DRAFT IECEC 2002 Paper No. xxxxx DISCRETE DYNAMICAL SYSTEM MODELS OF TURBULENCE– CHEMICAL KINETICS INTERACTIONS J. M. McDonough Departments of Mechanical Engineering and Mathematics University of Kentucky, Lexington, KY 40506-0108 Phone: (859)257-5164 Fax: (859)257-3304 jmmcd@uky.edu Sha Zhang Department of Mechanical Engineering University of Kentucky, Lexington, KY 40506-0108 Phone: (859)257-2368 szhan1@engr.uky.edul ABSTRACT A new approach to subgrid-scale (SGS) modeling for large-eddy simulation of turbulent non-premixed combustion is proposed and tested against experimental data. The model is composed of three specific factors: an amplitude, an anisotropy correction and a temporal fluctuation to be evaluated at each discrete point, during each time step, of resolved-scale calculations. We employ discrete dynamical systems (DDSs) for the third factor and in the present work focus on construction of these for a reduced kinetic mechanism and compare results with experimental data from the Technische Universität Darmstadt H2/N2−air jet diffusion flame H3. The DDS model is derived as a single-mode Galerkin approximation (with the mode left arbitrary) of the governing partial differential equations, but with the mode number and normalization(s) incorporated into bifurcation parameters. Such algebraic systems are capable of producing the full range of temporal behaviors of the original differential equations (while being very efficient to evaluate) and, in particular, can exhibit the chaotic behavior of fractal (strange) attractors that can be associated with turbulence. Moreover, they are able to mimic specific reaction pathways for any given kinetic mechanism on the subgrid scales. We compare computed results from the SGS model with the above mentioned data, both qualitatively (appearance of the time series) and quantitatively (rms fluctuation levels) and show reasonable agreement, especially for the former. INTRODUCTION It is by now widely accepted that direct numerical simulations (DNS) of turbulent combustion will be viable only as a research tool in the immediate future, and that large-eddy simulation (LES) in some form presents a feasible direction for near-term calculations. It is important, however, to recognize that although much progress in LES has been made in recent years, it is still far from being a completely reliable tool in the context of turbulent combustion ― especially if any but the simplest kinetic mechanisms are used. Most recent research in LES has focused on the subgrid-scale (SGS) models, and these are especially problematic in the context of finite-rate chemistry. Various approaches have been implemented ranging from laminar flamelets, Cook et al. (1997) through PDF models, e.g., Pope (1985), and Cook and Riley (1994), and including extension of scale similarity ideas to reactive scalars as in Germano et al. (1997). Each has its strengths and weaknesses as described in some detail by Peters (2000), and for the sake of brevity we will avoid an exhaustive review. It is worth commenting that application to combustion of notions from the theory of fractals has declined rather considerably in recent years with a few notable exceptions (e.g., Giacomazzi et al., 1999; 2000). Indeed, one of the main criticisms of such theories has been their inability to produce reliable, efficient predictive tools for turbulence in general, and turbulent combustion in particular. They have mainly provided ways to characterize experimental data of premixed flames as in Gouldin et al. (1988), although some success has been shown with modeling premixed turbulent flame speeds in Gouldin (1987). In the present paper we will revisit use of fractals in the guise of attractors of discrete dynamical systems (DDSs) to introduce a different approach to constructing SGS models for LES of turbulent combustion. These models will prove to be similar in some respects to those proposed in Giacomazzi et al. (1999; 2000), but they are rather different in detail. In particular, the form of LES considered here differs in three distinct ways from usual approaches. First, we choose to filter solutions rather than equations. Much