On the Comparison between Presumed and Full PDF Methods for Turbulent Precipitation

Chemical and Biological Engineering Publications Chemical and Biological Engineering 2001 On the Comparison between Presumed and Full PDF Methods for Turbulent Precipitation Daniele L. Marchisio Politecnico di Torino Rodney O. Fox Iowa State University, rofox@iastate.edu Antonello A. Barresi Politecnico di Torino See next page for additional authors Follow this and additional works at: http://lib.dr.iastate.edu/cbe_pubs Part of the Process Control and Systems Commons The complete bibliographic information for this item can be found at http://lib.dr.iastate.edu/ cbe_pubs/100. For information on how to cite this item, please visit http://lib.dr.iastate.edu/ howtocite.html. This Article is brought to you for free and open access by the Chemical and Biological Engineering at Iowa State University Digital Repository. It has been accepted for inclusion in Chemical and Biological Engineering Publications by an authorized administrator of Iowa State University Digital Repository. For more information, please contact digirep@iastate.edu.

On the Comparison between Presumed and Full PDF Methods for Turbulent Precipitation Abstract Turbulent precipitation is an important topic in chemical reaction engineering because of its numerous industrial applications. Several modeling approaches have been used in the past, but in recent years, computational fluid dynamics (CFD) coupled with micromixing models has been successfully applied to predict the influence of mixing on the crystal size distribution (CSD). The micromixing model is generally based on the presumed probability density function (PDF) approach, such as finite-mode PDF or beta PDF, and the aim of this work is to compare presumed PDF predictions and full PDF predictions with experimental data. The experimental data were obtained from a tubular reactor in which turbulent precipitation of barium sulfate is carried out from aqueous solutions of barium chloride and sodium sulfate. The implementation of the presumed PDF model was done using FLUENT user-defined subroutines, whereas the full PDF calculations were carried out with an in-house code based on Monte Carlo methods using the flow field prediction from FLUENT. Disciplines Chemical Engineering Process Control and Systems Comments This article is from Industrial & Engineering Chemistry Research 40 (2001): 5132-5139, doi: 10.1021/ ie0010262. Posted with permission. Authors Daniele L. Marchisio, Rodney O. Fox, Antonello A. Barresi, and Giancarlo Baldi This article is available at Iowa State University Digital Repository: http://lib.dr.iastate.edu/cbe_pubs/100

5132 Ind. Eng. Chem. Res. 2001, 40, 5132-5139 On the Comparison between Presumed and Full PDF Methods for Turbulent Precipitation Daniele L. Marchisio,*, Rodney O. Fox, Antonello A. Barresi, and Giancarlo Baldi Dipartimento Scienza dei Materiali ed Ingegneria Chimica, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129, Torino, Italy, and Department of Chemical Engineering, Iowa State University, 2114 Sweeney Hall, Ames, Iowa 50011-2230 Turbulent precipitation is an important topic in chemical reaction engineering because of its numerous industrial applications. Several modeling approaches have been used in the past, but in recent years, computational fluid dynamics (CFD) coupled with micromixing models has been successfully applied to predict the influence of mixing on the crystal size distribution (CSD). The micromixing model is generally based on the presumed probability density function (PDF) approach, such as finite-mode PDF or beta PDF, and the aim of this work is to compare presumed PDF predictions and full PDF predictions with experimental data. The experimental data were obtained from a tubular reactor in which turbulent precipitation of barium sulfate is carried out from aqueous solutions of barium chloride and sodium sulfate. The implementation of the presumed PDF model was done using FLUENT user-defined subroutines, whereas the full PDF calculations were carried out with an in-house code based on Monte Carlo methods using the flow field prediction from FLUENT. 