A ‘Synthetic Scalar’ Subgrid-Scale Model for Large-Eddy Simulation of Turbulent Combustion J. M. McDonough∗ Departments of Mechanical Engineering and Mathematics University of Kentucky, Lexington, KY 40506-0108 Abstract In this paper we present an alternative formalism for large-eddy simulation that avoids the problems arising from filtering the governing equations. In particular, we instead propose to filter solutions leading to the ability to directly model subgrid-scale primitive variables; i.e., model physics rather than statistics. We then demonstrate construction of subgrid-scale models of scalar quantities within this framework. Introduction It is by now widely accepted that large-eddy simulation (LES) in some form probably offers the best hope of solving practical engineering problems involving turbulent flows in the near to intermediate future. Direct numerical simulation (DNS) is useful for theoretical studies at low Reynolds number, Re, but even for simple incompressible flows in the absence of scalar transport the total arithmetic requirements scale as Re3 . Such calculations are not feasible in engineering practice on current or even near-term foreseeable machines when “over-night turnaround” is needed unless Re is less than ∼ O(104 ). At the same time, Reynoldsaveraged Navier–Stokes (RANS) procedures have repeatedly exhibited an inability to be truly predictive (again, even for relatively simple flows). So although they are currently widely used, it is recognized that they embody fundamental flaws, and it is only a lack of viable alternatives that leads to their present ubiquitous application. It can easily be argued that the structure of LES implies a total arithmetic requirement for incompressible flow that is roughly proportional to Re2 . This suggests that in the context of currently-available parallel supercomputers simulation of flows with Re at least as large as 106 should be nearly feasible. On the other hand, LES possesses certain shortcomings that will need to be remedied before it will provide a completely reliable practical simulation tool. In particular, it has been understood almost from the earliest studies of LES that both the equation filtering process and the subgrid-scale (SGS) models lead to significant problems (see Ferziger [1] for an early review, and Lesieur and Métais [2] for more recent discussions). Indeed, as might be expected, these problems are exacerbated by the need to treat detailed chemistry of combustion processes, in part explaining a somewhat heavier emphasis placed on DNS in combustion than elsewhere (see, e.g., Vervisch and Poinsot [3]). Filtering the governing equations results in the same types of closure problems encountered with RANS, and these are often treated similarly in LES. But unlike RANS, the basic structure of LES is such that its results do converge to those of the Navier– Stokes (N.–S.) equations as discretization step sizes are refined. This is its significant advantage over RANS. However, within the framework of the usual LES formalism it has been difficult to obtain high-quality results unless resolution is approaching that of DNS, precluding its use for practical problems. In the present paper we provide an alternative formalism that specifically addresses these issues. We remark that the ideas to be presented are not completely new, and much of what we will describe has evolved from work first presented by McDonough et al. [4,5], McDonough and Bywater [6,7], McDonough et al. [8] and Hylin and McDonough [9]. It is the last two of these that are most directly related to the present work, and this approach has recently been concisely summarized by Sagaut [10]. The technique we discuss here arises from a somewhat more general formalism than LES (and includes LES as a special case), which we have typically described as additive turbulent decomposition (ATD), and which has been subjected to rigorous mathematical analysis by Brown et al. [11]. References [8–10] describe what we believe is the most useful form of ATD for practical calculations, and in general such procedures embody three main ideas that distinguish them from typical LES: i) filter solutions rather than equations, ii) model primitive variables (e.g., temperature, species concentrations, etc.