work still is in progress concerning proper filtering of the governing equations in cases of nonuniform gridding and complex geometry, while filtering the solution reduces the problem to a relatively simple (but not altogether trivial) one of “signal processing.” At the same time it significantly alters (simplifies, we contend) the requirements for the SGS model. Second, we directly model fluctuating primitive variables ― velocity components, temperature, species concentrations ― on the subgrid scales, rather than modeling their statistics (correlations). The ability to do this has a major impact on treatment of advective terms in all governing equations, and also on that of terms containing reaction rates in the thermal energy and species concentration equations; the closure problem becomes far more manageable than is the case for all but the simplest Reynoldsaveraged Navier–Stokes (RANS) methods, and for typical LES SGS closures. (There are many fewer primitive variables than correlations, hence, much less to model.) Finally, we directly employ the modeled fluctuating quantities to augment the computed large-scale solution, which by the nature of LES is under resolved. Recall that in usual LES, and in RANS approaches, the models are used only to indirectly change the resolved part of the solution; but they are never used to construct an approximation to the complete solution. The modeling approach we are proposing here does approximate the complete solution and in that sense moves LES closer to DNS, even when performed on relatively coarse grids. We note that this approach was first studied by McDonough et al. (1995), and presented in detail by Hylin and McDonough (1999); Sagaut (2001) provides a useful overview of the method. The basic idea is to employ a typical LES-like decomposition of solution variables, say Q (x,t ) = q (x,t ) + q (x,t ) * x∈R , d d = 2,3, (1) and substitute this into the transport equation(s) for Q: (q + q ) + ∇ ⋅ F (q + q ) = ∇ ⋅ G (q + q )+ S (q + q ) * * * t * q = Aiζ i M i , i = 1,2,..., N v , ). T Then Eqs. (2) can be directly solved for q (followed by filtering to remove aliasing * due to under resolution, as needed), and q is added to the result as suggested by the form of Eq. (1). In Eqs. (3) the Ais are amplitudes derived from scaling laws of Kolmogorov (see, e.g., Frisch, 1995), and the ζis are anisotropy corrections calculated using scale similarity as in the dynamic SGS models first introduced by Germano et al. (1991). Details of these constructions can be found in McDonough et al. (1995), Hylin and McDonough (1999), Sagaut (2001). Until only recently little attention was given to the Mis. These represent the temporal fluctuations of the SGS quantities, and they were initially constructed as linear combinations of independent realizations of a modified logistic map (a chaotic map with fractal strange attractor described by May, 1976) for each qi*. This turns out to be inadequate for treatment of passive scalars, and we have recently introduced an alternative procedure (described below) for obtaining the Mis. This has previously been carried out for the 2D Navier–Stokes (N.–S.) equations in McDonough and Huang (2001), for a Boussinesq approximation to 2-D thermal convection in McDonough and Joyce (2002) and for a simple (and somewhat unrealistic) kinetic mechanism for H2–O2 combustion in McDonough and Huang (2000) and McDonough and Zhang (2002). The purpose of the present paper is to explore the behavior of such DDSs with a more realistic kinetic mechanism for H2–air combustion, and in particular provide comparisons with extant experimental results. The remainder of the paper consists of the following sections. We begin by presenting governing equations and assumptions, and then outline the derivations of the corresponding DDS for a specific reduced mechanism. We follow this with a section containing computed results and discussions of these, and end the paper with some conclusions. ANALYSIS In this section we introduce the governing equations, and from these derive a general discrete dynamical system that can be used to model any desired kinetics (including