Available online at www.sciencedirect.

com

Procedia Computer Science 4 (2011) 781790

Stochastic Droplet-Fiber Collisions

Simon Schrodera , Ferdinand Olawskyb , Martin Hering-Bertramc , Hans Hagena
a Computer Graphics and HCI Group, University of Kaiserslautern, Germany
b Fraunhofer Institute for Industrial Mathematics ITWM, Kaiserslautern, Germany
c Department of Technology and Bionics, Rhine-Waal University of Applied Sciences, Germany

Abstract
We present a novel method to simulate bers in ows transporting droplets. We discuss how to model the collisions
of droplets and bers. Since droplet-ber collision is computationally expensive we use a new stochastic approach
modeling the collision probability of droplets and bers. From a continuous spatial distribution of the particle masses
we can calculate the impact on a specic ber at any arbitrary point. This gives an average mass and momentum of
the droplets which is transferred to the ber. We integrate our model in a ber dynamics simulation tool (FIDYST)
used to simulate industrial manufacturing processes.
Keywords: ow simulation, ber, droplet, particles, stochastic process, collision

1. Introduction

Simulation of industrial production processes is still an open problem in many application areas. In this paper
we show fundamental work for simulation of interaction between droplets and bers in an air ow. Commonly,
an air ow is used to inuence the bers movements and the pattern how they arrange themselves on the ground.
E. g. in production processes of glass wool some binder has to be injected as droplets. Having a simulation for this
production process provides a way to plan new production machines or to improve existing ones. Simulating dierent
congurations is less expensive than modifying and testing existing machines.
Up to this point, separate research has been done in each area. First of all, there is our own simulation tool
FIDYST1 (Fiber Dynamics Simulation Tool) from the Fraunhofer ITWM2 for simulation of bers in an air ow. The
ow simulation is done with ANSYS FLUENT and then used as input for FIDYST. Further, there is an old model
for droplet simulation by ORourke [1] and some enhancements to this model (cf. [2, 3, 4]). We use FLUENT for a
combined ow-droplet simulation which employs the original model of ORourke.
There is also some basic research of droplet-ber interaction for stationary bers, e. g. wires. Most papers on this
topic discuss the shape of droplets on bers or the droplets behavior while interacting with the ber (cf. [5, 6, 7, 8, 9]).
But, there is no research on interaction of droplets with multiple moving bers.

Email addresses: simon.schroeder@itwm.fraunhofer.de (Simon Schroder), ferdinand.olawsky@itwm.fraunhofer.de

(Ferdinand Olawsky), Martin.Hering-Bertram@hochschule-rhein-waal.de (Martin Hering-Bertram), hagen@informatik.uni-kl.de
(Hans Hagen)
1 http://www.itwm.fraunhofer.de/en/departments/transport-processes/products/fidyst-fiber-dynamics-simulation-tool.

html
2 http://www.itwm.fraunhofer.de/en/

18770509 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of Prof. Mitsuhisa Sato and Prof. Satoshi Matsuoka
doi:10.1016/j.procs.2011.04.083
782 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790

In this paper, we will provide a new model for simulation of bers and droplets in an air ow. A simple approach
would simulate each droplet and each ber separately, calculating their collisions geometrically. But, even when
using an accelerating structure like an octree or kd-tree this approach would be computationally expensive. Instead,
we calculate a collision probability based on a spatial particle density distribution. The bers are still fully simulated
as their visualization is important for improvement of production processes.
In section 2 we introduce existing work related to our algorithm. This mainly includes the models for droplet
collisions and the ber model used in FIDYST. In the following section, the ber model is extended to describe
collisions of droplets with bers. Section 4 describes the simplest simulation of droplets and bers where we have
a constant droplet density distribution. Additionally, throughout the paper we assume a constant velocity of the ow
and the particles in the whole domain. In section 5 we lift the restriction on the constant particle density. As a
consequence, it must be possible to evaluate the particle density at each point in space individually. This is ensured
by resampling the particle information on a regular grid. In the sections 6 and 7 we discuss the error of our algorithm
to explicit calculations and the application of this algorithm in FIDYST. Finally, we discuss our method and suggest
areas for further research.

