2.1 Model for the breakup probability

Models for the breakup probability function $P(d_{i},d_{j})$ (or $\unicode[STIX]{x1D6FD}(d_{i},d_{j})$ ), can broadly be classified as statistical, phenomenological or empirical (see Lasheras et al. Reference Lasheras, Eastwood, Martínez-Bazán and Montaes2002; Liao & Lucas Reference Liao and Lucas2009). In this study we use the phenomenological model proposed by Tsouris & Tavlarides (Reference Tsouris and Tavlarides1994) that leads to a ‘U-shaped’ distribution. We keep in mind, however, that experiments for bubble breakup (Martínez-Bazán, Montañés & Lasheras Reference Martínez-Bazán, Montañés and Lasheras1999b ) have led to other possible shapes for $P(d_{i},d_{j})$ and that there remains considerable uncertainty about the best model to use. Here we proceed with the model of Tsouris & Tavlarides (Reference Tsouris and Tavlarides1994) because it is based on a relatively simple physical reasoning as shown below.

The breakup is considered to be binary, and $P(d_{i},d_{j})$ is formulated based on the formation energy required to form the daughter droplets of size $d_{i}$ and a complementary droplet to ensure volume conservation (Tsouris & Tavlarides Reference Tsouris and Tavlarides1994). The formation energy is proportional to the difference in initial and final surface areas according to

(2.4) $$\begin{eqnarray}E_{f}(d_{i},d_{j})=\unicode[STIX]{x03C0}\unicode[STIX]{x1D70E}[(d_{j}^{3}-d_{i}^{3})^{2/3}+d_{i}^{2}-d_{j}^{2}],\end{eqnarray}$$

where $\unicode[STIX]{x1D70E}$ is the interfacial tension between the dispersed and continuous phase. It can be shown using (2.4) that the breakup of a parent droplet into two equal size daughter droplets is a maximum energy process. Substituting $d_{i}=d_{j}/2^{1/3}$ in (2.4) we get a maximum formation energy equal to

(2.5) $$\begin{eqnarray}E_{f,max}=\unicode[STIX]{x03C0}\unicode[STIX]{x1D70E}d_{j}^{2}(2^{1/3}-1).\end{eqnarray}$$

Equation (2.4) is minimized when $d_{i}=0$ , that is no breakup of the parent droplet. To allow for breakup, a minimum diameter $d_{min}$ is specified and the corresponding surface formation energy is

(2.6) $$\begin{eqnarray}E_{f,min}=\unicode[STIX]{x03C0}\unicode[STIX]{x1D70E}[(d_{j}^{3}-d_{min}^{3})^{2/3}+d_{min}^{2}-d_{j}^{2}],\end{eqnarray}$$

where $d_{min}=1~\unicode[STIX]{x03BC}\text{m}$ in this study. Making the crucial assumption that the probability of breakup of a drop of size $d_{j}$ leading to a droplet of size such that it falls in a bin around $d_{i}$ decreases linearly with the required formation energy and remains within the bounds specified above, the discrete breakup probability $P(d_{i},d_{j})$ can be written as

(2.7) $$\begin{eqnarray}P(d_{i},d_{j})=\frac{[E_{f,min}+(E_{f,max}-E_{f}(d_{i},d_{j}))]}{\displaystyle \mathop{\sum }_{k=1}^{j-1}[E_{f,min}+(E_{f,max}-E_{f}(d_{k},d_{j}))]},\end{eqnarray}$$

where $E_{f}(d_{i},d_{j})$ is the surface formation energy defined in (2.4). Also, we assume that the bin sizes are logarithmically distributed. Thus (2.7) is meant to model the discrete probability that a particle of size $d_{j}$ breaks up into a particle inside a bin centred at $d_{i}$ with a width of $\unicode[STIX]{x1D6FF}(\log (d_{i}))$ , and its complement $d_{c}$ (to conserve volume). This distribution is U-shaped, with a minimum probability for the formation of two equally sized daughter droplets (when $E_{f}(d_{i},d_{j})=E_{f,max}$ which leads to a maximum of required energy), and probability maxima at the two ends (which have formation energy minima).

Martínez-Bazán et al. (Reference Rodríguez-rodríguez, Deane, Montañés and Lasheras2010) derived constraints that apply to the droplet size probability density function $\unicode[STIX]{x1D6FD}(d_{i},d_{j})$ for the breakup process to be volume conserving. The discrete probability of forming a droplet in bin $d_{i}$ must be equal to the probability of formation of the complement in bin $d_{c}$ . The discrete breakup probability in (2.7) conserves volume, since $P(d_{i},d_{j})=P(d_{c},d_{j})$ (we note that expressing this probability in terms of a universal density $\unicode[STIX]{x1D6FD}(d_{i},d_{j})$ presents further challenges (Martínez-Bazán et al. Reference Rodríguez-rodríguez, Deane, Montañés and Lasheras2010) that are left for future analysis, while here we use the discrete version).

2.2 Model for breakup frequency

Modelling breakup based on encounter rates of turbulent eddies and their characteristic fluctuations with droplets of a certain size has been a popular method in the literature. The phenomenological model by Coulaloglou & Tavlarides (Reference Coulaloglou and Tavlarides1977) postulates that a droplet in a liquid–liquid dispersion breaks up when the kinetic energy transmitted from droplet–eddy collisions exceeds the surface energy. Many other papers have pursued this approach, mostly in the RANS context (e.g. Narsimhan, Gupta & Ramkrishna Reference Narsimhan, Gupta and Ramkrishna1979; Chatzi & James Reference Chatzi and James1987). Here we follow the approach of Prince & Blanch (Reference Prince and Blanch1990) and Tsouris & Tavlarides (Reference Tsouris and Tavlarides1994), where the droplet–eddy collisions are treated akin to the of collisions between molecules in kinetic theory of gases. The breakup frequency is computed as an integral over the product of a collision frequency and a breakup efficiency according to

(2.8) $$\begin{eqnarray}g(d_{i})=K\int _{0}^{d_{i}}\frac{\unicode[STIX]{x03C0}}{4}(d_{i}+d_{e})^{2}u_{e}(d_{e})\unicode[STIX]{x1D6FA}(d_{i},d_{e})\,\text{d}n_{e}(d_{e}).\end{eqnarray}$$

Here $d_{i}$ is the diameter of the droplet, $d_{e}$ is the eddy size, $n_{e}(d_{e})$ is the number density of eddies of size $d_{e}$ , $u_{e}(d_{e})$ is the characteristic fluctuation velocity of eddies of size $d_{e}$ (in a frame moving with the advection velocity caused by larger eddies), $\unicode[STIX]{x1D6FA}(d_{i},d_{e})$ is a breakup efficiency and the integral is evaluated over all eddies, up to the size of the droplet (i.e. for $d_{e}$ up to $d_{e}=d_{i}$ ). A crucial assumption of the model is that eddies larger than the scale of the droplet are assumed to be only responsible for advection of the droplet, not contributing to collisions with the droplet that require relative velocity. One could develop a ‘smoother’ model in which the lack of deformation due to eddies larger than $d_{i}$ is included as an additional cutoff behaviour in the function $\unicode[STIX]{x1D6FA}(d_{i},d_{e})$ . Here we choose to include that cutoff behaviour by following earlier work (Tsouris & Tavlarides Reference Tsouris and Tavlarides1994; Luo & Svendsen Reference Luo and Svendsen1996) as a sharp cutoff, while lumping any possible dependencies on the exact cutoff scale into the unknown model parameter $K$ , expected to be of order unity.

The number density of eddies, $n_{e}(d_{e})$ , can be estimated from the energy spectrum (Azbel Reference Azbel1981; Tsouris & Tavlarides Reference Tsouris and Tavlarides1994; Solsvik, Maaß & Jakobsen Reference Solsvik, Maaß and Jakobsen2016), or more simply by assuming the eddies to be space filling, i.e.  $n_{e}(d_{e})\propto d_{e}^{-3}$ . The latter argument leads to $\text{d}n_{e}(d_{e})=C_{1}d_{e}^{-4}\text{d}(d_{e})$ , where $C_{1}$ is a constant of order $1$ .