1. Introduction The main advantage of using probability density function (PDF) methods lies in the fact that the chemical reaction term can be treated exactly, whereas most of the other methods require a closure assumption. PDF methods can be classified as full and presumed. Several differences between the two types of models exist, in terms of the discretization of the composition space, the spatial transport, and the numerical approach adopted. For full PDF methods, for example, spatial transport is treated as a random diffusion process, whereas presumed PDF methods are based on an assumption concerning scalar conditional mean concentrations. In the case of an inert scalar with uniform mean gradients, the full PDF method would predict that the joint scalar PDF is Gaussian (as indicated by direct numerical simulation results, as seen in the Experimental Setup and Numerical Simulation section), whereas the presumed PDF method cannot predict the correct limiting PDF. Moreover, unlike presumed PDF methods, full PDF methods do not require a priori knowledge of the shape of the joint composition PDF, but they are more CPU-intensive. In recent years, the presumed PDF method coupled with computational fluid dynamics (CFD) has been used to investigate turbulent precipitation in different devices. 1,2 The investigation of turbulent precipitation requires an adequate description of the population balance for the solid phase. This can be done by using discretization-based methods 3 or the moment method. 4 Unlike the former, the moment method requires only a small number of scalars, and recent results 5 show that is also applicable to complex problems such as sizedependent growth. * Corresponding author. E-mail: marchis@athena.polito.it. Fax: +390115644699. Politecnico di Torino. Iowa State University. In general, using CFD coupled with a presumed PDF, good agreement with experimental data is found, but the uncertainty in kinetic expressions and the poor reproducibility of the experimental data highlight the need for other validation techniques. Thus, a comparison with full PDF predictions provides a useful means of examining the approximation of choosing the shape of the PDF a priori and its influence on the final crystal size distribution (CSD). Previous results in a Taylor- Couette reactor 6 and in stirred vessels 7-9 showed that mixing has a strong influence on the development of supersaturation and thus on particle nucleation and growth. In this work, barium sulfate precipitation from aqueous solutions in a tubular reactor with coaxial feeds is investigated. 2. Full and Presumed PDF Methods The full PDF description involves a set of notional particles that obey stochastic differential equations and move in the computational domain mimicking PDF transport in physical and composition space. A detailed description of PDF formulation and solution using Monte Carlo methods can be found in ref 10, and here, we limit the discussion to the transport equation of the composition PDF f φ t + u j f φ + [S x j ψ R (ψ)f φ ] ) R - [ u x i ψ f φ ] + [ D i ψ R 2 φ R ψ f φ ] (1) R where f φ (ψ;x,t) is the joint composition PDF of all scalars, u i is the fluid velocity, φ R represents the scalars under consideration (i.e., concentration, temperature, moments of the CSD), ψ R represents the corresponding variables in phase space, D R is the molecular diffusivity of φ R, S R (ψ) is the corresponding source term, and indicates the Reynolds average. 10.1021/ie0010262 CCC: $20.00 2001 American Chemical Society Published on Web 06/19/2001

Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 5133 The first term on the right-hand side of eq 1 represents the convective transport due to the scalar conditioned velocity fluctuations. It can be closed by using a gradient-diffusion model that introduces a turbulent diffusion term u i ψ f φ )-Γ t f φ x i (2) where Γ t is the turbulent diffusivity defined as Γ t ) C µ k 2 Sc t ɛ Here, C µ is a constant equal to 0.09, Sc t is the turbulent Schmidt number equal to 0.7, k is the turbulent kinetic energy, and ɛ is the turbulent energy dissipation. The second term on the right-hand side of eq 1 represents the micromixing term that can be closed by using different approaches. In the case of nonreacting scalars, it has been shown 10 that the micromixing model must fulfill some constraints concerning the shape of the PDF: the mean values must be constant, the variance decay must be exponential, and the asymptotic shape of the PDF must be Gaussian. In this work, a simple interaction-with-the-mean (IEM) model was used. This model is also known as the linear mean square estimation (LMSE) model and is based on a linear relaxation of the scalars toward their mean values 11 (3) D R 2 φ R ψ ) 1 2τ φ ( φ R - ψ R ) (4) where φ R is the local scalar mean and τ φ is the local micromixing time. The IEM approach gives poor performance in the so-called partially segregated CSTR, where the combined effects of macro-, meso-, and micromixing are described by the micromixing model alone. 