Project Number: MQP ERT JK10

A PARTICLE-BASED METHOD FOR THE SIMULATION OF COMPLEX FLUIDS

AND POLYMER SOLUTIONS
Major Qualifying Project Report completed in partial fulllment
of the Bachelor of Science degree at
Worcester Polytechnic Institute
James Kingsley
April 28, 2011
Professor Erkan T uzel, Advisor
Department of Physics
Abstract
Stochastic Rotation Dynamics (SRD) is a particle-based simulation method that can be used
to model uids. It has many of the benets of direct simulation methods, while still being
computationally inexpensive. With hydrodynamic interactions and thermal uctuations
built-in, it can be used in mesoscale simulations of complex uids. Starting with a simple,
ideal uid, the method was veried for theoretical consistency and extended to support non-
ideal equations of state. As expected, the non-ideal uid demonstrated freezing behavior
at low temperatures. It has been shown that by employing these repulsive interactions, the
model can be applied to binary mixtures. Here, we demonstrate this technique by studying
the phase separation of two uids and the formation of droplets.
Contents
1 Introduction 5
1.1 Top Down Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.1 Navier-Stokes Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.2 Lattice Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Bottom Up Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Quantum Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Molecular Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Mesoscale Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 Dissipative Particle Dynamics . . . . . . . . . . . . . . . . . . . . . . 7
1.3.2 Direct Simulation Monte Carlo . . . . . . . . . . . . . . . . . . . . . 8
1.3.3 Stochastic Rotation Dynamics . . . . . . . . . . . . . . . . . . . . . . 8
1.3.4 Dimensionless Parameters . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Stochastic Rotation Dynamics 10
2.1 Description of Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Streaming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Collision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Energy and Momentum Conservation . . . . . . . . . . . . . . . . . . 13
2.2.4 Shifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.1 Periodic Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . 16
2.4.2 Slip and No-Slip Walls . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.3 Thermal Walls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.4 Ghost Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Equilibrium Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.1 Momentum and Energy . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.2 Velocity Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.3 Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.4 Diusive Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5.5 Viscous Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1
3 Modeling Polymers in an SRD Solvent 28
3.1 Modeling of Polymers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.1 Continuum Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.2 Discrete Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.3 Stretching Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.4 Resulting Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.5 SRD Heat Bath . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.6 Velocity-Verlet Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Equilibrium Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 Overall Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.2 Radius of Gyration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 Non-ideal and Binary Fluids 35
4.1 Description of Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.1.1 Super-cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.1.2 Modication of SRD Rules . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Test of Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2.1 Conservation Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2.2 Freezing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3 Binary Fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3.1 Binary Collisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3.2 Phase Separation and Droplet Formation . . . . . . . . . . . . . . . . 39
5 Conclusions and Future Directions 40
A Code Samples 41
A.1 Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
A.2 SRD Collisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
A.3 Eulerian Streaming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
A.4 Thermal Walls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
A.5 Periodic Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Bibliography 45
2
List of Figures
1.1 Dierent length and time scales that span various simulation methods. . . . 5
2.1 Example SRD grid with particles . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Example of an SRD collision with two particles . . . . . . . . . . . . . . . . 13
2.3 A shifted collision grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 An example of one dimensional periodic boundary conditions . . . . . . . . . 16