The eddy fluctuation velocity $u_{e}(r)$ written in terms of the two-point separation distance, $r$ , is assumed to be expressed based on the second-order longitudinal structure function $S_{2}(r)$ as $u_{e}(r)\sim [S_{2}(r)]^{1/2}$ . The structure function is defined according to (Pope Reference Pope2011),

(2.9) $$\begin{eqnarray}S_{2}(r)=\langle [u_{L}(\boldsymbol{x}+r\boldsymbol{e}_{L})-u_{L}(\boldsymbol{x})]^{2}\rangle ,\end{eqnarray}$$

where $u_{L}$ is the fluid velocity component in the direction of unit vector $\boldsymbol{e}_{L}$ and the angular brackets represent statistical averaging. In previous models (Tsouris & Tavlarides Reference Tsouris and Tavlarides1994; Bandara & Yapa Reference Bandara and Yapa2011; Zhao et al. Reference Zhao, Torlapati, Boufadel, King, Robinson and Lee2014b ) a Kolmogorov scaling valid in the inertial range of turbulence was used for $S_{2}(r)$ , leading to $u_{e}(r)\sim (\unicode[STIX]{x1D716}r)^{1/3}$ . However, this expression cannot be used if the size of the droplet is near the viscous range of turbulence. In order to capture both inertial and viscous ranges, as well as a smooth transition between the two ranges, we use the approach of Batchelor (Reference Batchelor1951) with a blending function. In this approach, the structure function is given by

(2.10) $$\begin{eqnarray}S_{2}(r)=C_{2}\unicode[STIX]{x1D716}^{2/3}r^{2/3}\left[1+\left(\frac{r}{\unicode[STIX]{x1D6FE}_{2}\unicode[STIX]{x1D702}}\right)^{-2}\right]^{-2/3},\end{eqnarray}$$

where $\unicode[STIX]{x1D702}$ is the Kolmogorov length scale. We choose the usual value for the Kolmogorov coefficient $C_{2}\approx 2.1$ (Pope Reference Pope2011). The parameter $\unicode[STIX]{x1D6FE}_{2}=(15C_{2})^{3/4}\approx 13$ sets the cross-over scale between the inertial and viscous range. We note that while most prior models are for use in a RANS framework using the average energy dissipation, $\unicode[STIX]{x1D716}$ , in LES we can use a local value of the instantaneous rate of dissipation averaged over the grid scale, modelled as the subgrid-scale (SGS) dissipation rate. As a result, even though (2.10) is based on K41 theory (Kolmogorov Reference Kolmogorov1941), in LES we only assume K41 scaling for the scales below the grid scale while intermittency in the resolved range of scales can be explicitly computed, and its effects on breakup rates taken into account in the LES model.

The breakup efficiency $\unicode[STIX]{x1D6FA}(d_{i},d_{e})$ in (2.8) is the probability that a given eddy interacting with the droplet has sufficient energy to overcome the resistive forces in the system, namely surface tension and viscosity. It is assumed to be given by the usual formation potential in terms of an exponential (Coulaloglou & Tavlarides Reference Coulaloglou and Tavlarides1977; Prince & Blanch Reference Prince and Blanch1990)

(2.11) $$\begin{eqnarray}\unicode[STIX]{x1D6FA}(d_{i},d_{e})=\exp \left(-\frac{E_{\unicode[STIX]{x1D70E}}(d_{i})+E_{\unicode[STIX]{x1D708}}(d_{i})}{E_{e}(d_{e})}\right),\end{eqnarray}$$

where $E_{\unicode[STIX]{x1D70E}}(d_{i})$ is resistive energy associated with a droplet of size $d_{i}$ due to surface tension, $E_{\unicode[STIX]{x1D708}}(d_{i})$ is the viscous resistive energy and $E_{e}(d_{e})$ is the kinetic energy of the turbulent eddy at scale $d_{e}$ . The resistive surface tension energy $E_{\unicode[STIX]{x1D70E}}$ is defined as the integral of the formation energy $E_{f}(d^{\prime },d_{i})$ multiplied by a measure of the breakup probability (see (2.7)):

(2.12) $$\begin{eqnarray}E_{\unicode[STIX]{x1D70E}}(d_{i})=\int _{0}^{d_{i}}c[E_{f,min}+(E_{f,max}-E_{f}(d^{\prime },d_{i}))]E_{f}(d^{\prime },d_{i})\,\text{d}(d^{\prime }),\end{eqnarray}$$

where $c$ is a normalization constant so that the integral of the probability between 0 and $d_{i}$ is unity. Using (2.4), changing the integration to $\unicode[STIX]{x1D709}^{\prime }=d^{\prime }/d_{i}$ and evaluating the integral numerically, we obtain

(2.13) $$\begin{eqnarray}E_{\unicode[STIX]{x1D70E}}(d_{i})=0.0702\unicode[STIX]{x03C0}\unicode[STIX]{x1D70E}d_{i}^{2},\end{eqnarray}$$

where $\unicode[STIX]{x1D70E}$ is the surface tension of the droplet. The viscous resistive energy of the droplet at steady state can be expressed as (Calabrese et al. Reference Calabrese, Wang and Bryner1986; Skartlien et al. Reference Skartlien, Sollum and Schumann2013; Zhao et al. Reference Zhao, Torlapati, Boufadel, King, Robinson and Lee2014b )

(2.14) $$\begin{eqnarray}E_{\unicode[STIX]{x1D708}}(d_{i})=\unicode[STIX]{x1D6FC}\frac{\unicode[STIX]{x03C0}}{6}\unicode[STIX]{x1D716}^{1/3}d_{i}^{7/3}\unicode[STIX]{x1D707}_{d}\sqrt{\frac{\unicode[STIX]{x1D70C}_{c}}{\unicode[STIX]{x1D70C}_{d}}},\end{eqnarray}$$

where $\unicode[STIX]{x1D6FC}\approx 2$ , $\unicode[STIX]{x1D70C}_{c}$ and $\unicode[STIX]{x1D70C}_{d}$ are the carrier and droplet phase density, and $\unicode[STIX]{x1D707}_{d}$ is the dynamic viscosity of the dispersed droplet phase. For $E_{e}(d_{e})$ , the kinetic energy of the turbulent eddy, we use the longitudinal structure function $S_{2}$ , defined in (2.10) applied to all three coordinate directions for the eddy fluctuation velocity, $u_{e}$ . Assuming the volume of the eddy to be equal to that of a sphere, with density equal to the carrier phase density, the total energy contained in an eddy can be written as

(2.15) $$\begin{eqnarray}E_{e}(d_{e})=\frac{3}{2}\left(\frac{\unicode[STIX]{x03C0}}{6}\unicode[STIX]{x1D70C}_{c}d_{e}^{3}\right)S_{2}(d_{e}).\end{eqnarray}$$

In order to formulate a parameterization for the breakup frequency, here we identify the important non-dimensional numbers of the system. The breakup frequency can be rewritten as a function of the Reynolds number ( $Re_{i}$ ) based on droplet diameter and a velocity scale defined as $u_{d_{i}}=(\unicode[STIX]{x1D716}d_{i})^{1/3}$ , the Ohnesorge number ( $Oh_{i}$ ) of the dispersed phase controlling the relative importance of viscosity to surface tension of the droplet, and the density and viscosity ratio of droplet to carrier flow fluid. These non-dimensional numbers are defined below,

(2.16a-c ) $$\begin{eqnarray}Re_{i}=\frac{\unicode[STIX]{x1D716}^{1/3}d_{i}^{4/3}}{\unicode[STIX]{x1D708}};\quad Oh_{i}=\frac{\unicode[STIX]{x1D707}_{d}}{\sqrt{\unicode[STIX]{x1D70C}_{d}\unicode[STIX]{x1D70E}d_{i}}};\quad \unicode[STIX]{x1D6E4}=\frac{\unicode[STIX]{x1D707}_{d}}{\unicode[STIX]{x1D707}_{c}}\left(\frac{\unicode[STIX]{x1D70C}_{c}}{\unicode[STIX]{x1D70C}_{d}}\right)^{1/2}.\end{eqnarray}$$

After some manipulation, equation (2.8) can be rewritten as

(2.17) $$\begin{eqnarray}g_{i}=K^{\ast }\frac{1}{\unicode[STIX]{x1D70F}_{b,i}}\int _{0}^{1}r_{e}^{-11/3}(r_{e}+1)^{2}\left(1+\left(\frac{r_{e}Re_{i}^{3/4}}{\unicode[STIX]{x1D6FE}_{2}}\right)^{-2}\right)^{-1/3}\unicode[STIX]{x1D6FA}(Oh_{i},Re_{i},\unicode[STIX]{x1D6E4};r_{e})\,\text{d}r_{e},\end{eqnarray}$$

where, $r_{e}=d_{e}/d_{i}$ is the eddy size normalized by the droplet diameter, $\unicode[STIX]{x1D70F}_{b,i}=\unicode[STIX]{x1D716}^{-1/3}d_{i}^{2/3}$ is the breakup time scale for an eddy of size equal to that of the droplet as if it were in the inertial range (it does not have to be) and $\unicode[STIX]{x1D6FA}(Oh,Re,\unicode[STIX]{x1D6E4};r_{e})$ is the non-dimensional breakup probability,

