MODELING DROPLET HEAT AND MASS TRANSFER IN SPRAYS: PROGRESS AND CHALLENGES Shankar Subramaniam Department of Mechanical Engineering Iowa State University, Ames, Iowa, USA The predictive capability of CFD calculations of sprays depends on many factors, principally: (i) adequacy of the underlying mathematical formulation, (ii) accuracy of the closure relations, especially those for the physical sub-models representing droplet heat and mass transfer, (iii) numerical accuracy of the solutions to the model equations, and (iv) the availability of experimental data or analytical solutions for comparison and validation. Outstanding challenges and recent progress in each of these areas are discussed. These recent developments motivate an integrated program for predictive spray model development, which is outlined in this work. INTRODUCTION The ultimate goal of spray modeling is to capture those essential, and hopefully universal, principles of spray dynamics which can successfully represent phenomena that are observed in the wide variety of spray applications, including: automobile fuel sprays, agricultural sprays, and pharmaceutical sprays for tablet coating and inhalation drug delivery. The underlying theme of this paper is that progress in four principal areas, namely (i) adequacy of the underlying mathematical formulation, (ii) accuracy of the closure relations, especially those for the physical sub-models representing droplet heat and mass transfer, (iii) numerical accuracy of the solutions to the model equations, and (iv) validation of models in canonical problems with analytical solutions or experimental data, is critical for the development of predictive spray models. Of all the modeling and simulation approaches that are currently in use, the statistical approach used in CFD is the most useful for complete spray calculations. This statement is justified by considering the feasibility of two other simulation approaches that are recently being applied to the spray problem, namely: Direct Numerical Simulation (DNS) and Large Eddy Simulation (LES). DNS is currently not a practical approach for complete spray calculations for two principal reasons. The first reason is that DNS of even fully dispersed droplets in a turbulent gas involves the physical processes of droplet coalescence and breakup which impose a dynamically changing resolution requirement that is considerably more demanding than simply resolving all the continuum scales in single-phase turbulence. In fact molecular dynamics studies of droplet collision1 show that there are sub-continuum scale physics can affect the outcome of collisions, thereby indicating that DNS of even dispersed two-phase flows is not a model-free simulation. However, DNS of certain canonical twophase flows can be useful in developing statistical models. The second reason why DNS of full sprays is currently impractical is because simulation of practical spray systems involves a primary atomization process. A rough estimate1 of the computational cost involved in performing DNS of just the turbulent liquid pipe flow in the nozzle prior to atomization indicates that such simulations are about 20 years in the future, with current growth in computing power. The scaling of grid points (N) DNS of turbulent pipe flow with Reynolds number for L/D=5 given by Friedrich et al2 is N=0.004826 Re21/8, and Wu and Faeth3 report turbulent pipe flow Reynolds numbers for spray atomization experiments in the range of 8400 to 850,000. Currently DNS of turbulent pipe flow for Reynolds numbers up to 8900 are reported in the literature. The scaling estimate for the number of grid points required for a DNS reveals that a hundred-fold increase in Reynolds number would require the number of grid points to increase by a factor of approximately 175000. Even if Moore’s law continues to hold, and we assume computing power doubles every year, this leads to the conclusion that DNS of realistic turbulent pipe flow will be feasible in about 17 years. DNS of primary breakup is more computationally demanding because the resolution requirement will increase with interface pinch-off and formation of small drops. The LES approach for simulation of single-phase turbulent flow was developed on the basis of two important assumptions concerning high Reynolds number turbulent flow. One is the Kolmogorov hypothesis of local isotropy, which states that at sufficiently high Reynolds number the small-scale turbulent motions are statistically isotropic4. The second is the fact that most of the energy in a singlephase turbulent flow is contained in the large scales, and a simple model for momentum transfer between the large and small scales can perhaps describe the evolution of the energy containing (largeeddy) motions. The assumption of local isotropy is borne out by experimental observation of high Reynolds number single-phase turbulent flow4, 5. Although not all sub-grid scale models are simple, it is now established that LES is useful in simulating single-phase turbulence. However, neither of these assumptions is shown to be valid for dispersed two-phase flows. LES formulations are not rigorously developed for two-phase flows, and it appears that significant enhancements or changes to the standard single-phase LES approach will be needed before it can be applied to two-phase flows. The statistical approach on the other hand, is useful in its own right as a basis for CFD calculations of sprays, and its development can also be useful in interpreting and using DNS results. Of the statistical formulations used for spray modeling, the Lagrangian-Eulerian approach based on Williams’ spray equation is the most promising for developing predictive spray models6. The underlying physics always dictate the mathematical representation that must be chosen to model any problem. For instance, while second-moment closures (RANS models) are often deemed adequate models for single-phase turbulent flows, a higher level of representation based on the velocity-composition probability density function (PDF) is used for turbulent reactive flows7. This is because a highly nonlinear dependence of the reaction rate on composition (mass fraction of chemical species and enthalpy) implies that the mean reaction rate is not well approximated by the reaction rate function evaluated at mean composition. Similarly, we need to examine what are the essential physical phenomena that determine spray dynamics, and how these may differ from singlephase turbulent flows. The first noteworthy difference is that in two-phase flows such as sprays, the essential dynamics is governed by non-local effectsa). For example, the acceleration of a spray droplet’s center-of-mass is determined by the pressure and viscous shear exerted by the ambient gas on the droplet surface, and the same is true for the droplet’s thermal and vaporization dynamics. Another unique feature of sprays is the presence of drop-drop interactions, which we expect to become more apparent at higher volume fractions. a) It is true that non-local effects due to pressure also arise in single-phase turbulent flows, but pressure is notoriously difficult to account for in purely Lagrangian PDF methods. These “action-at-a-distance” phenomena are difficult to capture in traditional single-point PDF models, which are appropriate for single-phase turbulent reactive flows2. Therefore, if the singlepoint PDF representation is directly applied to spray problems, it is not going to address the issue of capturing the essential physics. However, a multiphase PDF formulation can be very useful—as is shown later in this paper—provided the representation is capable of capturing the essential physics. Here we explore the adequacy of the LE statistical approach to modeling sprays. Although the LE approach is the most suitable of the existing modeling approaches, in its current form it is incapable of representing droplet-droplet interactions. Droplet-droplet interactions play an important role in spray heat and mass transfer, and the essence of spray modeling is the successful description of the collective behavior of the droplets comprising the spray8. While considerable work has been done on the modeling of isolated droplet heat and mass transfer in a convective environment9, the issue of modeling drop-drop interactions has received less attention10. We summarize our recent progress in developing an extended Lagrangian-Eulerian formalism (ELE), which enables the inclusion of dropdrop interaction effects within the existing LE framework. This extended