2.5 Slip and no-slip walls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Total energy with respect to time . . . . . . . . . . . . . . . . . . . . . . . . 19
2.7 Total x momentum with respect to time . . . . . . . . . . . . . . . . . . . . 19
2.8 Total y momentum with respect to time . . . . . . . . . . . . . . . . . . . . 20
2.9 Velocity distribution along x . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.10 Velocity distribution along y . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.11 Pressure as a function of area . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.12 Pressure as a function of temperature . . . . . . . . . . . . . . . . . . . . . . 22
2.13 Random walk of one particle . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.14 Expected and actual diusion . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.15 Accelerated diusion in a slow ow . . . . . . . . . . . . . . . . . . . . . . . 25
2.16 Velocity distribution across a channel . . . . . . . . . . . . . . . . . . . . . . 27
2.17 Temperature distribution across a channel . . . . . . . . . . . . . . . . . . . 27
3.1 An example of a continuous polymer . . . . . . . . . . . . . . . . . . . . . . 28
3.2 An example of a discretized polymer . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Coupling between the uid and a polymer . . . . . . . . . . . . . . . . . . . 31
3.4 Polymer Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Radius of Gyration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1 A super-cell on an SRD domain . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 Possible collision pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 A frozen non-ideal uid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 Binary mixture separating . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.5 A binary droplet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3
List of Tables
4.1 Non-ideal collision equations. . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4
Chapter 1
Introduction
Computational uid simulations are useful for many applications, and there are a number
of dierent techniques for it. However, for many of them, it is dicult to include arbitrary
structures in the uid, such as polymers, or model multiple types of uids with thermal
uctuations.
Figure 1.1: Dierent length and time scales that span various simulation methods.
In general, uid simulations can be created from the top down or from the bottom up.
The method chosen depends on the requirements of the specic situation. The length and
time scales of applicability of various simulation methods are shown in Figure 1.1.
5
1.1 Top Down Approach
Top down simulations typically start with dierential equations for the behavior of uids,
and simulate the uid as a continuous medium. If the domain under consideration is small
enough that the uid is no longer entirely continuous, it does not correctly model it. This
type of method is appropriate for many types of uid but it usually works best at a scale
large enough for the uid to behave deterministically. There are two popular approaches:
the Navier-Stokes solvers and lattice methods. In what follows, we will briey review these
methods.
1.1.1 Navier-Stokes Methods
The Navier-Stokes equations can be solved numerically in various ways. The nite volume
method breaks the experimental domain into a number of small, usually triangular subdo-
mains. Each subdomain is a controlled volume element, and is treated as a homogeneous
entity. The starting point for the simulation is a set of equations describing the behavior
of the uid. Usually, continuum mass and momentum equations are used. Using the diver-
gence theorem, any divergence terms can be converted into surface integrals, which are just
transport across boundaries to neighboring subdomains. From there, as expected, a system
of ordinary dierential equations representing the behavior of the system can be generated.
This system of dierential equations can then be numerically solved using methods such as
Eulers method or Runge-Kutta.[8]
1.1.2 Lattice Methods
The rst of the lattice methods was the lattice gas automata [9]. In lattice gas automata,
there are a series of particles which move along a discrete lattice. When they colocate
on a single point in the lattice, collision rules apply. With appropriate choice of collision
rules that conserve momentum and energy, accurate hydrodynamic behavior emerges at the
macroscopic scale. The lattice requires that both position and velocity are discretized, which
means that the simulation is inherently noisy.
An improvement on this is the lattice Boltzmann model [12, 18, 20]. This method solves
the discrete Boltzmann equation over a discrete lattice mesh. The improvement over lattice
gas automata is that lattice Boltzmann uses particle probability densities instead of entire
particles. This means that position and velocity can be approximately continuous. Even so,
both methods suer from Galilean invariance problems, since they both are on a discrete
grid.
1.2 Bottom Up Approach