(2.18) $$\begin{eqnarray}\unicode[STIX]{x1D6FA}(Oh,Re,\unicode[STIX]{x1D6E4};r_{e})=\exp \left(-\frac{\unicode[STIX]{x1D6E4}f_{2}}{Re}\left[1+\left(\frac{r_{e}Re^{3/4}}{\unicode[STIX]{x1D6FE}_{2}}\right)^{-2}\right]^{2/3}r_{e}^{-11/3}\right),\end{eqnarray}$$

where

(2.19) $$\begin{eqnarray}f_{2}=0.14\frac{\unicode[STIX]{x1D6E4}}{ReOh^{2}}+0.583.\end{eqnarray}$$

All the non-dimensional prefactors appearing when rewriting (2.8) have been absorbed into $K^{\ast }$ . Equation (2.17) provides a frequency for droplet breakup that depends on $Re_{i}$ , $Oh_{i}$ and $\unicode[STIX]{x1D6E4}$ . Note that if we had only considered an inertial-range scaling for the eddy fluctuation velocity, we can combine $Re$ and $Oh$ into a Weber number, defined as $We=Re^{2}~Oh^{2}$ . The breakup frequency would then only depend on $We$ and $\unicode[STIX]{x1D6E4}$ . The integral represents a correction to the frequency calculated by solely considering an eddy equal to the size of the droplet, by evaluating the effect of collisions of eddies smaller than the droplet. If $d_{i}$ falls in the viscous range, the integral cancels the inertial-range scaling assumed by the prefactor $\unicode[STIX]{x1D70F}_{b,i}=\unicode[STIX]{x1D716}^{-1/3}d_{i}^{2/3}$ so that this situation is also accounted for. We note that the value of the integral $g_{i}\unicode[STIX]{x1D70F}_{b,i}/K^{\ast }$ in (2.17) will inevitably depend on the assumed maximum eddy size interacting with the droplet, which is currently taken to be exactly the droplet size $d_{i}$ . However, if one were to take a different upper integration limit (still of order $d_{i}$ but not exactly $d_{i}$ ), the breakup frequency may not change much since the modified value of the integral will be largely (but not exactly) cancelled when fitting the prefactor $K$ to data. This behaviour is demonstrated in the next section.

In LES, $g_{i}$ needs to be evaluated on every grid point and timestep, depending on the local Reynolds number $Re_{i}$ and the local rate of dissipation. Evaluating numerically the integral in (2.17) at every timestep and grid point would be prohibitive in practice. Hence, we develop an empirical fit to prior numerical integrations. The speedup obtained from the fits is discussed in appendix A. We develop the parameterization for a wide range of Reynolds and Ohnesorge numbers, for a fixed value of $\unicode[STIX]{x1D6E4}$ . We define $g_{f}(Re,Oh,\unicode[STIX]{x1D6E4})$ as the integral in (2.17), i.e.  $g_{f}(Re,Oh,\unicode[STIX]{x1D6E4})=g_{i}(Re_{i},Oh_{i},\unicode[STIX]{x1D6E4})\,\unicode[STIX]{x1D70F}_{b,i}/K^{\ast }$ , and evaluate it numerically for a range of $Re$ and $Oh$ values for a fixed $\unicode[STIX]{x1D6E4}$ . Then, a fit can be developed in the following form,

(2.20) $$\begin{eqnarray}\log _{10}(g_{f})=ax^{b}+cx^{d}-e,\end{eqnarray}$$

where $x=\log _{10}Re$ , and $a,b,c,d,e$ are functions of $Oh$ . Further details of the functional form of the coefficients are provided in appendix A. The final model for the breakup frequency (for a given value of $\unicode[STIX]{x1D6E4}$ ) thus has the form

(2.21) $$\begin{eqnarray}\left.\begin{array}{@{}c@{}}\displaystyle g_{i}(Re_{i},Oh_{i};\unicode[STIX]{x1D6E4})=\frac{K^{\ast }}{\unicode[STIX]{x1D70F}_{b,i}}10^{G(Re_{i},Oh_{i})},\\ G(Re_{i},Oh_{i})=a[\log _{10}(Re_{i})]^{b}+c[\log _{10}(Re_{i})]^{d}-e,\end{array}\right\}\end{eqnarray}$$

where the fits for parameters $a$ – $e$ as functions of $Oh$ are provided in appendix A for a few representative values of $\unicode[STIX]{x1D6E4}$ .

The breakup frequency model is thus complete except for the prefactor $K^{\ast }$ appearing in (2.17). Its value is obtained by fitting results from an experiment (see next section), and the fitted value will then be used subsequently for comparisons with other data and for future applications.

3.1 Wave breaking experiment of Li et al. (Reference Li, Miller, Wang, Koley and Katz2017)

In order to fit a specific value for the parameter $K^{\ast }$ we use the data of a breaking wave experiment from Li et al. (Reference Li, Miller, Wang, Koley and Katz2017). The experiments were performed in an acrylic tank $6~\text{m}$ long, $0.6~\text{m}$ deep and $0.3~\text{m}$ wide. Breaking waves were generated mechanically using a piston-type wave maker consisting of a vertical plate that extends over the entire tank cross-section. The tank was filled with water up to a height of $0.25~\text{m}$ . The wave height and characteristics were controlled by varying the frequency and stroke of the vertical plate. Oil was placed on a patch at the surface. The wave impingement and subsequent breakup processes were recorded using 3 high-speed cameras. The droplet size distribution was measured using digital inline holography. A sketch of the set-up is depicted in figure 1. The oil patch on the surface was broken up into droplets by the plunging wave. The size distribution generated due to this process was recorded at a depth of $11.1~\text{cm}$ from the free surface. A simplified sketch of the evolution of the concentration of a particular droplet size is shown in figure 1(b).

Figure 1.

(a) Sketch of the wave breaking experiment of Li et al. (Reference Li, Miller, Wang, Koley and Katz2017). (b) Schematic dependence of $\overline{n}_{i}(z,t)$ on time and height for a particular droplet size, starting from a step-function initial condition that is assumed to be well mixed initially down to a depth of 13 cm below the surface and having zero concentration below. At increasing times, turbulent diffusion smooths the step and droplets rise towards the surface at different speeds depending on their size. The dotted horizontal line at $z=-11.1$  cm is where the experimental data are available.

As shown by Li et al. (Reference Li, Miller, Wang, Koley and Katz2017), the time evolution of the droplet size distribution at the measurement location could be represented well by a simple model that includes the effects of turbulent diffusion and droplet buoyancy only. Since the dissipation rate was quite low at the measurement location, Li et al. (Reference Li, Miller, Wang, Koley and Katz2017) neglected the effect of droplet breakup in their model. Consequently, for the case of crude oil with dispersants, the model under-predicted the number of smaller size droplets generated. This is due to the fact that with the effect of dispersants, the surface tension of the oil droplets was significantly lowered, resulting in droplet breakup despite of the weak turbulent dissipation rate. The Weber number ( $We=2\unicode[STIX]{x1D70C}(\unicode[STIX]{x1D716}d)^{2/3}d/\unicode[STIX]{x1D70E}$ ) based on the droplet diameter for the case with dispersants is approximately $We=3$ , confirming that the effects of droplet breakup are important. Our goal is to expand the model of Li et al. (Reference Li, Miller, Wang, Koley and Katz2017) by including breakup and select a value of $K^{\ast }$ that can achieve improved agreement with their experimental data.

Adding the effect of breakup and following Li et al. (Reference Li, Miller, Wang, Koley and Katz2017) in considering only vertical diffusion and droplet rise velocity, the ensemble-averaged number density $\bar{n}_{i}(z,t)$ for a droplet of size $d_{i}$ obeys

(3.1) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}\bar{n}_{i}(z,t)}{\unicode[STIX]{x2202}t}+w_{r}(d_{i})\frac{\unicode[STIX]{x2202}\bar{n}_{i}(z,t)}{\unicode[STIX]{x2202}z}=D(t)\frac{\unicode[STIX]{x2202}^{2}\bar{n}_{i}}{\unicode[STIX]{x2202}z^{2}}+\mathop{\sum }_{j=i+1}^{n}P(d_{i},d_{j})g(d_{j})\bar{n}_{j}(z,t)-g(d_{i})\bar{n}_{i}(z,t),\end{eqnarray}$$

where $z$ is the vertical coordinate, $w_{r}(d_{i})$ is the buoyancy-induced rise velocity of droplets of size $d_{i}$ and $D(t)$ is the turbulent diffusion coefficient. The daughter droplet probability function $P(d_{i},d_{j})$ and the breakup frequency $g$ are evaluated following the model presented in § 2. The droplet rise velocity is calculated as a balance between the drag and buoyancy force acting on a droplet.