Lagrangian-Eulerian (ELE) formalism is based on a concept of assumed point fields (APF). The ELE formalism shows promise of extending the existing LE mathematical formulation to a level that is adequate to describe important spray physics. We then turn to the next principal issue, which is the development of accurate closures for heat and mass transfer in sprays. Considerable work has already been done on the modeling of isolated droplet heat and mass transfer in a convective environment9, and a plethora of models of varying levels of detail are available to the modeler. The choice of the appropriate model is ultimately dictated by the droplet vaporization regime, and this is determined by the spray application under consideration. Our understanding of the physics of heat transfer in droplet clouds and collective vaporization of droplets can benefit greatly from DNS, which can be a very useful research tool for developing better physical sub-models of droplet heat and mass transfer that account for drop-drop interactions. Direct numerical simulations of scalar transport based on a discrete-time immersed-boundary method (DTIBM)11 are presented for particle-laden flow. The fluid velocity and scalar evolution are governed by the extended Navier-Stokes equations. Exact boundary conditions are imposed at the surface of each spherical particle. These simulations are useful for developing better physical sub-models for heat transfer in particle-laden and droplet-laden flows that account for particle-particle or drop-drop interactions. In order for a spray model to be predictive, it is critical to have a numerically convergent model implementation that also computes accurate solutions to the spray model equations. The importance of validation tests—whether these are based on comparison with analytical solutions or experiments—as necessary first step to gain confidence in spray model performance is underscored. The issue of model accuracy, and criteria drawn from a closely related field of turbulence modeling are adapted to assess and appraise spray model development. In summary this paper attempts to provide a perspective on spray model development that is based on the concept of balancing uncertainties/errors arising from (i) the mathematical representation, (ii) the closure models, (iii) numerical methods, and (iv) experimental uncertainty. ADEQUACY OF THE UNDERLYING MATHEMATICAL REPRESENTATION The principal spray modeling approaches currently used in computational fluid dynamics (CFD) calculations of sprays are based on either the Eulerian-Eulerian (EE) formulation for two-phase flows12-17, or the Lagrangian-Eulerian (LE) formulation18, 19. The point-process statistical representation20, which forms the basis for the mathematical formulation of the LE spray modeling approach, provides a more suitable framework than the EE approach to represent the droplet size distribution19. Also the representation of droplets in the Lagrangian frame of reference in the LE approach is more suitable for model development. In spite of these advantages, we show here that the LE approach based on Williams’ spray equation is incapable of representing drop-drop interactions. (The same is also true of other first-order statistical formulations such as the EE twofluid formulation.) This inadequacy of the LE formulation is apparent on reviewing the relevant basics of stochastic point process theory. The LE formulation is based on Williams’ droplet distribution function (ddf) f (x, v, r , t ) , which is the number density of droplets in position-velocityradius space20 . The ddf implies an average number of droplets in physical space n(x, t ) , which is obtained from the ddf by integrating over velocity-radius space: n(x, t ) f (x, v, r , t )dvdr. (1) It is important to note that the ddf and number density are first-order statistics that contain no information regarding drop-drop interactions. It is also important to note that the same number density (or average number of drops in a unit volume) can correspond to different spatial arrangements of droplets in physical space, as shown in Figure 1. Figure 1. Two monodisperse droplet fields corresponding to the same average number density, but different spatial arrangements of the droplets. In the left panel the droplet centers can be as close as one drop diameter, but on the right panel the droplet centers are not allowed to be closer than 1.72 drop diameters. Isosurfaces of a passive scalar are shown for the same scalar value of 0.5 in both panels. For the same average number density the left panel shows considerable drop-drop interaction with merging of the isosurfaces which envelop neighboring droplets, whereas the right panel shows less drop-drop interaction with distinct scalar isosurfaces. The natural question to then ask is: What are the implications of different spatial arrangements of droplets in physical space with same number density for heat and mass transfer modeling in sprays? Clearly one way in which different spatial arrangements can affect models is because different spatial arrangements affect LE models for individual droplet heat and mass transfer. A typical LE model21 for time evolution of the droplet radius r is ( D) g Y fs Y f dr (2) Shd , dt 2 d r 1 Y fs which states that the rate of droplet regression is determined by the transfer number Bd , defined as Y fs Y f Bd . (3) 1 Y s f The transfer number Bd characterizes the vaporization driving force arising from spatial inhomogeneity of the vapor fuel mass fraction Y f , and is the nondimensional difference between Y fs , the vapor fuel mass fraction at the surface of the droplet, and Y f , the vapor fuel mass fraction far away from the droplet surface in the gas phase. In Eq (1.2) d is the thermodynamic density of the liquid fuel droplet, D g is the mass diffusivity of fuel vapor in the gas phase, and Shd is the Sherwood number characterizing the rate of mass transfer. The Frossling correlation is often used for the Sherwood number, and this accounts for convective effects through a droplet Reynolds number dependence of the form ln(1 Bd ) Shd 2.0 0.6 Re1/d 2 Sc1/ 3 . (4) Bd A typical LE model21 for time evolution of the droplet temperature Td is dT 4 dr d r 3Cl d d 4 r 2 L 4 r 2 qd , (5) 3 dt dt which states that the total surface heat transfer rate 4 r 2 qd goes into either heating up the droplet (first term on the left hand side of above equation), or into supplying energy required for phase change (second term on the left hand side of above equation). In the above equation Cl is the specific heat of the liquid droplet and L is the latent heat of vaporization. The surface heat flux is calculated in terms of the droplet Nusselt number through the expression k (T T ) qd g g d Nud , (6) 2r and the Nusselt number is calculated using the Ranz-Marshall correlation which accounts for convective effects through the following dependence on the droplet Reynolds number ln(1 Bd ) Nud 2.0 0.6 Re1/d 2 Pr1/ 3 , (7) Bd where Pr is the Prandtl number in the gas phase. Many modeling assumptions underlie the typical LE models21 for droplet regression rate and heat transfer that are presented here. The important assumptions are a spherically symmetric droplet undergoing unsteady heating but internal circulation, turbulence and radiative effects are neglected. The physical effects of internal circulation, turbulence and radiation can be accounted for by incorporating more complex model expressions for droplet heat and mass transfer9 within the existing LE framework. Relaxing the spherical droplet assumption is nontrivial, but for small droplets this assumption is valid. For automotive and gas turbine fuel sprays the droplet sizes are small. However, there is one principal assumption underlying LE models that cannot be relaxed within the existing LE framework. This assumption is that all LE models are for an isolated droplet in a gas flow and drop-drop interactions are neglected. As noted earlier, accounting for drop-drop interactions is the essential part of describing the collective behavior of droplets in a spray. In this sense, the existing LE modeling approach is not adequate to describe the physics of drop-drop interactions in sprays. In order to further probe the effects of drop-drop interactions, we attempt to quantify the importance of drop-drop interactions on individual droplet heat and mass transfer by performing direct numerical simulations. The influence of drop-drop interactions on physical submodels of droplet heat and mass transfer are described in