investigation of discrete population balance models and...
TRANSCRIPT
Published: August 07, 2011
r 2011 American Chemical Society 11358 dx.doi.org/10.1021/ie2006033 | Ind. Eng. Chem. Res. 2011, 50, 11358–11374
ARTICLE
pubs.acs.org/IECR
Investigation of Discrete Population Balance Models and BreakageKernels for Dilute Emulsification SystemsPer Julian Becker, Franc-ois Puel, Reynald Henry, and Nida Sheibat-Othman*
Universit�e de Lyon, Universit�e Lyon 1, CNRS, CPE Lyon, UMR 5007, Laboratoire d’Automatique et de G�enie des Proc�ed�es (LAGEP),43 Bd du 11 Novembre 1918, F-69622 Villeurbanne, France
ABSTRACT: A novel in situ video probe with automated image analysis was used to develop a population balance model for abreakage-dominated liquid�liquid emulsification system. Experiments were performed in a 2 L tank, agitated by an axial flowpropeller. The dispersed phase (ethylene glycol distearate) concentration was varied from 0.2 to 1.0% (w/w), and agitation rateswere varied from 0.2 to 0.5W/kg, in the presence of excess surfactant. Three numerical discretizationmethods were compared: fixedpivot, cell average, and finite volumes. The latter was then chosen for the subsequent simulations due to its rapidity and higherprecision. An investigation of the different theories for bubble/droplet breakage was done and the frequencies (or breakage ratekernels) were compared. Four models were found applicable: the models developed by Coulaloglou and Tavlarides (Coulaloglou,C. A.; Tavlarides, L. L. Chem. Eng. Sci. 1977, 32, 1289); Sathyagal and Ramkrishna (Sathyagal, A. N.; Ramkrishna, D. Chem. Eng. Sci.1996, 51, 1377); Alopaeus, Koskinen, and Keskinen (Alopaeus, V.; Koskinen, J.; Keskinen, K. I.Chem. Eng. Sci. 1999, 54, 5887); andBaldyga and Podgorska (Baldyga, J.; Podgorska,W.Can. J. Chem. Eng. 1998, 76, 456). The one by Sathygal and Ramkrishna includedthe daughter size distribution. A log-normal daughter size distribution was chosen for the models by Coulaloglou and Tavlarides andAlopeus et al. Also, a normal distribution was used in the model by Baldyga and Podgorska. These models were compared with theexperimental data to allow parameter identification. The model by Baldyga and Podgorska was found to give the best prediction ofthe shape of the distribution, its mean diameter, and standard deviation.
1. INTRODUCTION
Many processes used across the chemical, food, cosmetics, andpharmaceutical industries involve two-phase interactions. Thesecan be gas�liquid, used, for example, in bubble columns for ab-sorption processes, solid�liquid (e.g., crystallization and emul-sion/suspension polymerizations), or liquid�liquid for pharma-ceutical, cosmetic, or alimentary preparations.
In this work, we will be interested in liquid�liquid emulsifica-tions for cosmetic and alimentary preparations. In such pro-cesses, the quality of the emulsion is importantly related to thedroplet size distribution (DSD) which, in many cases, is non-Gaussian. Therefore, the emulsion cannot be characterized by amean diameter or a few moments. This is the case, for instance, ofhigh oil in water content emulsions (>80%). Since irregular pack-ing of uniform spheres cannot exceed a density limit of 63.4% (andmaximal regular packing is 74%), it can easily be seen in this casethat theDSD is either bimodal or very large. Therefore, populationbalance equations (PBEs) should be used to keep track of the fullnumber density distribution of the droplet size.1,2 The PBE canalso be embedded within flow fields to take into account spatialdistribution of the droplets.
A number of different processes have to be taken into accountfor themodeling of birth, death, and growth of droplets in emulsi-fication systems. These are diffusion, coagulation, and breakage.Diffusion, i.e., mass transfer between the dispersed and the con-tinuous phases, resulting in Ostwald ripening depends on thesolubility of the dispersed phase in the continuous phase. It isnegligible if the dispersed phase is insoluble in the continuousphase. The breakage and the coalescence rates as well as thedaughter droplet size distribution are related to the reactorhydrodynamics and properties of the dispersed phase.
Interesting reviews and analysis of breakage kernels in turbu-lent flows are given by Lasheras et al.,3 Patruno et al.,4 and Liaoand Lucas.5 It appears that the main causes for breakage arerelated to the turbulence where different models were developedbased on different assumptions each.6�20 Liao and Lucas5 alsoclassified the daughter size distribution resulting from breakageas empirical, physical (Bell-shape,15 U-shape,20 M-shape16) andstatistical models (normal distribution,6 uniform distribution13).They recommend using physical models and postulate that theM-shape daughter size distribution is more reasonable. Thesemodels will be investigated more deeply in sections 4.2 and 4.3.Similarly, Liao and Lucas21 proposed a good review of coales-cence kernels. In physical models, the aggregation kernel is givenby the product of the collision frequency and the coalescenceefficiency. The collision frequency can be induced by viscousshear, body forces, turbulence wake entrainment, and capture inturbulent eddies, with turbulent collision as a dominant phe-nomena in turbulent dispersions (see Liao and Lucas21 andreferences therein). The coalescence efficiency can be obtainedeither by the film drainage model13 or energy model.22 Liao andLucas21 again recommend modeling based on physical observa-tions (droplet size, liquid properties, and turbulent parameters)and including all potential mechanisms in the model.
These breakage and coalescence models are in theory valid forfluid particles: droplets and bubbles as they are comparable innature. However, the assumption that drops behave like gas
Received: March 25, 2011Accepted: August 6, 2011Revised: August 3, 2011
11359 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
molecules has some limitations. Droplets have higher density andviscosity than bubbles which slow down the breakage processcompared to bubbles (resulting in longer times to reach equilibrium).In addition, the size of bubbles is generally of the order to fewmillimeters while interesting emulsions are micrometer sizedor smaller. As a result, higher energy levels are needed to breakdroplets (that are smaller, more viscous, and denser). However,models that are initially developed for gas�liquid breakage donot account for the viscosity of the dispersed phase. Therefore,they are not directly applicable to liquid�liquid breakage.
Another challenge in using the PBE resides in the resolution ofthis equation. With the above reviewed breakage and aggregationkernels, the PBE does not have an analytical solution. In fact, itcan be solved analytically (for instance by the method of char-acteristics) only for very simple forms of breakage or coalescence.23
Different numerical methods were therefore used to solve the PBEincluding moment methods,24 stochastic methods such as MonteCarlo simulations,25 and discretization methods such as finite ele-ment methods,26 finite volumes, and sectional methods.27 Kostoglouand Karabelas28 note that, even after 35 years of development,there is still a large scope for further innovation in the area andthat numerical methods have a limited range of validity. Themoment methods allow reducing the calculation time but giveonly certain integral properties of the distribution. Stochasticmethods are known to be less computationally expensive formultidimensional PBE. Discretization methods allow calculatingthe full distribution and give the possibility to use linear or in-homogeneous grids. Using geometric grids has the advantage ofreducing the computational effort while ensuring accuracy.29
In finite element methods, the solution is approximated aslinear combinations of piecewise basis functions, which moti-vates their use over complicated domains. Their implementationis not straightforward, unless using specific commercial softwares.The finite volumes (FV) method is widely used to solve the PBE.It was adapted by Fibert and Laurenc-ot30 to solve aggregationproblems. In their formulation, the number density of the PBE istransformed to a mass conservation law. Therefore, the method isconsistent with respect to the first moment but does not ensuregood predictions of the zerothmoment. The fixed pivot (FP) tech-nique proposed by Kumar and Ramkrishna29 is consistent with thefirst two moments of the distribution in aggregation and breakageproblems (without growth or nucleation terms). It is simple toimplement and is computationally attractive. The authors notehowever that the method over predicts the particle number in thelarge size range due to sharp variations in the density functions. Theypropose themoving pivot technique to overcome the over predictionproblem, but this method is more difficult to implement and tosolve.31More recently, Kumar et al.32 proposed the cell average (CA)technique for aggregation problems. It was extended to aggregationand breakage problems by Kumar et al.33 In this method, the averagesize of the newborn particles in a cell is calculated and particles areassigned to neighboring nodes such that the properties of interest arepreserved. It was found to give less over prediction in the numberdensity distribution compared to the FP technique but is supposed tobe more computationally expensive. Note that the CA techniquereduces to FP for linear grids. Many other varieties of disretizationmethods were developed.Making a comprehensive review of numer-ical techniques is however out of the scope of this work.
In this work, we consider a particular process of emulsificationand study the validity of differentmodel kernels and different numer-ical resolutionmethods. An in situ video probewith automated imageanalysis is used to measure online the DSD in a 2 L stirred tank
reactor. The dispersed phase concentration (ϕ) is varied from0.2 to 1.0 wt % and the average energy dissipation rate (ε) from0.2 to 0.5 W/kg, resulting in emulsions made up of droplets withdiameters between 20 and 100 μm. Being turbulent, the processis modeled by different kernels adapted to turbulent dispersions,and the modeling results are compared. Also, three differentnumerical solution schemes are used to discretize the PBE:a finite volumes scheme,30 the fixed pivot,29 and cell average32
techniques. Ultimately, a combination of breakage and coales-cence kernels with a numerical scheme providing accurate pre-dictions at moderate computational cost is desired. Such a modelcan then be used for integration into a computational fluiddynamics simulation to account for local variations of energydissipation and shear rates in more complex geometries and thusprovide an in-depth understanding of the emulsification system.
The paper is organized as follows: The materials and experi-mental setup are presented in section 2. Section 3 presents theexperimental results. The relevant PBE is introduced in section4.1. Models for the breakage and the coalescence rates as well asthe daughter droplet size distribution are reviewed in sections4.2�4.4, and a preselection of the most applicable kernels ismade. The available numerical methods and their applicability tothis work are briefly discussed in section 5. Comparison toexperimental data is done in section 6.