(3.2) $$\begin{eqnarray}w_{r}(d)=\left\{\begin{array}{@{}ll@{}}w_{r,S}\quad & \text{if }Re_{d}<0.2,\\ w_{r,S}(1+0.15Re_{d}^{0.687})^{-1},\quad & 0.2<Re_{d}<750,\end{array}\right.\end{eqnarray}$$

where

(3.3) $$\begin{eqnarray}w_{r,S}=\frac{(\unicode[STIX]{x1D70C}_{0}-\unicode[STIX]{x1D70C}_{d})g~d^{2}}{18\unicode[STIX]{x1D707}_{f}},\end{eqnarray}$$

and $Re_{d}=\unicode[STIX]{x1D70C}_{0}w_{r}d/\unicode[STIX]{x1D707}_{f}$ is the droplet rise velocity Reynolds number (not to be confused with the eddy Reynolds number $Re_{i}$ used to express the breakup frequency). The time-dependent diffusion coefficient can be estimated using $D(t)=k_{D}u^{\prime }(t)L(t)$ , where $u^{\prime }(t)$ is the time-dependent turbulent fluctuation (root-mean-square) velocity as measured in the experiment and $L(t)$ is the corresponding integral length scale, also measured. The constant $k_{D}$ is known to be between $0.23$ and $0.6$ for diffusion of droplets in isotropic turbulence (Sato & Yamamoto Reference Sato and Yamamoto1987; Gopalan, Malkiel & Katz Reference Gopalan, Malkiel and Katz2008). We chose a value of $k_{D}=0.3$ in accordance with Li et al. (Reference Li, Miller, Wang, Koley and Katz2017) for this study. Li et al. (Reference Li, Miller, Wang, Koley and Katz2017) fit the values of $u^{\prime }(t)$ (in $\text{m}~\text{s}^{-1}$ ) and $\unicode[STIX]{x1D716}$ (in $\text{m}^{2}~\text{s}^{-3}$ ) with a power law in time. The data can be represented as $(\unicode[STIX]{x1D716}/\unicode[STIX]{x1D716}_{0})=(t/t_{0})^{p}$ and $(u^{\prime }/u_{0}^{\prime })=(t/t_{0})^{q}$ where $\unicode[STIX]{x1D716}_{0}\approx 0.2~\text{m}^{2}~\text{s}^{-3}$ , $u_{0}^{\prime }\approx 0.2~\text{m}~\text{s}^{-1}$ and $t_{0}\approx 7~\text{s}$ . The exponents $p$ and $q$ can be related by $p=2q-1$ , and the data were fitted with $q=-0.89$ (Li et al. Reference Li, Miller, Wang, Koley and Katz2017). The integral length scale $L(t)$ is then calculated as $u^{\prime }(t)^{3}/\unicode[STIX]{x1D716}(t)$ .

We solve (3.1) numerically for the number density of the oil droplets. We discretize the size range into $N_{d}=70$ bins and assume that at the initial time all the concentrations are spatially homogeneous in the $z$ direction down to an initial intrusion depth of $z=13$ cm (see figure 1). The concentration equations are solved for each droplet size using a second-order Crank–Nicholson temporal discretization method. The boundary condition at the bottom of the domain, at $z=-25~\text{cm}$ is that of no flux, i.e.  $\bar{n}_{i}w_{r}-D(\unicode[STIX]{x2202}\bar{n}_{i}/\unicode[STIX]{x2202}z)=0$ . A Neumann boundary condition is applied at the top surface, i.e.  $(\unicode[STIX]{x2202}\bar{n}_{i}/\unicode[STIX]{x2202}z)=0$ . We initialize the concentration of each diameter bin with the measured concentration at $z=-11.1~\text{cm}$ which was recorded after $5~\text{s}$ of impingement in the experiment. We integrate the model using different values of $K^{\ast }$ ranging from $0$ to $1$ .

Figure 2.

Time evolution of the droplet size distribution for two values of $K^{\ast }$ , at the measurement location. The symbols correspond to the experimental data, $t=5~\text{s}$  (○) represents the initial condition where experiment and simulation are matched, $t=15~\text{s}$  (*),  $t=35~\text{s}$  (▫),  $t=55~\text{s}$  (♢), while the lines correspond to the simulation, $t=15~\text{s}$  (- - -),  $t=35~\text{s}$  (- $\cdot$ - $\cdot$ ) and  $t=55~\text{s}$  ( $\cdots$ ). The green stars $(\star )$ in (a) correspond to the size distribution at $t=55~\text{s}$ using $N_{d}=15$ bins.

In figure 2 we compare the modelled size distributions (lines) to the experimental data (symbols) at various times, for $K^{\ast }=0.2$ and $K^{\ast }=1$ , at the measurement location $z=-11.1~\text{cm}$ . Since the initial condition (black circles in figure 2) already includes the effects of significant initial breakup of the oil, a size distribution is already formed. Since the energy dissipation at the point of measurement, $z=-11.1~\text{cm}$ is relatively low, the rate of further breakup is not very large and thus the effect of $K^{\ast }$ in the model is subtle. Nevertheless, close inspection shows that there is too much breakup effect for the large droplets for $K^{\ast }=1$ , as we can see that the number density of the larger droplets is lower than the experimental data, especially for later times. Qualitatively, it appears that $K^{\ast }=0.2$ captures the distribution slightly better, for both small and large droplets at the various time instants. We also calculate the size distribution in (3.1) using a coarser discretization of $N_{d}=15$ bins. The resulting number density $\bar{n}$ is shown in figure 2(a) at $t=55~\text{s}$ using green stars. We see that for this coarser resolution there is a good agreement with the $N_{d}=70$ case.

In order to make a quantitative comparison we define an error measure ${\mathcal{E}}$ as the integrated squared difference between the logarithmic experimental and modelled size distributions. The error is calculated for each droplet size, and integrated over all bins (the bin size varies logarithmically),

(3.4) $$\begin{eqnarray}{\mathcal{E}}=\mathop{\left\langle \mathop{\sum }_{i_{min}}^{i_{max}}[\log _{10}(\bar{n}_{expt}(d_{i}))-\log _{10}(\bar{n}_{mod}(d_{i}))]^{2}\frac{\unicode[STIX]{x1D6FF}d_{i}}{d_{i}}\right\rangle }\nolimits_{\!t},\end{eqnarray}$$

where $\unicode[STIX]{x1D6FF}d_{i}$ is the bin width, $\bar{n}_{expt}$ refers to the experimental size distribution and $\bar{n}_{mod}$ is the modelled one. The maximum diameter at which the experimental data are reported is $d\approx 500~\unicode[STIX]{x03BC}\text{m}$ . Therefore, we select $i_{max}=52$ corresponding to $d=505~\unicode[STIX]{x03BC}\text{m}$ . And we use $i_{min}=4$ corresponding to $d=86~\unicode[STIX]{x03BC}\text{m}$ . The error is averaged over the three available times, $t=15~\text{s}$ , $35~\text{s}$ and $55~\text{s}$ .

Figure 3.

Average square error between predicted and measured logarithms of number densities averaged over three times during the evolution, at the measurement position, as a function of the breakup constant $K^{\ast }$ assumed in the model.

As can be seen from figure 3, the absolute value of the error is smallest at $K^{\ast }\approx 0.2$ and hence this value is chosen as the fitted parameter for future applications of the model. Note that $K^{\ast }=0$ corresponds to the case without breakup. We can see from figure 3 that the error is larger, showing the improved agreement when including the breakup term in the model.

As stated earlier in the text, previous breakup models use an inertial-range scaling for the eddy fluctuation velocity for the entire range of scales,

(3.5) $$\begin{eqnarray}u_{e}^{2}=2.1(\unicode[STIX]{x1D716}d_{e})^{2/3}.\end{eqnarray}$$

This approach results in overestimated velocities of eddies in the viscous range. In order to illustrate the net effects of this overestimation of turbulence at small scales in the overall model predictions, we can use (3.5) in (2.8) to compute the breakup frequency using $K^{\ast }=0.2$ for this scaling of eddy velocity (through numerical integration). We solve (3.1) and plot the resulting size distributions in figure 4. Figure 4(a) compares the computed size distributions with the experimental data. We see that there is too much breakup effect resulting in too few of the larger droplets and an increase in the concentration of the intermediate size droplets. In order to obtain a better agreement with the experiment we would need to set $K^{\ast }\sim O(10^{-2})$ . Thus, we conclude that in order to maintain reasonable range of value for the parameter $K^{\ast }$ (which is expected to be of order unity), it is important to capture both the inertial and viscous range scalings as in (2.10) for the eddy fluctuation velocity.