the following section. There is another effect of drop-drop interactions that must be considered in two-way coupled models for sprays. This is the effect of drop-drop interactions on the mass, momentum and energy source terms that arise in the gas-phase conservation equations due to the presence of the spray droplets. As an example, the mass source term due to droplet vaporization that appears in the gasphase mass conservation equation in the LE approach is dr s f (x, v, r , Td , t ) d 4 r 2 dvdrdTd . (8) dt The vaporization source term s can be calculated if the model for the droplet radius regression rate dr dt and the ddf f (x, v, r , Td , t ) 3 are known. For droplet radius regression rate models that assume an isolated droplet vaporizing in a gaseous stream, the form of the vaporization source term implies that in the case of a homogeneous spray with monodisperse droplets having identical initial velocity and temperature, the source term is linear in the number density. This follows from the decomposition of the ddf f (x, v, r , Td , t ) into the product of a number density n(x, t ) , and the conditional joint probability density function of velocity, radius and temperature f VcRTd ( v, r, Td ; x, t ) , that has been established by Subramaniam19, 20. For the case of a homogeneous spray with monodisperse droplets having identical initial velocity v 0 , radius r0 , and temperature Td 0 , the ddf decomposition simplifies to (9) f (x, v, r, Td , t ) n(x, t ) f VcRTd ( v, r, Td ; x, t ) n(t ) ( v v0 ) (r r0 ) (Td Td 0 ) , from which we deduce the linear dependence of the vaporization source term on the number density for isolated droplet vaporization models. The second important question that needs to be addressed is how different spatial arrangements of droplets corresponding to the same average number density affect the source terms that appear in the gas-phase conservation equations. Specifically, it is important to know under what conditions the linear dependence of the source terms on the number density is justified. Another limitation of existing LE formulations is that they are restricted to dilute sprays with low liquid volume fraction (in fact, recent analysis shows that the restriction is even stronger, and it limits the gradient of the volume fraction even in dilute flows). The extended LE (ELE) approach that is outlined in this paper describes how the existing LE approach can be applied to sprays that are not dilute. In this case the gas-phase conservation equations also need to be modified to account for volume displacement effects in the mass, momentum and energy conservation equations. In summary, we have shown in this section that although the LE approach is the most suitable of the statistical models for complete spray simulations, in its current form it is an inadequate The ddf as defined here includes the droplet temperature property, which is an extension of the form of the ddf defined earlier. 3 mathematical representation of spray physics because it does not account for different spatial arrangements of droplets that correspond to the same average number density. The effect of these different spatial arrangements needs to be accounted for because drop-drop interaction effects can influence both the models for individual droplet heat and mass transfer, as well as the source terms in the gas-phase conservation equations that represent the influence of the spray droplets. The principal challenges in developing an adequate mathematical representation are: (i) a statistical characterization of the different spatial arrangements of droplets (ii) an understanding of the effect of these different spatial arrangements of droplets on the physics of heat and mass transfer in sprays (iii) a formalism to incorporate these effects in statistical models, if possible by extending the existing LE framework itself. Second-Order Statistics: A second-order statistical description of droplets in a homogeneous spray using the pair-correlation function provides a characterization of different spatial arrangements of droplets10, 22, 23. The pair correlation function g (r ) represents the relative frequency of occurrence of all possible values of the inter-droplet spacing r . The simplest possible point process model for the spray droplets is that the droplet centers are distributed according to a homogeneous Poisson process in physical space23. The independence property of the Poisson process is reflected in its pair correlation function being unity for all possible values of the inter-droplet spacing r . Since the location of every point in a Poisson point process is assigned independent of all the other points (the presence of any Poisson point does not influence the location of any other point), all inter-point spacing values are equally likely and g ( r ) 1 . The Poisson process is not an appropriate model for droplets because it allows for overlapping droplets. (The centers of two droplets can be arbitrarily close to each other, and with finite probability they can lie closer than the droplet radius.) A simple modification to the Poisson process using a technique called dependent thinning results in a Matérn process that has no overlapping droplets23. In fact, based on a parameter h , the hard-core distance, one can define a spherical exclusion volume centered at each droplet within which no other droplet centers are allowed. For a homogeneous spray of monodisperse droplets with radius r0 , the minimum value of the hard-core distance h 2r0 guarantees non-overlapping droplets (see Figure 2). The Matérn process has an analytic pair correlation function that is shown in Figure 2 (the blue line is for a hard-core distance h 2r0 ). The hard-core distance parameter h can be varied in the Matérn process, thereby generating droplets with different spatial arrangements and known statistics. For a different hard-core distance value of h 3.46r0 , the droplet field is shown in Fig 1b, and the corresponding pair-correlation function is shown in Fig 2 (red line). As the hard-core distance is increased for the same average number density, the droplets are constrained to be farther from each other. There are obviously limits to the maximum hard-core distance that can be specified for a given average number density23. However, the hard-core distance parameter in the Matérn process enables us to investigate the effects of different spatial arrangements of droplets, independent of the average number density (or in other words, the volume fraction). Figure 2. Two monodisperse droplet fields generated using the Matérn process that removes points from a Poisson process using the technique of dependent thinning. Both droplet fields have the same average number density (and volume fraction), but the hard-core distance parameter for the field shown above is chosen to be twice the droplet radius, while it is 3.46 times the droplet radius for the field shown below. The corresponding pair-correlation functions (which are analytically known for the Matérn process) are also shown. Note that the pair-correlation function goes to zero for droplet center separation distances below the hard-core distance, and falls off to the Poisson value of one for separation larger than twice the hard-core distance. Other researchers have attempted to characterize drop-drop interaction effects starting with the early work of Tishkoff24 where a drop-in-bubble model was proposed. More recently Chiu10 has shown that second-order statistics can be used to describe drop-drop interactions. Chiu’s proposal for the paircorrelation function10 is more typical of the pair-correlation of molecules in liquid phase, and more than one parameter is needed to specify this pair-correlation function. Furthermore, it is non-trivial to generate a point process that corresponds to a specified pair-correlation function, even if the paircorrelation function has an analytic form. The principal advantage of the Matérn point process model22, 23 is that it is easy to simulate a realization of this point process, and this process has a known analytic pair-correlation function. The main reason why the Matérn point process model in this work is that if the point process that corresponds to a specified pair-correlation function can be generated, then this point process can be used to initialize a droplet field with known statistics for DNS. With this statistical description of different spatial arrangements to address the first challenge, we now investigate the effect of these different spatial arrangements