12 However, when the IEM model is properly used as a model for only micromixing and macro- and mesomixing due to spatial transport are handled separately, it yields acceptable predictions, once the micromixing time is chosen correctly. Several approaches have been proposed for evaluating this characteristic time. Baldyga 13 proposed a cascade model for the variance decay prediction, whereas more recently, Fox 14 proposed a spectral relaxation model in which the micromixing time was evaluated by solving the transport equations of the scalar variance and the scalar variance dissipation rate. However, these multiscale models are based on the idea of fully developed turbulence spectra (high Reynolds numbers), and to our knowledge, no fully validated models exist for lower Reynolds numbers. As explained below, this work was carried out at moderate Reynolds numbers, relatively close to the transition region. Under these conditions, the use of a simple large-scale-motion-dominated model, 10 seems to be more reasonable 1 ) C φ ɛ τ φ k (5) where C φ is a constant that can be adjusted to take into account the above-mentioned moderate Reynolds number effects. In the presumed PDF approach, the functional form of the PDF is assumed a priori. In this work, we make use of the finite-mode-pdf (FM-PDF) model. Using this model, the PDF is represented by a finite set of delta functions 15 where p n (x,t) is the probability of mode n, φ R n (x,t) is the value of scalar R corresponding to mode n, N e is the total number of modes, and m is the total number of scalars. The total number of modes affects the model s ability to approximate the real PDF. When this number is increased, the description becomes more accurate, but the computational cost increases. The advantage of this model is its ability to accurately describe nonpremixed reacting flows with a small number of modes. Previous results 16 showed that N e ) 3 is sufficient to obtain good accuracy. A detailed description of this approach can be found in ref 2. Without loss of generality, we take the first scalar to be the mixture fraction. The evolution of the mode probabilities is defined once a transport equation for the probability vector p n (x,t) has been provided. To ensure the conservation of probability (probabilities sum to unity) and the conservation of mass, the transport equations can be written in terms of p n (x,t) and the weighted mixture fraction in mode 3 ( s ξ 3 ) ξ 3 p 3 )as and N e m f φ (ψ;x,t) p n (x,t) δ(ψ R - φ R n (x,t)) (6) n)1 R)1 p 1 t + ( u x i p 1 ) ) i x i( Γ p 1 t x i) + γ s p 3 - γp 1 (1 - p 1 ) (7) p 2 t + ( u x i p 2 ) ) i x i( Γ p 2 t x i) + γ s p 3 - γp 2 (1 - p 2 ) (8) s ξ 3 t + ( u x i s ξ 3 ) ) i x i ( Γ s ξ 3 t x i ) - γ s p 3 ( ξ 1 + ξ 2 ) + γp 1 (1 - p 1 ) ξ 1 + γp 2 (1 - p 2 ) ξ 2 (9) In these equations, p 3 ) 1 - p 1 - p 2 ; Γ t is the turbulent diffusivity; γ s and γ are the spurious dissipation rate and the micromixing rate, respectively; and ξ 1 and ξ 2 are the mixture fraction values in modes 1 and 2, respectively, which are constant and equal to their inlet values (i.e., ξ 1 ) 1 and ξ 2 ) 0). Note that the terms involving γ correspond to molecular mixing and, thus, generate the scalar dissipation rate in the transport equation for the mixture fraction variance, whereas the terms involving γ s are required to eliminate spurious scalar dissipation resulting from the finite-mode representation of turbulent diffusion. When the full PDF method is used, this spurious dissipation term is eliminated at the price of requiring a large ensemble of notional particles, as will be explained below. In the FM-PDF model, this term can be eliminated by setting γ s as follows γ s ) 2Γ t ξ 3 ξ 3 (10) 1-2 ξ 3 (1 - ξ 3 ) x i x i The term γ can be calculated by forcing the mixture