a full mechanism). We then provide a reduced mechanism for H2–air combustion and give the specific DDS corresponding to this. (2) Here, the subscript t denotes partial differentiation with respect to time, and ∇· is the divergence operator. F and G are, respectively, advective and diffusive fluxes, and S is a general nonlinear source term. We model q*, the fluctuating part of Eq. (1), as * i ( q* = q1* , q2* ,..., q*N v (3) where Nv is the total number of dependent variables: Governing Equations The general equations describing fluid flow, heat transfer and chemical reactions are well known and can be found in any standard reference. We present them here in the following form. (4a) ρt + ∇ ⋅ (ρU ) = 0 , ρ DU = −∇p + ∇ ⋅ (µ∇U ) + ρg , Dt ρc p (4b) Ns Ns ρY DT = ∇ ⋅ (λ∇T ) + ∑ c pi DiWi∇ i ⋅ ∇T − ∑ hiω! i , Dt i =1 i =1 Wi (4c) IECEC 2002 Paper No. 20163 - 2 D(ρYi ) = ∇ ⋅ (ρDi∇Yi ) + ω! i Dt i = 1,⋅ ⋅ ⋅, N s . (4d) Here, ω! i = Wi ∑ (ν i′′, j − ν i′, j )ω j , Nr (5) j =1 with Ns ρY ω j = k f , j ∏ l l =1 Wl ν l′, j Ns ρY − kb , j ∏ l l =1 Wl ν l′′, j . (6) In these equations U is the velocity vector, (u, v)T in 2D; D/Dt is the substantial derivative; ∇ is the gradient operator, and g is the body-force acceleration vector; ρ and p are density and pressure, respectively, and T denotes temperature. Yi is the mass fraction, and hi is specific enthalpy of species i. The transport properties are denoted as µ for (dynamic) viscosity, λ for thermal conductivity and Di for the binary diffusion coefficient of species i in the ambient background gas. The c pi and Wi are, respectively, specific heat and molecular are weight of species i, and ν′i,j and ν″i,j stoichiometric mole numbers for reactants and products, respectively, corresponding to species i in reaction j. Ns denotes the number of species, and Nr is the number of reactions. Finally, kf,j and kb,j are the forward and backward specific reaction rates for the jth reaction, typically given in the form of an Arrhenius expression, −E j k j = B jT nj exp RoT , (7) in which Bj is the pre-exponential factor, nj is the temperature exponent, and Ej is the activation energy; R0 is the universal gas constant. We would assume for a complete LES that in addition to these equations, sufficient initial and boundary data would be provided to constitute a mathematically well-posed problem for the chosen spatial domain. But because the DDSs (and also the corresponding complete SGS models) are local, this is not a particular concern in the present study. The Discrete Dynamical System The approach we employ for modeling the Mis in Eq. (3) was first introduced in McDonough and Huang (2000) and studied in detail for the 2-D N.–S. equations in McDonough and Huang (2001). It provides a systematic technique for deriving DDSs that are directly related to the partial differential equations (PDEs) they are to model. The premise employed to start the procedure is that all solution variables possess (generalized) Fourier series representations. For convenience we make the following additional assumptions associated with the basis sets {ϕk} used to construct the Fourier representations: i) {ϕk} is complete (in an appropriate norm); ii) the set is orthonormal; iii) each ϕk has compact support; iv) the combination of ϕ k s used to represent the velocity field is divergence free, and v) the basis set “behaves like” complex exponentials (which do not exhibit all of the desired properties) with regard to differentiation. We next substitute these into the governing equations (4), and then construct the Galerkin ordinary differential equations (ODEs) by forming the inner product of each equation with every basis function, utilizing the preceding simplifications (see McDonough and Huang, 2001 for complete details). This leads to a countable system of ODEs in place of each of the original PDEs, and from the standpoint of arithmetic complexity this is no more tractable than would be any typical form of DNS. In order to produce a model of minimal arithmetic complexity, we here consider the extreme shell model (see, e.g., Bohr et al., 1998 for a thorough treatment of shell models) ― one consisting