2. MATERIALS AND METHODS
2.1. Materials. A model oil-in-water (O/W) emulsion, madeup of ethylene glycol distearate C38H74O4 (EGDS) and distilledwater as dispersed and continuous phases, respectively, is used inthis study. EGDS was supplied byWako Chemicals. It is a cosmeticingredient which is generally used to enhance aspects such aspearlescence, transparency, or color in a wide range of personalcare formulations.34 It is almost insoluble inwater. A summary of therelevant physical properties of EGDS is shown in Table 1.The emulsions are stabilized using the surfactant tricosaethy-
lene glycol dodecyl ether C12E23 (Brij 35) supplied by Fluka. Itsmelting point is 38�41 �C. Its HLB is 16.9 at ambient tempera-ture (which ensures the formation of oil-in-water emulsions at70 �C). The EGDS to surfactant weight ratio was fixed at 2:1.2.2. Experimental Setup. Emulsifications were carried out at
a controlled temperature of 70 �C, in a jacketed 2 L vessel. Thereactor is equipped with a condenser cooled with water to pre-vent evaporation of the water. Agitation was provided by an axialflow profiled three blade Mixel TT propeller (Np = 0.8); whilefour equally spaced baffles were used to avoid vortex formation.The speed of the impeller (ω) was adjusted to provide the
Table 1. Physical Properties of Model Emulsification System,i.e., EGDS-in-Water with Excess Surfactant (Brij 35) at 70 �C
parameter symbol value
EGDS density Fd 858.2 kg/m3
water density Fc 977.7 kg/m3
EGDS dynamic viscosity a μd 0.01 Pa s
water dynamic viscosity μc 0.0004 Pa s
EGDS�water interfacial tensionb σ 0.005 62 N/m
EGDS melting pointc 58�65 �CaMeasured by Couette rheometer. bMeasured by the pendent droptechnique with a Kr€uss DSA10 tensiometer. cMeasured with a TA100differential scanning calorimeter.
11360 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
required specific average energy dissipation rate (ε) according tothe well-known power number equation:
ε ¼ Pm0
¼ Npω3DI5
m0
Table 2 shows the most important reactor dimensions. A detaileddescription of the experimental setup can be found in Khalil et al.35
2.3. In Situ Video Monitoring. An in situ video probe EZProbe-D25 L1300 with automatic image analysis was used tomonitor in situ the transient DSD (50 images/s). It was located at5 cm above the stirrer, close to the stirring shaft, with a vertical angleof about 30�. At this point, the flow moved downward before beingagitated by the stirrer. This position is one of the probe locationsrecommended in the literature for this impeller.36 Image analysis isbased on a circular Hough transform.37 Its application on dropletsize measurement is described in detail by Khalil et al.35 A sampleimage with the detected circles is shown in Figure 1. This probeallowed real time acquisition of 2D images of the droplets generatedduring the emulsification, whereas the automatic image analysistreatment was performed in delayed time.2.4. Operating Conditions. The EGDS was first melted and
dispersed into water containing the surfactant under low agitation(0.03W/kg) during reactor heating to the operating temperature of70 �C. Then it was left to rest for 30 min, after which an initialdistribution of the emulsion was generated by subjecting the systemto a short burst (5�10 s) of high agitation (5W/kg). Agitation wascontinued then at the desired rate. A total of 9 emulsificationexperiments with EGDS concentrations between 0.2 and 1.0% andagitation rates between 0.2 and 0.5 W/kg were performed. Treat-ments of the in situ video camera images were conducted atincreasing intervals of 5�15 min up to 60 min and intervals of
20�40 min up to 300 min. See Table 3 for a summary of theexperimental operating conditions.
3. EXPERIMENTAL RESULTS
The evolution of the DSD of a typical run, n(t,x), is shown inFigure 2. The Kolmogorov microscale, de,min, which defines thesize of the smallest eddy and thus the size of the smallest stabledroplet,38 is given by the following equation:
de, min ¼ μc3
Fc3ε
!1=4
ð1Þ
The DSD is consistently bell-shaped and monomodal, as alreadyreported in the literature.39 Thus, a log-Gaussian distribution isused for interpolation of the experimental data to obtain theinitial condition in the subsequent simulations. The distributionmoves to lower sizes over time, due to droplet breakage, alwaysstaying above the Kolmogorov microscale. The DSD is alsofound to become narrower with time.
Results for the mean and standard deviation of the dropletdiameter are shown in Figures 3 and 4, respectively. The first threemoments were used to estimate the mean diameter (also known asd10) and the standard deviation of the experimental data:
d10 ¼ μ1μ0
ð2Þ
stdev ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiμ2μ0
� μ1μ0
!2vuut ð3Þ
Table 2. Summary of Key Dimensions of the 2 L AgitatedReaction Vessel Used for the Emulsification Experiments
description factor dimension
internal vessel diameter T 0.15 m
vessel working volume VT 2 L
impeller diameter DI = 3/5T 0.088 m
blade thickness DI/50 1.8 mm
number of blades 3
height of emulsion ≈ T ≈ 0.15 m
impeller location (from bottom) T/3 0.05 m
number of baffles 4
baffle width T/10 0.015 m
Figure 1. Experimental image with circles detected by the Hough transform (left), where the scales are in pixels (1px =1.84 μm) and effective measurednumber DSD (right).
Table 3. Design of Experiments for Energy Dissipation Rate(ε) and Dispersed Phase Concentration (O)
run ε [W/kg] ϕ [%w/w]
1 0.2 0.2
2 0.2 0.5
3 0.2 1.0
4 0.35 0.2
5 0.35 0.5
6 0.35 1.0
7 0.5 0.2
8 0.5 0.5
9 0.5 1.0
11361 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
where the moments of order k (μk) are given by24
μk ¼Z ∞
0dknðt, dÞ∂d≈ ∑
N
i¼ 0diknðt, diÞ ð4Þ
The initial DSDs and thus the values for d10 and the standarddeviation are very similar for all experiments. The spread of 55�65μm inmean diameter is due to the fact that the firstmeasurementwas taken at 5 min after agitation was started. The profiles rapidlydecrease at the beginning and flatten out, asymptotically approach-ing an equilibrium value. At high agitation (ε = 0.35 and 0.5W/kg),61% of the final mean diameter and 72% of the final standarddeviation is reached after 60 min. This progression is much slowerfor the low agitation case (ε = 0.2W/kg), where only 31% and 52%of the final mean diameter and standard deviation, respectively, arereached after 60 min.
As expected, increased agitation rates produce smaller drop-lets and narrower size distribution. The variation of mean dia-meter and standard deviation due to a change in EGDS con-centration was found to be much less visible for the range ofconcentrations used in this study. Any variation is most likelydue to experimental or image treatment errors. However, more
visible influence is expected for more elevated dispersed phaseconcentrations as aggregation phenomena gain importance.These data sets will be used to assess the validity of differentmodel kernels with emulsions made up of micronic droplets andto test several numerical resolution methods.
4. POPULATION BALANCE MODELING
4.1. Population Balance Equation. The PBE for a liquid�liquid system (i.e., absence of nucleation) in a homogeneousbatch reactor is given by eq 5.1,2 This equation is continuous forthe number density distribution, n(t,x), where x is the dropletsize in volume or diameter. In fact, the variable x can denote anyrelevant variable as a function of the application.
∂nðt, xÞ∂t
¼ ∂½Gðt, xÞnðt, xÞ�∂x
þ BBrðt, xÞ �DBrðt, xÞ
þ BAgðt, xÞ �DAgðt, xÞ ð5Þwith
BBrðt, xÞ ¼Z ∞
xbðx, εÞSðt, εÞnðt, εÞ dε ð6Þ
DBrðt, xÞ ¼ SðxÞnðt, xÞ ð7Þ
BAgðt, xÞ ¼ 12
Z x
0βðx, x� εÞnðt, x� εÞnðt, εÞ dε ð8Þ
DAgðt, xÞ ¼ nðt, xÞZ ∞
0βðx, εÞnðt, εÞ dε ð9Þ
As mentioned in section 2, EGDS has a negligible solubility inwater. Therefore, mass transfer between the dispersed and thecontinuous phases is considered insignificant. The growth (ordissolution) (∂[Gn]/∂x) term, (∂[Gn]/∂x), is therefore equal tozero. Droplet breakage and coalescence are the only phenomenawith a significant influence on theDSD. The birth and death ratesdue to the aggregation of particles, BAg and DAg, are governed bythe aggregation rate kernel, β(xi,xj), which is the product of thefrequency of interparticle collisions and their efficiency. Thebirth and death rates due to breakage of particles, BAg and DAg,
Figure 3. Measured mean diameter evolution for all experimental runs.
Figure 4. Measured standard deviation evolution for experimental runs.Figure 2. DSD evolution for one experimental run with ϕ = 0.5% EGDSand ε = 0.35W/kg, including the Kolmogoroffmicroscale (vertical line).
11362 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
are governed by the breakage rate kernel, S(xi); i.e., the breakagefrequency and the daughter droplet size distribution, b(xi,xj).4.2. Breakage Kernels. A large number of different theories
describing droplet and/or bubble breakage rates have beendeveloped during the last few decades; the major part of whichare discussed in detail in the review by Liao and Lucas.5 Fivemainclasses of breakage theories for turbulent regimes are defined inthis review: (category 1) energy transmitted to particle > criticalvalue; (category 2) velocity fluctuations around particle > criticalvalue; (category 3) bombarding eddy energy > critical value;(category 4) inertial force of eddy > interfacial force of smallestdaughter particle; and (category 5) combination of categories3 and 4.Other theories for breakage processes in systems dominated
by viscous shear are also discussed in Liao and Lucas.5 However,they are not applicable in this work and thus not discussed. Aselection of the most recent and popular kernels falling into oneof each of the five classes, which were considered as potentialcandidates for the system under investigation, is summarized inTable . The equation (or set of equations) constituting thebreakage rate function is given together with associated daughtersize distribution, a list of the empirical parameters, as well as the
conditions (e.g., bubbles/droplets) for which the models wereoriginally developed for.All of the theories presented below assume a locally isotropic
turbulent flow field (i.e., Re > 104) and droplet\bubble sizeswithin the inertial subrange.5,6 With the exception of the modeldeveloped by Baldyga and Podgorska9 (that considers a multi-fractal approach to describe the intermittent nature of turbulentenergy dissipation), all of the models presented in Table arebased on Kolmogorov’s theory of turbulence; i.e., they do nottake fluctuations of ε about its mean value into account.38 It isalso assumed that only eddies with a size less than or equal to thedroplet itself can cause breakage. The lower limit of the inertialsubrange (de,min) is defined as the Kolmogorov microscale(eq 1). The upper limit (Li) is of the order of the impeller size(Di). Baldyga et al.