Figure 4.

Time evolution of the droplet size distribution for $K^{\ast }=0.2$ . Panel (a) uses an inertial scaling of $u_{e}$ . In (b) we plot the effect of using a normal distribution proposed by Coulaloglou & Tavlarides (Reference Coulaloglou and Tavlarides1977). The symbols correspond to the experimental data, $t=5~\text{s}$  (○) represents the initial condition where experiment and simulation are matched, $t=15~\text{s}$  (*),  $t=35~\text{s}$  (▫),  $t=55~\text{s}$  (♢), while the lines correspond to the simulation, $t=15~\text{s}$  (- - -),  $t=35~\text{s}$  (- $\cdot$ - $\cdot$ ) and $t=55~\text{s}$  ( $\cdots$ ).

We can also study the effect of using a different breakup probability model $\unicode[STIX]{x1D6FD}(d_{i},d_{j})$ . We use a truncated normal distribution of Coulaloglou & Tavlarides (Reference Coulaloglou and Tavlarides1977) for simplicity. In this case it is assumed that the daughter droplet sizes for a parent drop of diameter $d_{j}$ are normally distributed about a mean value, $\bar{d}=d_{j}/2^{1/3}$ , i.e.

(3.6) $$\begin{eqnarray}\unicode[STIX]{x1D6FD}(d_{i},d_{j})=\frac{1}{s\sqrt{2\unicode[STIX]{x03C0}}}\exp \left[-\frac{(d_{i}-\bar{d})^{2}}{2s^{2}}\right],\end{eqnarray}$$

where $s=d_{j}/(3\times 2^{1/3})$ . This gives us a maximum probability for equal volume breakup. We plot the resulting size distributions for $K^{\ast }=0.2$ in figure 4(b). We observe a bump in the size distribution at $d\approx 300~\unicode[STIX]{x03BC}\text{m}$ and a more rapid cutoff of the large droplets as compared to figure 2(a). Additionally, we do not find an optimum value of $K^{\ast }$ that minimizes (3.4), i.e. the error grows with increasing $K^{\ast }$ . Hence, we may conclude that the form of the particular droplet breakup probability distribution is also important, though it plays a weaker role as compared to the effect of including the viscous range for this particular wave breaking case.

We demonstrate the effect of an increase in the assumed maximum eddy size allowed to break the droplets (upper limit of integration in (2.12)) in figure 5. We find that for a maximum eddy size 20 % larger than the droplet size ( $d_{e,max}=1.2d_{i}$ ), the fitted prefactor is reduced to $K^{\ast }=0.1$ . Using this value of $K^{\ast }$ we see from figure 5(a) that the steady-state size distribution is in good agreement with case where $d_{e,max}=d_{i}$ and $K^{\ast }=0.2$ . Clearly, however, the resulting breakup frequency shows some dependence on the upper cutoff of the integral (figure 5 b).

Figure 5.

We depict the effect of changing the maximum diameter size on the size distribution and breakage frequency. In (a) the symbols correspond to the experimental data, $t=5~\text{s}$  (○) represents the initial condition where experiment and simulation are matched, $t=15~\text{s}$  (*),  $t=35~\text{s}$  (▫),  $t=55~\text{s}$  (♢), while the lines correspond to the simulation using $K^{\ast }=0.2$ , $t=15~\text{s}$  (- - -, red), $t=35~\text{s}$  (- $\cdot$ - $\cdot$ ) and  $t=55~\text{s}$  ( $\cdots$ ). The green thick dashed line (- - -, green thick) corresponds to the case where $d_{e,max}=1.2d_{i}$ and $K^{\ast }=0.1$ was used to calculate $g(d_{i})$ . In (b) we compare the resulting breakup frequency for the two cases, $d_{e,max}=1.2d_{i}$  (- - -, black) and $d_{e,max}=d_{i}$  (——).

3.2 Breakup of oil in a turbulent jet (Eastwood et al. Reference Eastwood, Armi and Lasheras2004)

In order to begin testing the model when applied to a system with different flow properties than the case for which $K^{\ast }$ was fitted, we consider the experiment by Eastwood et al. (Reference Eastwood, Armi and Lasheras2004). Oil droplets of varying density, viscosity and interfacial tension are injected continuously at the centreline in the fully developed region of a turbulent water jet. The downstream evolution of the centreline velocity and dissipation rate was well characterized and found to obey classic scalings of a turbulent round jet,

(3.7a,b ) $$\begin{eqnarray}\frac{U_{0}}{U(x)}=\frac{1}{C_{u}}\left(\frac{x}{D_{j}}-\frac{x_{0}}{D_{j}}\right),\quad \frac{\unicode[STIX]{x1D716}D_{j}}{U_{0}^{3}}=C\left(\frac{x}{D_{j}}-\frac{x_{0}}{D_{j}}\right)^{-4},\end{eqnarray}$$

where $C_{u}=4.08$ and $C=36$ are empirical constants. The virtual origin $x_{0}/D_{j}=5.47$ was found by fitting the experimental data with (3.7). The breakup and downstream evolution of oil droplets were tracked using digital image processing techniques (Eastwood et al. Reference Eastwood, Armi and Lasheras2004). The authors defined a characteristic droplet size $d_{max}$ whose number density $n(d_{max})$ can only change due to its own breakup but cannot increase due to breakup of other (larger) droplets. This condition isolates the effect of the breakup frequency on the evolution of the number density. Mathematically the evolution of the size distribution can be tracked using (2.1), where for the size $d_{max}$ we can drop the first term in (2.2). Additionally for the quasi-one-dimensional steady-state jet flow considered, we can write the PBE for the largest size as

(3.8) $$\begin{eqnarray}\unicode[STIX]{x1D735}_{x}\boldsymbol{\cdot }[\boldsymbol{v}(d_{max},\boldsymbol{x})n(d_{max},\boldsymbol{x})]=-g(d_{max},\boldsymbol{x})n(d_{max},\boldsymbol{x}),\end{eqnarray}$$

where $n(d_{max},x)=N(d_{max},x)/V_{w}$ , with $V_{w}$ the volume of the interrogation window and $N(d_{\max },x)$ the total number of droplets measured in the window at $x$ . The droplet velocity $\boldsymbol{v}(d_{max},\boldsymbol{x})$ can be approximated by the local mean velocity of the turbulent jet, $U(x)$ . Equation (3.8) can then be written for the number of droplets $N(d_{max},x)$ as,

(3.9) $$\begin{eqnarray}\frac{\text{d}}{\text{d}x}[U(x)N(d_{max},x)]=-N(d_{max},x)g(d_{max},x).\end{eqnarray}$$

Table 1.

Dispersed fluid properties.

The maximum diameter $d_{max}$ represented the size where at least $80\,\%$ of the volume of the distribution was contained in droplets smaller than $d_{max}$ . The overall decay of $N(d_{max},x)$ with downstream distance was found to be similar when this criterion was enforced, thereby ensuring that the evolution of the largest size class is captured. In order to validate our model with the experiments, we solve (3.9) using a fourth-order explicit Runge–Kutta method (ode45 in MATLAB). Equation (2.17) is used to calculate the breakup frequency $g(d_{max},x)$ with $K^{\ast }=0.2$ for all the oils considered. A summary of the physical properties of the different oils used for the simulation along with the corresponding size $d_{max}$ is provided in table 1.

Figure 6.

Evolution of $N(d_{max},x)$ with downstream distance for the four different oils in table 1. The symbols correspond to the experimental data of the different oils; 50 cSt silicone oil (○), olive oil (♢), 10 cSt silicone oil (▫) and heptane (*). The different lines correspond to the model; 50 cSt silicone oil (——), olive oil (- $\cdot$ - $\cdot$ ), 10 cSt silicone oil ( $\cdots$ ) and heptane (- - -) with $K^{\ast }=0.2$ . Also shown are the model for heptane () with $K^{\ast }=0.1$ and 10 cSt silicone oil () with $K^{\ast }=0.15$ .