2. State of the Art

2.1. Models
There are some models we need in order to describe our droplets and bers. Here, we discuss existing models.
The droplet model was suggested by ORourke [1] and is commonly used to calculate collisions in spray simulations.
For the simulation of the bers in the ow a simple line strip model is used. Data such as mass, velocity, radius, etc.
is given at each sampling point of the line.

2.1.1. Droplet Model

ORourkes droplet model dates back to 1981 and tries to be computationally feasible in calculating droplet col-
lisions. This model is still commonly used in spray simulations. In order to get a signicant simulation of particles,
one would have to inject thousands of particles in each timestep. But, it is too expensive to calculate collisions in such
a model.
Instead, ORourke suggested computational particles which were later named parcels. A parcel represents several
particles with similar properties, like position, velocity, density, mass, temperature, etc. These parcels are used in
particle simulation to calculate their collisions and the interaction with the ow. We will shortly discuss the particle
collision model in the following.
ORourkes model restricts particle collisions to parcels in the same comput ational grid cell with the cell volume
V. As an example, we consider two parcels P1 and P2 . Imagine P1 to contain the larger particles called collector
drops. Particles in P2 are still called droplets. When a collision occurs the smaller droplets are collected by the bigger
collector drops. Suppose, that the parcels P1 and P2 contain n1 and n2 particles respectively, and the relative velocity
between the two parcels is vrel . For a xed timestep t there is the following average of droplets colliding with a
single collector drop:
n2 (r1 + r2 )2 vrel t
n = . (1)
V
Visually, the collision volume, which is the numerator of the fraction above, is a cylinder (see gure 1).

2.1.2. Line Strip Fiber Model

This model describes a ber as a set of points forming a line strip. By this approach the ber is discretized with
a ne resolution. Each of the points holds information about the radius rF , its discretization length l, its density F ,
the velocity vF including the direction, and several other properties. According to internal and external forces, the
position and velocity are updated in every timestep of the simulation (see gure 2). This is described by the governing
system of partial dierential equations used in FIDYST:

tt x(s, t) = s (T (s, t) s x(s, t)) S b ssss x(s, t) + f gravity (s, t) + f air (s, t) + f contact (s, t) (2)
 s x(s, t) =1 t 0. (3)
 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790 783

Vcoll
r2

vrel t
r1 + r 2 air
contact

r1
gravity

Figure 1: Collision probability of Figure 2: Line strip model for bers and external
two droplets described by their col- forces inuencing the ber.
lision volume Vcoll . The larger col-
lector drop p1 captures the droplet
p2 .

The function parameters s and t are the ber parametrization and the time respectively. On the right hand side of
the rst equation (2), the rst two terms describe internal forces due to the tractive force T and the bending stiness
S b . The last three terms represent the external forces like gravity and air drag. The second equation (3) requires the
inextensibility of the ber. The impact of droplets on the ber will be represented in the contact forces f contact . While
the internal forces are discretized with higher order nite dierences the contact models consider the ber as a line
strip. Further information on the equations can be found in [10], [11], and [12].
As this model does not include information for collisions with particles it is extended in the following.

3. Stochastic Droplet-Fiber Collision Model

Similarly as in the droplet model by ORourke, we model the collision volume of a droplet and a ber more
precisely of a droplet and a linear segment of our ber model.

3.1. Collision Volume

Fiber and droplet can only collide when they are close enough and their relative
 movement intersects within the
time interval t. This results in a rectangle with side lengths l and 2 rF + r p as a basis for the box (see gure 3).
Intuitively, this rectangle will be orthogonal to the plane spanned by the ber discretization l and the relative
velocity vrel between the droplet and the ber. Hence, the normal of the rectangle is denoted by n:
(l vrel ) l
n= . (4)
 (l vrel ) l

The height of the box is the path vrel t projected to n. Thus, the height equals vrel t n.
Putting everything together the formula for the collision volume reads:
   