of only a single Fourier mode in each original PDE. Then, as in McDonough and Huang (2000, 2001), McDonough and Joyce (2002), McDonough and Zhang (2002), we use very simple numerical integrators (forward or backward Euler methods) to perform the temporal discretization. After possible rescaling (to account for a specific selected wavenumber and the absence of those modes with which it would strongly interact ― all included in a bifurcation parameter) we obtain the following DDS model of Eqs. (4): ( ) ( ) (1 − b )− γ a( )b( ) + α c( ) a (n +1) = β u a (n ) 1 − a (n ) − γ u a (n )b (n ) b (n +1) = β vb (n ) n n n n v T c (n +1) = ∑ α Td i d i(n +1) − γ uT a (n +1) − γ vT b (n +1) c (n ) i =1 (8a) (8b) Ns − ∑ H iω! i (1 + βT ) + c0 i =1 Ns ( ) d (n +1) = − βYi + γ uYi a (n +1) + γ vYi b (n +1) d i(n ) + ω! i + d i ,0 (8c) (8d) i = 1,2,⋅ ⋅ ⋅, N s , with Ns Ns Nr ν′ ν ′′ ω! i = ∑ C f ,ij ∏ d l j ,l − Cb ,ij ∏ d l j ,l . j =1 l =1 l =1 Here, superscripts n denote time step (or map iteration) index; a, b, c and the di s can be viewed (heuristically) as Fourier coefficients of the two velocity components, temperature and the species concentrations, respectively; the subscripted α s , β s , etc., are bifurcation parameters of the DDS, all of which are related to the various physical bifurcation parameters. For example, βu and βv are (the same) functions of the flow Reynolds number; αT is related to the Rayleigh or Grashof number; α Td i contains Schmidt and Lewis number information, and the Hi are associated with specific enthalpies for each species i; the Cf,ij, Cb,ij can be related to Kolmogorovscale Damköhler numbers. The various γ s IECEC 2002 Paper No. 20163 - 3 correspond to velocity, temperature and species concentration gradients (as would be available from resolved-scale results). For example, γu ∼ uy, γuT ∼ Tx (subscripts x and y indicate partial differentiation), etc. The d i , 0 s and c0 are high-pass filtered species concentrations and temperature, respectively (obtained from the resolved-scale calculation), about which the subgrid-scale behavior fluctuates. Finally, we remark that in the present form of the model we calculate fluctuating density from the equation of state, rather than from a discrete form of Eq. (4a). each product appearing in each elementary reaction. Each iterated map is of the general form of Eq. (8d) and the formula for ω! i following it. But now in this latter expression Nr = 1, and backward reactions are treated separately (as forward reactions with the backward reaction rate). Thus, the individual maps are quite simple. To fix notation we make the following identifications: Reduced Mechanism for H2–Air Combustion In this subsection we present the reduced kinetic mechanism to be studied and specialize Eqs. (8) to this case. We employ a nine-step mechanism for the H2–air reaction consisting of the following: We now carry out the details for the initiation reaction Eq. (9a). We first observe that the reaction (9a) yields two products. Thus, there must be an iterated map for each of these. The first product in reaction (9a) is HO2 so the corresponding DDS is d1 ∼ H2 , d 2 ∼ O2 , d 3 ∼ H2O , d 4 ∼ OH , d 5 ∼ H , d 6 ∼ O , d 7 ∼ HO2 , d8 ∼ N2 , ( H2 + O2 → HO2 + H (9a) O2 + H → OH + O (9b) H2 + O → OH + H (9c) with HO2 + H → OH + OH (9d) ω! 7 = C f , j ,1d1d 2 , OH +M → H +O + M (9e) OH + O → O2 + H (9f) OH + H + M → H2O + M (9g) H2 +OH → H2O + H (9h) H2O + H → H2 + OH (9i) These reactions have been selected from among the 17 H2–O2 elementary reactions listed by Peters (1993). We have chosen these according to the following constraints. First, we have chosen reaction (9a) as the initiation step following discussions in Glassman (1996), and we have attempted to minimize the number of reactions requiring a third body, denoted as M. Second, to the extent possible within the context of constructing a reasonable reaction mechanism, we have tried to impose the constraint that overall consumption nearly equals overall production of species other than reactants (H2 and O2) and product (H2O). With regard to this, we note that this does not imply such a detailed balance actually exists in