comparison of FDS Model Predictions with FM/SNL Fire Test Data. NIST GCR 01-810, National Institute of Standards and Technology, Gaithersburg, Maryland, April 2001. 6 [20] A. Bounagui, N. Benichou, C. McCartney, and A. Kashef. Optimizing the Grid Size Used in CFD Simulations to Evaluate Fire Safety in Houses. In 3rd NRC Symposium on Computational Fluid Dynamics, High Performance Computing and Virtual Reality, pages 18, Ottawa, Ontario, Canada, December 2003. National Research Council, Canada. 6 [21] R.L. Alpert. SFPE Handbook of Fire Protection Engineering, chapter Ceiling Jet Flows. National Fire Protection Association, Quincy, Massachusetts, 3rd edition, 2003. 6 [22] A. Bounagui, A. Kashef, and N. Benichou. Simulation of the Dynamics of the Fire for a Section of the L.H.-La Fontaine Tunnel. IRC-RR- 140, National Research Council Canada, Ottawa, Canada, K1A0R, September 2003. 6 [23] Y. Xin. Assessment of Fire Dynamics Simulation for Engineering Applications: Grid and Domain Size Effects. In Proceedings of the Fire Suppression and Detection Research Application Symposium, Orlando, Florida. National Fire Protection Association, Quincy, Massachusetts, 2004. 6 [24] J.A. Ierardi and J.R. Barnett. A Quantititive Method for Calibrating CFD Model Calculations. In Proceedings of the CIB-CTBUH International Conference on Tall Buildings, pages 507514. International Council for Research and Innovation in Building and Construction (CIB), 2003. 6 [25] G. Heskestad. SFPE Handbook of Fire Protection Engineering, chapter Fire Plumes, Flame Height and Air Entrainment. National Fire Protection Association, Quincy, Massachusetts, 3rd edition, 2002. 6 36

[26] N.M. Petterson. Assessing the feasibility of reducing the grid resolution in fds eld modeling. Fire Engineering Research Report 2002/6, University of Canterbury, Christchurch, New Zealand, March 2002. 6 [27] A. Musser, K. B. McGrattan, and J. Palmer. Evaluation of a Fast, Simplied Computational Fluid Dynamics Model for Solving Room Airow Problems. NISTIR 6760, National Institute of Standards and Technology, Gaithersburg, Maryland, June 2001. 6 [28] K.B. McGrattan, S. Hostikka, J.E. Floyd, H.R. Baum, and R.G. Rehm. Fire Dynamics Simulator (Version 5), Technical Reference Guide. NIST Special Publication 1018-5, National Institute of Standards and Technology, Gaithersburg, Maryland, October 2007. 7 [29] W. Zhang, A. Hamer, M. Klassen, D. Carpenter, and R. Roby. Turbulence Statistics in a Fire Room Model by Large Eddy Simulation. Fire Safety Journal, 37:721752, 2002. 7 [30] J. Smagorinsky. General Circulation Experiments with the Primitive Equations. I. The Basic Experiment. Monthly Weather Review, 91(3):99164, March 1963. 7 [31] J.W. Deardorff. Numerical Investigation of Neutral and Unstable Planetary Boundary Layers. Journal of Atmospheric Sciences, 29:91115, 1972. 7 [32] M. Germano, U. Piomelli, P. Moin, and W.H. Cabot. A Dynamic Subgrid-Scale Eddy Viscosity Model. Physics of Fluids A, 3(7):17601765, 1991. 7 [33] D.K. Lilly. A Proposed Modication of the Germano Subgrid-Scale Closure Method. Physics of Fluids A, 4(3):633635, 1992. 7 [34] J. Hietaniemi, S. Hostikka, and J. Vaari. FDS Simulation of Fire Spread Comparison of Model Results with Experimental Data. VTT Working Papers 4, VTT Building and Transport, Espoo, Finland, 2004. 8 [35] C. Lautenberger, G. Rein, and C. Fernandez-Pello. The application of a genetic algorithm to estimate the material properties for re modeling from bench-scale re test data. Fire Safety Journal, 41:204 214, 2006. 8 [36] J.C. Adams, W.S. Brainerd, J.T. Martin, B.T. Smith, and J.L. Wagener. Fortran 95 Handbook: Complete ISO/ANSI Reference. MIT Press, Cambridge, Massachusetts, 1997. 9 [37] R. McDermott. A nontrivial analytical solution to the 2-d incompressible Navier-Stokes equations. http://randy.mcdermott.googlepages.com/NS_exact_soln.pdf, 2003. 11 [38] G. Comte-Bellot and S. Corrsin. Simple Eularian time correlation of full- and narrow-band velocity signals in grid-generated, isotropic turbulence. J. Fluid Mech., 48:273337, 1971. 15 [39] Stephen M. de Bruyn Kops. Numerical simulation of non-premixed turbulent combustion. PhD thesis, The University of Washington, 1999. 15 [40] R. McDermott, A. Kerstein, R. Schmidt, and P. Smith. Characteristics of 1D spectra in nite-volume large-eddy simulations with one-dimensional turbulence subgrid closure. In American Physical Society, Division of Fluid Dynamics, Chicago, IL, http://randy.mcdermott.googlepages.com/implied_lter.pdf, Nov. 2005. 15, 16 [41] Y. Morinishi, T. S. Lund, O. V. Vasilyev, and P. Moin. Fully conservative high order nite difference schemes for incompressible ow. J. Comp. Phys., 143:90124, 1998. 15 37