 (l vrel ) l 
Vcoll = vrel  t l 2 rF + r p . (5)
 (l vrel ) l

4. Collision Modeling for Constant Droplet Density Distribution

In this simple model we require that the particle density  is constant over the whole domain. The particle density
is dened as the sum over all particles in an area times their masses m(r p ). This restriction on constant particle density
will help evaluating the new formula with a simple example.
784 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790

1.0

0.8

vrel 0.6
ber segment
0.4

vrel tn 0.2

l 2(rF + r p )
0.002 0.001 0.000 0.001 0.002

Figure 3: Collision volume of a ber and a Figure 4: Gaussian used for MLS with
droplet. n particles = 1000, and r p = 105 m.

4.1. Collision Probability

First, we dene the collision probability:
 
 (l vrel ) l  
Pc (r) = vrel  t l 2 (rF + r) (6)
 (l vrel ) l  m(r)
with m(r) being the mass of a particle with radius r (there is a direct correlation between the mass and the radius by
the formula m(r) =  p 34 r3p as the density of a particle is assumed to be constant). The rst part of the equation is

the collision volume Vcoll as described by equation (5). The second term m(r) describes the number of particles per
volume.
In this case, r is a variable which can be any
 possible particle radius. The distribution of dierent radii is given by
a probability density function Pr (r) satisfying Pr (r)dr = 1.

4.2. Calculation of Mass and Momentum

Since we are interested in the impact of particles on bers, we need to calculate their momentum which will be
transferred during a collision. As for the momentum I holds that

I = mcoll v (7)

and we required the velocity v of the particle to be xed, only the transferred mass has to be calculated in this case.
Given the collision probability Pc (r) and the radius distribution Pr (r), we nd the following to hold:

mcoll = Pc (r) Pr (r) m(r) dr. (8)

In all cases, integration will be over the whole range of r which basically means integration from to +. Inserting
the collision probability of equation (6) into equation (8) we get the following result:


mcoll = vrel n tl2 (rF + r) Pr (r) m(r)dr
m(r)

= vrel n t l 2  Pr (r) (rF + r) dr
   
= vrel n t l 2  rF Pr (r)dr + Pr (r) rdr

= vrel n t l 2  (rF + ) . (9)

where denotes the expected value for r.

 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790 785

In our case, we only have a discrete represention of particles and hence there is only a discrete radius distribution.
Therefore, the expected value is described by
1
= ni ri (10)
n i
with n the total number of particels and ni the number of particles with the same radius ri .

5. Collision Modeling with Varying Droplet Density

Getting a step closer to real simulation, we allow for a varying particle density. This may be inuenced either by
droplets with dierent radii or by an inhomogeneous spatial distribution of the droplets. To be compatible with the
current implementation of the ber simulation in FIDYST the particle density information and the expected value for
the radius should be available on a regular grid. Thus, this information has to be extracted from the particle simulation
for every grid point.

5.1. Density Grid Generation from Parcels

In the particle simulation droplets are injected as parcels in several particle streams. Each stream has its own
injection position, but they are far away from each other compared to the particles radius. Therefore we need to
calculate each parcels inuence to each grid point. The simple approach where a parcel inuences only one grid
point leads to aliasing artifacts. Hence, we look for a solution where the parcels inuence more grid points such that
the result is smoother.
In this paper we use a moving least squares (MLS) approach in which each parcel scatters its inuence according
to a bell curve. Additionally, the parcels movement is also taken into account.
Unfortunately, ORourkes droplet model does not describe the spacial extent of the parcel. Thus, we need to
dene reasonable properties which are used in the MLS. First of all, MLS calculates a value v at a specic position p
by
v= (p pi ) f . (11)
i
f is a function value at point p which in this case are the parameters describing the inuence on the particle density.
The point pi are all grid points, and () is a weighting function dependent on the distance between the points p and
pi . Here, we use a truncated gaussian (see gure 4):
 
d2
(d) = exp 2 exp(). (12)
r
and r are the parameters which have to be chosen for our purposes. r is the maximum radius where (r) = 0.
Outside this radius we set () to zero.