40 give the relation Li = 0.05Di for Rushtonturbines and state that the upper limit is much higher for axialflow propellers, similar to the one used here, but acknowledgethat this is much less well understood. On the basis ofthis reasoning, the upper limit is estimated as Li = 0.4Di.The expression for the multifractal scaling exponent (α) canbe modified to eq 10 if the dispersed phase viscosity is notnegligible.9,19,40
αd ¼ 3
ln 20:16μd
Fcε1=3Li1=3dþ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi0:16μd
Fcε1=3Li1=3d
!2
þ 0:35σ
ε2=3Li2=3Fcd
vuut264
375�1
8>><>>:
9>>=>>;
lnLid
� � ð10Þ
It is important to note that the breakage kernels for categories 1and 2 have been developed independently from the daughter sizedistribution and generally use simple statistical models which areto an extent interchangeable (see section 4.3). The more recentmodels, developed for categories 3, 4, and 5, on the other hand,were developed together with a specific daughter distribution.Some of these models such as the ones developed by Luo andSvendsen,14 Lehr et al.,16 Zhao andGe18 derive the total breakagerate, S(xi), and daughter distribution, b(xi,xj), from a partialbreakage rate, S(xi,xj), which is directly describing the breakagefrequency of particle of size xi into two daughter particles, withone of size xj and can therefore not be separated from thedaughter size distribution. The models by Luo and Svendsen,14
Wang et al.,17 and Zhao and Ge18 have been developed directlyfrom theoretical considerations and thus do not rely on empiricalparameters. In fact, it has been noted by Lasheras et al.3 that thechoice of integration limits used in these models has a significantinfluence on the model and can thus be considered as tuningparameters.Figure 5 shows some of the breakage kernels, covering all five
categories using the physical parameters listed in Table 1, withthe originally proposed empirical parameters where applicable,ε = 0.35 W/kg and ϕ = 0.5%, up to a droplet size of 1 mm on alogarithmic scale. The predicted breakage rates differ by a num-ber of orders of magnitude for the different models. Even thoughthe models of Luo and Svendsen, Zaho and Ge, and Lehr et al.give higher breakage frequency than Coulaloglou and Tavlaridesand Alapaeus et al. for high sizes, these models rapidly becomeinsignificant for smaller droplets. It clearly shows that the fourmodels in categories 1 and 2, i.e., Coulaloglou and Tavlarides,Alopaeus et al., Sathyagal and Ramkrishna, and Baldyga and
Podgorska are the most applicable ones for the system underinvestigation. These four models are to be tested against the ex-perimental data. Note that these models were originally devel-oped for O/W droplets of micrometer size, except the model ofCoulaloglou and Tavlarides where the droplets varied from 0.1 to1 mm. The models that are originally developed for bubbles donot correctly represent the breakage in our system. Even thoughthe model of Luo and Svendsen was originally developed andvalidated for bubbles of size 3�6mm andO/W emulsions of size200 μm, it is not applicable in our system if we introduce theaverage energy dissipation rate in the model. A very high energydissipation rate, as present in the impeller region, is required intheir model tomake the breakage rate significant in liquid�liquidsystems. While the mean energy dissipation rate is easily acquired,an in-depth analysis of the hydrodynamics of the reactor is nec-essary to obtain an estimate for the maximum energy dissipationrate and the concerned zone volume, which would allow using theabove models. Sathyagal and Ramkrishna8 also noticed that themodel of Tsouris andTavlarides20 (and thus all themodels that aresubsequently derived from it) requires using the energy dissipa-tion rate in the impeller regionwhile themodel ofCoulaloglou andTavlarides uses the average dissipation energy to give breakagerates of the same order of magnitude. Coulaloglou and Tavlaridesnote however that breakage is predominant in the impeller regionwith a dissipation rate of 70ε but using this value requires correctlyevaluating the impeller region. The model by Baldyga and Pod-gorkska seems to be the only one to take geometry into account, byincluding the macroscale of turbulence (Li) as well as two param-eters directly related to the impeller zone where most of the break-age is expected to happen: B1 and B2. In fact, if it is assumed thatall kinetic energy is dissipated in the impeller zone, this is reduced
11363 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
Table4.
Summaryof
BreakageRateKerneland
AssociatedDaughterDistributionMod
els(D
ueto
Turbulent
FluctuationandCollision)
a
ref/assumptions
breakage
ratekernel
associated
daughtersize
distrib
ution
developedfor
empiricalparameters
(1)e particle>e c
Coulaloglou
andTavlarid
es6
SðdÞ¼
C1d
�2=3
ε1=3
1þ
ϕexp
C2σ
F dE2
=3d5
=3
"#
βdistrib
ution:
e.g.,K
onno
etal.41:
bðdi,d jÞ¼ 1 d j
Γða
þbÞ
ΓðaÞ
ΓðbÞ
d i d j !ðα
�1Þ=
3
1�
d i d j !1=3
0 @1 Ab
�1
(aandb=parameters)
O/W
droplets
(0.1�1
mm)
turbineimpeller
ω=190�
310rpm
ϕ=2.5�
15%
VT=12
L
C1,C2
Baldyga
andPo
drgorska
9
basedon
interm
ittentn
ature
ofturbulence
SðdÞ¼
0:0035B 1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
lnL i d��
sB2ε
ðÞ1=
3
d2=3
Z αd
0:12
d L i��ðα
þ2�
3fðαÞÞ=
3
dα
O/W
droplets
(100�5
00μm)
turbineimpeller
ε=0.1�
0.6W
/kg
ϕ=1%
VT=2.7�
21.2L
B 1,B
2
with
themultifractalscalingexponent
fornegligibleμd:
αd¼
5 2ln
L iðB
2εÞ0
:4F c
0:6
0:23σ0:6
!
lnL i d��
norm
aldistrib
ution:
e.g.,V
alentasetal.(1966)
bðVi,VjÞ
¼2c
Vjffiffiffiffiffiffi 2πp exp
�ðV
i�0:5V
jÞ2ð2c
Þ22V
j2
!
(c=tolerance)
andtheparabolic
approximationforthemultifractalspectrum
:c
fðαÞ¼
1�ðα
�1:117Þ2
0:468
log-norm
aldistrib
ution
(2)υ>υc
Alopaeusetal.48
taking
surfaceforce(σ)and
viscousforce(μ
d)into
account.
SðdÞ¼
A1ε1
=3erfc
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
A2
σ
F cε2
=3d5
=3þ
A3
μd
ffiffiffiffiffiffiffiffiffi
F cF d
pε1
=3d4
=3
s2 43 5
bðVi,VjÞ
¼1
Viffiffiffiffiffiffiffiffi
ffiffiffiffi2π
σg2
p
exp
�lnðV
i�ln
Vj 2�� þ
σg2
�� 2
2σg2
2 6 6 6 6 43 7 7 7 7 5
(σg=geom
etric
standard
deviation)
O/W
droplets(μm)
turbineimpeller
ω=378�
765rpm
ϕ=40%
VT=50
L
A1,A2,A3,a,b
Sathyagaland
Ram
krishna
(1994)
self-similarityinverseproblem
basedon
Narsimhanetal.
(1984)
SðVÞ¼
S 1
ffiffiffiffiffiffiffiffi σ F cV
rexp
�S 2
ln2
We
V DR3
�� 5=9
μc
μd ! 0:2
2 43 5
8 < :þS 3
lnWe
V DR3
�� 5=9
μc
μd ! 0:2
�2 4
9 = ;
bell-shaped
(cum
ulative):
b cumðV
i,VjÞ
¼
SðVjÞ
SðViÞ
! α
1�α 4
�� þ
α 4
SðVjÞ
SðViÞ
! 4
with
lnðαÞ¼
S 4ln
μc
μd ! �
S 5
O/W
droplets(μm)
Rushton
turbine
ω=350—
700rpm
ϕ=0.58%
VT=1.7L
S 1,S
2,S 3,S
4,S 5
11364 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
Table4.
Con
tinu
edref/assumptions
breakage
ratekernel
associated
daughtersize
distrib
ution
developedfor
empiricalparameters
(3)e eddy>e c
Martin
ez-Bazan
etal.15,
42
kinematicsof
fully
developed
turbulentflow
s
SðdÞ¼
Kg
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
βεdðÞ2=
3�12
σ F cd
rd
bell-shaped:
bðdi,d jÞ¼
Sðdi,d jÞ
Z d j 0Sðλ
,djÞd
λ
bubbles(0.1�3
mm)
(extendedto
L/L
system
sby
Eastwood
etal.2000)
turbulentwaterjet
ε=25�3
000W/kg
Kg,β
TsourisandTavlarid
es20
idealgastheory,based
on
Prince
andBlanch1
3
taking
into
accountthe
increase
inthesurface
energy
tocalculateE c
SðdÞ¼
0:0118
DFðjÞε1
=3Z 2=
λ min
2=d i
2 kþ
d
�� 2
8:2k
�2:3þ
1:07d2
=3�
� 1=2k2
exp
�e cðdÞ
1:3eðλÞ
�� dk
U-shaped:
bðdi,d jÞ¼
ε minþ
ε max�εðd
Þ�
�Z d j 0
ε min
ðÞþ
ε max�εðd
Þ�
� δd
O/W
droplets
(0.3�0
.5mm)
Rushton
turbine
ω=270�
330rpm
ϕ=10�3
0%
VT=0.75
L
integrationlim
its,
ε min,ε
max
turbulence
damping
factor:
DFðϕÞ¼
1þ
2:5ϕ
μdþ
0:4μ
c
μdþ
μc
!
"# 2
andcriticalenergy:
e cðdÞ
¼πσ 2
2d 21=3
�� 2
þd m
ax2þ
d min2�2d
2
!
LuoandSvendsen
14
idealgastheory,takinginto
accounttheincrease
inthe
surfaceenergy
tocalculateE c
daughterdistrib
ution
derived
from
partialbreakup
frequency
SðdjÞ
¼0:5Z d j 0
Sðdi,d jÞd
d iwith
Sðdi,d jÞ¼
0:923ð1
�ϕÞ
ε d j2 ! 1=3
Z d j λ min
1þ
λ d j
! 2
λ d j !11=3
Pðd j,λÞd
λ
U-shaped:
bðdi,d jÞ¼
Sðdi,d jÞ
SðdjÞ
O/W
droplets
(200
μm)
pipelineflow
ε=70�3
50W/kg
v cont=2.22
m/s
and
airbubbles(3�6
mm)
pipelineflow
ε=0.5�
1W/kg
v cont=3.98
m/s
integrationlim
its
Pðd j,λÞ¼
exp
�e cðd i
,djÞ
eðλÞ
!
e cðd i
,djÞ
¼πd j2σc f
b
11365 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
Table4.