5134 Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 Table 1. Source Terms for Each Scalar mixture fraction (ξ) 0 jth moment of the CSD (m j) (0) j B(c A,c B) + jg(c A,c B)m j-1 reaction progress variable (Y) F3k vm 2G(c A,c B)/[ξ stc AoM] fraction variance to follow the exact transport equation. Then, according to eq 5 γ ) C φ ɛ k ξ 2 [p 1 (1 - p 1 )(1 - ξ 3 ) 2 + p 2 (1 - p 2 ) ξ 3 2 ] (11) where ξ 2 is the Reynolds-averaged mixture fraction variance and C φ was kept equal to 1. Concerning the choice of C φ, the effect of this constant on the CSD was discussed in ref 16, and this value was chosen with particular regard to the problem of dealing with nonfully developed turbulence. Moreover, it should be highlighted that, because the purpose of this work is the comparison of the two approaches, it is important to use the same value of C φ in both models. Concerning the other scalars, generally for mode n, it is possible to define a local concentration denoted by φ R n, whereas the weighted concentration s R n is defined by s R n ) φ R n p n (12) The transport equation for the weighted concentration of scalar R for mode 3 is s R 3 + ( u t x i s R 3 ) ) i x i( Γ s R 3 t x i ) - γ t p 3 ( φ R 1 + φ R 2 ) + γ s p 1 (1 - p 1 ) φ R 1 + γ s p 2 (1 - p 2 ) φ R 2 + p 3 S R ( φ 3 ) (13) where φ R 1 and φ R 2 are the local concentrations in modes 1 and 2, respectively, and S R ( φ 3 ) is the source term for scalar R. 3. Moment Method Approach As highlighted in the Introduction, the population balance must be solved in order to describe particle nucleation and growth. Using the moment method, the system can be solved in terms of the moments of the CSD. The first five moments are of particular interest, as they are related to the total particle number density (N t ) m 0 ), the total particle area (A t ) k a m 2 ), and the total solid volume (V t ) k v m 3 ) by shape factors (k a, k v ) that depend on particle morphology. Using this approach, the mean crystal size (L 43 ) and the solid concentration (c C ) can be written as L 43 ) m 4, c m C ) Fk v m 3 3 M (14) where F is the crystal density, k v is the volume shape factor, and M is the molecular weight of the crystal. Using this approach, a set of five transport equations must be solved. Depending on the approach used (i.e., full or presumed PDF), the implementation is slightly different. In Table 1, the source term that should be used for each scalar is reported. These terms must be included in eq 1 or eq 13. Note that, for the FM-PDF model, the transport equation for the mixture fraction has already been written (see eq 9). In Table 1, the nucleation rate [B(c A,c B )] and the growth rate [G(c A,c B )] appear, and they can be calculated once the reactant concentrations are known using the equations where c A c Ao ) ξ - ξ s Y, c B c Bo ) (1 - ξ) - (1 - ξ s )Y (15) ξ s ) c Bo c Ao + c Bo (16) and Y is the reaction progress variable whose source term is defined in Table 1. Note that Y ) 0 when ξ ) 0 and ξ ) 1. The nucleation rate is taken from literature 7 {2.83 10 10 c 1.775 [1/(m 3 s)] for c e 10 mol/m 3 (heterogeneous) B(c A,c B ) ) 2.53 10-3 c 15 [1/(m 3 s)] for c > 10 mol/m 3 (homogeneous) (17) where c ) c A c B - k s and k s is the solubility product of barium sulfate (at room temperature, k s ) 1.14 10-4 mol 2 /m 6 ). For the growth rate, a two-step diffusion-adsorption model 17 was used, which gives the following expression G(c A,c B ) ) 5.8 10-8 c s 2 ) k d (c A - c As ) ) k d (c B - c Bs ) (m/s) (18) where c s ) c As c Bs - k s and k d is the mass transfer coefficient. According to ref 18, this coefficient is sizedependent but remains nearly constant for particles smaller than 10 µm, notwithstanding the difference in solute. As a consequence, k d was fixed to fall in the range between 10-7 and 10-8 (m/s)(m 3 /mol). 7 More recent results 19 have shown that the limit of the Sherwood number Sh ) 2 is also valid for microparticles. Accordingly, the mass transfer coefficient can be calculated from k d ) ShDM L 43 F (19) For average particle sizes (between reactor inlet and outlet) of about 0.1 and 1.0 µm, k d is equal to 10-6 and 10-7 (m/s)(m 3 /mol), respectively. Concerning the shape factors, for the operating conditions under consideration, the crystals are flat, and the mean crystal size measured by a laser particle sizer is very close to the width of the crystal. In previous work, 20,21 shape factors using the width as the characteristic dimension were determined, and their effects on the CSD were studied. Nevertheless, the population balance works in terms of a unique characteristic dimension (i.e., L 43 ), and thus, equivalent shape factors for a system with equidimensional growth along each particle axis must be used. 22 Using this approach, it was found that k v ) 5.0 and k a ) 30.0. 23 Because the comparison with experimental data must be made in consistent units, the characteristic dimension of the population balance has to be converted to the dimension measured by the particle sizer, that is, the characteristic

Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 5135 dimension used in the former definition of shape factors, by using d 43 ) φ c L 43 (20) where φ c is a function of shape factors and was found to be equal to 3. 4. Experimental Setup and Numerical Simulation Simulations and experiments were carried out in a single-jet tubular reactor consisting of a small tube centered on the reactor axis with a 1-mm inner diameter, a 1.5-mm outer diameter, and a 0.168-m length. The tubular reactor has an inner diameter of 1 cm and a length of 2.1 m. The velocity ratio (VR) between the jet and the coaxial flow was kept equal to 1, and the macroscopic Reynolds number was near 10 000. Reactants were fed at several concentrations, with the reactant concentration ratio (R )c Bo /c Ao ) varied in the range 0.01-3 for c Ao ) 34.101 mol/m 3. The index A denotes that the reactant fed in the inner tube for which ξ ) 1. Likewise, B denotes the reactant fed in the annular region for which ξ ) 0. A detailed description of the experimental setup can be found in ref 23, and here, we limit our discussion to the final results. Barium chloride and sodium sulfate were alternatively fed in the inner tube and in the annular region, keeping the concentrations constant. Hereinafter, these two configurations are denoted by AB and BA, respectively. The flow field was modeled by using the k - ɛ model 24 in FLUENT (version 5.2), and standard wall functions were used for the wall boundary layer. The grid for the flow field simulation had 131 nodes in the axial direction and 35 nodes in the radial direction. This grid included the annular region before the injection zone (x o ) 168 mm with 20 nodes) so that the velocity profile in the inlet zone could be predicted correctly. This region was not considered in the full PDF simulation because no reaction occurs. Thus, the final grid for the full PDF simulation had 111 nodes in the axial direction and 35 nodes in the radial direction. The full PDF code was initialized by reading the axial and radial coordinates used in FLUENT and the steadystate solution of the flow field, including the mean velocities (u x and u y ), k, ɛ, the mean pressure, and the fluid density. A number of particles equal to n p n cell, where n p ) 100, were positioned in the computational domain, and each particle s flow properties were fixed by linear interpolation depending on its location. Particles were initialized at the inlet and were moved according to the velocity vector and the turbulent diffusivity assigned to them. Then, particles interacted with each other as a result of micromixing (see eq 4), where the local mean values were derived by taking the ensemble average of the scalar in each cell. (For details, see ref 25.) The changes in scalar composition due to chemical reaction, nucleation, and crystal growth were calculated for each particle by direct integration using a routine based on a fourth-order Runge-Kutta method. Concerning the FM-PDF simulations, transport equations for the mode probabilities p 1 and p 2, the weighted mixture fraction s ξ 3, the reaction progress variable s Y 3, and the five moments of the CSD s mj 3 (j ) 0,..., 4) were introduced in the CFD code as user-defined scalars. Once the flow field had converged, this set of equations was solved. Appropriate under-relaxation Figure 1. Full PDF prediction of concentration of reactant A (c Ao ) 34.101 mol/m 3, R)1). Figure 2. Full PDF prediction of concentration of reactant B (c Ao ) 34.101 mol/m 3, R)1). factors were used depending on the stiffness of the problem (i.e., initial reactant concentration). Note that, unlike the full PDF model (which is based on Monte Carlo methods), the FM-PDF model was solved in the CFD code itself based on a finite-volume discretization. With this approach, the grid had to be fine enough to avoid numerical diffusion. To ensure that the solution was grid-independent, different grids and different discretization schemes were tested, but for these operating conditions, no substantial differences were detected. 5. Results and Discussion In Figures 1 and 2, the reactant concentrations (c A, c B ) obtained with the full PDF code are reported. As one can see, reactant A is fed in the small tube, whereas reactant B is fed in the annular region. As a result of turbulent diffusion, the reactants mix together and react. Reactant B is in great excess, and thus, its concentration remains almost constant, whereas reactant A disappears, first because of mixing and later because of chemical reaction. The plot of the nucleation rate (see Figure 3) shows a peak near the injection point, which is due to very high supersaturation in this region.