5.1.1. Determination of the Minimum and Maximum Range of Inuence

Assuming a very sparse distribution of particles inside a parcel we choose the maximum radius to be
rmax = r p n particles . (13)
There has also to be a minimum radius where all particles are optimally packed inside a spherical representation of
the parcel. According to the circle packing problem in 2D the volume fraction of equally sized small circles to one
big circle is 2 3 . Correspondingly in 3D, this volume fraction equals 18 . Throughout this paper we consider only
the two-dimensional case, but with this knowledge all formulas can be generalized for three dimensions.
With this information we can establish an equation for the minimum radius:

n particles r2p = rmin
2
. (14)
2 3
Solving this equation for rmin leads to

2 3
rmin = n particles r p
2 . (15)

Obviously, the inuence of the parcel cannot be smaller than rmin .
786 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790

5 cm

injection range: -2.49 cm to +2.49 cm

max radius

old

min radius

7 cm
inection point vrel

3 cm
y
new
x

Figure 5: Inuence of a moving parcel on Figure 6: Simulation setting in 2d show-

surrounding grid points. ing the simulation range, the injection range
and the ber described as circle.

5.1.2. Identication of the Parameter

Since our parameter r is xed to be rmax , only the parameter can determine the inuence of the particle. One
way to do this is to x the inection point of the bell curve. For this, we calculate the root of the second derivative of
(d) and get:
rmax
d= . (16)
2
Now, we have to set the inection point such that the main inuence is inside the minimum radius rmin . Testing
has shown that locating the inection point directly at rmin does not give satisfactory results. The reason for this could
be that the minimum and maximum radius dier largely in their magnitude. Accounting for this, the inection point
should be between these two values (see gure 5). This leads the following new equation:
rmax
rmax + (1 )rmin = . (17)
2
Solving this for gives
2
rmax
= . (18)
2 (rmax + (1 )rmin )2
In the examples we used = 2.5% as testing showed this to be a good choice.

5.1.3. Minimal Particle Distance

Because the particle is moving we have to take care of this, too. For this, we introduce a distance function based
on the particles path and a grid point:
dist pg () = (p + tv) g (19)
with p the particles position, g the grid point, v the particles velocity, and therefore tv the travelled length in this
timestep. Together with the parameter the rst term describes a straight line.
In a rst step, we determine the minimal distance satisfying
 
min dist pg () < rmax . (20)
 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790 787

The for this can be obtained by setting the rst derivative to zero. The formula for the minimial distance in the
two-dimensional case then reads  
v x t (g x p x ) + vy t gy py
min =  2 . (21)
(v x t)2 + vy t
If min r the inuence of this particular particle to the grid point g is zero, and thus we do not need to continue the
calculation.

5.1.4. Finding the Integration Range

In a next step, we calculate the intersections of the distance function with the maximum radius by solving the
equation
!
dist pg () = rmax . (22)
We set 1 and 2 to these intersection points with 1 < 2 and restrict these to the range [0, 1]. The impact of a particle
with position p to a grid point gi is then calculated by integratian over the path:
 2  
i = dist pg () d. (23)
1

5.1.5. Calculating a Parcels Inuence

From these i we can calculate weights wi for each grid point. E. g. if a parcel has mass m then the mass in the
 
grid point gi would be wi m. Certainly, it must hold that i wi = 1 as we have to ensure that i wi m = m. This is
guaranteed by setting
i
wi =  (24)
j j

where j sums over all grid points inuenced by the particle i.

Then, the mass of one parcel in a grid point is calculated by

mi = wi n particles m p (25)

with n particles the number of particles in this particular parcel. We sum this information over all particles. For better
results, we use the particle information from several timesteps. But, only the average mass in one timestep is of
interest. Thus, this value has to be divided by the number of timesteps which is calculated by tmaxttmin . tmax and tmin
represent the maximum and minimum simulation time, respectivly. Finally, we can calculate the particle density in a
grid point by mi
= (26)
i
h x h y tmaxttmin
with h x and hy being the discretization step size of our grid in x- and y-direction.
Similarly, we can calculate the average radius in each grid point. With this, we can bilinearly trilinearly in 3D
interpolate these values for any arbitrary point inside our grid. Then, the interpolated density  and radius can be
used with the formulas from section 4 to calculate collisions of droplets with bers.