of droplets on the physics of heat and mass transfer in sprays. In the following section we quantify the influence of different spatial arrangements on individual droplet heat and mass transfer by performing DNS of passive scalar transport over spherical particles. We also show how DNS can also be used to quantify the influence of drop-drop interactions on the source terms that arise in the gas-phase conservation equations. PHYSICAL SUB-MODELS: DROPLET HEAT AND MASS TRANSFER Accurate modeling of heat and mass transfer between the droplet liquid phase and the vapor phase is critical for predictive spray computations. Even for simple thermodynamically sub-critical conditions, this is a serious challenge. While considerable work has been done on the modeling of isolated droplet heat and mass transfer in a convective environment, the issue of modeling drop-drop interactions has received less attention. Although more sophisticated LE models than the typical LE model presented in the previous section can be used, they all assume an isolated droplet vaporizing in a gaseous stream. The principal challenge for statistical models is describing the collective behavior of spray droplets from our knowledge of heat and mass transfer in single droplets, droplet arrays, and droplet groups. In this work we show that direct numerical simulations based on solution of the Navier-Stokes equations with exact boundary conditions on the surface of spherical particles are a promising approach to understanding the physics of drop-drop interactions on heat and mass transfer in sprays. Direct Numerical Simulation Approach: The coupling of the fluid and particles is accomplished through the DTIBM algorithm of Yusof11, which imposes the exact no-slip and no-penetration boundary conditions at the particle-fluid interface. This new capability allows us to perform simulations of large-scale polydisperse fluid-particle systems over the entire range of volume fraction—from dense granular assemblies and dilute fluid-particle flow—thus giving us reliable, direct access to details of the complex physics underlying these systems. The spectral NS/DTIBM is a numerical method for solving the incompressible, forced Navier-Stokes (NS) equations of the form: u u u P 2u g (x, t ), (10) t u 0 where g(x,t) is a body force term which may be a function of space and time. In the DTIBM approach g(x,t) is the force on the fluid due to the presence of solid particles. These equations are solved using a pseudo-spectral method that employs Fourier transforms in the y- and z- directions, and space-centered finite differences in the x- direction. Time stepping is performed using a Crank-Nicholson scheme for the viscous terms and a second-order Adams-Bashforth scheme for all other terms. A second-order velocity-pressure predictor scheme is used, followed by a pressure correction step that ensures that the flow is incompressible (i.e., divergence free). The Discrete Time Immersed Boundary Method couples the solid and fluid flow through the spatially and temporally varying force on the fluid, g(x,t). For a no-slip boundary condition between the fluid and a given particle the fluid velocity, u, must equal the velocity of the surface of a particle, v. Using the discrete-time Navier-Stokes equations and noting that at time n+1 u n1 v On the particle surface this gives 1 g(, t ) u u P 2u ( v u n ) (11) t for the force, g on the fluid due to the particle surface Ω. Finally, for the no-slip condition, the smoothest local tangential velocity is achieved by simply reversing the tangential velocities between flow reversal pairs on either side of the solid boundary, and scaling the internal tangential velocity to get no-slip at the desired location. This method allows for many complex moving boundaries to be included within the solution of the NS equations because the computational cost scales very favorably with the number of boundaries (computational cost is almost independent of number of boundaries). Therefore DTIBM is well-suited for simulation of a large number of particles in flows at Reynolds numbers O(100) . We have extended the NS/DTIBM approach to incorporate the transport of a passive scalar field , which obeys a transport equation u 2 h(x, t ) , (12) t where is the scalar diffusivity, and h(x,t) is a scalar forcing term which may be a function of space and time. For the scalar DTIBM, h(x,t) is the scalar forcing on the fluid due to the presence of solid particles. Currently the code is capable of imposing the Dirichlet boundary condition on all the particle surfaces. It is possible to impose Neumann and mixed boundary conditions on the particle surfaces, and we plan to implement these in the near future. Using the discrete-time equations and noting that at time n+1 n 1 s On the particle surface this gives 1 h(, t ) u 2 (s n ) (13) t for the scalar forcing h(x,t) on the fluid due to the particle surface Ω. Droplet Field Initialization: The effects of different spatial arrangements of droplets on heat and mass transfer are quantified by performing DNS of passive scalar transport over monodisperse spherical particles that are generated by simulations of the Matérn process. In one set of simulations the average number density of the particles is kept constant (volume fraction is constant) and the hard-core distance is systematically varied to investigate the effect of drop-drop interactions. Several such sets of calculations over a range of number density (volume fraction) values were performed to investigate the effect of number density (volume fraction). The droplet field and initial conditions are chosen to be representative of typical gasoline direct injection spray conditions with 60 m diameter droplets vaporizing in air at a temperature of 1200K. A droplet’s region of influence is first determined by performing DNS of steady flow with passive scalar transport over a single sphere with the particle Reynolds number Rep=20, and Pr=0.7 (in this case the passive scalar is taken to be temperature with the thermal diffusivity equal to the scalar diffusivity, ). The ambient gas temperature is initially uniform with value , and the sphere is maintained at a temperature s for all time. The nondimensional temperature field is defined as s T , s so that it is equal to one in the ambient gas and zero on the sphere. The drop region of influence is defined by the length of the T 0.5 contour, which is denoted d in the longitudinal (streamwise) direction, and d in the transverse (cross-stream) direction. Figure 3 shows two cut-planes through the flow which reveal d 3.37 r0 and d 2.77r0 for Rep=20, and Pr=0.7, where r0 is the radius of the spherical droplet. Estimates of the drop’s region of influence are used to determine the range over which the hard-core distance parameter h is varied. Figure 3. Contours of a passive scalar (denoted T) obtained from direct numerical simulation of steady flow at Reynolds number of 20 (based on sphere diameter) past a sphere with Pr=0.7. Two cut-planes are shown to characterize the drop region of influence in streamwise and cross-stream directions based on the dimensions of the T=0.5 contour. Computational requirements for accurate resolution of the flow around each sphere require a minimum of 8 grid points on the diameter of each droplet. The total number of grid points ( 256 128 128 ) is limited by processor memory, which in turn determines the size of the computational box. The average number density value is chosen to simulate around 100 droplets in the computational domain, resulting in a number density of 175739 drops/cc. (It should be noted that the computational cost increases very little with increase in number of droplets, therefore simulating higher number densities is not a limitation of the DTIBM method.) For droplet radii of 31.5 microns this results in a volume fraction of 0.023, which corresponds to a fairly dilute spray regime. Initial and Boundary Conditions: The steady flow past the droplet field is computed at a nominal Rep=20 (based on an isolated sphere). Then the scalar field is introduced with T 1 in the ambient gas and T 0 on the sphere surface. The inflow condition is at x 0 and outflow at the end of the computational domain into a buffer region to prevent artificial pressure wave reflection. DNS Results: The T 0.5 isosurfaces (orange), as well as the T 0 isosurface (in violet denoting the sphere surfaces) are shown for a volume fraction of 0.02 and various hard-core distance values in Figure 4. For the lowest hard-core distance of h 2r0 there is