the context of the resulting DDSs; it is merely an heuristic leading to a well-defined process for selecting the reactions, and other approaches could be used. Next, because we are considering combustion of H2 with N2 dilution, the third body M is taken to be N2. Fourth, the order of evaluation of the equations comprising the DDS model is consistent with the ordering of the above elementary reactions. In particular, we are able to at least in part model a reaction pathway by requiring any intermediate species to already be present before it can be used in subsequent elementary reactions. To construct the DDS corresponding to this reduced mechanism we derive an iterated map for (10) ) d 7(n +1) = − βY7 + γ uY7 a (n +1) + γ vY7 b (n +1) d 7(n ) + ω! 7 + d 7 , 0 , C f , 7,1 = ν 7′′,1 (11) W7 k f ,1 . W1W2 We note that introduction of the species molecular weights arises from the form of Eqs. (5,6). Similarly, the second product of reaction (9a) is atomic hydrogen corresponding to d5. Thus, the DDS is ( ) d5(n +1) = − βY5 + γ uY5 a (n +1) + γ vY5 b (n +1) d 5(n ) + ω! 5 + d 5, 0 , (12) with ω! 5 = C f ,5,1d1d 2 , C f ,5,1 = ν 5′′,1 W5 k f ,1 . W1W2 One can check that of the nine elementary reactions, Eqs. (9), seven lead to two products each, and only two result in a single product. In addition, two reactions contain the third body M whose fluctuations also must be modeled. Hence, there must be a total of 18 iterated maps in the complete DDS for the chemical reactions. These are all obtained as just demonstrated for the first reaction, and for the sake of brevity we will not present details for the remaining ones. There are several items to mention with regard to the overall DDS. First, for reactions involving a third body M, as already noted we have taken that to always be N2. Thus, we must use efficiencies corresponding to this species. But once an efficiency is found, there is still the question of how to incorporate it in the present formulation. There appear to be several possibilities, but the most straightforward is simply to use it as a multiplier on the corresponding specific reaction rate. This is the approach followed for the results reported herein. The second concern regarding the overall system is evaluation of the terms in the thermal energy IECEC 2002 Paper No. 20163 - 4 equation that arise from enthalpies of the chemical species. We note that Hi appearing in Eq. (8c) is just a scaled (due to normalization of Galerkin inner products) instance of the physical enthalpy. Although we include the possibility to input a scale factor in our computer codes, we have always computed with that factor set to unity. Finally, it is important to consider the number of iterations undergone by each map of the overall system. One of the features of the present approach is its ability to directly account for different reaction rates of the individual elementary reactions comprising a kinetic mechanism and relate this to the time scale of the hydrodynamic turbulence. This is done by calculating the Kolmogorov-scale Damköhler number DaK for each elementary reaction using its specific reaction rate to obtain a chemical time scale, and employing an input hydrodynamic turbulence time scale. Then (in the case of DaK >1 which we consider here) for each velocity time step (map iteration) a number of iterations proportional to DaK is performed for each reaction (i.e., for the species maps corresponding to products of that reaction). In fact, in a complete LES employing the present approach, this would be done at each discrete point of the large scale, and for each resolved time step; moreover, the local turbulence time scale would be estimated from the resolved-scale calculations (see Hylin and McDonough, 1999). Here, we will consider only a single spatial location. RESULTS AND DISCUSSION In this section we present results of running the DDSs discussed above. We begin by describing the problem setup; we then present computed results and compare these with a portion of the experimental coflow data, denoted H3, from Meier et al. (1996). We will, for the sake of brevity, consider only one location within the non-premixed 