Con
tinu
edref/assumptions
breakage
ratekernel
associated
daughtersize
distrib
ution
developedfor
empiricalparameters
(4)F e
ddy>F σ
Lehr
etal.16
expressing
integralassum
of
incompletegammafunctio
ns
SðdÞ¼
0:5d5
=3ε1
9=15F c
7=5
σ7=5
exp
�ffiffiffi 2pσ9=5
d3F c
9=5ε6
=5
!
M-shaped:
Bðd i,d
jÞ¼
6
πd i3ffiffiffi πp
exp
�9 4ln
222
=5d jF c
3=5 ε
2=5
σ3=5
"#
!
1þ
erf3 2ln
21=15d jF c
3=5 ε
2=5
σ3=5
!
"#
()
bubbles(m
m�c
m)
bubblecolumn
v gas=0.02�0
.1m/s
none
(5)e eddy>e candF e
ddy>F σ
Wangetal.17
extensionof
energy
constraintsmodelfrom
LuoandSvendsen
14by
adding
capillary
constraint
from
Lehr
etal.16
SðdiÞ
¼0:5Z d i 0
Sðdi,d jÞd j
with
Sðdi,d jÞ¼
0:923ð1
�jÞε1
=3Z d j λ m
in
d jþ
λ
2λðÞ11
=3
Pðd j,λÞd
λ
M-shaped:
bðdi,d jÞ¼
Sðdi,d jÞ
SðdjÞ
airbubbles(m
m)
pipelineflow
ε=0.5�
1W/kg
v=3.98
m/s
integrationlim
its
Pðd j,λÞ¼
Z ∞ 01
f bv,min�f bv,max
1eðλ
Þexp
�e c eðλÞ
�� d
e c
andminimum
andmaximum
breakage
fractio
ns:
f bv,min
¼πλ3σ
6eðλÞd i
!
and
f bv,max
¼fðc
f,maxÞ
with
maximum
increase
insurfaceenergy:
c f,m
ax¼
min
21=3�1
�� ,
eðλÞ
πd i2σ
��
ZhaoandGe1
8
extensionof
Wangetal.17
byintroducingeddy
efficiency,C
ed
daughterdropleto
fsam
esize
ascolliding
eddy
SðdiÞ
¼0:923ð1
�jÞε1
=3Z d i λ m
in
d iþ
λð
ÞλðÞ11
=3
exp
�e cðd i
λ
e iðλÞ
�� dλ
M-shaped:
bðdi,d jÞ¼
Sðdi,d jÞ
SðdjÞ
airbubbles(m
m)
pipelineflow
ε=0.5W/kg
ϕ=5%
integrationlim
its
with
criticalenergy:b
e cðd j
,λÞ¼
max
c fπd j2σ
Ced
,πσλ
3djmin
f bv,1
�f bv
�� 1=3
0 @1 A
aAllmodelsassumedropletsizeto
bewith
intheinertialsubranged e
,min<d<L i.A
llmodelsassumeaspatially
homogeneous
energy
dissipationrate(hom
ogeneous
andisotropicturbulence).Meaneddy
energy
ise iðλÞ¼
πβL
12F cε2 3λ11 3
.bThe
coeffi
cientofsurface
increaseduetobreakage
isc f=fbv2/3+(1
�f bf)2/3�1,with
thevolumeratio
ofthedaughterdropletitothebreaking
dropletjisf bv=
(di3 /d j3 ).cFo
rotherexam
ples
ofmultifractalspectra,seeBaldyga
andBourne.19b
11366 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
to a single parameter19 by setting B1B2 = 1 and makes this modelparticularly well adapted for scale up.40
4.3. Daughter Size Distribution. Similarly to the breakagerate, a number of daughter droplet/bubble distributions havebeen proposed, which are equally reviewed by Liao and Lucas.5
Statistical models consider the daughter distribution as a randomvariable, of which the probability distribution can be described bysimple relations such as uniform distributions,9 normal distribu-tions,6 or β functions.41
The more evolved, phenomenological models take into ac-count empirical observations as well as theoretical considera-tions. The model developed by Martinez-Bazan et al.42 takes abell-shape, similar to the statistical models. Others are based onthe observation that contrary to the previously proposed bell-shaped models, breakage into two equally sized daughter parti-cles is energetically unfavorable and breakage into large and smalldaughters has been observed experimentally. U-shapedmodels, aminimal probability of forming daughter droplets of equal sizeand maximal probability as the smaller daughter size tends tozero, have been developed by Tsouris and Tavlarides20 and Luoand Svendsen.14 The most recent models, which are similar tothe U-shaped models but have zero probability as the smallerdaughter size approaches zero, take an M-shaped, e.g., Wanget al.17 or Zhao and Ge.18
The U- and M-shaped daughter distributions were considerednot appropriate for the emulsification system studied here be-cause the associated breakage kernels were found to be inapplic-able in the region of interest (see section 4.2). On the basis of thebell-shape of the DSD and the detection of a negligible numberof particles with size below the Kolmogorov microscale (seeFigure 2), it is reasonable to assume that binary breakage intoapproximately equal sized daughter particles is the predominantmechanism. Thus a normal distribution was chosen to be usedwith the breakage kernels by Coulaloglou and Tavlarides, Alopaeuset al., and Baldyga and Bourne. Baldyga and Bourne advise on usingbreakup into two equal size droplets if dispersed phase viscosity ishigh. As shown in Table 1, the μd = 0.01 Pa s, which is much higherthan the continuous phase viscosity and can thus be consideredsignificantly high. As will be discussed in the simulations, the normaldistribution will be changed to log-normal distribution to betterrepresent the daughter size distribution with these breakage kernels.
Themodel by Sathyagal andRamkrishna also provides a bell-shapeddaughter size distribution based on the inverse problem.4.4. Coagulation Kernels. The coagulation kernel β(xi,xj) is
essentially the product of the collision frequency, h(di,dj), be-tween two particles of diameters di and dj, and the coalescenceefficiency of such collisions, λ(di,dj), in forming a new particle:
βðdi, djÞ ¼ hðdi, djÞλðdi, djÞ ð11ÞCollisions induced by fluctuations of turbulent velocity in thecontinuous phase liquid are the dominant mode for turbulentflow regimes, where the collision frequency of two particles can beexpressed as11,12
hðdi, djÞ ¼ C3π
4ðdi þ djÞ2ðdi2=3 þ dj
2=3Þ1=2ε1=3 ð12Þ
Two different physical theories exist for the coalescence efficiency:the film drainagemodel and the energymodel. The former assumesthat a liquid film is formed between two colliding droplets, whichthen drains out from in-between them. The probability that thecollision will then form a new particle can be expressed as a functionof the ratio of the characteristic film drainage time (tdrain) and thecontact time (tcont):
42
λðdi, djÞ ¼ exp � tdraintcont
� �ð13Þ
A number of different theories are available for tdrain and tcont, whichare not discussedhere.Oneof themost popularmodels, basedon theone developed by Coulaloglou and Tavlarides43 assuming de-formable particles with immobile surfaces, is given by Tsourisand Tavlarides:20
λðdi, djÞ ¼ exp �C4μcFcε
σ2ð1 þ ϕÞ3didj
di þ dj
!40@
1A ð14Þ
The energymodel, on the other hand, is based on the assumptionthat high-energy collisions result in immediate coalescence.22
The coalescence efficiency is thus related to the kinetic collisionenergy and the surface energy of the droplets. It can be expressedas
λðdi, djÞ ¼ exp �C5σðVi
2=3 þ Vj2=3Þ
Fdε2=3ðVi11=9 þ Vj
11=9Þ
!ð15Þ
Given the low power input (e1 W/kg) and in the presence ofexcess surfactant in such a diluted dispersion, coalescence is ex-pected to be of vanishing significance compared to breakage. There-fore, aggregation was omitted from any subsequent simulations.
5. NUMERICAL SIMULATIONS
The PBE (eq 5), including breakage and aggregation terms, isa linear partial integro-differential equation, which is character-ized by computationally intensive terms involving a number ofintegrals and double integrals, depending on the kernels (seeTable ). This equation has a considerable level of stiffness due tosignificant differences in the time constants of the individualsubprocesses.44 A large number of solution techniques have beendeveloped, many for specific applications with some focusing onthe correct prediction of specific moments of the DSD, others onthe distribution itself. Reviews of numerical solution methodscan be found in refs 2, 45, and 46. Discretizing the PBE and thus
Figure 5. Comparison of the selection of breakage rate kernels pre-sented in Table using physical properties presented in Table 1 withε = 0.35 and ϕ = 0.5 for up to 1 mm.
11367 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
transforming the PBE into a set of ordinary differential (ODEs),which can then be solved using standard solution methods andoff-the-shelf ODE solvers, is a very common technique. It is usedfor the simulations in this work, despite the fact that it is muchmore computationally demanding than the method of momentsbecause it allows the direct simulation of the evolution of theDSD and thus retains more information about the distribution.
The continuous domain, the DSD n(t,x), is divided into anumber of cells Λi = [xi-1/2,xi+1/2], each of which is representedby a common size, the so-called pivots xi. The discrete numberdistribution N(t,xi) is then given by (i = 1,..., I)
Nðt, xiÞ ¼Z xiþ1=2
xi�1=2
nðt, xÞ dx ð16Þ
Here, geometric discretization by droplet volume was chosen.Three popular methods were chosen to be tested for the emulsi-fication system of this study: the finite volume method, Fibertand Laurenc-ot;30 the cell average technique, Kumar et al.;32 andthe fixed pivot technique, Kumar and Ramkrishna.29 The threetechniques are discussed briefly below.5.1. Finite Volumes Scheme (FV). The finite volumes dis-
cretization scheme for the Smoluchowski equation for pure co-agulation systems, based on a conservative form of the PBE47
(eq 17) for the mass distribution g(t,x) = xn(t,x), was first devel-oped by Fibert and Laurenc-ot.30 This scheme uses the massfluxes between the individual cells to conserve the total mass ofthe system. It was adapted by Kumar et al.33 to the combinedbreakage and aggregation case.