considerable overlap of the T 0.5 isosurfaces, which in some cases enclose several neighboring spheres indicating considerable dropdrop interaction. On the other hand, at the same volume fraction but with higher hard-core distances up to h 3.46r0 there is decreasing overlap of the T 0.5 isosurfaces indicating minimal drop-drop interaction at the highest hard-core distance. Figure 4. Isosurfaces of T=0.5 from DNS of steady flow past different spatial arrangements of spheres initialized using a Matérn point process. All cases are for an average volume fraction of 0.02. As the hardcore distance is increased, the drop-drop interaction effects are less pronounced. The surface scalar flux was calculated for each sphere, and an average surface scalar flux was determined for the entire droplet field. The variation of this average surface scalar flux (scaled by the isolated sphere’s surface scalar flux) with increasing hard-core distance is examined at a fixed volume fraction of 0.02. See Figure 5. For the lowest hard-core distance where maximum drop-drop interactions are expected, it is found that the average surface scalar flux can be about 18% less than the isolated sphere’s surface flux. As the hard-core distance is increased, the effects of drop-drop interactions are weaker, and the ratio tends towards one. These preliminary results are obtained for a single realization of the droplet field, and need to be repeated over multiple independent initializations to obtain reliable averages and confidence intervals. Average values formed from multiple independent simulations may also revise the non-monotonic trend seen in this figure. Figure 5. The variation of the ratio of average surface scalar flux (obtained by averaging over all the spheres in the domain) to the surface scalar flux for an isolated sphere, with increasing hard-core distance. The volume fraction is 0.02 for all the cases. The interphase scalar source term from the droplet cluster that would appear in the gas-phase scalar conservation equations is also calculated, and it is scaled by the interphase scalar source term that would appear for a single isolated droplet. If the source terms are indeed linear in the number density for this homogeneous problem (which is an assumption inherent in the LE approach), then the ratio of these interphase scalar source term for the droplet cluster should just be the number of droplets (102, in this case) times the interphase scalar source term for a single isolated droplet. However, Figure 6 shows that the effect of drop-drop interactions even at this low volume fraction (2%) results in a scaling of the interphase source terms that is not linear in the number density. Figure 6. Ratio of the scalar interphase source term calculated from the droplet cluster composed of 102 droplets, to the source term arising from an isolated droplet, as a function of hard-core distance. The source terms computed in the current LE approach assume a linear scaling with the droplet number density for homogeneous droplet clusters, which corresponds this ratio being a constant value of 102. Finally the effects of varying the volume fraction on the average surface flux for a single sphere are shown in Figure 7. The effects of hard-core distance become more apparent as the volume fraction increases, and they are non-negligible even at a volume fraction of 0.02 for a hard-core distance of h 2r0 . Figure 7. The ratio of average surface scalar flux (obtained by averaging over all the spheres in the domain) to the surface scalar flux for an isolated sphere, with increasing hard-core distance for different values of the volume fraction. Conclusions from DNS study: The principal conclusions from the DNS study are that drop-drop interactions can be important even at relatively low volume fraction, and they can affect the LE model for individual droplet heat transfer that assumes an isolated droplet. Also the interphase source terms are influenced by drop-drop interaction effects, and depending on the spatial arrangement of the droplets they may not scale linearly with number density even for a homogeneous spray. The form of the interphase source terms in an LE approach assumes a linear dependence on the number density for a homogeneous spray. Therefore, even these preliminary DNS results point to the importance of accounting for drop-drop interactions in spray models of heat and mass transfer. Extensions to current DNS study: The DNS results presented here are very preliminary and there are several restrictions and assumptions inherent in the simulations. Many of these restrictions can be easily removed by code development and extension, while others are fundamental to the formulation and are considerably harder to overcome. The extensions that are relevant to this work are itemized below: 1. Multiple Independent Simulations (MIS): The DNS need to be repeated for different initializations of the droplet field for each hard-core distance specification in every set of calculations performed at a fixed volume fraction. Such MIS are trivial to perform and will enable the calculation of converged statistics with confidence intervals. 2. Thermal Boundary Condition and Coupling: The thermal boundary condition can be extended from pure Dirichlet to specified heat flux (including the zero heat flux insulated condition). The thermal coupling to the droplet can be improved by first considering unsteady heating of the droplet assuming a lumped capacitance approach that is valid for small Biot number. 3. 4. 5. 6. 7. 8. Solution of the conduction equation for the temperature field within the sphere can also be incorporated based on semi-analytical procedures. Effects of upstream turbulence: This is easily incorporated in the pseudo-spectral DTIBM code that was designed to study particle-turbulence interactions. Testing and validation of the turbulence calculation in underway and the upstream turbulence effects will be incorporated in the code in the near future. Droplet motion: This is already incorporated in the code. Combining droplet motion with turbulence will enable us to study the effect of dispersion of the droplets in a more natural way, than forcing them to occupy fixed locations based on a point-field model. The evolution of the pair-correlation function due to droplet-turbulence interactions can be studied using these enhancements to the current DNS capability. Collisions: An important issue is how to treat collisions. If the spheres are treated as solid particles undergoing elastic collisions, then this can be easily incorporated. Droplet coalescence: If the DNS is to truly represent droplets, then the phenomenon of droplet coalescence or bounce-back after collisions needs to be accounted for. As noted previously, MD calculations indicate that the physical outcome of coalescence or bounce-back after droplet collision may depend on sub-continuum phenomena. These outcomes can be incorporated in a somewhat ad hoc fashion, but this will undermine the model-free characteristic of the true DNS calculations presented here. Droplet internal flow and droplet deformation: The DTIBM technique generates a fictitious flow within the sphere. Extending the hydrodynamic boundary condition at the droplet surface from no-slip to a stress boundary condition is a non-trivial extension, and possibly one of the principal challenges in extending the DTIBM technique to realistic droplet calculations. Droplet vaporization: Currently the DTIBM technique implements a no-penetration hydrodynamic boundary condition at the surface of the sphere, corresponding to a solid surface. In order to represent vaporization, this boundary condition needs to be extended to include non-zero normal velocity corresponding to droplet surface regression due to mass transfer. Also the gas-phase mass conservation equations need to be extended to include variable-density effects. Having established through preliminary DNS calculations that drop-drop interactions can be important even at relatively low volume fraction, we show in the next section how the LE approach may be extended using the assumed point field (APF) formalism to incorporate drop-drop interaction effects. THE ASSUMED POINT FIELD FORMALISM We have already noted that other works by Chiu10 have already demonstrated the use of second-order statistics such as the pair-correlation function to characterize drop-drop interactions. Here we extend this characterization to augment existing LE models using the concept of an assumed point field. Just as the mean and variance