5136 Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 Figure 3. Full PDF prediction of nucleation rate (c Ao ) 34.101 mol/m 3, R)1). Figure 5. Comparison between the two PDF approaches for the mean mixture fraction (ξ) and the intensity of segregation I s at (x - x o)/d ) 0.0. Figure 4. Full PDF prediction of mean crystal size (c Ao ) 34.101 mol/m 3, R)1). The nucleation rate reaches very high values in this region and then decreases rapidly. The plot of the mean crystal size (L 43 ) is shown in Figure 4. As can be seen, the mean crystal size increases until the final value is reached. It should be highlighted that the wavelike fluctuations in the scalar fields are caused by statistical errors, which can be improved by using more notional particles per cell or by averaging over several time steps. To compare results provided by different models that are based on different methods, it is useful to compare the mixture fractions and the intensities of the segregation profiles in different sections of the reactor. The mixture fraction is a nonreacting scalar, and thus, its mean value ( ξ ) and its variance ( ξ 2 ) are affected only by convection and turbulent diffusion. The intensity of segregation I s is defined by I s ) ξ 2 ξ (1 - ξ ) (21) and is an index of the degree of mixing; in fact, it is equal to 1 when the system is perfectly segregated and 0 when the system is perfectly micromixed. In Figures 5-7, the mean mixture fraction and the intensity of segregation versus the radial direction at several axial Figure 6. Comparison between the two PDF approaches for the mean mixture fraction (ξ) and the intensity of segregation I s at (x - x o)/d ) 1.0. positions [(x - x o )/D] are reported. A comparison is made between the time-averaged full PDF model predictions and the FM-PDF model predictions, and as can be seen, the predictions are quite similar. Concerning the reacting scalars, contour plots of the mean reactant concentrations, nucleation rate, and mean crystal size look similar, but differences were detected. For example, the FM-PDF model seems to underpredict the nucleation rate (as compared to the full PDF results) and, thus, the final particle number density. The value of the total number density is extremely sensitive to numerics because of the strong nonlinearity of the nucleation rate. However, as mentioned before, by testing different grids, the solution was shown to be grid-independent. Likewise, by using different discretization schemes, the approach used was shown to be adequate to handle the stiffness of the problem. In Figure 8, the total particle number density along the reactor axis is reported for c Ao ) 34.101 mol/ m 3 and R)1. The peak of m 0 after the injection zone predicted by all models is initially very high but decreases sharply. The comparison of the mean crystal sizes (see Figure 9) highlights the fact that, without any closure, this property is detectably underestimated.

Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 5137 Figure 7. Comparison between the two PDF approaches for the mean mixture fraction (ξ) and the intensity of segregation I s at (x - x o)/d ) 10.0. Figure 9. Comparison between model predictions for mean crystal size along the center line (c Ao ) 34.101 mol/m 3, R)1). Figure 8. Comparison between model predictions for total number density along the center line (c Ao ) 34.101 mol/m 3, R) 1). Notice that the full PDF predictions presented in the last two figures are time-averaged values. However, for other values of R, agreement between the FM-PDF and full PDF predictions was better. Radial-averaged results at the end of the reactor and a comparison with experimental data are presented below. Concerning the comparison with experimental data, the choice of the mass transfer coefficient (k d ) is crucial. In Figure 10, the FM-PDF predictions with different k d values are compared with experimental data for the two configurations (AB and BA). The values of k d reported in the legend can be obtained by using eq 19 and assuming average mean crystal sizes (i.e., L 43 ) of about 1, 0.1, and 0.01 µm, respectively. As one can see, the two configurations (AB and BA) result in different experimental CSDs. This has been explained by some authors in terms of activity coefficients 26 that should be included in the