∂gðt, xÞ∂t
¼ ∂
∂xðZ x
0
Z xmax
x � uuβðt, uÞnðt, uÞnðt, vÞ du dvÞ
þ ∂
∂xðZ ∞
0
Z x
0ubðu, vÞSðvÞnðt, vÞ du dvÞ ð17Þ
The evolution of the mass distribution can then be described interms of the mass flux across the cell boundaries (Ji+1/2 and Ji�1/2);for cell i at time t, this is given by
dnidt
¼ ðJiþ1=2, Co þ Jiþ1=2, Br � Ji�1=2, Co � Ji�1=2, BrÞðxiþ1=2 � xi�1=2Þ
ð18Þ
With the mass fluxes
Jiþ1=2, Co ¼ ∑i
k¼ 1∑I
j¼αi, k
ZΛj
βðu, xkÞu
dugj
þZ xα
i, k�1=2
xiþ1=2�xk
βðu, xkÞu
dugαi, k�1
!ð19Þ
Jiþ1=2, Br ¼ � ∑I
k¼ i þ 1gk
ZΛk
SðvÞv
dvZ xiþ1=2
0ubðu, xkÞ du
ð20Þwhere αi,k is the index of each cell, such that xi+1/2 � xk ∈ Λαi,k�1
.5.2. Fixed Pivot Technique (FP). One of the most popular
discretization techniques for breakage and coalescence problemsis the fixed pivot technique, introduced by Kumar and Ramkrishna.29
This method works on an arbitrary grid by redistributing newlyformed particles to the adjoining nodes such as to preserve anytwo moments of the DSD. Here the moments to be conservedare the zeroth and first, representing total number and total
diameter of all droplets in the system (or total volume x re-presents droplet volume), resulting in the formulation shown inthe following equations.
dNi
dt¼ ∑
j g k
j, kxi�1 e ðxj þ xkÞ e xiþ1
1� 12δj, k
� �ηβðxj, xkÞNðxjÞNðxkÞ
�NðxiÞ ∑I
k¼ 1βðxi, xkÞNðxkÞ þ ∑
I
k¼ 1ni, kSðxkÞNðxkÞ
� SðxiÞNðxiÞ ð21ÞWith the Kroenecker delta, δj,k = 1 for j = k, 0 otherwise
η ¼xiþ1 � vxiþ1 � xi
, xi e v e xiþ1
v� xi�1
xi � xi�1, xi�1 e v e xi
8>><>>: ð22Þ
and
ni, k ¼Z xiþ1
xi
xiþ1 � vxiþ1 � xi
bðv, xkÞ dv þZ xi
xi�1
v� xi�1
xi � xiþ1bðv, xkÞ dv
ð23ÞKumar and Ramkrishna29 note that this technique tends to overpredict the DSDwhen fast moving fronts are present in the distri-bution with coarse grids.5.3. Cell Average Technique (CA). A new technique, which
particularly addresses the over prediction in the FP technique,the so-called cell average technique was introduced by Kumaret al.32 This technique does not redistribute each newly formedparticle individually but uses the average of all incoming particlesinto the adjacent cells and thus retains more information aboutthe distribution. The discrete formulation for combined breakageand aggregation was taken from Kumar et al.33
dNi
dt¼ BCACo þ Br, i �DCA
Co þ Br, i ð24Þ
With the distribution scheme for the death and birth terms
DCACo þ Br, i ¼ DCo, i þ DBr, i ð25Þ
BCACo þ Br, i ¼ BCoþBr, i�1λ�i ðvi�1ÞHðvi�1 � xi�1Þ
þ BCoþBr, iλ�i ðviÞHðxi � viÞ
þ BCoþBr, iλþi ðviÞHðvi � xiÞ
þ BCoþBr, iþ1λþi ðviþ1ÞHðxiþ1 � vi�1Þ ð26Þ
where H(x) denotes the Heaviside step function.Volume average of the incoming particles:
vi ¼ VCo, i þ VBr, i
BCo, i þ BBr, ið27Þ
where the discrete death terms and the discrete and volumetricbirth terms (denoted D, B, and V, respectively) for aggregationand breakage are given by
DBr, i ¼ SðxiÞNðxiÞ ð28Þ
BBr, i ¼ ∑k g i
NðxkÞSðxkÞZ pik
xi�1=2
bðv, xkÞ dv ð29Þ
11368 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
VBr, i ¼ ∑k g i
NðxkÞSðxkÞZ pik
xi�1=2
xbðv, xkÞ dv ð30Þ
DCo, i ¼ NðxiÞ ∑I
k¼ 1βðxi, xkÞNðxkÞ ð31Þ
BCo, i ¼ ∑j g k
j, kxi�1=2 e ðxj þ xkÞ e xiþ1=2
1� 12δj, k
� �βðxj, xkÞNðxjÞNðxkÞ
ð32Þ
VCo, i ¼ ∑j g k
j, kxi�1=2 e ðxj þ xkÞ e xiþ1=2
1� 12δj, k
� �βðxj, xkÞNðxjÞNðxkÞðxj þ xkÞ
ð33Þ
5.4. Comparison of Numerical Techniques.The three num-erical methods were used to obtain simulations of the DSDwith ageometric discretization with I = 30, 60, and 120 grid points usingthe Coulaloglou and Tavlarides breakage kernel with a normaldaughter distribution based on an experimental initial distribution.Figure 6 shows the results for the evolution of the zeroth and
first moments (see eq 4) for I = 60 grid points. The error for thefirst moment was found by comparing the simulations and theknown value for the total dispersed phase volume.Uponpreliminaryinspection, a simulation with I = 300 grid points was found to resultin practically indistinguishable results for all three methods. Theerror for the zerothmoment was thus determined by comparing thesimulation result to the simulation with I = 300 grid points.A summary of the percentage errors in the first two moments
together with the calculation times for the three methods isshown in Table 5. It is important to note that the calculation timedoes not include the once-off calculation of terms that are inde-pendent of n(x) and thus independent of t, which are calculatedbefore the numerical solution of the PBE.It can be seen that the fixed pivot technique is themost accurate
technique in terms of zeroth moment for very coarse discretiza-tion. However, as the number of grid points is increased to 60 and
120, the other twomethods becomemore accurate. In addition, itserror in total volume (i.e., firstmoment) is an important number ofmagnitudes larger than for the other two techniques. The error inthe zeroth moment is higher for the cell average technique withrespect to the finite volumes scheme by 40% for I = 30, 16% forI = 60, and 21% for I = 120. The accuracy of the latter is also muchhigher for the first moment of the distribution. However, thedifference can be considered negligible because the error for bothtechniques is of a very low order (10�7 to 10�10).The fixed pivot and finite volumes techniques are roughly
equivalent in terms of computation time, while the cell averagetechnique takes about twice as long to perform the calculations.It is important to note that the above calculation times do notinclude the preparation of terms or parts of terms which are time-independent and can be calculated before the actual simulation,at the time when the discretization is being determined. Becauseof a number of additional integrals in the time-independent terms,the fixed pivot technique is overall far more time-consuming thanthe other two schemes.Considering the data presented above, the finite volumes
technique is used in the simulations in the last part of this work,due to its good combination of accurate prediction of themomentsand low computation time. A discretization of I = 30 can beconsidered sufficiently accurate in terms of the first two moments.Enhancing accuracy is likely to be far below the experimental errorand thus unnecessary. However, a relatively fine discretization ofI = 120 was chosen to be used to minimize errors due to thenumerical method and thus allow direct comparison of thebreakage kernels.
6. RESULTS AND DISCUSSION
6.1. Adjustment of Breakage Model Parameters. Whenfour breakage kernels chosen in section 4.2 were used in simula-tions with the empirical parameters proposed in the originalpublications, they were found to give unsatisfactory predictionsof the DSDwhen compared to the experimental data. The modelCoulaloglou and Tavlarides (CT) model results in a slight underprediction of the breakage rate. The model by Alopaeus et al.(AP) was found to extremely over predict the breakage rate withthe original parameters. Note that the authors used a multi-block model48 (with different energy dissipation rates and flows
Figure 6. Comparison of prediction of moments of order zero with and prediction at I = 300 (left) and order one with the theoretical value (right)by three numerical methods at I = 60.