50% N2-diluted H2–air flame of this experiment. This corresponds to an axial location of x/D = 20 with D = 8mm being the H2/N2 jet exit diameter; the measurement location is at a radial distance of 10.5 mm from the jet centerline. DDS Model Problem Setup The full system of Eqs. (8) has been coded in Fortran 90 and for the present study was specialized to equations of the form (11) and (12) corresponding to the nine-step mechanism (9). We have made the reported runs with constant density (i.e., ρ* = 0) to more simply elucidate the ability of the DDS to model effects of turbulence on combustion (and not vice versa). The model is intended to provide instantaneous fluctuations about local in time and space mean temperature and species concentrations; the reported mean quantities from the H3 data set have been employed for this, with some modification. In particular, no minor species concentrations are reported in Meier et al. (1996), so we have slightly altered the mean values of those reported data to allow inclusion of mass fractions of minor species while maintaining the overall average (in time) sum of mass fractions equal to unity. The velocity time scale used in the code to set the Damköhler number was equated to the data sampling 4 frequency of the H3 experiment (10 kHz), and 10 velocity time steps were computed using Eqs. (8a,b). During each such time step a number (≥1) of chemical time steps was computed for each reaction with the number being calculated during each velocity time step using local in time temperature results to set the reaction rate, and thus DaK, for each reaction. Following calculation of all fluctuating species, the fluctuating temperature is updated via Eq. (8c). Clearly, this is not the only possibility for managing the time stepping procedure. Computed Results, Comparisons with Data Figures 1 (experimental, Meier et al., 1996) and 2 (computational, present model) provide a qualitative comparison of our results with those from a wellestablished data base that was selected as a “standard flame” at the International Workshop on Measurements and Computation of Turbulent Nonpremixed Flames, Naples, 1996. We first observe that the qualitative appearance is very similar between corresponding parts of these figures. This demonstrates the general ability of the DDS model to provide physically realistic turbulent fluctuations in a nontrivial situation. We note, however, that with the exception of temperature, the fluctuation amplitudes are not extremely accurate. Considerable effort was made to adjust the bifurcation parameters for temperature because of its direct influence on reaction rates, but this was done only to a lesser extent for the species mass fractions for several reasons. First, our main objective in the present study was to demonstrate the model's ability to mimic physical qualitative behavior; the fluctuation amplitudes in a complete LES would be set by the factors Ai in Eq. (3) described above, while the DDS results would be normalized to unity. Second, the velocity field for the experiments was not available in detail beyond the fact that it corresponds to a 4 Reynolds number of 10 ; so a somewhat arbitrary fluctuating velocity was generated via Eqs. (8a,b). This enters in the advection terms of all remaining equations and clearly has a direct effect as shown in McDonough and Huang (2000), McDonough and Zhang (2002). Finally, we had no data for the minor species H, O, OH, and HO2 about which to center the fluctuations. As noted above, somewhat arbitrary (but small) values for their mass fractions were employed; moreover, it can be expected that a different reduced mechanism might also alter results somewhat. All of these factors can lead to inaccuracies, but if the DDS behavior is stable they should have only relatively minor effects on qualitative computed results. Comparison of Figs. 1 and 2 suggests this is the case. In particular, it can be seen that the general features of the physical turbulent fluctuations are preserved in each of the cases of model results IECEC 2002 Paper No. 20163 - 5 although there appears to be a mild discrepancy in oscillation frequency