Alternatively, bottom-up methods can be used to model the behavior of uids. In this
case, the lowest-level physical rules are combined and applied, giving rise to the macroscopic
behaviors of uids. Below, we briey describe some examples of bottom up techniques.
6
1.2.1 Quantum Simulation
Starting at the very bottom, there is quantum-mechanical simulations[3, 17]. The most
powerful and expensive simulation type, it involves calculation and direct manipulation of
wavefunctions. This is often done by solving Schr odingers equation with some amount of
approximation[4]. This provides the most complete and accurate description of a uid, but
in most cases is prohibitively expensive. Additionally, there are multiple methods of taking
the approximations, and choosing the appropriate one for a given situation is dicult. As
such, a full quantum mechanical model is often only used in cases where quantum eects are
known to be important to the behavior of a system.
1.2.2 Molecular Dynamics
To avoid having to solve the Schr odinger equation, molecular dynamics simulations treat
the situation classically. Each individual atom in the system is modeled using Newtons
laws under the inuence of given potentials, without including quantum mechanical eects.
Instead of wavefunction interactions, interactions are all done as inter-particle potentials.
These potentials usually are electrostatic, though they can in theory be anything, such as
Lennard-Jones, etc. This is signicantly less expensive than a quantum-mechanical simu-
lation, but still is untenable for very large systems. Current supercomputing systems can
handle molecular dynamics simulations for hundreds of thousands of atoms, but that is still
too little for many applications. Molecular dynamics is best for modeling complex atomic
things like proteins, in which case a more coarse-grained model cannot simulate the system
properly. However, for cases such as large volumes of a simple solvent, this is unnecessarily
complex and time consuming.
1.3 Mesoscale Models
Mesoscale models have some similar properties to molecular dynamics models, but they use
particles which represent a region of the uid. They do not resolve the full microscopic details,
as in molecular dynamics, but they still have uctuations and produce correct hydrodynamics
hydrodynamic behavior. In what follows, we will give a few examples of such models.
1.3.1 Dissipative Particle Dynamics
In order to avoid some of the artifacts inducted by lattice models, dissipative particle dy-
namics was developed[6, 7, 13]. It is a particle-based model, with each particle representing
a uid region. The interactions between individual particle pairs are a sum of three forces.
The rst interaction force is a conservative one. This force, as per its name, conserves en-
ergy as the particles interact, and is responsible for most of the behavior of the uid. The
dissipative force normalizes energy in the system. This allows the entire system to be kept
close to a constant temperature. The last force is the random force. This is what gives rise
to random behavior, and is what makes this method not deterministic. All of the forces are
all antisymmetric as per Newtons third law, which means that momentum is conserved in
7
interactions. While the conservative force does conserve energy, the dissipative and random
forces, do not. On average, however, the system stays at a constant (set) temperature.
1.3.2 Direct Simulation Monte Carlo
Direct Simulation Monte Carlo (DSMC) is the oldest of the mesocale models[1, 2, 10]. It
models uids using aggregate representative particles. These particles, instead of modeling
a single atom or simple molecule, are representative of an entire region of the uid. The
model runs in discrete timesteps with continuous position and velocity vectors. During each
timestep, the particles undergo a streaming and a collision step. In the streaming step,
the particles advance their positions through time ballistically. In the collision step, the
particles collide with each other via a set of collision rules. The system is divided into a set
of boxes and only particles within a given box have a chance to collide. These collisions are
calculated using probabilistic models, with collisions happening between individual particle
pairs. Since collisions are elastic (unless intentionally set otherwise), DSMC conserves energy
and momentum.
1.3.3 Stochastic Rotation Dynamics
Stochastic Rotation Dynamics (SRD) is very similar to DSMC, but it uses a dierent collision
scheme [11, 14, 16]. It still has the same type of particles, and the same streaming and
collision steps. The dierence lies in the method by which the collision step is implemented.
Instead of having randomly chosen particle pairs collide with each other, it collides groups of
particles with each other simultaneously. These groups are picked by dividing the simulation
domain into a set of boxes. All the particles within a box collide with each other at once, using
the simplest available collision rule that conserves momentum and energy. Since particle pairs