We compare the model predictions with the experimental data in figure 6. We see that the model does a good job of capturing the decay of the number of droplets for the 50 cSt silicone oil and the olive oil cases. For heptane and 10 cSt silicone oil the predicted decay is too rapid and $K^{\ast }=0.2$ appears not to be the optimal value, while $K^{\ast }=0.1$ and $K^{\ast }=0.15$ are seen to give better agreement with the data. We have not found any obvious parameter dependencies that could explain the different $K^{\ast }$ . So based on the argument that $K^{\ast }=0.2$ gives a good match for a majority of the applications considered, we proceed to use this value for the LES test case in the next section.

4.1 LES equations for Eulerian description of scalar transport

We adapt the code used by Yang, Meneveau & Shen (Reference Yang, Meneveau and Shen2014) and Yang et al. (Reference Yang, Chen, Chamecki and Meneveau2015) for simulations of hydrocarbon plume dispersion in ocean turbulence to simulate here a turbulent round jet with an imposed cross-flow and with oil droplets being released near the source of the jet. Let $\boldsymbol{x}=(x,y,z)$ with $x$ and $y$ the horizontal coordinates and $z$ the vertical direction, and let $\boldsymbol{u}=(u,v,w)$ be the corresponding velocity components. The jet and surrounding fluid are governed by the three-dimensional incompressible filtered Navier–Stokes equations with a Boussinesq approximation for buoyancy effects,

(4.1) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}\boldsymbol{\cdot }\tilde{\boldsymbol{u}}=0, & \displaystyle\end{eqnarray}$$

(4.2) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}\tilde{\boldsymbol{u}}}{\unicode[STIX]{x2202}t}+\tilde{\boldsymbol{u}}\boldsymbol{\cdot }\unicode[STIX]{x1D735}\tilde{\boldsymbol{u}}=-\frac{1}{\unicode[STIX]{x1D70C}_{c}}\unicode[STIX]{x1D735}\tilde{P}-\unicode[STIX]{x1D735}\boldsymbol{\cdot }\unicode[STIX]{x1D749}^{d}+\left(1-\frac{\unicode[STIX]{x1D70C}_{d}}{\unicode[STIX]{x1D70C}_{c}}\right)\mathop{\sum }_{i}(V_{d,i}{\tilde{n}}_{i})g\boldsymbol{e}_{3}, & \displaystyle\end{eqnarray}$$

(4.3) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}{\tilde{n}}_{i}}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }(\tilde{\boldsymbol{v}}_{i}{\tilde{n}}_{i})+\unicode[STIX]{x1D735}\boldsymbol{\cdot }\unicode[STIX]{x1D745}_{i}=\widetilde{S_{b,i}},\quad i=1,2\cdots N_{d}. & \displaystyle\end{eqnarray}$$

A tilde denotes a variable resolved on the LES grid, $\tilde{\boldsymbol{u}}$ is the filtered fluid velocity, $\unicode[STIX]{x1D70C}_{d}$ is the density of the droplet, $\unicode[STIX]{x1D70C}_{c}$ is the carrier fluid (sea water) density, $V_{d,i}=\unicode[STIX]{x03C0}d_{i}^{3}/6$ is the volume of a spherical oil droplet of size $d_{i}$ , $\unicode[STIX]{x1D749}=(\widetilde{\boldsymbol{u}\boldsymbol{u}}-\tilde{\boldsymbol{u}}\tilde{\boldsymbol{u}})$ is the subgrid-scale stress tensor with deviatoric part, $\unicode[STIX]{x1D749}^{d}=\unicode[STIX]{x1D749}-[tr(\unicode[STIX]{x1D749})/3]\unicode[STIX]{x1D644}$ where $\unicode[STIX]{x1D644}$ is the identity tensor, $\tilde{P}=\tilde{p}/\unicode[STIX]{x1D70C}_{r}+tr(\unicode[STIX]{x1D749})/3+|\tilde{\boldsymbol{u}}|^{2}/2$ is the pseudo-pressure, with $\tilde{p}$ the resolved dynamic pressure, ${\tilde{n}}_{i}$ is the resolved number density of the droplet of size $d_{i}$ and $\boldsymbol{e}_{3}$ is the unit vector in the vertical direction. The filtered version of the transport equation for the number density ${\tilde{n}}_{i}(\boldsymbol{x},t;d_{i})$ is given by (4.3). The term $\unicode[STIX]{x1D745}_{i}=(\widetilde{\boldsymbol{v}_{i}n_{i}}-\tilde{\boldsymbol{v}}_{i}{\tilde{n}}_{i})$ is the subgrid-scale concentration flux of oil droplets of size $d_{i}$ (no summation over $i$ implied here).

Closure for the SGS stress tensor $\unicode[STIX]{x1D749}^{d}$ is obtained from the Lilly–Smagorinsky eddy viscosity model (Smagorinsky Reference Smagorinsky1963), $\unicode[STIX]{x1D70F}_{ij}^{d}=-2\unicode[STIX]{x1D708}_{\unicode[STIX]{x1D70F}}\tilde{\unicode[STIX]{x1D61A}}_{ij}=-2(c_{s}\unicode[STIX]{x1D6E5})^{2}|\tilde{\unicode[STIX]{x1D64E}}|\tilde{\unicode[STIX]{x1D61A}}_{ij}$ , where $\tilde{\unicode[STIX]{x1D61A}}_{ij}=(\unicode[STIX]{x2202}\tilde{u} _{i}/\unicode[STIX]{x2202}x_{j}+\unicode[STIX]{x2202}\tilde{u} _{j}/\unicode[STIX]{x2202}x_{i})/2$ is the resolved strain rate tensor, $\unicode[STIX]{x1D708}_{\unicode[STIX]{x1D70F}}$ is the SGS eddy viscosity and $\unicode[STIX]{x1D6E5}$ is the grid (filter) scale. The Smagorinsky coefficient $c_{s}$ is determined dynamically during the simulation using the Lagrangian averaging scale-dependent dynamic (LASD) SGS model (Bou-Zeid, Meneveau & Parlange Reference Bou-Zeid, Meneveau and Parlange2005), which accounts for spatial inhomogeneity. The SGS scalar flux $\unicode[STIX]{x1D745}_{i}$ is modelled using an eddy-diffusion SGS model. We use the approach of prescribing a turbulent Schmidt and Prandtl number, $Pr_{\unicode[STIX]{x1D70F}}=Sc_{\unicode[STIX]{x1D70F}}=0.4$ (Yang et al. Reference Yang, Chen, Socolofsky, Chamecki and Meneveau2016). The SGS flux can then be parameterized as $\unicode[STIX]{x03C0}_{n,i}=-(\unicode[STIX]{x1D708}_{\unicode[STIX]{x1D70F}}/Sc_{\unicode[STIX]{x1D70F}})\unicode[STIX]{x1D735}{\tilde{n}}_{i}$

The LES code with the LASD model has been used in a number of prior LES studies (see Porté-Agel, Meneveau & Parlange Reference Porté-Agel, Meneveau and Parlange2000; Tseng, Meneveau & Parlange Reference Tseng, Meneveau and Parlange2006; Yang et al. Reference Yang, Meneveau and Shen2014). With the evolution of oil droplet concentrations being simulated, their effects on the fluid velocity field are modelled and implemented in (4.2) as a buoyancy force term (the last term on the right-hand side of the equation) using the Boussinesq approximation. A basic assumption for treating the oil droplets as an active scalar field being dispersed by the fluid motion is that the volume and mass fractions of the oil droplets are small within a computational grid cell.

The droplet transport velocity $\tilde{\boldsymbol{v}}_{i}$ is calculated by an expansion in the droplet time scale $\unicode[STIX]{x1D70F}_{d,i}=(\unicode[STIX]{x1D70C}_{d}+\unicode[STIX]{x1D70C}_{c}/2)d_{i}^{2}/(18\unicode[STIX]{x1D707}_{f})$ (Ferry & Balachandar Reference Ferry and Balachandar2001). The expansion is valid when $\unicode[STIX]{x1D70F}_{d,i}$ is much smaller than the resolved fluid time scales, which requires us to have a grid Stokes number $St_{\unicode[STIX]{x1D6E5},i}=\unicode[STIX]{x1D70F}_{d,i}/\unicode[STIX]{x1D70F}_{\unicode[STIX]{x1D6E5}}\ll 1$ , where $\unicode[STIX]{x1D70F}_{\unicode[STIX]{x1D6E5}}$ is the turbulent eddy turnover time at scale $\unicode[STIX]{x1D6E5}$ . The transport velocity of droplets of size $d_{i}$ , $\tilde{\boldsymbol{v}}_{i}$ , is given by (Ferry & Balachandar Reference Ferry and Balachandar2001)

(4.4) $$\begin{eqnarray}\tilde{\boldsymbol{v}}_{i}=\tilde{\boldsymbol{u}}+w_{r,i}\boldsymbol{e}_{3}+(R-1)\unicode[STIX]{x1D70F}_{d,i}\left(\frac{\text{D}\tilde{\boldsymbol{u}}}{\text{D}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }\unicode[STIX]{x1D749}\right)+O(\unicode[STIX]{x1D70F}_{d,i}^{3/2}),\end{eqnarray}$$

where $w_{r,i}$ is the droplet terminal (rise) velocity, $\boldsymbol{e}_{3}$ is the unit vector in the vertical direction and $R=3\unicode[STIX]{x1D70C}_{c}/(2\unicode[STIX]{x1D70C}_{d}+\unicode[STIX]{x1D70C}_{c})$ is the acceleration parameter. A more detailed discussion of the droplet rise velocity in (4.4) can be found in Yang et al. (Reference Yang, Chen, Socolofsky, Chamecki and Meneveau2016).