of a random variable are only incomplete statistical characterizations whereas the probability density function (PDF) is needed to completely characterize the random variable, similarly the average number density is an incomplete statistical characterization of a stochastic point process whereas the complete characterization of a point process20 requires specification of additional information. It is common in assumed PDF models7 to assume a Gaussian or normal PDF because the mean and variance completely specify the Gaussian PDF. By analogy, we may ask what is the simplest point process that is completely specified by the average number density. This is the Poisson point process, but it is unsuitable for describing the location of droplet centers because the complete independence property of the Poisson process allows for overlapping droplets, which results in an unphysical model. One can remove overlapping drops in the Poisson model by using a thinning procedure on a Poisson point process22, 23, which is the procedure to simulate a realization of a Matérn point process. The Matérn point process model has an analytic expression for the pair correlation that is completely specified by the hard-core distance parameter. The advantage of using a Matérn point process model is that it is easy to simulate this point process and it has an analytic pair-correlation function that is specified in terms of the hard-core distance and number density of the base Poisson process. It is not obvious how to simulate point processes for arbitrary pair correlation functions. However, it is important to also note that there are limits on the volume fraction that can be simulated using the Matérn point process. In the next section we describe the extended LE (ELE) approach that consists of locally representing the spray as a realization of a Matérn process of non-overlapping droplets, that corresponds to the average number density obtained from the LE approach. This allows for a more realistic calculation of individual droplet heat and mass transfer as well as the interphase heat and mass transfer source terms, by accounting for drop-drop interaction effects through the pair-correlation function of the Matérn process. THE EXTENDED LE CONCEPT A full second-order closure would require a solution of the evolution equation of the pair-correlation function at each cell in an LE calculation. This approach is deemed to not be practical for computational and modeling reasons at this stage. The APF provides an alternative route to incorporating second-order information in the existing LE approach by assuming the underlying point field and pair-correlation in each cell in the LE calculation. The schematic in Figure 8 describes the ELE concept. The Matérn number density is required to match the local average number density calculated by the LE model. All that remains is the specification of the hard-core distance parameter. Figure 8. The extended LE concept. The quantities (g(r)) represent parameters or moments of the paircorrelation function. Here we propose a simple specification of the hard-core distance parameter h . This quantity has a lower bound of 2r0 arising from the requirement of non-overlapping droplets. We also note that the Matérn pair-correlation function takes the Poisson value of one for values of r 2h . See Figure 2.This implies that in the Matérn model, for droplet separations beyond this correlation distance the droplets behave practically independently of each other. We argue that a pair of droplets that are initially very close to each other will remain correlated over a length scale l d characteristic of the turbulent eddy whose timescale is the same as d , the droplet’s momentum response time. Over this length scale l d we assume the pair of droplets is “trapped” in the same eddy, and that they evolve independently once their separation distance exceeds this length scale. This independence is reflected in their paircorrelation going to unity beyond this length scale. Assuming inertial range scaling of turbulent eddies in high Reynolds number turbulence4 u3 u 2 , l , (l ) we estimate the length scale of the eddy whose time scale is the same as the droplet’s momentum response time to be 2 1/ 2 2 ro d l d u d k (14) 9 g f The hard-core distance is then simply specified to be l h max 2r0 , d 2 . (15) The above specification is adequate for monodisperse droplets. For polydisperse sprays some modifications are needed. The dependent thinning procedure for generating a Matérn point process from an underlying Poisson point process can be extended to the polydisperse case quite easily. However, to our knowledge there is no analytic form for the pair-correlation function in the polydisperse case. Nevertheless, the pair-correlation function can be estimated by numerical simulation. EFFECTS OF TURBULENCE ON DISPERSION It is instructive to examine what affects the evolution of the pair-correlation function. Clearly turbulence in the ambient gas plays a central role in pair- or relative- dispersion because it affects the relative velocity between droplet pairs. It is important to note that this is an effect that manifests itself at the level of second-order statistics, and has been denoted preferential concentration by Eaton and co-workers25-28. However, turbulence in two-phase flows, and its effect on dispersion of droplets pose significant modeling challenges even at the level of first-order statistics. The effects of turbulence on first-order statistics of droplet dispersion are non-trivial to reproduce since it is coupled to the dynamics of interphase turbulent kinetic energy transfer. The problem is particularly challenging because it is difficult to simultaneously reproduce the disparate trends of turbulent kinetic energy decay and droplet dispersion with variations in droplet Stokes number. A recently developed dual-timescale Langevin model29 has been shown to be capable of reproducing these effects simultaneously. The question of droplet relative dispersion is an important one, and it will need careful DNS calculations to resolve the question of how the pair-correlation function evolves under the effects of turbulence. NUMERICAL CONVERGENCE AND ACCURACY Devising a numerically convergent scheme to solve the system of equations in the LE approach, and validating the accuracy of the numerical solution is a necessary step for predictive spray modeling. Many studies have demonstrated the non-convergence of popular numerical implementations of the LE approach to spray modeling30-32. A systematic investigation of the numerical convergence characteristics of the KIVA implementation of the LE approach led to the conclusion that even global spray characteristics such as the spray penetration length did not show any trend towards convergence to an asymptotic value as the numerical parameters such as grid size, time step and computational particles were varied toward their limiting values. Similar concerns regarding the convergence of numerical implementations of the LE approach have also been reached by works of Abraham31, and Schmidt33. Numerical implementations that do not converge, and that are not accurate, can seriously undermine the predictive capability of a spray model. In the numerical convergence study a systematic approach to establishing convergence was also outlined30. Considerable success has been attained in establishing numerical convergence for the vaporization sub-model of the LE implementation34. In recent work34 we show how the numerical implementation of vaporization in the LE approach can be improved. A simple vaporization test problem is proposed, that admits analytic solutions to the spray equation, and is useful for testing the accuracy of numerical solutions. This study shows that a simple particle method solution using uniform sampling of the ddf yields an accurate solution to the simple vaporization test problem. However, many spray codes such as KIVA use importance sampling of the mass-weighted ddf on the grounds that this is more computationally efficient. The implementation of importance sampling in KIVA results in an inaccurate numerical solution of the spray equation that does not converge to the analytic solution for the simple vaporization test, even for a very large number of computational particles. We show that importance sampling can be accurate and computationally efficient if statistical weights are correctly assigned to match the initial radius distribution. Simulations also reveal that the discontinuous evolution