are not considered, the simulation scales much better than DSMC. This means that SRD
can simulate larger systems more eciently.
1.3.4 Dimensionless Parameters
Due to the mesoscale nature of these simulations, directly setting parameters based on real
systems does not adequately set the behavior of a system. This is due to the unphysical na-
ture of the units used in the simulation. While quantum mechanical and molecular dynamics
simulations can have parameters with physical meaning, the aggregate pseudo-particles used
in these mesoscale methods have no direct physical signicance. It is necessary to represent
the system with dimensionless parameters. As dimensionless parameters, their value is in-
dependent of the dimensions of the uid they are in, and as such can be used to compare
the behaviors of dierent uids. Examples of commonly used dimensionless numbers are the
Reynolds, Schmitt, Peclet, and Prandtl numbers.
The Reynolds number is a measure of the ratio of inertial forces to viscous forces in a
uid [5]. In low Reynolds number uids, viscous forces dominate, and turbulence is minimal
as a result. In high Reynolds number uids, however, inertial forces are more signicant.
The Schmidt number measures the ratio of momentum transport to mass transport [19].
For uids like water, it is on the order of 10
2
10
3
, while for gases it is close to 1.
8
The Peclet number is the ratio of transport of a quantity to the diusion of the same
quantity [15]. The Peclet number is the product of the Reynolds and Prandtl numbers.
The Prandtl number is the ratio of kinematic viscosity to thermal diusivity[22].
1.4 Summary
There are a number of methods available for simulating uids. At one end of the spec-
trum there are continuum methods such as Navier-Stokes solvers. These are very good for
large-scale simulation of macroscopic situations. This simulation type is commonly used in
industry to model uid ows in and around systems ranging from airplanes to windmills.
However, they are dicult to implement for small scales. Lattice methods are better at
handling uctuations, but also can suer from some instabilities due to discrete positions
and velocities. This is particularly an issue in the case of multi-component mixtures. At
the other end of the granularity spectrum, there are high-accuracy atomic simulations using
quantum mechanics or molecular dynamics. These are very expensive and used in research
applications to get results that no other method is capable of providing. In the following
chapters, we will focus on the coarse grained method SRD and its extensions.
9
Chapter 2
Stochastic Rotation Dynamics
The SRD simulation method was originally proposed by Malevanets and Kapral in 1999[16].
As briey described earlier, it uses a set of particles, which move around and collide with each
other in a coarse-grained fashion. This is easy and fast to simulate, and has the advantage
of making it easy to include other objects into the simulation, since the interaction between
a particle and an arbitrary object is comparatively simple to simulate. The mechanics of the
SRD method are as follows.
2.1 Description of Model
In a typical SRD simulation, there are N particles on a eld of side length L. The eld is
divided into boxes of side length a, as shown in Figure 2.1. This gives an average of
M =
Na
2
L
2
(2.1)
particles per cell. Each particle has a mass m, a position r, and a velocity v.
The SRD method has two main parts that it performs in each timestep. The rst step
is the streaming step, where particles increment forward in time. At this point, boundary
conditions are applied, and any additional forces are added. This is followed by the collision
step, where the particles collide with each other, as described in Section 2.2.2.
10
Figure 2.1: Example SRD grid with particles.
11
2.2 Dynamics
2.2.1 Streaming
The rst part of the core simulation is the simple streaming step. The particles take a step
through time, advancing their positions. Note that for notational convenience, when referring
to the properties of a single particle, the subscript will be dropped, since the equation applies
to all particles individually . In its most basic form, streaming is done with a simple Euler
integration,
x(t + ) = x(t) +v(t + ) , (2.2)
where t is the time, and is the timestep duration.
During the streaming step (depending on the eect, applied before, instead of, or after, as
appropriate), additional forces can be applied to the particles. For example, a gravitational
acceleration term could be applied to all particles, to create a gravity-driven ow.
2.2.2 Collision
The streaming step is followed by the collision step. In order to collide the particles, a grid of
boxes is placed across the eld. All of the particles are assigned into whichever box they are
in, and total momentum is summed for each box. The collision is a rotation of the velocities
of all the particles in each box, relative to the overall momentum of the box. This is done as
a matrix multiplication of a rotation matrix R for an angle on the relative velocity vector
v u, and given by
v(t + ) = R(v(t) u(t)) +u(t) , (2.3)
where u is given by
u =
M