11369 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
between individual zones) for their 50 L reactor. The Sathyagaland Ramkrishna (SR) model predicts a breakage rate of thecorrect order of magnitude but results in a distorted daughterdistribution (with respect to its effect on the evolution of theDSD), which results in the formation of very small droplets. Themodel by Baldyga and Podgorska (BP) gives a very reasonablebreakage rate when the average energy dissipation is assumed(i.e., B1 = B2 = 1).This difference is most likely due to differences in the
impeller type: in this study, a Mixel TT propeller was used,while the parameters were originally fitted using experimentaldata from a system using a Rushton turbine. The former isclassed as an axial flow and the latter as a radial flow impeller.SR and AP used a batch system, while CT used a CSTR. It hasbeen shown that the impeller type can have a profoundinfluence on the DSD in emulsification systems.49�51 In fact,Pacek et al.51 found that axial flow, low power number impel-lers produced smaller droplets and more narrow distributionsthan a Rushton turbine at a similar energy dissipation rate.While the model by BP is the only one that has parameters thatare explicitly linked to the system geometry, it can be assumedthat the parameters in the other three models have somedependency on system geometry that is unaccounted for inthe original models.The empirical parameters for the three models were there-
fore adjusted for each of the nine experimental runs using the
least-squares criteria of the DSD at all available time intervals, ti
F ¼ ∑i∑jðnexpðti, djÞ � nmodðti, djÞÞ2 ð34Þ
The averages of the parameters obtained from all runs were thenused in the subsequent simulations (Table 6).The estimated breakage rates for the four models with the
identified parameters for the range of energy dissipation ratesused in this study are shown in Figure 7. It can be seen that thekernels from AP, BP, and CT are relatively close to each other forε = 0.2, 0.35, and 0.5 W/kg, with the rate predicted by AP deviat-ing from the other two at ε = 0.2 W/kg. The kernel developedby SR follows the same pattern of increasing the breakage rateat a higher energy input but gives much lower breakage than theother models for the same agitation rate. The breakage rate pre-dicted by BP is the highest of the four at large droplet sizes.Using a normally distributed daughter kernel (see Table ) for
the CT and AP breakage kernels was found to result in an underprediction of the standard deviation of the distribution and in theappearance of a significant number of droplets below the mini-mum observed size. A log-normal distribution with a geometricstandard deviation adjusted to σg = 0.5 was found to giveimproved results. The BP model was found to give satisfactoryresults with a normal distribution.6.2. Model Predictions vs Experimental Data. Out of the
total nine experimental runs that were performed, the simula-tion results for three runs with the same dispersed phase massfraction (ϕ = 0.5%) and covering the range of energy dissipationrates (ε = 0.2, 0.35, and 0.5 W/kg) are presented in the follow-ing. Dispersed phase concentration dependence was omittedfrom the discussion because the concentration of EGDS wasfound to have a much smaller influence on the DSD (seesection 3). The initial distribution was taken at t = 10 min afterthe start of the agitation at the required rate. This was done toeliminate the influence of the preparation of the emulsion by ahigh agitation burst (see section 2.4). Figures 8�13 showcomparisons of modeling results with the experimental data.Figures 8�10 show intermediate number density distributionsat four time steps, and Figures 11�13 show evolution of thenumber mean diameter and number standard deviation of thedistribution.It can be seen that the SRmodel gives a good prediction for the
DSD after 300 min for high energy dissipation rates of 0.35 and0.5 W/kg. For the lowest agitation rate of 0.2 W/kg, it over
Table 6. Original and Adjusted Parameters for Coulaloglou and Tavlarides, Alopaeus et al., and Sathyagal and RamkrishnaBreakage Kernels
kernel daughter distribution parameter original value adjusted value
Coulaloglou and Tavlarides log-normal: σg = 0.5 C1 4.87 � 10�3 3.4 � 10�4
C2 0.0552 0.0403
Baldyga and Podroska normal: c = 3 B1 0.56
B2 = 1/B1 1.78
Alopaeus et al. log-normal: σg = 0.5 A1 0.986 0.657
A2 0.892 � 10�3 0.021
A3 0.2 0.402
Sathyagal and Ramkrishna bell-shaped daughter distribution S1 0.422 0.515
S2 0.247 0.232
S3 2.154 2.107
S4 0.0577 �0.177
S5 0.558 0.318
Table 5. Comparison of Calculation Time and Errors in theFirst Two Moments at Three Different Discretizations forThe Finite Volumes (FV), Fixed Pivot (FP), and Cell Average(CA) Techniques
method N =
absolute %
error in μ0
absolute %
error in μ1
calculation
time [s]
fixed pivot 30 0.2501 0.4569 0.1060
60 0.3007 0.3509 0.1637
120 0.3493 0.3515 0.4970
cell average 30 �2.0952 �0.72 � 10�5 0.1869
60 �0.2417 �0.73 � 10�7 0.3682
120 �0.1789 �0.73 � 10�7 1.1831
finite volumes 30 �1.1810 0.04 � 10�10 0.1220
60 �0.2030 �0.58� 10�10 0.1806
120 �0.1406 0.02 � 10�10 0.4372
11370 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
predicts the final distribution. However, the SRmodel under pre-dicts the intermediate DSDs for all three agitation rates; parti-cularly with respect to the position of the peak. In fact, the otherthree models give better predictions of the intermediary dis-tributions. The AP model tends to under predict the DSDs at allenergy dissipation rates. The CT and BPmodels give very similarpredictions of the DSD, relatively close to the experimental data.This similarity is mainly due to the two models being based onthe same theory and their breakage rates being characterizedby the d�2/3ε1/3 dependency on droplet diameter and energydissipation rates.In terms of the evolution of themean diameter (Figures 11�13),
the AP, CT, and BP kernels tend to give a very similar shape(curvature) of the mean diameter curve obtained from theexperimental data, with the CT and BP kernels relatively close
to the experimental data. The APmodel under predicts the meandiameter significantly for ε = 0.2W/kg (Figure 11). The extent ofthis under prediction becomes progressively less with increasedagitation rates (Figure 12) and ends in a slight over prediction forε = 0.2 W/kg. The SR model over predicts the mean diameter atfirst, before crossing the experimental curve. The crossing pointis found to be lower for higher agitation rates, i.e., at 220 min forε = 0.2 W/kg and 150 min for ε = 0.35 W/kg. The prediction ofthe standard deviation for the lowest agitation rate (ε = 0.2W/kg,Figure 11) is poor for the four models, all of which predict a
Figure 7. Coulaloglou and Tavlarides (CT), Sathyagal and Ramkrishna(SR), Alopaes et al. (AP), and Baldyga and Podgorska (BP) breakage ratekernels with adjusted parameters at relevant agitation rates (ϕ = 0.5%).
Figure 8. Comparison of experimental (circles) data with modelingresults (lines) for the DSD at four time steps, including the Kolmogoroffmicroscale (vertical line) for ϕ = 0.5% EGDS and ε = 0.2W/kg, resolvedby the FV technique with I = 120.
Figure 9. Comparison of experimental (circles) data with modelingresults (lines) for the DSD at four time steps, including the Kolmogoroffmicroscale (vertical line) for ϕ = 0.5% EGDS and ε = 0.35 W/kg,resolved by the FV technique with I = 120.
Figure 10. Comparison of experimental (circles) data with modelingresults (lines) for the DSD at four time steps, including the Kolmogoroffmicroscale (vertical line) for ϕ = 0.5% EGDS and ε = 0.5W/kg, resolvedby the FV technique with I = 120.
11371 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
narrower distribution than experimentally observed. The predic-tion of standard deviation becomes better for increased agitationrates. As for the mean diameters, a crossing of the experimentalcurve is observed for the SR model at ε = 0.5 W/kg (Figure 13).
This suggests that the variation of the breakage rate of the SRmodel is not enough nonlinear with respect to time. However,the only parameter which influences the breakage rate that changeswith time is the droplet size. This means that the nonlinearity of
Figure 11. Comparison of experimental (circles) data with modeling results (lines) for the mean diameter evolution (left) and standard deviation(right) for ϕ = 0.5% EGDS and ε = 0.2 W/kg, resolved by the FV technique with I = 120.
Figure 12. Comparison of experimental (circles) data with modeling results (lines) for the mean diameter evolution (left) and standard deviation(right) for ϕ = 0.5% EGDS and ε = 0.35 W/kg, resolved by the FV technique with I = 120.
Figure 13. Comparison of experimental (circles) data with modeling results (lines) for the mean diameter evolution (left) and standard deviation(right) for ϕ = 0.5% EGDS and ε = 0.5 W/kg, resolved by the FV technique with I = 120.
11372 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
the breakage rate prediction with the droplet size in the models ofCT and BP is closer to the observation than in the SR model. Thiscan be confirmed by inspection of the curvature of the breakagerate kernels presented in Figure 7. With the AP model, predictingthe correct shape of the breakage rate does not seem to be able tocorrectly predict its variation with energy dissipation and, hence,results in a large difference in the mean diameter prediction.The above analysis of the modeling results shows that models
that are based on the amount of turbulent energy transferred tothe particle (i.e., category 1, CT and BP) give the most consistentand accurate predictions of the DSD, mean diameter, and stan-dard deviation. The model by SR, which is based on an inverseproblem, does not give a correct prediction of the nonlinearity ofthe breakage rate with droplet size, which could be corrected byan adjustment of the powers in the equation for S(V), presentedin Table . Such an adjustment would be somewhat arbitrary andwould certainly result in a model with no physical relevance andno capability of generalization to different systems.The BP model is to be preferred over the model by CT be-
cause the multifractal approach to intermittence represents animprovement over the older model based on the Kolmogorovtheory. This results in a number of advantages of BP over CT: theconstants in the CT model (C1 and C2) had to be modified toobtain acceptable results, even though Coulaloglou and Tavlaridespostulate these constants to be universal, while the only variablewhich was adjusted in the BP model (B1) is one that is explicitlyrelated to system geometry. Furthermore, the BP model takessystem scale into account by including the integral length scale Liand allows for dispersed phase viscosity (eq 10), both of which areignored by CT (that uses the density of the dispersed phase). Thecorrection for high dispersed phase concentrations ε = ε (1� ϕ)3,which is used in the CTmodel can be easily incorporated into BP.
7. CONCLUSIONS
A set of oil-in-water emulsification experiments with EGDS asthe model substance, in a stirred tank, was analyzed with a novelin situ video probe coupled with an automated image analysis toobtain a number of intermediate DSDs for times up to 300 min.The dispersed phase concentration was varied from ϕ = 0.2 to 1%,and themean energy dissipation rate was varied from ε = 0.2� 0.5W/kg. The bell-shaped, monomodal DSD was found to be in theregion of 20�80 μm.
A thorough review of breakage rate models was performed,and four models were found appropriate to the system studied,with respect to type (bubbles/droplets), energy dissipationrate, and bubble/droplet size. They were the models developedby Coulaloglou and Tavlarides, Alopaeus et al., Sathyagal andRamkrishna, and Baldyga and Progorska. The coagulation ratewas judged insignificant because of the dilute EGDS concentra-tion and the use of excessive surfactant.
Three discretization schemes, finite volumes, fixed pivot, andcell average, were implemented and compared. The CA and FVtechniques were found to provide better prediction of themoment of order 1 (i.e., total volume/mass conservation) andbetter prediction of the moment of order 0 (i.e., total numberof droplet) than the FP technique. The FV scheme was foundto be less computationally intensive than CA for comparableaccuracy; FV was therefore chosen to be used in the subse-quent simulations.
The parameters of the breakage kernels were identified torepresent the system used in this study. On comparison of
the experimental results with the model simulation, it was foundthat the kernels based on category 1 (eparticle > ec) gave the bestmodeling results; this included the oldest of the models byCoulaloglou and Tavlarides and an adaptation of this model witha multifractal approach to intermittence of turbulence by Baldygaand Progorska. The latter of which is to be preferred because ittakes system scale and geometry into account and has thus onlyone adjustable parameter.
’AUTHOR INFORMATION
Corresponding Author*Phone: +33 (0)4.72.43.18.50. Fax: +33 (0)4.72.43.16.82. E-mail:[email protected].
’ACKNOWLEDGMENT
The work leading to this invention has received funding fromthe European Union Seventh Framework Program (FP7/2007-2013) under Grant Agreement No. 238013.