growth and nucleation rates or by invoking different growth mechanisms, as suggested in ref 27. It should be also highlighted that the models presented here are perfectly symmetric and, thus, are unable to recognize the difference between excess Figure 10. Comparison between FM-PDF predictions and the experimental mean crystal size (c Ao ) 34.101 mol/m 3 ). barium versus excess sulfate. Although the results show good agreement, they highlight the need for using a sizedependent mass transfer coefficient. In fact, the best agreement is obtained by using a different k d value for each point depending on the corresponding mean crystal size. In Figure 11, model predictions for the three models are compared with experimental data. As can be seen, the overall agreement is quite good, but in general, for these operating conditions, the effect of micromixing is small. At higher concentrations, the effect of micromixing should be greater, but aggregation begins to play an important role, at which point all of the models definitively underpredict the mean crystal size. However, at higher concentrations, the use of a micromixing model should be even more important, as it is needed for correctly predicting the local interactions between microcrystals close to the injection zone because of local segregation. 6. Conclusion Turbulent precipitation of barium sulfate in a tubular reactor was studied. Reactants were fed in a coaxial single-jet configuration. The effect of the reactant

5138 Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 Figure 11. Comparison between FM-PDF predictions and the experimental mean crystal size [c Ao ) 34.101 mol/m 3, k d ) 10-6 (m/s)(m 3 /mol)]. concentration ratio on the CSD was studied, and the predictions of the FM-PDF and full PDF approaches were compared with experimental data. Results show that the population balance treatment needs to be improved; in fact, the importance of using a sizedependent mass transfer coefficient was highlighted (i.e., size-dependent growth rate.) This can be handled by using the quadrature of moments method as explained in ref 5. Moreover, at higher concentrations both models fail in predicting the final mean crystal size because aggregation is neglected. However, the ability of this approach to describe turbulent precipitation appears to be extremely promising, and the use of a complete description of solid evolution (including aggregation) linked with the PDF description represents the natural development of this work. Acknowledgment This research was partially supported by an Italian National Research Project (MURST 40%sMultiphase reactors: Hydrodynamics analysis and solid-liquid analysis). Notation B ) nucleation rate, number/(m 3 s) c A ) barium chloride concentration, mol/m 3 c B ) sodium sulfate concentration, mol/m 3 c C ) solid concentration, kg/m 3 C µ ) turbulent constant C φ ) micromixing constant d 43 ) mean crystal size, m f φ (ψ;x,t) ) joint probability density function G ) crystal growth rate, m/s k ) turbulent kinetic energy, m 2 /s 2 k a ) surface shape factor k d ) mass transfer coefficient, (m/s)(m 3 /mol) k s ) barium sulfate solubility product, mol 2 /m 6 k v ) volume shape factor L 43 ) ratio of the fourth and third moments of the CSD, m m j ) jth moment of the CSD, m j-3 M ) barium sulfate molecular weight, kg/mol n cell ) number of cells in the computational domain n p ) number of notational particles per cell N e ) number of modes for the description of the FM-PDF model p n ) probability of mode n s ξ n ) weighted mixture fraction in mode n s R n ) weighted concentration of scalar R in mode n Sc t ) turbulent Schimdt number u i ) Reynolds averaged velocity in the ith direction, m/s Y ) reaction progress variable Greek Letters R)reactant concentration ratio Γ t ) turbulent diffusivity, m 2 /s γ ) micomixing rate, 1/s γ s ) spurious dissipation rate, 1/s ɛ ) turbulent dissipation rate, m 2 s 3 ɛ ξ ) scalar dissipation rate, 1/s ν ) kinematic viscosity, m 2 /s ξ n ) local mixture fraction in mode n ξ 2 ) mixture fraction variance ξ s ) stoichiometric mixture fraction F)crystal density, kg/m 3 τ φ ) micromixing time, s φ R n ) local concentration of scalar R in mode n φ c ) length shape factor Subscript o ) value in the inlet stream Operators ) expected value ) conditioned expected