The term $\widetilde{S_{b,i}}$ in (4.3) represents the rate of change of droplet number density due to breakup and is modelled using (2.7) and (2.21) described in § 2. In implementing the model, when evaluating the filtered source term, we use the filtered parameters (e.g. grid-scale dissipation rate, etc.), that is to say, we assume $\widetilde{gn}\approx \tilde{g}{\tilde{n}}$ , and further that $\widetilde{g(\unicode[STIX]{x1D716},\ldots )}\approx g(\tilde{\unicode[STIX]{x1D716}},\ldots )$ . This means that we neglect the subgrid correlations between locally fluctuating dissipation rates at scales smaller than the grid scale and subgrid fluctuations in the concentration field. The neglect of such subgrid-scale contributions to the modelled source terms must be kept in mind, especially for applications at very high Reynolds numbers when the Kolmogorov scale is much smaller than the grid scale where further refinements and new subgrid models may be required.

The breakup rate $g_{i}$ is evaluated using the fits that depend on the local Reynolds number expressed in terms of the local rate of dissipation. From the SGS model, the local rate of dissipation at the LES grid scale is given by

(4.5) $$\begin{eqnarray}\unicode[STIX]{x1D716}(\boldsymbol{x},t)=2(c_{s}\unicode[STIX]{x1D6E5})^{2}|\tilde{\unicode[STIX]{x1D64E}}|\tilde{S}_{ij}\tilde{S}_{ij}.\end{eqnarray}$$

In order to capture a range of sizes the number density is discretized into $N_{d}=15$ bins for droplets between $d_{1}=20~\unicode[STIX]{x03BC}\text{m}$ up to $d_{N_{d}}=1~\text{mm}$ . The droplet size range is discretized according to

(4.6) $$\begin{eqnarray}d_{i+1}=2^{p}d_{i},\quad i=1,2,\ldots ,N_{d},\end{eqnarray}$$

where, $p=0.403$ . We then solve $N_{d}$ separate transport equations for the number densities ${\tilde{n}}_{i}(\boldsymbol{x},t;d_{i})$ with $i=1,2,\ldots ,N_{d}$ .

The equations (4.1) and (4.2) are discretized using a pseudo-spectral method on a collocated grid in the horizontal directions and using centred finite differencing on a staggered grid in the vertical direction (Albertson & Parlange Reference Albertson and Parlange1999). The velocity field is advanced using a fractional-step method with a second-order Adams–Bashforth scheme for the time integration. The transport equations for the droplet number densities, equation (4.3), are discretized as in Chamecki, Meneveau & Parlange (Reference Chamecki, Meneveau and Parlange2008), Yang et al. (Reference Yang, Meneveau and Shen2014) and Yang et al. (Reference Yang, Chen, Chamecki and Meneveau2015), by a finite-volume algorithm with a bounded third-order upwind scheme for the advection term (Gaskell & Lau Reference Gaskell and Lau1988). This approach prevents the appearance of negative local concentrations that tend to occur when using spectral methods. Information between the finite-volume and pseudo-spectral grids is exchanged using a conservative interpolation scheme developed by Chamecki et al. (Reference Chamecki, Meneveau and Parlange2008).

Figure 7.

Sketch of the simulation domain and dimensions.

4.2 Simulation set-up

A sketch of the simulation domain is shown in figure 7. We simulate a turbulent jet with imposed cross-flow aiming to reproduce the experiments of Murphy et al. (Reference Murphy, Xue, Sampath and Katz2016), who studied a turbulent oil jet in cross-flow and measured droplet size distributions using a submersible digital inline holography technique. The experiments were carried out in a $2.5~\text{m}\times 0.9~\text{m}\times 0.9~\text{m}$ acrylic tank. The injection nozzle was connected to a carriage, driven by a stepper motor to generate desired cross-flow speeds. The injection nozzle had a $D_{expt}=4~\text{mm}$ diameter orifice and was located at a distance of $0.14~\text{m}$ from the bottom of the tank. The oil was injected at a flow rate of $Q=1.9~\text{l}~\text{min}^{-1}$ , (i.e. an injection velocity of $U_{j,expt}=2.5~\text{m}~\text{s}^{-1}$ ) and the carriage was towed at a speed of $U_{c}=0.15~\text{m}~\text{s}^{-1}$ . In the experiments performed, the number of droplets was measured using a holocam fixed at the centre of the tank in the horizontal plane, and at a vertical distance of $0.47~\text{m}$ from the nozzle exit. It thus sampled different sections of the jet in the cross-stream direction in the course of its evolution. Numbers of droplets in various size bins were measured at two times, at $t=3.4~\text{s}$ and $t=6.9~\text{s}$ . Additionally, the total oil distribution calculated by summing over five time measurements was recorded. As the nozzle in the experiments moved with a constant speed of $0.15~\text{m}~\text{s}^{-1}$ , this corresponds to a downstream location of measurement at $x=0.76~\text{m}$ and $x=1.3~\text{m}$ respectively, in a frame moving with the jet nozzle (as will be done in the simulation). The total oil concentration at the measurement location height corresponds to the sum of the number of droplets measured as a function of downstream distance. In the experiments, the droplet size distributions were measured in three realizations of the experiments, hence the resulting size distribution and droplet concentrations were not fully statistically converged but the shape of the size distribution (relative size distribution) was well captured in the measurements.

As shown in figure 7, the simulations are performed in a rectangular box of size $(L_{x},L_{y},L_{z})=(2,0.781,1)~\text{m}$ in which the jet is stationary but a constant inflow (cross-flow) velocity is prescribed. The domain size is chosen to mimic the experimental set-up and the length is sufficiently large to capture the complete turning of the jet. The cross-flow is imposed along the $x$ direction while the jet is pointed in the $z$ direction. In order to handle the inflow and outflow conditions at the two boundaries in the $x$ direction within the code that uses a Fourier-series-based pseudo-spectral method, we specify a fringe zone that starts at $x=1.666~\text{m}$ , which damps out the velocity towards the inflow value. The simulations use a grid with $N_{x}\times N_{y}\times N_{z}=384\times 150\times 384$ points for spatial discretization, and a timestep $\unicode[STIX]{x0394}t=0.0002~\text{s}$ for time integration. The resolution in the horizontal directions is $\unicode[STIX]{x0394}x=\unicode[STIX]{x0394}y=5.2~\text{mm}$ , and is $\unicode[STIX]{x0394}z=2.6~\text{mm}$ in the vertical direction for the finite difference method. Since the latter can be regarded as requiring approximately twice as much resolution for the same accuracy, we regard the overall numerical resolution to be approximately $5.2~\text{mm}$ .

The injected jet is modelled in the LES using a locally applied vertically upward pointing body force, since at the LES resolution used in the simulation it is not possible to resolve the small-scale features of the injection nozzle. The applied force is spatially smoothed in a region over three grid points in $x,y$ and $z$ using a super-Gaussian smoother (of order 5 and width $\unicode[STIX]{x1D70E}_{G}=\unicode[STIX]{x0394}x$ ) centred at $(x,y,z)=(0.25,0.39,0.85)~\text{m}$ .