6. Evaluation

In this section we compare some simulation results with other approaches. We changed all parameters in our
simulation several times. But, in this paper we only show common properties of all tests by using results which have
the most prominent manifestations of these features. More peculiarities are not found.
All test cases are done in 2D as this is sucient for signicant comparisons. The simulation takes place in a
rectangular area with x [2.5 cm, +2.5 cm] and y [0 cm, 7 cm] (see gure 6). The ber is described by a circle with
its center at (0 cm, 3 cm) and a radius r = 1 cm if nothing else is mentioned.
788 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790

1.4e-10 0.04
algebraic formula
1.38e-10 formula 0.035 grid 100x100
grid 100x100 0.03 grid 50x50
1.36e-10 grid 50x50 geometric

Relative Error
geometric
Mass [kg]

1.34e-10 0.025

1.32e-10 0.02
0.015
1.3e-10
0.01
1.28e-10
0.005
1.26e-10
0
500 450 400 350 300 250 200 150 100 50 500 450 400 350 300 250 200 150 100 50
Number of Streams Number of Streams

(a) Mass hitting the ber. (b) Relative error of the dierent methods.

Figure 7: Injection of particles with constant ow rate but variable number of streams.

4.5e-10 0.05
algebraic formula
4e-10 formula grid
3.5e-10 grid 0.04 geometric
geometric

Relative Error
3e-10
Mass [kg]

0.03
2.5e-10
2e-10
0.02
1.5e-10
1e-10 0.01
5e-11
0 0
0 2 4 6 8 10 12 14 16 0 100 200 300 400 500 600 700 800 900 1000
Particles per Parcel Particles per Parcel

(a) Mass hitting the ber. (b) Relative error of the dierent methods.

Figure 8: Injecting particles with dierent ow rates controlling the number of particles per parcel. 100 streams and
a grid with 100x100 points are used.

For the geometric calculation of collisions we count the particles hitting the ber per timestep and multiply with
their mass. This provides us with the mass per timestep colliding with the ber. Since we xed the velocity this is all
the information we need to calculate the momentum transferred from particles to the ber.
At rst, we do some tests for the rst model as described in section 4. In the rst test we x the total ow rate of
all streams and vary the number of particle streams (see gure 7). This results in 1-10 particles per parcel. Calculating
the actual particle density and using this value with our formula eq. (9) we get a relative error of 1. We can also
see that the density grids resolution has an impact on the error, and that the accuracy for the algorithm calculating
collisions geometrically decreases for fewer particles per parcel.
The next test still uses constant density investigating the inuence of the number of particles per parcles on the
result (see gure 8). As before, the relative error of our formula is still 1. Additionally, we nd that the relative
error of the geometric method is approximately 2.8% in all cases. Between 700 and 800 particles per parcels the mass
of the grid method becomes larger than the correct value for the rst time. This increases the error for our formula
from this point on.
In the following we take a look at simulations with varying droplet density. This is achieved by randomly changing
the particles radii at injection. The rst example (see gure 9) shows only few particles per parcel (6-100) with
200 particle streams to demonstrate the drawbacks of the geometric method. First, there is the staged line for the
geometric method (see gure 9(a)). While changing the ber radius additional particle streams hit the ber only at
certain intervals. Another disadvantage of the geometric method is that for small ber radii the impact is always
zero. Instead, our approach extrapolates for small radii. Looking at the error diagrams we see that the relative error
decreases while the absolute error increases. But, as we know from gure 8(b) the geometric method has an error on
its own. Accounting for this the error for our approach will be even less.
In addition, we found the resolution of the underlying density grid having an impact on the precision of the
result (see gure 10). Accounting for the error found in gure 8(b) we lowered the values for the adjusted geometric
calculation accordingly. The diagrams here show a smoothed version of the relative error demonstrating the trend.
 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790 789