—hence the term “synthetic scalar”—in the present case), instead of modeling their correlations, and iii) directly use these * Corresponding author: jmmcd@uky.edu Proceedings of the 2002 Technical Meeting of the Central States Section of the Combustion Institute modeled results to impart improved resolution to the (deliberately, by the nature of LES) under-resolved large-scale part of the solution. We comment that all of these have been utilized together beginning with the work reported in [8], but formal analysis and further development is still ongoing. In the present study we focus on ii), specifically the SGS models, per se, and we restrict the analysis to passive scalars. Details for incompressible velocity fields can be found in [9] and [10]. The remainder of this paper is organized as follows. In the next section we provide a brief overview of all three of the ideas mentioned above to establish the setting into which the SGS models will be incorporated. We follow this with a section devoted to derivation of the synthetic scalar models, and end the paper with a short summary. Overview of ATD Structure In this section we will discuss the three abovementioned notions that distinguish the current method from usual LES, emphasizing advantages of this approach for combustion simulations. The first difference between the proposed approach and LES as usually constructed is choosing to filter solutions rather than the governing equations. Filtering equations presents no difficulties for linear terms, but it is well known that averaging and filtering operators do not commute with nonlinear operators of a typical transport equation. In the context of combustion, probably the best-known example of this is the Arrhenius reaction rate formula for which it is clear that µ ¶ µ ¶ E E α α T exp 6= T exp , (1) R0 T R0 T the consequence (and treatment) of which is discussed in, e.g., Liñán and Williams [12]. Here, overbars denote application of a filtering (or averaging) operator. T is temperature; E is activation energy; R0 is the universal gas constant, and α is an exponent which is given for each reaction (as is E). It is clear that this is just the beginning of troubles arising from filtering, and this has been one of the main motivations for development of the PDF methods by Pope [13] and others. An additional difficulty arising from filtering is that of treating problems posed on complicated geometric domains. If structured grids are to be used, the differential operators will now contain variable coefficients that in general prevent commutation of the filtering operator with even linear differential operators. If unstructured grids are employed, special grid-dependent filters must be constructed. Even in the absence of the above geometric problems, the consequence of Eq. (1) is that statistical correlations representing the commutators of the differential and filtering operators must be included in the governing equations. For example, a somewhat simpler-to-treat case than that of Eq. (1) is ∂uρYi ∂u ρ Yi 6= , ∂x ∂x (2) which requires introduction of the correlation ρuYi , in which ρ is density, u is the x-component of velocity, and Yi is mass fraction of species i. It is essential to recognize that there are no fundamental equations for these correlations, and that modeling such quantities entails modeling statistics rather than physics. Such models can tell us essentially nothing regarding the instant-to-instant behavior of a turbulent combustion process on the scales where chemical reactions occur and interact with the turbulence. The alternative we propose is to avoid filtering the governing equations. This removes the need to model statistics because no correlations arise, and it permits direct modeling of physics. Furthermore, if the solution itself is being filtered, the filtering process is reduced to one of signal processing. On structured grids this is an almost trivial process although some difficulties can arise when unstructured grids are employed. For the sake of brevity we will not provide further details regarding this here; the interested reader is referred to Yang and McDonough [14] for more information on filtering solutions. The form of subgrid-scale model proposed in [8] and treated in detail in [9] is qi∗ = Ai ζi Mi , (3) where qi∗ is the ith component of the small-scale (SGS) part of a solution vector Q ≡ (Q1 , Q2 , . . . , QM )T of M dependent variables that has undergone the usual LES decomposition Qi (x, t) = qi (x, t)+qi∗ (x, t) , i = 1, 2, . . . , M . (4) Here, qi is the large- or resolved-scale part that is directly computed via some discrete approximation, and qi∗ is the unresolved, small-scale part; x is the spatial coordinate, and t is time. We can (informally, but not rigorously) view qi as the first terms of a Fourier representation of Qi , with the number of terms depending on the level of resolution, and qi∗ can be considered to be the remainder. This is constructed to directly model primitive variables at the level of the unresolved scales. In Eq. (3) Ai is the amplitude of the ith SGS solution component; this is derived from Kolmogorov scalings (see Frisch [15] for a basic treatment, [9] for details in the present context and Giacomazzi et al. [16] for a somewhat similar treatment). ζi is an anisotropy correction, the need for which arises from the isotropy assumption of the Kolmogorov theories, and the Mi are temporal fluctuations. The latter have received considerable attention recently after it was demonstrated by McDonough and Huang [17,18] that they could be obtained as discrete dynamical systems (DDSs) derived directly from essentially any governing equation(s). Examples of this specifically related to combustion include [17], McDonough and Zhang [19,20] and Zhang et al. [21]. The final aspect of the modeling approach under consideration here is direct use of model results. Any transport equation can be expressed in the general form above, it is now known how to calculate the Mi s directly from the governing equations, and we shall see that the new formulation for the Ai s given here for the first time obviates the need for the ζi s. Thus, under the assumptions we will make as we proceed, we will produce a complete SGS model of passive scalars for LES of finite-rate kinetics descriptions of combustion. The basic idea underlying construction of formulas for the Ai s is that qi∗ represents behavior not resolved in the large-scale calculation, and thus the corresponding amplitude should be the square root of some measure of the “energy” summed over all wavenumbers higher than the highest one supported by the largescale discretization. We denote this wavenumber by kh for a grid of spacing h. Then Qt + ∇·F (Q) = ∇·G(Q) + S(Q) , k=kh +1 A2i ∼ = (5) with Q as given above. The subscript t denotes partial differentiation with respect to time, and ∇· is the divergence operator. The functions F and G are associated with, respectively, advection and diffusion. S is a general nonlinear source term which in the present case corresponds to species production and the resulting heat release. In usual LES Eq. (5) would be filtered, and (4) would be used along with models of terms related to (1) and (2) to produce equations for the qi s. Thus, except for the adjustment of qi due to these models, information associated with the qi∗ s would be lost. No direct effect of small-scale turbulent fluctuations can be imparted to the large-scale part of the solution, and it is precisely on the small scales that important interactions between turbulence and combustion occur. In the modeling technique treated here, we substitute Eq. (4) into (5) to obtain (q + q ∗ )t + ∇·F (q + q ∗ ) = ∇ · G(q + q ∗ ) + S(q + q ∗ ) , (6) similar to usual LES—but we do not filter this equation. Instead, we use (3) to evaluate q ∗ , substitute this into (6), and solve for q. This result will be under resolved, and probably aliased; so we filter it, that is, filter the solution q. Then we again use the qi∗ s in Eq. (4) with the filtered qi s to obtain an approximation to the complete solution Q. This sequence of steps is executed for each discrete time step of the numerical integration of Eq. (5). The Primitive-Variable SGS Model In the remainder of this paper we will concentrate on deriving formulas for the Ai s of Eq. (3). As noted ∞ X Ei (k) , (7) where Ei is the abovementioned characterization of energy associated with the ith dependent variable. We remark that Eq. (7) is actually global in space because it has arisen from Fourier analysis of the energy. Namely, by the Parseval identity, (7) is simply the square of the (spatial) L2 norm of the small-scale part of the ith dependent variable. But in applications we will make of Ai , this value must be local. If we consider a local (re-)expansion of, say Yi (= yi + yi∗ ), on the interval [0, 2h] (or any translate of this interval) then wavenumbers of this expansion can be related to those of the global expansion through the scaling N kh where N is the number of points used in discretizing the physical domain. Thus, the size of the local wavenumbers increases more rapidly than does that of the global ones, and on a 2h-interval we should not need very many wavenumbers to adequately represent yi∗ . If we denote these local wavenumbers as knloc , we have that the lowest wavenumber in the Fourier representation of yi∗ should be k0loc = kh , and these are related to global wavenumbers via knloc = nN kh . If we denote restriction