The resulting injection velocity is controlled by the strength of the imposed body force. Since the details of the nozzle cannot be resolved, the body force is applied at a location downstream of the real nozzle, where the jet in the experiment is expected to have grown to a scale at which the numerical grid is sufficient to resolve at least the mean velocity profile. Using a classical round jet without cross-flow for calibration, the force is chosen such that the resulting centreline velocity in the LES matches the mean centreline velocity expected for the experiment at that location. More details are provided in appendix B. The location where the body force is applied is found to be $38.2~\text{mm}$ downstream (i.e. above) of the experimental nozzle’s virtual origin (see appendix B), while the injection velocity in the simulation is determined to be $U_{j,sim}=1.6~\text{m}~\text{s}^{-1}$ , i.e. lower than that in the experiment owing to the fact that the centreline velocity has already decayed at the simulated injection point. A uniform cross-flow of $U_{cross}=0.15~\text{m}~\text{s}^{-1}$ is imposed at the inflow boundary in order to simulate a jet in cross-flow. The droplet number density fields are initialized to zero everywhere. Oil droplets of size $d=1~\text{mm}$ are injected at the jet source after a delay of $1~\text{s}$ to allow the flow to be established. The number density transport equation for the bin corresponding to the largest droplets ( $i=N_{d}$ with $d_{i}=1~\text{mm}$ ) contains a source term on the right-hand side which represents injection with a specified volume flow rate that matches the experimental value of $Q=1.9~\text{l}~\text{min}^{-1}$ as in Murphy et al. (Reference Murphy, Xue, Sampath and Katz2016). The source is distributed over two grid points in the $z$ direction with weights $0.25$ and $0.75$ and over three grid points in the $x$ and $y$ directions centred at $x=0.245~\text{m}$ , $y=0.385~\text{m}$ with weights $0.292$ for the centre and $0.177$ for the neighbouring points. The physical properties of the oil and the simulation parameters are given in tables 2 and 3. The Weber number ( $We$ ) has been calculated using the near-nozzle dissipation $\langle \unicode[STIX]{x1D716}\rangle =30~\text{m}^{2}~\text{s}^{3}$ and the injection droplet size $d=1~\text{mm}$ .

Table 2.

Physical properties of fluids used in the simulation.

Table 3.

Simulation parameters.

The jet in cross-flow is simulated for a total of $26~\text{s}$ , corresponding to $13\times 10^{4}$ timesteps. The oil droplets are released after $t=1~\text{s}$ giving sufficient time for the jet to become fully turbulent. Starting from $t=12~\text{s}$ , 350 three-dimensional snapshots of the entire simulation domain are recorded for statistical analysis with an interval of $0.04~\text{s}$ (200 timesteps) between each snapshot.

4.4 Analysis of droplet size distribution variability in LES

As mentioned earlier, LES enables us to diagnose variability of the droplet size distributions and number density transport that RANS cannot obtain, since the latter only predicts time or ensemble average values. In order to illustrate this capability of LES, we now ask what is the inherent variability of typical droplet sizes as well as that of other practically relevant quantities.

Figure 16.

Representative time signals (left panels) and histograms (right panels) for the Sauter mean diameter, $d_{32}$ at two downstream locations on the centreline. Dotted lines denote mean values.

Figure 17.

Representative time signals (left panels) and histograms (right panels) of $\log _{10}(\unicode[STIX]{x1D716})$ at two downstream locations on the centreline. $\unicode[STIX]{x1D716}$ is in $\text{m}^{2}~\text{s}^{-3}$ . Dotted lines denote mean value.

We plot a time signal and histograms of the $d_{32}$ diameters at different plume locations in figure 16. We can see that there is a high variability of the diameter about the time-averaged mean value. This variability can be observed in LES since we are solving for three-dimensional time-dependent number density fields for each bin and so $d_{32}$ can be evaluated at any grid point at any timestep. At $x=0.75~\text{m}$ , where the averaged $d_{32}$ diameter is approximately $800~\unicode[STIX]{x03BC}\text{m}$ , we see a variability from  $300~\unicode[STIX]{x03BC}\text{m}$ to $1000~\unicode[STIX]{x03BC}\text{m}$ . The probability density function (PDF) also shows non-Gaussianity, with two peaks, clearly showing that the average values of $d_{32}$ do not provide a complete description. The peak near $900~\unicode[STIX]{x03BC}\text{m}$ is affected by the discrete bins. This scale corresponds to the first bin considered in the summation corresponding to $d=1~\text{mm}$ .

Recall that in LES the breakup time scale depends upon the local value of dissipation, which is also known to be highly intermittent in turbulent flows. In order to illustrate the (grid-scale-averaged) dissipation intermittency, in figure 17 we show time signals of the logarithm of dissipation as well as histograms. The histogram of the logarithm of dissipation is reminiscent of Gaussian (log-normality) but with a non-Gaussian highly asymmetric tail and some outliers at very low dissipation, corresponding possibly to laminar regions outside the plume. This highly variable quantity then determines the local time scale of droplet breakup in the LES model.

Next, we consider a property that is crucial in determining reaction rates for processes that occur on the droplet surface such as bio-degradation. The total rate of bio-degradation will depend on the total surface area of the oil available for microorganisms to act upon. Given the instantaneous concentration of droplets in each bin, the instantaneous total surface area available for surface reactions can be evaluated according to

(4.10) $$\begin{eqnarray}A_{tot}(\boldsymbol{x},t)=\mathop{\sum }_{i=1}^{N_{d}}{\tilde{n}}_{i}(\boldsymbol{x},t)(\unicode[STIX]{x03C0}d_{i}^{2}).\end{eqnarray}$$

Representative signals and histograms of $A_{tot}(\boldsymbol{x},t)$ are shown in figure 18, again at the two locations $x=0.75~\text{m}$ and $x=0.9~\text{m}$ at $z=0.56~\text{m}$ and the plume centre in the transverse direction.

Figure 18.

Representative time signals (left panels) and histograms (right panels) of total surface area of the oil per cubic metre of fluid at two downstream locations on the centreline. $A_{tot}$ is in $\text{m}^{2}~\text{m}^{-3}$ . Dotted lines denote mean values.

We can see from the panels in figure 18 that there is a high variability of the total area about the mean of approximately $30~\text{m}^{2}$ per cubic metre of water at $x=0.75~\text{m}$ and about $16~\text{m}^{2}$ per cubic metre of water at $x=0.9~\text{m}$ (even though one may expect smaller droplet sizes to be associated with an increase in total surface area, further downstream the total oil concentration has also decreased due to turbulent transport thus leading to the smaller area there). The root mean square of the surface area distribution is quite significant, of similar order of magnitude to the mean area.

It is also instructive to examine time signals and statistics of the breakup source term for each droplet size, $\tilde{S}_{b,i}$ , normalized by the concentration. This normalization can be interpreted as an inverse time scale for the droplet breakup, i.e. it tells us the inverse of the time taken for the number of droplets in any given bin to change appreciably over its existing value, at any given scale. In figure 19 we show representative signals of $\tilde{S}_{b,i}(\boldsymbol{x},t)/{\tilde{n}}_{i}(\boldsymbol{x},t)$ in logarithmic scale, as well as its histograms at two locations. As can be seen from the right panel of figure 19(a), the average values are around $0.5~\text{s}^{-1}$ , with very large variability about this value. It means that it takes about $2$ seconds for the local breakup to appreciably change the local concentration of droplets of size $20~\unicode[STIX]{x03BC}\text{m}$ but occasionally the breakup can be far more rapid.

Figure 19.

Representative time signals (left panels) and histograms (right panels) of $\log _{10}(\tilde{S}_{b,i}/{\tilde{n}}_{i})$ for $d=20~\unicode[STIX]{x03BC}\text{m}$ plotted at two downstream locations of the plume at a fixed height. The values of $\tilde{S}_{b,i}/{\tilde{n}}_{i}$ are given in $1/s$ . Dotted lines denote mean values.

Finally, we document the breakup source term by plotting it in linear units for different droplet sizes, at two different locations as shown in figure 20. We can see that the time signals for the source term are highly intermittent, with the largest size (i.e. 15th bin) acting as a source for the smaller ones (negative source term in its transport equation). Further downstream at $x=1.56~\text{m}$ , figure 20(b) shows much smaller frequencies indicating a decreased breakup of the largest droplets. Some of the intermediate bins display both positive and negative values, as some intermediate droplet sizes act as both sources and sinks at different locations along the plume (for example see panel for $\tilde{S}_{b,7}/{\tilde{n}}_{7}$ in figure 20 c).

We still see a significant variance in the volume median diameter ( $d_{32}$ ) at this location despite the magnitude of the normalized source terms for the larger droplets being small. Clearly, the turbulent nature of the flow prevents us from solely relying on the averaged quantities to provide us with a complete view of the droplet size distribution in this flow, while LES contains significant amount of new information regarding the fluctuations, at least down to the grid scale.