1.8e-10 0.7 1.8e-11

grid grid vs. geometric grid-geometric
1.6e-10 geometric 0.6 linear trendline 1.6e-11 linear trendline
1.4e-10 1.4e-11

Absolute Error [kg]

0.5

Relative Error
1.2e-10 1.2e-11
Mass [kg]

0.4
1e-10 1e-11
0.3
8e-11 8e-12
0.2
6e-11 6e-12
4e-11 0.1 4e-12
2e-11 0 2e-12
0 -0.1 0
0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014 0.0016 0.0018 0.002 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014 0.0016 0.0018 0.002 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014 0.0016 0.0018 0.002
Fiber Radius Fiber Radius Fiber Radius

(a) Mass hitting the ber. (b) Relative Error. (c) Absolute Error.

Figure 9: Simulation of 200 particle streams on a 100x100 grid with only about 6-100 particles per parcel. The
diagrams show the impact of changing the ber radius.

0.3 0.045
100x100 grid vs. geometric 100x100 grid vs. geometric
250x250 grid vs. geometric 0.04 250x250 grid vs. geometric
0.25
100x100 grid vs. adjusted geom. 0.035 100x100 grid vs. adjusted geom.
250x250 grid vs. adjusted geom. 250x250 grid vs. adjusted geom.
Relative Error

Relative Error
0.2 0.03
0.025
0.15
0.02
0.1 0.015
0.01
0.05
0.005
0 0
0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.004 0.006 0.008 0.01 0.012 0.014
Fiber Radius Fiber Radius

(a) Full diagram. (b) Close-up for relevant details.

Figure 10: Comparing smoothed results for dierent grid resolutions. The adjusted geometric approach takes into
account that itself has an error of approximately 2.8%.

The actual error is more jagged and alternates around this curve.

7. Results

The simulation is implemented in our ber simulation tool FIDYST. An example output from a simulation can be
seen in gure 11.
In this example droplets are simulated using Fluent. They are injected as parcels from the upper right corner using
ve particle streams with a spray angle depicted by the two green lines. The red grid is the grid used for the simulation
and for the interpolation of the particle density. Hence, grid nodes outside the spray cone can have a droplet count
dierent from zero. Fibers are injected from the upper left corner using four streams and no air drag is used. We can
clearly see the change in the bers direction due to the droplets and the increase of the ber mass shown by their
color.

8. Conclusion and Future Work

We introduced a robust algorithm for modeling of droplet-ber collisions. Using a reasonable number of parcels
for the simulation of our droplets our approach provides even better results than a geometric simulation for small
ber radii because of the discrete nature of the geometric approach. Instead, our density grid provides continuous
information about particles by means of interpolation.
There are some limitations of our approach. First, we restricted the ow and the particles to have constant ve-
locity on the whole domain. This restriction has to be lifted because in real production process the air ow includes
turbulences. A further limitation may be that we get only the expected value of the mass and momentum from our
calculation. This means that in every simulation step the collision will occur with the expected mass and momentum
transferred to the ber. This can be enhanced by introducing a probability distribution allowing to vary the collision
probability. The last drawback is that the collision probability is not changed anywhere after a collision occured.
Intuitively, after a droplet hit a ber its path should be discontinued and the droplet should not have an impact on a
ber anymore. It has to be investigated if this is a problem or if it is save to make our simplied assumption.
790 Simon Schrder et al. / Procedia Computer Science 4 (2011) 781790

Figure 11: Simulation output from FIDYST (post-processed with ParaView). sigma is the mass per ber segment
length.