of Ei to [0, 2h] by Ei∗ , we have A2i ∼ = ∞ X Ei∗ (knloc ) . (8) n=1 As noted above, we expect to need only a few terms in this series to obtain an adequate representation because knloc increases very rapidly with n. Our next task is to determine the Ei∗ s. There are several ways by which this might be done, and we will present only one herein. The present approach will be based on the second-order structure function arising in the Kolmogorov-Oboukov-Corrsin (KOC) treatment of passive scalars, as described by Warhaft [22], and it will utilize a form of the scale similarity hypothesis widely used in dynamic SGS models (see Germano et al. [23] and Domaradzki and Saiki [24]). We will not here give details of this latter aspect of the model beyond explaining the manner in which it enters. The second-order structure function of a scalar yi∗ is defined as S2 (yi∗ , `) = h(δyi∗ )2 i ≡ h(yi∗ (x + `) − yi∗ (x))2 i , (9) where h · i denotes a chosen form of averaging (local spatial in the present case), and ` is a spatial displacement from the point x. Here, we will average over a selected number of points in the `-neighborhood of x. For example, on a uniform finite-difference grid in 2D we would define this q neighborhood as those points such that ` ≡ |`| = h2x + h2y , where hx and hy are grid spacings in the indicated directions. Then for any given point x we calculate (δyi∗ )2 for each of the nearest neighbors at a distance ` from this point and average the results. It is clear from (9) that S2 is directly related to the energy of yi∗ . Indeed, for high-Re, turbulence the wellknown k −5/3 inertial subrange scaling follows directly from this and the Kolmogorov 23 -law (see [15]). Then the KOC theory gives 2/3 S2 (yi∗ , `) = C2,Yi ε−1/3 εYi `2/3 , (10) where C2,Yi is a universal constant, ε is dissipation rate of turbulence kinetic energy, and εYi is dissipation rate of the scalar (in this case, mass fraction Yi ). We have and ε = νk∇u∗ k2 , (11) εYi = DYi k∇yi∗ k2 (12) for the dissipation rates. In (11) u∗ is the small-scale velocity vector, and ν is kinematic viscosity; in (12) DYi is molecular mass diffusivity of the species whose mass fraction is Yi . Several remarks regarding these formulas are in order. First, Eq. (10) holds for high-Re flows in the inertial-convective subrange with respect to the scalar, i.e., in an inertial range in which the scalar is merely convected. There are other inertial subranges (see, e.g., Tennekes and Lumley [25]), so (10) is not general. On the other hand, Eqs. (11) and (12) are general, and simpler forms are widely used. Furthermore, in combustion studies εYi is usually expressed in terms of mixture fraction. To obtain a suitable formula for the Ai s under general flow conditions we will work locally to remove needed assumptions associated with homogeneity, calculate separate Ai s for every variable (especially, the velocity components) to remove the need for isotropy, and use the expression S2 (yi∗ , `) = C2,Yi ε−α (εYi `) β (13) obtained as a generalization of (10). In this formula α will have been determined from an independent analysis of the velocity field similar to what we present here for passive scalars. It is important to note that ε and εYi contain small-scale information, and this is not immediately available—it is what is being modeled. So to estimate these quantities we will employ the scale similarity hypothesis as indicated above. Scale similarity utilizes the assumption that flow and scalar behavior at the lowest modes of the small, unresolved scales is the same as that of the highest modes of the resolved scales. In dynamic SGS models [23] use of this idea is embodied in a double filtering procedure that ultimately leads to local in space and time values for the Smagorinsky constant. In [24] scale similarity is used to “estimate” subgrid-scale synthetic velocities similar to what we do here. But it is important to recognize the fundamental flaw in this reasoning. While the indicated relationship between behaviors on opposite sides of the maximum grid-scale wavenumber is reasonable, applications of this in highly under-resolved regions of the flow do not provide the correct information. Namely, at high Re the small-scale behavior predicted by scale similarity might not even correspond to that in the inertial subrange; yet, it would be employed as if it were at the Kolmogorov scales. This problem is addressed by Domaradzki and various co-workers, e.g., [24] and Domaradzki and Loh [26]. The basic idea is to employ an extrapolation of wavenumber information to inexpensively gain higher