of statistical weights corresponding to vaporization in existing particle methods results in numerical estimates of spray statistics that do not unconditionally converge to a continuous asymptotic limit as the time step is decreased. A continuously-evolving-weights algorithm is developed that yields numerically convergent results that also match the analytic solution very well. These improvements to the particle method solution of the spray equation, which result in an excellent match of numerical predictions with the analytical solution in the test problem, can be expected to dramatically improve the accuracy of complex spray calculations at minimum computational expense. This success indicates that if proper attention is paid to testing and validation of the numerical implementation following the systematic approach outlined in Ref.30, it is possible to obtain accurate, numerically convergent solutions using particle methods. The next step in the approach to establishing numerical convergence of LE implementations is accurate coupling of the particle method used to solve the spray equation, to the finite-volume solution to the Eulerian gas-phase conservation equations. This coupling requires numerically convergent and accurate estimation of the number density and spray source terms. In earlier work where particle methods were combined with finite-volume calculations, it is shown that kernel estimation methods35 offer a grid-free estimation approach for the number density and spray source terms. Some ideas for developing test problems that can be used to validate the numerical convergence and accuracy of spray computations using such estimation methods are now discussed. In our efforts to demonstrate convergence of interphase source terms obtained from particle data, we are currently investigating three different approaches for estimation of mean statistical quantities on an Eulerian grid from Lagrangian particle data. These three algorithms are: (i) Lagrangian Polynomial Interpolation (LPI), (ii) Cubic Spline Interpolation (CSI), and (iii) a Two-Stage Kernel Estimation (TSE) algorithm based on local least-squares proposed by Dreeben & Pope (details of which are given in Ref.35) The three algorithms are tested by specifying an analytic number density function from which Lagrangian particle data is generated. See Figure 9. The algorithms then estimate the number density from the Lagrangian particle data, and the error with respect to the analytic number density function is calculated on a finer test grid. For the number density estimation our first finding is that a nonnegativity constraint for number density has to be imposed on the TSE algorithm that uses local leastsquares. For number density estimation the CSI and TSE algorithms perform better than the LPI algorithm, as seen in Figure 10. If estimates of the number density gradient are considered (not shown here), then the TSE algorithm performs better than the other two algorithms. Figure 9. A test for number density estimation where particles are distributed according to a specified analytic number density function that is representative of sprays. Errors incurred by different numerical algorithms in estimating the number density from particle data can be compared. Figure 10. Estimates of the number density from particle data using three different algorithms are compared with the analytic specification for the number density (top left panel). The two-stage estimation (TSE) algorithm based on local least-squares with an imposed non-negativity constraint on the number density (top right panel), and the Cubic Spline Interpolation (CSI) yield good results. The Lagrangian Polynomial Interpolation (LPI) is subject to large fluctuations in the number density estimate, especially near the edge of the spray. DISCUSSION: A PERSPECTIVE ON SPRAY MODEL DEVELOPMENT If the objective of spray modeling is to develop robust and predictive models of sprays for use in engineering problems, it is useful to speculate on what determines model accuracy. When a model prediction is compared with experimental data, there are many reasons why differences can arise between the two. Even the perfect spray model will not give exact agreement with experimental data due to experimental uncertainty. In general, differences between model prediction and experimental data can arise due to (i) inadequacies in the mathematical representation, (ii) inaccurate closure models for physical phenomena such as droplet heat and mass transfer, (iii) numerical inaccuracy due to inadequate grid resolution, finite number of computational particles, time step not being small enough to resolve relevant time scales, and finally, (iv) experimental uncertainty. From an engineering perspective, the optimal model should balance the uncertainties and errors arising from each of these sources, so that modeling and computational efforts are efficiently focused on those sources that contribute the most to the predictive capability. Another perspective on what constitutes a good spray model can be gained by applying Pope’s criteria4 for assessing turbulence models to the spray problem. The principal criteria advanced in that analysis, and their relevant interpretation in the spray context are: (a) Level of description: in the spray context this means does the model represent the SMD, or the entire distribution of droplet sizes. A higher level of representation usually requires more complex models and more computation time (b) Completeness: in the spray context this means the number of additional inputs required by the model, e.g., is the spray angle a required input to the model, or is that quantity predicted by the model itself (c) Cost and ease-of-use: Is the model used for repetitive calculations in industry, or is it used in large-scale one-of-a-kind research computations (d) Range of applicability: Can the model be applied equally to internal combustion engine sprays as it is to pharmaceutical inhalers? (e) Accuracy: Model accuracy, as discussed earlier in this section. The use of such criteria can be very useful in performing legitimate comparisons of models. It is of particular importance to not expect models with different levels of representation to perform equally well in predicting sprays to the same level of detail. With these perspectives, the extended LE modeling approach proposed in this work may be viewed as striking a balance between incorporating adequate mathematical representation to model important spray physics, while retaining the cost and ease-of-use for repetitive spray calculations in industrial applications. The use of DNS to improve physical sub-models in the LE approach is also an attempt in this direction. SUMMARY AND CONCLUSIONS Significant progress has been made in areas that are critical to improving the predictive capability of spray models, such as (a) the mathematical formulation of spray theory10, 18-20, 36, (b) improving accuracy of the closure relations for the physical sub-models representing droplet heat and mass transfer9, 37, and (c) improving the accuracy of numerical solution to the model equations31, 33, 34. Specifically, in this work we describe recent advances in developing an extended LE modeling approach, the use of DNS to improve closure relations for physical sub-models representing heat and mass transfer, and the use of importance-sampling and kernel estimation methods to improve the accuracy of numerical solutions to the model equations. The extended LE approach not only provides a mathematical formalism for incorporating drop-drop interactions, but it also permits the extension of existing LE formulations which are restricted to dilute sprays with low liquid volume fraction to dense sprays. In fact, even collision modeling in sprays can be improved using the extended ELE approach, although those details are not presented here. In spite of this progress, the ultimate goal of developing a truly predictive and robust CFD spray simulation remains a challenging, unsolved problem. While most research efforts have focused on developing accurate closure relations for droplet heat and mass transfer9, the importance of other factors, such as numerical accuracy, and adequacy of the underlying mathematical formulation, is also now being recognized. An important conclusion of this work is that integration of knowledge in the specialty sub-areas is critical if we are to make progress in solving the spray problem, even for the simple test case of a single-component sub-critical vaporizing spray. An