9. Acknowlegdements

This research has been supported by the Innovationszentrum Applied System Modeling which is funded by the
Technische Universitat Kaiserslautern (University of Kaiserslautern), the Fraunhofer-Gesellschaft, and the German
Federal Land Rheinland-Pfalz (Rhineland-Palatinate).
[1] P. J. ORourke, Collective drop eects on vaporizing liquid sprays, Ph.D. thesis, Los Alamos National Lab, New Mexico (November 1981).
[2] G. H. Ko, H. S. Ryou, Droplet collision processes in an inter-spray impingement system, Journal of Aerosol Science 36 (11) (2005) 1300
1321. doi:DOI: 10.1016/j.jaerosci.2005.02.005.
URL http://www.sciencedirect.com/science/article/B6V6B-4FTS33Y-2/2/e43fa978e084cefcc664baed9998b902
[3] G. H. Ko, H. S. Ryou, Modeling of droplet collision-induced breakup process, International Journal of Multiphase Flow 31 (6) (2005) 723
738. doi:DOI: 10.1016/j.ijmultiphaseow.2005.02.004.
URL http://www.sciencedirect.com/science/article/B6V45-4FX23NX-1/2/0360ab6dae53960d6524103ae21cc677
[4] S. Kim, D. J. Lee, C. S. Lee, Modeling of binary droplet collisions for application to inter-impingement sprays, International Journal of
Multiphase Flow 35 (6) (2009) 533549. doi:DOI: 10.1016/j.ijmultiphaseow.2009.02.010.
URL http://www.sciencedirect.com/science/article/B6V45-4VPM5HT-5/2/1e3eebe2ad7017253883bb59e290cf65
[5] K. Piroird, C. Clanet, E. Lorenceau, D. Quere, Drops impacting inclined bers, Journal of Colloid and Interface Science 334 (1) (2009)
7074. doi:DOI: 10.1016/j.jcis.2009.03.004.
URL http://www.sciencedirect.com/science/article/B6WHR-4VY2C94-9/2/147cdf646a5697967b48c7e5fc5e64a5
[6] E. Lorenceau, C. Clanet, D. Quere, Capturing drops with a thin ber, Journal of Colloid and Interface Science 279 (1) (2004) 192197.
doi:DOI: 10.1016/j.jcis.2004.06.054.
URL http://www.sciencedirect.com/science/article/B6WHR-4CXTW75-2/2/424a5e2197e6d3c192b80a4c264e7bf6
[7] L. S. Hung, S. C. Yao, Experimental investigation of the impaction of water droplets on cylindrical objects, International Journal of Multiphase
Flow 25 (8) (1999) 15451559. doi:DOI: 10.1016/S0301-9322(98)00085-8.
URL http://www.sciencedirect.com/science/article/B6V45-3YB1RG0-2/2/8f145d0437f7b126ef2ea1f9a1df1257
[8] S. Dawar, G. Chase, Drag correlation for axial motion of drops on bers, Separation and Purication Technology 60 (1) (2008) 613. doi:DOI:
10.1016/j.seppur.2007.07.027.
URL http://www.sciencedirect.com/science/article/B6THJ-4P778B1-K/2/ab5aa8a937a7f52411be27433612fd89
[9] G. McHale, M. I. Newton, Global geometry and the equilibrium shapes of liquid drops on bers, Colloids and Surfaces A: Physicochemical
and Engineering Aspects 206 (1-3) (2002) 7986. doi:DOI: 10.1016/S0927-7757(02)00081-X.
URL http://www.sciencedirect.com/science/article/B6TFR-45DF869-8/2/c8d47f05c9123837f3da5b10d2bdbe14
[10] N. Marheineke, R. Wegener, Fiber dynamics in turbulent ows: General modeling framework, SIAM Journal on Applied Mathematics 66 (5)
(2006) 17031726. doi:10.1137/050637182.
URL http://link.aip.org/link/?SMM/66/1703/1
[11] A. Klar, N. Marheineke, R. Wegener, Hierarchy of mathematical models for production processes of technical textiles, ZAMM - Journal of
Applied Mathematics and Mechanics 89 (2009) 941961. doi:DOI: 10.1002/zamm.200900282.
[12] N. Marheineke, R. Wegener, Modeling and application of a stochastic drag for bers in turbulent ows, International Journal of Multiphase
Flow In Press, Corrected Proof (2010) . doi:DOI: 10.1016/j.ijmultiphaseow.2010.10.001.
URL http://www.sciencedirect.com/science/article/B6V45-51726WP-2/2/47eb35982e5c2746aae60eb35620ec3e