resolution for application of scale similarity. In [26] this is modified to utilize interpolation in physical space. The methods employed in [24], [26] and elsewhere provide only a factor of two increase in wavenumbers in each spatial direction. Hence, at very high Re, or from the standpoint of scalars, high Péclet number, P e, this will not be sufficient. Moreover, because the high-wavenumber information produced in this way is not actually a solution of the governing equations, recursive application of this is not possible. The present author and co-workers have recently developed a physical-space interpolation procedure based on the discrete solution operator of the governing equations which produces interpolants on any desired scale that are solutions to the discrete governing equations (see Xu et al. [27]). This permits local arb- itrarily high-wavenumber representations (exactly what is needed in the present context), and thus valid application of the scale similarity approach. Hence, we can suppose that (11) and (12) are approximated with reasonable accuracy. Then calculation of S2 (yi∗ , `) as a function of ` at several values of ` using these approximate values in the right-hand side of (9) provides sufficient information to determine C2,Yi and β in (13). We then directly calculate the required energy ¢β−1 ¡ EYi (k loc ) = C2,Yi ε−α εβYi k loc , (14) and this is used in (8) to predict AYi . There are several important observations to be made regarding use of Eq. (14) in (8) to obtain AYi . The first is that this must be done for every dependent variable: all species concentrations, temperature and the velocity components. Second, it is done at every large-scale grid point during every time step, except when local subgrid-scale Reynolds and/or Péclet numbers indicate that turbulence cannot be present (see [9] for details of this). Despite this, the amount of arithmetic is not excessive. Indeed, it scales as O(N ) where N is the total number of grid points of the largescale calculation. But in addition, these computations are highly parallelizable. Our experience with earlier implementations of models with similar structure has shown that even without parallelization, run times for executions utilizing the turbulence model do not exceed those for laminar flow on the same grid by more than a factor of two. Yet, the details seen in the fluctuating flow fields approach those of DNS. Finally, we note that there are no adjustable constants in the construction of the amplitude factors. This differs from the model presented in [9]. We have already indicated that the fluctuating factors, the Mi s, of Eq. (3) have received considerable attention in recent studies, but for the sake of completeness we will provide a brief overview of this important aspect of the SGS models; more details are available in the cited references. We noted above that these coupled DDSs can be derived from essentially any governing equations. This is done via a Galerkin procedure applied locally in space (in the same spirit as construction of the Ai s), and thus with the same expectation that only a few terms in a Fourier representation will be needed to accurately represent small-scale temporal behavior. In fact, Yang et al. [28] have shown that a linear combination of only three realizations of the DDS modeling the 2-D N.–S. equations [18] is sufficient to provide a very good model of an experimental turbulent flow field with Re = 105 . Furthermore, in [19] and [20] it is shown that a single realization of the complete DDS presented below is sufficient to closely mimic temporal behavior of major species mass fraction fluctuations in H2 –O2 and H2 –air, respectively, reactions. Here we will present the general form of this DDS and indicate how it is incorporated into the overall model of the qi∗ s. As is fairly common, we assume low Mach number and require the hydrodynamic pressure to be divergence free. Then we recover density from an equation of state. Within this context the DDS leading to the fluctuations Mi is the following. ³ ´ a(n+1) = βu a(n) 1 − a(n) − γu a(n) b(n) , (15a) ³ ´ b(n+1) = βv b(n) 1 − b(n) − γv a(n) b(n) + αT c(n) , c(n+1) = "Ã N s X (15b) (n+1) αT di di − γuT a(n+1) − i=1 (n+1) γvT b ´ (n) c − Ns X # Hi ω̇i /(1 + βT ) + c0 , i=1 (n+1) di ³ (n+1) = − βYi + γuYi a + γvYi b (n+1) ´ (15c) (n) di + ω̇i + di,0 i = 1, 2, . . . , Ns , (15d) with ω̇i = Nr X j=1 " Cf,ij Ns Y `=1 ν0 d` j,` − Cb,ij Ns Y # ν 00 d` j,` . `=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 αs, βs, 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; αT di 