integrated approach to a CFD spray simulation requires assessing the impact of (i) the adequacy of mathematical formulation, (ii) the accuracy of closure relations for droplet heat and mass transfer, and (iii) the numerical accuracy of the computational scheme, on prediction of spray characteristics that are relevant to design engineers. Such an assessment would help identify where spray modelers should focus their efforts. Systematic studies of the sensitivity of predicted of global spray characteristics to numerical parameters, model constants, and underlying mathematical formulation can be very useful in guiding a program of spray modeling development. Preliminary studies indicate that spray predictions using Lagrangian-Eulerian codes exhibit strong sensitivity to numerical parameters31. It has been conclusively demonstrated that these problems can be addressed through using better numerical schemes and algorithms34, 35. A similar systematic assessment of sensitivity to the model constants has not yet been reported in the spray literature. It is considerably more difficult to assess the sensitivity of model predictions to the underlying mathematical formulation. It is anticipated that the issue of obtaining accurate numerical solutions to the spray problem will be resolved in time, given the increasing availability of inexpensive computational power, and the development of better algorithms for solving the model equations. The same is true for the development of more accurate closures for droplet heat and mass transfer using direct numerical simulations and experimental data. What is likely to remain a more long-term challenge is the development of better mathematical formulations to address the complex multiscale, nonlocal and nonlinear interactions in sprays. A rigorous mathematical formulation that bridges the internal flow in the nozzle through breakup and atomization into dispersed spray droplets of a given size distribution and initial spray angle, would be a significant breakthrough in spray modeling and simulation. In summary, the predictive capability of CFD calculations of sprays will continue to improve with robust numerical implementations, accurate sub-models, and significant enhancements to the underlying mathematical formulation, examples of which have been described in this paper. REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. Murad, S. and C.K. Law, Molecular simulation of droplet collision in the presence of ambient gas. Molecular Physics, 1999. 96(1): p. 81-85. Friedrich, R., et al., Direct numerical simulation of incompressible turbulent flows. Computers & Fluids, 2001. 30(5): p. 555-579. Wu, P.K. and G.M. Faeth, Onset and End of Drop Formation Along the Surface of Turbulent Liquid Jets in Still Gases. Physics of Fluids, 1995. 7(11): p. 2915-2917. Pope, S.B., Turbulent flows. 2000, Cambridge ; New York: Cambridge University Press. xxxiv, 771 p. Saddoughi, S.G. and S.V. Veeravalli, Local Isotropy in Turbulent Boundary-Layers at High Reynolds-Number. Journal of Fluid Mechanics, 1994. 268: p. 333-372. Subramaniam, S. Relationship between Lagrangian Statistical Spray Models and Eulerian Statistical Models of Two-Phase Flow. in ILASS Americas, 14th Annual Conference on Liquid Atomization and Spray Systems. 2001. Dearborn, MI. Pope, S.B., PDF Methods for Turbulent Reactive Flows. Progress in Energy and Combustion Science, 1985. 11(2): p. 119-192. O'Rourke, P.J., COLLECTIVE DROP EFFECTS ON VAPORIZING LIQUID SPRAYS, in Mechanical Engineering. 1981, Princeton University: Princeton. p. 329. Sirignano, W.A., Fluid dynamics and transport of droplets and sprays. 1999, Cambridge ; New York: Cambridge University Press. xvii, 311 p. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. Chiu, H.H., Advances and challenges in droplet and spray combustion. I. Toward a unified theory of droplet aerothermochemistry. Progress in Energy and Combustion Science, 2000. 26(4-6): p. 381-416. Yusof, J.M., Interaction of massive particles with turbulence, in Aerospace Engineering. 1996, Cornell University: Ithaca. p. 154. Beck, J.C. and A.P. Watkins, The droplet number moments approach to spray modelling: The development of heat and mass transfer sub-models. International Journal of Heat and Fluid Flow, 2003. 24(2): p. 242-259. Beck, J.C. and A.P. Watkins, On the development of a spray model based on drop-size moments. Proceedings of the Royal Society of London Series a-Mathematical Physical and Engineering Sciences, 2003. 459(2034): p. 1365-1394. Beck, J.C. and A.P. Watkins, On the development of spray submodels based on droplet size moments. Journal of Computational Physics, 2002. 182(2): p. 586-621. Lain, S. and R. Aliod, Study on the Eulerian dispersed phase equations in non-uniform turbulent two-phase flows: discussion and comparison with experiments. International Journal of Heat and Fluid Flow, 2000. 21(3): p. 374-380. Lain, S. and R. Aliod, Discussion on second-order dispersed phase Eulerian equations applied to turbulent particle-laden jet flows. Chemical Engineering Science, 2003. 58(20): p. 4527-4535. Iyer, V. and J. Abraham, An evaluation of a two-fluid Eulerian-liquid Eulerian-gas model for diesel sprays. Journal of Fluids Engineering-Transactions of the Asme, 2003. 125(4): p. 660-669. Williams, F.A., Spray Combustion and Atomization. Physics of Fluids, 1958. 1(6): p. 541545. Subramaniam, S., Statistical modeling of sprays using the droplet distribution function. Physics of Fluids, 2001. 13(3): p. 624-642. Subramaniam, S., Statistical representation of a spray as a point process. Physics of Fluids, 2000. 12(10): p. 2413-2431. Orourke, P.J., THE KIVA COMPUTER-PROGRAM FOR MULTIDIMENSIONAL CHEMICALLY REACTIVE FLUID-FLOWS WITH FUEL SPRAYS. Lecture Notes in Physics, 1985. 241: p. 74-89. Stoyan, D., W.S. Kendall, and J. Mecke, Stochastic geometry and its applications. 2nd ed. Wiley series in probability and statistics. Applied probability and statistics. 1995, Chichester ; New York: J. Wiley. xix, 436 p. Stoyan, D. and H. Stoyan, Fractals, random shapes, and point fields : methods of geometrical statistics. Wiley series in probability and mathematical statistics. Applied probability and statistics. 1994, Chichester ; New York: Wiley. xiv, 389 p. Tishkoff, J.M., Model for the Effect of Droplet Interactions on Vaporization. International Journal of Heat and Mass Transfer, 1979. 22(10): p. 1407-1415. Rouson, D.W.I. and J.K. Eaton, On the preferential concentration of solid particles in turbulent channel flow. Journal of Fluid Mechanics, 2001. 428: p. 149-169. Fessler, J.R., J.D. Kulick, and J.K. Eaton, Preferential Concentration of Heavy-Particles in a Turbulent Channel Flow. Physics of Fluids, 1994. 6(11): p. 3742-3749. Eaton, J.K. and J.R. Fessler, Preferential Concentration of Particles by Turbulence. International Journal of Multiphase Flow, 1994. 20: p. 169-209. Squires, K.D. and J.K. Eaton, Preferential Concentration of Particles by Turbulence. Physics of Fluids a-Fluid Dynamics, 1991. 3(5): p. 1169-1179. 29. 30. 31. 32. 33. 34. 35. 36. 37. Pai, G.M. and S. Subramaniam. A new dual–timescale Langevin model for droplet–laden turbulent flows. in ILASS Americas, 18th Annual Conference on Liquid Atomization and Spray Systems. 2005. Irvine, CA. Subramaniam, S. and P.J. O'Rourke, Numerical Convergence of the KIVA-3 Code for Sprays and its Implications for Modeling. 1998, Los Alamos National Laboratory: Los Alamos. Aneja, R. and J. Abraham, How far does the liquid penetrate in a diesel engine: Computed results vs. measurements? Combustion Science and Technology, 1998. 138(1-6): p. 233255. Iyer, V. and J. Abraham, Penetration and dispersion of transient gas jets and sprays. Combustion Science and Technology, 1997. 130(1-6): p. 315-334. Schmidt, D.P. and C.J. Rutland, Reducing grid dependency in droplet collision modeling. Journal of Engineering for Gas Turbines and Power-Transactions of the Asme, 2004. 126(2): p. 227-233. Pai, G.M. and S. Subramaniam, Accurate numerical solution of the spray equation using particle methods. Atomization and Sprays, 2005(accepted). Subramaniam, S. and D.C. Haworth, A probability density function method for turbulent mixing and combustion on three-dimensional unstructured deforming meshes. International Journal of Engine Research, 2000. 1(2): p. 171-188. Edwards, C.F., Toward a comprehensive theory of dense spray flows. Atomization and Sprays, 2000. 10(3-5): p. 335-353. Law, C.K., Recent Advances in Droplet Vaporization and Combustion. Progress in Energy and Combustion Science, 1982. 8(3): p. 171-201.