i=1
m
i
v
i
m
t
. (2.4)
M

is the number of particles in a box at any given time. On average, M

= M. Equation
(2.4) also uses the total mass in the box, i.e.,
m
t
=
M

i=1
m
i
. (2.5)
R is given by
R =
_
cos() sin()
sin() cos()
_
. (2.6)
Since the rotation matrix R is an orthogonal matrix, it satises
RR
1
= RR
T
= I . (2.7)
In order to ensure detailed balance, the direction of rotation is randomly picked for each
box. This is implemented by randomly picking a plus or minus sign in Equation (2.6). The
collision process is demonstrated for two particles in Figure 2.2.2.
12
(a) Particles
come in with
velocities

V
1
and

V
2
(b) Relative velocities
are rotated by
(c) Initial and nal veloci-
ties
Figure 2.2: Example of an SRD collision with two equal mass particles. The collision angle
= 90

, and U is the mean velocity of the two particles.

2.2.3 Energy and Momentum Conservation
Since the only operation being applied to the particles is a rotation in a static frame, kinetic
energy and momentum are conserved within each box. This can be shown by explicitly
calculating the momentum before and after the collision
M

i=1
m
i
v
i
(t + ) =
M

i=1
m
i
R(v
i
(t) u(t)) + m
t
u(t) .
Using the denition of u from Equation (2.4) gives
M

i=1
m
i
v
i
(t + ) = m
t
u(t) +
M

i=1
Rm
i
(v
i
(t) u(t))
= m
t
u(t) +
M

i=1
Rm
i
(v
i
(t)) m
t
Ru(t)
= m
t
u(t) +
M

i=1
Rm
i
(v
i
(t))
M

i=1
m
i
Rv
i
(t)
=
M

i=1
m
i
v
i
(t) +
M

i=1
m
i
(R(v
i
(t) v
i
(t))
=
M

i=1
m
i
v
i
(t) .
13
Energy conservation can similarly be calculated by
M

i=1
1
2
m
i
|v
i
(t + )|
2
=
M

i=1
1
2
m
i
(R(v
i
(t) u(t)) +u(t))
2
=
M

i=1
1
2
m
i
_
u(t)
2
+ 2R(v
i
(t) u(t)) u(t) +RR
T
(v
i
(t) u(t))
2
_
which by Equation (2.7) becomes
M

i=1
1
2
m
i
|v
i
(t + )|
2
=
M

i=1
1
2
m
i
_
u(t)
2
+ 2R(v
i
(t) u(t)) u(t) + (v
i
(t) u(t))
2
_
,
and the expansion of the last term, by Equation (2.4) becomes
M

i=1
1
2
m
i
|v
i
(t + )|
2
=
1
2
m
t
_
2u(t)
2
2u(t)
2
+ 2R(u(t) u(t)) u(t)
_
+
M

i=1
1
2
m
i
v
i
(t)
2
=
M

i=1
1
2
m
i
v
i
(t)
2
.
Since momentum and energy are conserved within each individual box, they are also
conserved in the sum of all boxes.
2.2.4 Shifting
In cases where a simulation includes an arbitrarily imposed structure, there is a potential for
Galilean invariance problems. In SRD, diusive transport is aected (see Section 2.5.4)[14].
When there is a ow in one direction (or equivalently, the reference frame is moving), diusion
in the direction of the ow is faster than it would otherwise be (Figure 2.15). In order to
resolve this, the location of the whole grid is randomly shifted at each timestep. Instead of
every collision happening with the boxes in the same place, each timestep places them in a
dierent location, oset randomly in the range [a/2, a/2]. Since the range is equal to the
size of the boxes, this shifting means that the boxes are blurred across their entire size.
An example shift by a vector

b is given in Figure 2.3.
Even with shifting, the simulation still has a square structure to it. This shows up as a
higher order eect such as in the freezing transition demonstrated in Section 4.2.2.
14
Figure 2.3: A collision grid of side length a that is shifted by

b, overlaid on its unshifted
position.
2.3 Initialization
In a typical simulation, the particles are initialized with positions drawn from a uniform
distribution across the entire eld. Their velocities are also drawn from a uniform distribu-
tion, but a uniform distribution does not guarantee zero overall momentum and a specic
energy or temperature. As such, the starting distribution is arbitrary and irrelevant as long
as it is uniform, because the velocities of all particles are then scaled to match the specied
momentum and temperature. First, the momentum is normalized by taking the average
velocity of the system
v
total
=
N

i=1
m
i
v
i
N

i=1
m
i
(2.8)
and subtracting it from the velocity of each individual particle. This velocity is similar to
the u of the particles in a single box, except averaged over all boxes. Next, the temperature
is normalized by calculating the actual temperature with
T
1
2N
N