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 Kolmogorov-scale Damköhler numbers, DaK . The various γs 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 di,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. For DaK > 1, the case considered in [17,19–21], an iteration step is performed with Eqs. (15a,b) to obtain fluctuating velocity components. Then Eqs. (15d) are iterated a number of times based on their respective Damköhler numbers, and finally Eq. (15c) is evaluated. This process is repeated a number of times determined by the ratio of integral-scale to Kolmogorovscale times, tL /tK , for the velocity field. The results from the final such step are normalized by the maximum value over all steps calculated during the current large-scale time step, and these are inserted into Eqs. (3) as the Mi s. We remark that DaK ≤ 1 can also be treated, and Eqs. (15) extend to three space dimensions in a natural way. Summary In this paper we have presented a derivation of a SGS model for LES that represents a rather distinct departure from previous approaches, and we have indicated some of its potential advantages with regard to combustion simulations. Among these are its ability to directly simulate fluctuating behavior of physical variables, and thus interaction of turbulence with chemical kinetics on appropriate length and time scales, providing a distinct advantage over RANS methods and usual LES. Acknowledgements The author expresses his gratitude to the AFOSR for partial support of this work and to NASA/EPSCoR for the funding that permitted initiation of these studies. References 1. Ferziger, J. H., in Theoretical Approaches to Turbulence, Dwoyer et al. (eds), Springer-Verlag, New York, 1985, pp. 51–72. 2. Lesieur, M. and Métais, O., Annu. Rev. Fluid Mech. 28:45–82 (1996). 3. Vervisch, L. and Poinsot, T., Annu. Rev. Fluid Mech. 30:655–691 (1998). 4. McDonough, J. M., Bywater, R. J. and Buell, J. C., presented at AIAA 17 th Fluid Dynamics, Plasmadynamics and Lasers Conf., Snowmass, CO, June 24–28, 1984. 5. McDonough, J. M., Buell, J. C. and Bywater, R. J., ASME Paper 84-WA/HT-16, presented at ASME Winter Annual Meeting, New Orleans, LA, Dec. 9–14, 1984. 6. McDonough, J. M. and Bywater, R. J., AIAA J., Vol. 24, 1986, pp. 1924–1930. 7. McDonough, J. M. and Bywater, R. J., in Forum on Turbulent Flows – 1989, Bower & Morris (eds), ASME-FED Vol. 76, ASME, New York, 1989, pp. 7–12. 8. McDonough, J. M., Yang, Y. and Hylin, E. C., in Proc. 1st Asian Comput. Fluid Dyn. Conf. (Hui et al., eds), Hong Kong Univ. of Sci. and Tech. Hong Kong, 1995, pp. 747–752. 9. Hylin, E. C. and McDonough, J. M., Int. J. Fluid Mech. Res. 26:539–567 (1999). 10. Sagaut, P., Large Eddy Simulation for Incompressible Flows, Springer, Berlin, 2001, pp. 191–194. 11. Brown, R. M., Perry, P. and Shen, Z., SIAM Appl. Math. 59:139–155, (1998). 12. Liñán, A. and Williams, F. A., Fundamental Aspects of Combustion, Oxford University Press, Oxford, 1993, pp. 141–142. 13. Pope, S. B., Prog. Energy Combust. Sci. 11:119– 192 (1985). 14. Yang, T. and McDonough, J. M., in preparation, (2002). 15. Frisch, U., TURBULENCE The Legacy of A. N. Kolmogorov, Cambridge University Press, Cambridge, 1995, pp. 89–92. 16. Giacomazzi, E., Bruno, C. and Favini, B., Combust. Theory Modelling 4:391–412 (2000). 17. McDonough, J. M. and Huang, M. T., paper ISSM3-E8 in Proc. of 3rd Int. Symp. on Scale Modeling, Nagoya, Japan, Sept. 10–13, 2000. 18. McDonough, J. M. and Huang, M. T., submitted to Int. J. Numer. Meth. Fluids (2001). 19. McDonough, J. M. and Zhang, Sha, submitted for presentation at 32nd AIAA Fluid Dynamics Conf., St. Louis, June 24–27, 2002. 20. McDonough, J. M. and Zhang, Sha, submitted for presentation at 29 th Int’l Symp. on Combust., Sapporo, Japan, July 21–26, 2002. 21. Zhang, Sha, Slade, Jeremy D. and McDonough, J. M., to be presented at 2002 Tech. Mtg. of Central States Sec. of Combust. Inst., Knoxville, April 7– 9, 2002. 22. Warhaft, Z., Annu. Rev. Fluid Mech. 32:203–240 (2000). 23. Germano, M., Piomelli, U., Moin, P. and Cabot, W. H., Phys. Fluids A 3:1760–1765 (1991). 24. Domaradzki, J. A. and Saiki, E. M., Phys. Fluids 9:2148–2164 (1997). 25. Tennekes, H. and Lumley, J. L., A First Course in Turbulence, MIT Press, Cambridge, MA, 1972, pp. 282–286. 26. Domaradzki, J. A. and Loh, K. C., Phys. Fluids 11:2330–2343 (1999). 27. Xu, Ying, Yang, T., McDonough, J. M. and Tagavi, K. A., submitted for presentation at 8th AIAA/ASME Joint Physics and Heat Transfer Conference, St. Louis, June 24–27, 2002. 28. Yang T., McDonough, J. M. and Jacob, J. D., submitted to AIAA J. (2001).