of H2O concentration. FIGURE 1: EXPERIMENTAL DATA, TEMPERATURE AND MASS FRACTIONS; (a) TEMPERATURE, K, (b) H2, (c) O2, (d) H2O. are considering only a single location of the experimental flow field data in this brief communication we cannot establish whether the current model exhibits improved behavior in this respect, but we note that the data location was chosen a priori, and results presented herein are not a “best case” (nor are they necessarily a worst case); they are, we believe, representative. As already indicated, we have employed the mean quantities provided with data set H3 as reference levels about which fluctuations produced by the DDS model would take place. However, because of the spatial sparseness of these data it was not possible to perform the usual high-pass filtering (as would be done in our complete LES implementations) in order to directly obtain the di,0s of Eqs. (8d). In fact, as can be seen from Fig. 2 and Table 1 we did not accurately match the mean values for H2 and O2 mass fractions. On the other hand, the computed mean values for temperature and H2O concentration are reasonably accurate ― probably to within experimental discrepancies. Table 1 summarizes these results. As can be seen, the amplitudes for rms fluctuations produced by the model are generally low compared with the experimental observations; but as already emphasized, it is the temporal qualitative behavior that is more important for the factors Mi of the overall model. TABLE 1. MEAN AND RMS VALUES FOR TEMPERATURE AND MAJOR SPECIES MASS FRACTIONS T,K YH 2 YO2 YH 2 O YN 2 FIGURE 2: DDS MODEL RESULTS; (a) THROUGH (d), SAME AS IN FIG. 1. One of the main requirements of any turbulence model is to produce reasonably accurate statistics. Indeed, most models are not actually able to do this globally throughout a flow field; they may be accurate for some quantities at some locations, but never accurate for all quantities at all locations. Because we Experi- Mean ment rms 1639.5 0.0054 0.0376 0.1372 0.8198 288.0 0.0061 0.0500 0.0330 0.0220 DDS Mean Model rms 1548.1 0.0028 0.0213 0.1349 0.8190 252.0 0.0026 0.0221 0.0083 0.0033 One additional output from the model that we have monitored is the sum of mass fractions. We do not force this to be unity by calculating all but one species and then setting it to satisfy the required unity value. Rather, we directly calculate all species and then utilize the sum of their mass fractions as a measure of the success of the model simulations. For the runs presented here the time average of the sum of the mass fractions was 0.983, showing less than 2% discrepancy from the required value. Moreover, the instantaneous deviation from unity over the run of 4 10 time steps never exceeded ±6%. FURTHER DISCUSSION The results we have presented and the manner in which they have been calculated are somewhat different from those of usual LES SGS modeling techniques. There are some current models that are at least somewhat similar, and these should be briefly IECEC 2002 Paper No. 20163 - 6 compared. The two approaches that seem closest to the present one are those of Kerstein and various coworkers (see e.g., Kerstein, 1992 and Echekki et al., 2001), viz., the linear eddy models (LEMs) and the more recent “one-dimensional turbulence” (ODT) models, and those of Giacomazzi et al. (1999, 2000). But it should be noted that the former of these is mainly phenomenological and ultimately driven by a random number generator; moreover, physical turbulence is not one dimensional. The latter approach, while incorporating some of the ideas we employ to evaluate Ai and ζi (not carried out in the present study) of Eq. (3), nevertheless utilizes averaged or filtered equations of motion, and an eddy viscosity. The method we have presented, although two dimensional as given herein, extends to three dimensions in a straightforward way. It employs approximations of the physical equations themselves on the subgrid scales to produce turbulent fluctuations that are completely deterministic, just as would be the case in a DNS. Moreover, the equations are neither averaged nor filtered; problems associated with closure of