value Literature Cited (1) Baldyga, J.; Orciuch, W. Closure method for precipitation in inhomogeneous turbulence. Proceedings of the 14th Symposium on Industrial Crystallization, Sept 12-16, 1999; Institution of Chemical Engineers: Rugby, U.K., 1999; Paper 86, p 1069. (2) Marchisio, D. L.; Fox, R. O.; Barresi, A. A. Simulation of turbulent precipitation in a Taylor-Couette reactor using CFD. AIChE J. 2001, 47, 664. (3) Vanni, M. Approximate population balance equations for aggregation-breakage processes. J. Colloid Interface Sci. 2000, 221, 143. (4) Rivera, T.; Larson, M. A. A model for the precipitation of pentaerythritol tetranitrate (PETN). Ind. Eng. Chem. Process Des. Dev. 1978, 17, 183. (5) McGraw, R. Description of aerosol dynamics by quadrature methods of moments. Aerosol Sci. Technol. 1997, 27, 255. (6) Barresi, A. A.; Marchisio, D.; Baldi, G. On the role of microand mesomixing in a continuous Couette-type precipitator. Chem. Eng. Sci. 1999, 54, 2339. (7) Baldyga, J.; Podgorska, W.; Pohorecky, H. Mixing precipitation model with application to double feed semibatch precipitation. Chem. Eng. Sci. 1995, 50, 1281. (8) Fitchett, D. E.; Tarbell, J. M. Effect of mixing on the precipitation of barium sulphate in an MSMPR reactor. AIChE J. 1990, 36, 511. (9) Phillips, R.; Rohani, S.; Baldyga, J. Micromixing in a singlefeed semi-batch precipitation process. AIChE J. 1999, 45, 82. (10) Pope, S. Pdf methods for turbulent reacting flows. Prog. Energy Combust. Sci. 1985, 11, 119. (11) Dopazo, C. Probability density function approach for turbulent axisymmetric heated jet. Centerline evolution. Phys. Fluids 1975, 18, 397. (12) Baldyga, J.; Bourne, J. R. Comparison of the engulfment and the interaction-with-the-mean micromixing models. Chem. Eng. J. 1990, 45. 25. (13) Baldyga, J. Turbulent mixer model with application to homogeneous, instantaneous chemical reactions. Chem. Eng. Sci. 1989, 44, 1175. (14) Fox, R. O. The Lagrangian spectral relaxation model of the scalar dissipation in homogeneous turbulence. Phys. Fluids 1997, 9, 2364.

Ind. Eng. Chem. Res., Vol. 40, No. 23, 2001 5139 (15) Fox, R. O. On the relationship between Lagrangian micromixing models and computational fluid dynamics. Chem. Eng. Proc. 1998, 37, 521. (16) Piton, D.; Fox, R. O.; Marcant, B. Simulation of fine particle formation by precipitation using computational fluid dynamics. Can. J. Chem. Eng. 2000 78, 983. (17) Nielsen, A. E. Electrolyte crystal growth mechanisms. J. Crystal Growth 1984, 67, 289. (18) Nagata, S. Mixing Principles and Applications; Kodansha: Tokyo, Japan, 1975. (19) Armenante, P. M.; Kirwan, D. J. Mass transfer to microparticles in agitated system. Chem. Eng. Sci. 1989, 44, 2796. (20) Pagliolico, S.; Marchisio, D.; Barresi, A. A. Influence of operating conditions on BaSO 4 crystal size and morphology in a continuous Couette-type precipitator. J. Therm. Anal. Calorim. 1999, 56, 1423. (21) Marchisio, D. L.; Barresi, A. A.; Fox, R. O.; Baldi, G. Comparison of different modelling approaches to turbulent precipitation. In Proceedings of the 10th European Conference on Mixing, Delft, The Netherlands, July 2-5, 2000; van den Akker, H. E. A., Derksen, J. J., Eds.; Elsevier Science: New York, 2000; p 77. (22) Randolph, A. D.; Larson, M. A. Theory of Particulate Processes, 2nd ed.; Academic Press: New York, 1988. (23) Marchisio, D. L.; Fox, R. O.; Barresi, A. A.; Garbero, M.; Baldi, G. On the simulation of turbulent precipitation in a tubular reactor via Computational Fluid Dynamics. Proceedings of the 4th International Symposium on Mixing in Industrial Processes (ISMIP4), CD-ROM ed., Toulouse, France, May 14-16, 2001, paper 55; submitted to Chem. Eng. Res. Dev. (24) Fluent Inc. FLUENT 5 User s Guide; Fluent Inc.: Lebanon, NH, 1995. (25) Tsai, K.; Fox, R. O. PDF simulation of turbulent seriesparallel reaction in an axisymmetric reactor. Chem. Eng. Sci. 1994, 49, 5141. (26) Wei, H.; Garside, J. Application of CFD modelling of precipitation system. Chem. Eng. Res. Dev. 1997, 75A, 219. (27) Wong, D. C. Y.; Jaworski, J.; Nienow, A. W. Effect of ion excess on particle size and morphology during barium sulphate precipitation: An experimental study. Chem. Eng. Sci. 2001, 56, 1. Received for review December 1, 2000 Revised manuscript received April 4, 2001 Accepted April 26, 2001 IE0010262