’NOMENCLATUREa = parameter in β distributionA1�A3 = parameters in Alopaeus et al. breakage modelb = parameter in β distributionb(xi,xj) = dimensionless daughter size distribution for xi, from
breaking droplet of size xj [�]bcum(xi,xj) = cumulative dimensionless daughter size distribution
for xi, from breaking droplet of size xj [�]c = tolerance of normal distributionC1�C2 = parameters in Coulaloglou and Tavlarides break-
age modelC3�C5 = parameters used in coagulation modelsCed = eddy efficiency [�]cf = coefficient of surface increased = droplet diameter [m]d10 = number mean diameter (μm)DF(ϕ) = damping factorDI = impeller diameter [m]e(λ) = mean energy of eddy with size λ [W]ec(di, λ) = critical energy of eddy with size λ for droplet with size
di [W]erf(x) = error functionfBV = breakage volume ratioF = minimization criteriaG(t,x) = particle growth rateg(t,x) = volume density distribution of droplet variable x at time
t [m3 m�3]H(x) = Heaviside step functionh(xi,xj) = collision frequency of droplets with size xi and xjk = wave number [m�1]Kg = parameter in Martinez�Bazan breakage modelm0 = mass in reactor [kg]n(t,x) = number density distribution of droplet variable x at time t
[m�3]I = total number of grid cellsNi = N(t,xi) = discrete number density distribution of cell i at
time tNp = impeller power number [�]P = total impeller power input [W]P(di,λ) = breakage probability of a collision of droplet of size di
with eddy of size λ
11373 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
Re = impeller Reynolds number = (FωDI2)/μ
S(xi) = breakage frequency of droplet with size xi [s�1]
S(xi,xj) = breakage frequency of droplet with size xi into a dropletof size xj [s
�1]S1�S2 = parameters in Sathyagal and Ramkrishna break-
age modelstdev = number standard deviation of the droplet size distribu-
tion (μm)T = reactor diameter [m]t = time [s]tdrain = drainage time [s]tcont = contact time [s]vi = volume average of droplets incoming into cell i [m3]V = droplet volume [m3]VT = vessel working volume [m3]v = volumetric flow rate through pipe [m3/s]vgas = volumetric gas flow rate in bubble column [m3/s]We = impeller Weber number = (ω2DI
3F)/σx = population balance parameter (e.g., droplet diameter or
volume)η = redistribution variable for fixed pivot technique
Greek Symbolsα = multifractal scaling exponentμi = moment of order i of the DSDμ = viscosity [Pa s]ϕ = dispersed phase concentration [kg/kg]ε = mean specific energy dissipation rate [W/kg]ω = impeller rotation speed [rev/s]σ = surface tension [N/m]σg = geometric standard deviationλ = eddy diameter [m]λmin = minimum eddy diameter (i.e., Kolmogorov microscale)
[m]λ(di,dj) = coagulation efficiency of a collision between droplets of
size di and djβ = parameter in Martinez�Bazan breakage modelβ(di,dj) = coagulation frequency of particles with size di and dj [s
�1]Γ(x) = gamma functionF = density [kg/m3]
Subscriptsi, j, k = designation of pivots for cellsi ( 1/2 = upper/lower bound of cell with pivot of size ic = continuous phased = dispersed phase
’REFERENCES
(1) Ramkrishna, D. Population Balances�Theory and Applicationsto Particulate Systems in Engineering; Academic Press: San Diego, CA,2000.(2) Ramkrishna, D.; Mahony, A. W. Population Balance Modeling.
Promise for the Future. Chem. Eng. Sci. 2002, 57, 59.(3) Lasheras, J. C.; Eastwood, C.; Martinez-Bazan, C.; Montanes,
J. L. A Review of Statistical Models for Break-up of an Immiscible FluidImmersed into a FullyDeveloped Turbulent Flow. Int. J. Multiphase Flow2002, 28, 247.(4) Patruno, L. E.; Dorao, C. A.; Svendsen, H. F.; Jakobsen, H. A.
Analysis of breakage kernels for population balance modeling. Chem.Eng. Sci. 2009, 64, 501.(5) Liao, Y.; Lucas, D. A Litterature Review of Theoretical Models
for Drop and Bubble Breakup in Turbulent Dispersions. Chem. Eng. Sci.2009, 64, 3389–3406.
(6) Coulaloglou, C. A.; Tavlarides, L. L. Description of InteractionProcesses in Agitated Liquid-Liquid Dispersions. Chem. Eng. Sci. 1977,32, 1289.
(7) Alopaeus, V.; Koskinen, J.; Keskinen, K. I. Simulation of thePopulation Balance for Liquid-Liquid Systems in a Nonideal StirredTank. Part 1 Description andQualitative Validation of theModel.Chem.Eng. Sci. 1999, 54 (24), 5887.
(8) Sathyagal, A. N.; Ramkrishna, D. Droplet Breakage in StirredDispersions. Breakage Functions from Experimental Drip-size Distribu-tions. Chem. Eng. Sci. 1996, 51, 1377.
(9) Baldyga, J.; Podgorska, W. Drop breakup in intermittent turbu-lence. Maximum stable and transient sizes of drops. Can. J. Chem. Eng.1998, 76, 456.
(10) Narsimhan, G.; Gupta, J. P. A model for transitional breakageprobability of droplets in agitated lean liquid-liquid dispersions. Chem.Eng. Sci. 1979, 34, 257.
(11) Lee, C. H.; Erickson, L. E.; Glasgow, L. A. Bubble Breakup andCoalescence in Turbulent Gas-Liquid Dispersions. Chem. Eng. Commun.1987, 59, 65.
(12) Lee, C. H.; Erickson, L. E.; Glasgow, L. A. Bubble Breakup andCoalescence in Turbulent Gas-Liquid Dispersions. Chem. Eng. Commun.1987, 61, 181–195.
(13) Prince, M. J.; Blanch, H. W. Bubble Coalescence and Break-upin Air Sparged Bubble Columns. AIChE J. 1990, 31, 1485.
(14) Luo, H.; Svendsen, H. F. Theoretical Model for Drop andBubble Breakup in Turbulent Dispersions. AIChE J. 1996, 42 (5), 1225.
(15) Martinez-Bazan, C.; Montanes, J. L.; Lasheras, J. C. On theBreakup of an Air Bubble Injected into a Fully Developed TurbulentFlow. Part 1. Breakup Frequency. J. Fluid Mech. 1999, 401, 157.
(16) Lehr, F.; Millies, M.; Mewes, D. Bubble-Size Distribution andFlow Fields in Bubble Columns. AIChE J. 2002, 48 (11), 2426.
(17) Wang, T.; Wang, J.; Jin, Y. A Novel Theoretical Breakup KernelFunction for Bubbles/Droplets in a Turbulent Flow. Chem. Eng. Sci.2003, 58, 4629.
(18) Zhao, H.; Ge, W. A Theoretical Bubble Breakup Model forSlurry Beds or Three-phase Fluidized Beds under High Pressure. Chem.Eng. Sci. 2007, 62, 109.
(19) (a) Baldyga, J.; Bourne, J. R. Chapter 15: Further Applications,in Turbulent Mixing and Chemical Reactions; John Wiley & Sons:Chichester, U.K., 1999. (b) Baldyga, J.; Bourne, J. R. Chapter 5: Theoryof Turbulence and Models of Turbulent Flow, in Turbulent Mixing andChemical Reactions; John Wiley & Sons: Chichester, U.K., 1999.
(20) Tsouris, C.; Tavlarides, L. L. Breakage and CoalescenceModelsfor Drops in Tubrulent Dispersions. AIChE J. 1994, 40 (3), 395.
(21) Liao, Y.; Lucas,D.ALitteratureReviewonMechanisms andModelsfor the Coalescence Process of Fluid Particles.Chem. Eng. Sci. 2010, 65, 2851.
(22) Simon, M. Coalescence of drops and drop swarming, Koales-zenz von Tropfen und Tropfenschw€armen, PhD. Dissertation, Tech-nische Universit€at Kaiserslautern, Kaiserslautern, Germany, 2004.
(23) Diemer, R. B.; Olson, J. H. A moment methodology forcoagulation and breage problems: part 3-generalized daughter distribu-tion functions. Chem. Eng. Sci. 2002, 57 (19), 4187.
(24) Marchisio, D. L.; Vigil, R. D.; Fox, R. O. Implementation ofthe Quadrature Method of Moments in CFD Codes for Aggregation-Breakage Problems. Chem. Eng. Sci. 2003, 58, 3337.
(25) Bapt, P. M.; Talarides, L. L.; Smith, G. W. Mont Carlosimulation of mass transfer in liquid-liquid dispersions. Chem. Eng. Sci.1983, 38 (12), 2003.
(26) Mahoney, A. W.; Ramkrishna, D. Efficient solution of popula-tion balance equations with discontinuouties by finite elements. Chem.Eng. Sci. 2002, 57, 1107.
(27) Vanni, M. Approximate population balance equations foraggregation-breakage processes. J. Colloid Interface Sci. 2000, 221 (2), 143.
(28) Kostoglou, M.; Karableas, A. J. On Sectional Techniques for theSolution of the Breakage Equation. Comput. Chem. Eng. 2009, 33, 112.
(29) Kumar, S.; Ramkrishna, D. On the Solution of PopulationBalance Equations by Discretization � I. A Fixed Pivot Technique.Chem. Eng. Sci. 1996, 51 (8), 13311.
11374 dx.doi.org/10.1021/ie2006033 |Ind. Eng. Chem. Res. 2011, 50, 11358–11374
Industrial & Engineering Chemistry Research ARTICLE
(30) Fibert, F.; Laurenc-ot, P. Numerical Simulation of the Smolu-chowski Coagulation Equation. SIAM J. Sci. Comput. 2004, 25 (6), 2004.(31) Kumar, S.; Ramkrishna, D. On the Solution of Population
Balance Equations by Discretization � II. A moving pivot technique.Chem. Eng. Sci. 1996, 51 (8), 1333.(32) Kumar, J.; Peglow, M.; Warnecke, G.; Heinrich, S.; M€orl, L.
Improved Accuracy and Convergence of Discretized Population Balancefor Aggregation: The Cell Average Technique. Chem. Eng. Sci. 2006,61, 3327.(33) Kumar, J.; Warnecke, G.; Peglow, M.; Heinrich, S. Comparison
of Numerical Methods for Solving Population Balance EquationsIncorporating Aggregation and Breakage. Powder Technol. 2009,189, 218.(34) Crombie, R. L. Cold Pearl surfactant-based blends. Int. J.
Cosmet. Sci. 1997, 19, 205.(35) Khalil, A.; Puel, F.; Chevalier, Y.; Galven, J. M.; Rivoire, A.;
Klein, J. P. Study of Droplet Size Distribution during an EmulsificationProcess using in situ Video Probe Coupled with an Automatic ImageAnalysis. Chem. Eng. J. 2010, 165, 946.(36) Brown, D. A. R.; Jones, P. N.; Middleton, J. C. Part A:
Measuring tools and techniques for mixing and flow visualizationstudies. In Handbook of Industrial Mixing; Paul, E. L., Atiemo-Obeng,V. A., Kresta, S. M., Eds.; JohnWiley & Sons: Hoboken, NJ, 2004; p 145.(37) Illingworth, J.; Kittler, J. A Survey of the Hough Transform.