[42] F. E. Ham, F. S. Lien, and A. B. Strong. A fully conservative second-order nite difference scheme for incompressible ow on non-uniform grids. J. Comp. Phys., 177:117133, 2002. 15 [43] R. McDermott. Discrete kinetic energy conservation for variable-density ows on staggered grids. In American Physical Society, Division of Fluid Dynamics, Salt Lake City, UT, http://randy.mcdermott.googlepages.com/aps2007_notes.pdf, Nov. 2007. 15 [44] R. McDermott and S. B. Pope. A particle formulation for treating differential diffusion in ltered density function methods. J. Comp. Phys., 226(1):947993, 2007. 18 [45] M. Germano, U. Piomelli, P. Moin, and W. Cabot. A dynamic subgrid-scale eddy viscosity model. Phys. Fluids A, 3(7):17601765, 1991. 18 [46] M. Pino Martin, U. Piomelli, and G. Candler. Subgrid-scale models for compressible large-eddy simulation. Theoret. Comput. Fluid Dynamics, 13:361376, 2000. 18 [47] P. Moin, K. Squires, W. Cabot, and S. Lee. A dynamic subgrid-scale model for compressible turbulence and scalar transport. Phys. Fluids A, 3(11):27462757, 1991. 18 [48] T. S. Lund. On the use of discrete lters for large eddy simulation. Center for Turbulence Research Annual Research Briefs, 1997. 18 [49] R. McDermott. Variable density formulation of the dynamic smagorinsky model. http://randy.mcdermott.googlepages.com/dynsmag_comp.pdf, 2004. 18 [50] Stephen B. Pope. Turbulent Flows. Cambridge, 2000. 18 [51] S.B. Pope. Ten questions concerning the large-eddy simulation of turbulent ows. New Journal of Physics, 6:124, 2004. 19 [52] R. Siegel and J. R. Howell. Thermal Radiation Heat Transfer. Taylor & Francis, New York, 4th edition, 2002. 21, 22 [53] Y.B. Zeldovich and Y.P. Raizer. Physics of shock waves and high-temperature hydrodynamic phenomena. Dover Publications, New York, 2002. Translated from the Russian and then edited by W.D.Hayes and R.F. Probstein. 23, 24 [54] D. Drysdale. An Introduction to Fire Dynamics. John Wiley and Sons, New York, 2nd edition, 2002. 26 [55] H.S. Carslaw and J.C. Jaegar. Conduction of Heat in Solids. Oxford University Press, 2nd edition, 1959. 26 [56] K.W. Childs. HEATING 7: Multidimensional, Finite-Difference Heat Conduction Analysis Code System. Technical Report PSR-199, Oak Ridge National Laboratory, Oak Ridge, TN, 1998. 27

38