nonlinear terms are thus greatly reduced. CONCLUSIONS In this paper we have presented a discrete dynamical system model of a nine-step reduced mechanism for H2–air combustion derived directly from the governing equations which we propose for use as the temporal fluctuations in subgrid-scale models for large-eddy simulation of the type studied in McDonough et al. (1995), Hylin and McDonough (1999), Sagaut (2001). We have demonstrated that the fluctuations produced by this DDS are very similar to those observed in well-known experimental data (Meier, 1996) in a qualitative sense, but that turbulence statistics, especially second-order statistics, are not very accurate. We have, however, noted reasons for which this might be expected; viz., the conditions under which the DDS was run for the present study were lacking information that would generally be available in a complete LES. It is thus difficult to offer firm conclusions beyond noting that the qualitative behavior of the DDS is strikingly similar to that of all aspects of the experimental data for an arbitrarily chosen measurement location. We believe this is ample justification for more thorough studies of this formalism for constructing SGS models for LES. Moreover, as is the case for the fractal modeling procedure reported in Giacomazzi et al. (1999, 2000), it would in principle be possible to incorporate the present approach into RANS modeling methods, although we have not as yet attempted this. ACKNOWLEDGMENTS The authors gratefully acknowledge the support for these studies provided in part by AFOSR Grant #F49620-00-1-0258 and NASA/EPSCoR Grant #WKU 522635-00-10. REFERENCES Bohr, T., Jensen, M. H., Paladin, G. and Vulpiani, A., Dynamical Systems Approach to Turbulence, Cambridge University Press, Cambridge, 1998, pp. 78–88. Cook, A. W., Riley, J. J. and Kosály, G., Combust. Flame 109:332–341 (1997). Cook, A. W. and Riley, J. J., Phys. Fluids 6:2868– 2870 (1994). Echekki, T., Kerstein, A. R. and Dreeben, T. D., Combustion and Flame 125:1083–1105 (2001). Frisch, U., TURBULENCE The Legacy of A. N. Kolmogorov, Cambridge University Press, Cambridge, 1995, pp. 89–92. Germano, M., Maffio, A., Sello, S. and Mariotti, G., in Direct and Large Eddy Simulation II (Collet et al., eds), Kluwer Academic Publishers, Amsterdam, 1997, pp. 291–300. Germano, M., Piomelli, U., Moin, P. and Cabot, W. H., Phys. Fluids A3:1760–1765 (1991). Giacomazzi, E., Bruno, C. and Favini, B., Combust. Theory Modelling 3:637–655 (1999). Giacomazzi, E., Bruno, C. and Favini, B., Combust. Theory Modelling 4:391–412 (2000). Glassman, I., Combustion, Academic Press, San Diego, 1996, pp. 65–72. Gouldin, F. C., Hilton, S. M. and Lamb, T., Proc. Combust. Inst. 22:541–550 (1988). Gouldin, F. C., Combust. Flame 68:249–266 (1987). Hylin, E. C. and McDonough, J. M., Int. J. Fluid Mech. Res. 26:539–567 (1999). Kerstein, A. R., Comb. Sci. and Tech. 81:57–96 (1992). May, R. M., Nature 261:459–467 (1976). Meier, W., Prucker, S., Cao, M.-H. and Stricker, W., Combust. Sci. Technol. 118:293, (1996). McDonough, J. M., Yang, Y. and Hylin, E. C., in Proc. First Asian Comput. Fluid Dyn. Conf. (Hui et al., eds), Hong Kong University of Science and Technology: Hong Kong, 1995, pp. 747–752. McDonough, J. M. and Huang, M. T., submitted to Int. J. Numer. Meth. Fluids (2001). McDonough, J. M. and Huang, M. T., paper ISSM3-E8 in Proceedings of Third Int. Symp. on Scale Modeling, Nagoya, Japan, Sept. 10–13, 2000. McDonough, J. M. and Joyce, D. L., to be th presented at 8 AIAA Thermophysics and Heat Transfer Conference, St. Louis, June 24–27, 2002. McDonough, J. M. and Zhang, Sha, to be nd presented at 32 AIAA Fluid Dynamics Conference, St. Louis, June 24–27, 2002. Peters, N., in Reduced Kinetic Mechanisms for Applications in Combustion Systems (Peters and Rogg, eds), Lecture Notes in Physics m15, SpringerVerlag, Berlin, 1993, pp. 8–12. Peters, N., Turbulent Combustion, Cambridge University Press, Cambridge, 2000, pp. 57–65. Pope, S. B., Prog. Energy Combust. Sci. 11:119– 192 (1985). Sagaut, P., Large Eddy Simulation for Incompressible Flows, Springer, Berlin, 2001, pp. 191–194. IECEC 2002 Paper No. 20163 - 7