Comput. Vision, Graphics Image Process. 1988, 44, 87.(38) Kolmogorov, A. N. The local structure of turbulence in
incompressible viscous fluid for very large Reynolds numbers. Dokl.Akad. Nauk SSSR 1941, 30, 301.(39) Leng, D. E.; Calabrese, R. V. Immiscible liquid-liquid systems.
In Handbook of Industrial Mixing; Paul, E. L., Atiemo-Obeng, V. A.,Kresta, S. M., Eds.; John Wiley & Sons: Hoboken, NJ, 2004; p 639.(40) Baldyga, J.; Bourne, J. R.; Pacek, A. W.; Amanullah, A.; Nienow,
A. W. Effect of agitation and scale-up on drop size in turbulentdispersions: allowance for intermittency. Chem. Eng. Sci. 2001, 56, 3377.(41) Konno, M.; Aoki, M.; Saito, S. Scale Effects on Breakup Process
in Liquid-Liquid Agitated Tanks. J. Chem. Eng. Jpn. 1983, 16, 312.(42) Martinez-Bazan, C.; Montanes, J. L.; Lasheras, J. C. On the
Breakup of an Air Bubble Injected into a Fully Developed TurbulentFlow. Part 2. Size PDF of the Resulting Daughter Bubble. J. Fluid Mech.1999, 401, 183.(43) Coulaloglou, C. A.; Tavlarides, L. L. Drop size distributions and
coalescence frequencies of liquid-liquid dispersions in flow vessels.AIChE J. 1976, 22 (2), 289.(44) Sun, N.; Immanuel, C. D. Efficient Solution of Population
Balance Models Employing a Hierarchical Solution Strategy based on aMulti-level Discretization. Trans. Inst. Meas. Control 2005, 27 (5), 347.(45) Kiparissides, C.; Krallis, A.;Meimaroglou, D.; Pladis, P.; Baltsas,
A. From Molecular to Plant-Scale Modeling of Polymerization Pro-cesses: A Digital High-Pressure Low-Density Polyethylene ProductionParadigm. Chem. Eng. Technol. 2010, 33 (11), 1754.(46) Gunawan, R.; Fusman, I.; Braaz, R. D. High resolution algo-
rithms for multidimensional population balance equations. AIChE J.2004, 50 (11), 2738.(47) Tanaka, H.; Inaba, S.; Nakaza, K. Steady-State Size Distribution
for the Self-Similar Collision Cascade. ICARUS 1996, 123, 450.(48) Alopaeus, V.; Koskinen, J.; Keskinen, K. I. Simulation of the
Population Balance for Liquid-Liquid Systems in a Nonideal StirredTank. Part 2. Parameter Fitting and the Use id the MultiblockModel forDense Dispersions. Chem. Eng. Sci. 2002, 57 (10), 1815.(49) Zhou, G.; Kresta, S. M. Correlation of mean drop size and
minimum drop size with the turbulence energy dissipation and the flowin an agitated tank. Chem. Eng. Sci. 1998, 53, 2063.(50) Zhou, G.; Kresta, S. M. Evolution of drop size distributions in
liquid�liquid dispersions for various impellers. Chem. Eng. Sci. 1998,53, 2099.(51) Pacek, A. W.; Chamsart, S.; Nienow, A. W.; Bakker, A. The
influence of impeller type onmean drop size and drop size distribution inan agitated vessel. Chem. Eng. Sci. 1999, 54, 4211.
MULTIMOD is an Initial Training Network funded within the 7th Framework Programme of the European Commission
www.multimod.eu
MULTIMOD ITN Multi-Scale Computational Modeling of Chemical and Biochemical Systems
ITN
M
UL T
IM
O
DITN
M
UL T
IM
O
D
Coupling Population Balance Modelling to Computational Fluid Dynamics in a High-Pressure Homogenizer
Julian BECKERa, Francois PUELa, Arend DUBBELBOERb, Nida SHEIBAT-OTHMANa
a Université Lyon 1, LAGEP UMR5007, CPE-Lyon, F-69622 Villeurbanne, Franceb UNILEVER R&D, Vlaardingen, Netherlands
Developing a coupled framework for CFD & Discretized Population Balance Models in OpenFoam® for breakage-dominated emulsification.
The framework thus developed is to applied to an emulsification process of sunflower oil in a lab-scale High-Pressure Homogenizer (HPH).
Objectives
The Population Balance Equation in a Turbulent Flow Field
Accumulation of discrete volume distribution Gi = (V
i+1/2- V
i-1/2)g
i
Turbulent diffusion term with turbulent diffusion coefficient from turbulence model (k-ε)
Convective transport of droplets inside 3D flow field U = (u
x,u
y,u
z)
PBE source terms from discretization schemee.g. from breakage model
∂G i∂ t
+∇⋅(UG i )−∇ 2 (DTurbG i )=−( J i+1/2− J i−1/2 )
Source Terms for Breakage Model
J i+1/ 2=− ∑k=i+1
N P
g k ∫V i− 1/ 2
Vi+1/2
S (v )
vdv ∫
0
V i+ 1/ 2
ub (u,V k ) du
Finite Volume Method, Fibert & Laurencot (2003), Kumar et al. (2009)
Integrals can be evaluated without knowledge of DSD (gk)
Niro-Soave Twin Panda HPH (UNILEVER, Vlaardingen)
2D-wedge (5º) cut of axisymmetric simulation zone
Fixed inlet velocity (measured: 2.15 m/s)Fixed outlet pressure set to 0 bargGap height adjusted for ΔP = 200 barHexahedral mesh with 158,400 cellsMesh refined in gap & near wallsRNG k-ε turbulence model (low Re
formulation; no wall functions)SIMPLE solution algorithm for 6000
iterations
Pre-emulsionPressure sensor
Outlet
Valve stage 1
Valve stage 2(unused)
Volumetric pumps
Pressure adjustment screws
PBE-CFD Coupling Strategy
Conclusions & Perspectives
Simulated Sauter Mean Diameter(zoomed on gap)
Experimental vs. Simulated DSD
gap end
gap center
to outlet
to inlet
turbulent jet
Valve arrangementValve arrangement
Steady-state CFD:
inlet
to outlet
symm
etry axis
max |U| = 92 m/s(in gap)
recirculation recirculation zonezone
Breakage Kernel (S) & Gaussian Daughter Distribution (b)Coulaloglou & Tavlarides (1977)
Droplet Transport Mechanisms in Turbulent Flow
S (V i )=C1V i−2/9 ϵ
1+ϕexp(C 2σ (1+ϕ )2
ρdϵ2 /3V i
5/9 ) b (V i ,V j )=2 . 4V i
exp (−4 . 52V j−V i2
V i2 )
Droplets are transported by the bulk flow of the continuous phase U
=> Convective transport
Droplets are also transported within the eddies (λ), which are moving with a fluctuating velocity u
λ
=> Diffusive transport
Modelled as turbulent diffusion along a concentration gradient, with turbulent diffusion coefficient D
Turb
From k-epsilon turbulence model:
k = turbulent kinetic energy [m² s-2]
ε = turbulent energy dissipation rate [m2 s-3]
bulk fluid velocity U
Fluctuating eddy velocity(mean velocity:
Coupled PBE-CFD Simulation Results
Reference emulsification experiment:
5% sunflower oil, 0.5% calcium caseinate (surfactant) at 200 bar
Parameters in Coulaloglou & Tavlarides (1997) breakage model were adjusted to give “reasonable” breakage rate
Reduction of droplet diameter seen along gap
Negligible breakage before gap Negligible breakage in turbulent jet
and/or recirculation zones Bi-modal experimental DSD cannot
be reproduced with a Gaussian daughter distribution
Coulaloglou & Tavlarides breakage kernel is only valid in the turbulent inertial regime; turbulent viscous effects could be present for very small drops
DTurb=C μ√kϵ
25emes Entretiens du Centre Jacques CartierEmulsification: Modeling, Technologies, and Applications
References
Steady-state CFD model for Niro-Soave Twin Panda HPH was set up
CFD-PBE coupling Framework was implemented in the OpenFoam® CFD environment as a custom solver
Coupled Framework allows simulating complex geometry with highly inhomogeneous local conditions (e.g. energy dissipation rate ε)
Coupling was tested for emulsification of 5% sunflower oil at 200 bar
Correct representation of main peak in the range 10 – 40 µm
Main Drawback of the Model : Not capable to generate second fine peak
The applicability of Coulaloglou breakage model is questionable
Coupled framework to be used with confidence for the testing of different turbulent breakup models
Becker P.J., Puel F., Reynald H., Sheibat-Othman N., “Ind. & Eng. Chem, Res.”, Vol. 50, 2011, pp.11358-11374Coulaloglou C.A., Tavlarides L.L., “Chemical Engineering Science”, Vol. 32, 1977Filbert F., Laurencot P., “SIAM Journal of Scientific Computing”, Vol. 25, 2003, pp. 1289-1297Kumar J., Warnecke G., Peglow M., Heinrich S., “Powder Technology”, Vol. 189, 2009, pp. 218-229
recirculation recirculation zonezone
19-21 November 2012, LAGEP, Lyon, France
Author contact: [email protected]
u λ≈√2 ( λϵ )1/3
Inital & boundary conditions
MeshN
M cells
Solve turbulent
CFD
Face flux(volume conservative)
Energy dissipation rate
Discretized DSD (G
i) with N
P cells
Physical PBE properties
Calculate PBE integrals & store in matrices:
(Np x N
p x N
M & N
P x N
M)
Calculate (NP x N
M)
PBE source terms
Solve NP transport
eqn. with source terms on mesh (N
M)
Create NP scalar
fields for DSDUpdated
DSD
CFD
Population Balance
write to fileload
Write output to file
(one-way coupling)
load
load
yes
No
load
load
Converged to SS?Or max. iterations / time steps reached?
Time-stepping / Steady-state convergence loop
Vi VVi+1/2Vi-1/2
Ji-1/2 Ji+1/2
∂G i∂ t
+∇⋅(UG i )−∇ 2 (DTurbG i )=−( J i+1/2− J i−1/2 )
Mass flow in and out of cell