i=1
m
i
_
v
2
ix
+ v
2
iy
_
, (2.9)
and multiplying each velocity by the factor
_
Ttarget
T
actual
. Finally, any desired starting velocity for
the system is directly added to the velocity of each particle. This is done after the calculation
of the kinetic energy for the temperature, because the coherent motion of all particles is
15
equivalent to a Galilean transformation, and has no eect on internal temperature. This
could be done, for example, to have a ow in a channel (Section 2.5.5) with slip walls
(Section 2.4). If it was applied before the temperature normalization, it would articially
cool the system.
2.4 Boundary Conditions
Since free particles will eventually leave the eld, it is necessary to somehow constrain them to
the designated area. To deal with this, boundary conditions are applied after the streaming
step. There are a few types of boundary conditions: periodic, slip, no-slip, and thermal.
2.4.1 Periodic Boundary Conditions
Periodic boundary conditions are an implementation of an innite area. This is done by
taking particles on one edge of the eld and transporting them to the opposite edge. In
order to be able to make measurements using the total displacement however, it is necessary
to preserve that parameter and apply the boundary condition to another one. This is eected
by assigning
x
simulation
= x L x . (2.10)
The reason why x
simulation
is separate from x is so that while all simulation interactions use
the simulation coordinate, measurements can be done on the system which require knowing
total displacement. For example, when measuring diusion, if a particle went a distance L
in one direction, x
simulation
would report that it had not moved, while x would show that it
had. An example of periodic boundary conditions in one dimension is shown in Figure 2.4.
This is a simple boundary method that strictly conserves momentum and energy. Since
the wrapping mimics an innite space (topologically it is actually the surface of a toroid),
it is also good for tests. It does, however, have the disadvantage of allowing the eects of a
particle to propagate all the way around the simulation and come back to interact with the
same particle. For small system sizes, such self-interactions can cause problems.
2.4.2 Slip and No-Slip Walls
Slip and no-slip walls are very similar, in that they both take an incoming particle and
reverse its direction in the direction normal to the wall. The dierence is that while a slip
Figure 2.4: An example of one dimensional periodic boundary conditions.
16
(a) No-slip (b) Slip
Figure 2.5: Slip and no-slip walls.
wall does nothing to the tangential component of the velocity, a no-slip wall inverts it. To
deal with the fact that the particle will have a position outside of the eld, it is reected
back inside. This makes the particle act as if it had participated in the collision part way
through the timestep, and had nished the available time with its new parameters. Both
types of walls strictly conserve energy, but only conserve momentum on average.
2.4.3 Thermal Walls
Thermal walls entirely reset a particles velocity in both directions. The normal component
of the velocity, v

, is drawn from a biased exponential distribution

P (v

) =
mv

k
B
T
e

mv
2

2k
B
T
, (2.11)
while the tangential component of the velocity, v
||
, is drawn from a normal distribution
P
_
v
||
_
=
_
m
2k
B
T
e

mv
2
||
2k
B
T
. (2.12)
This acts both as a no-slip boundary, as well as a thermostat at the wall [11].
For all wall types, the actual collision of the particle and the wall does not happen
precisely at the end of a timestep. For the previous types of walls, this could be accounted
for simply with algebra. However, in the case of thermal walls, the post-collision velocity
is not a function strictly of the pre-collision velocity, and thus extra care is required. To
deal with for this, the collision is re-calculated, with the timestep split into two parts. The
intersection time of the particle and the wall is calculated, and the timestep is broken into
two portions at that moment. The rst portion of the timestep is used to bring the particle
right up to the wall, and then the particles velocity is assigned from the above distributions.
Now that the particle has its new trajectory, it uses the second portion of the timestep to
continue away from the wall.
2.4.4 Ghost Particles
One of the problems that arises when shifting and walls are used in conjunction with each
other is that of shifting across a boundary. If one is not careful, shifting can make the system
act (during the collision step) as if it had periodic boundary conditions, when it in fact does
17
not. This is because the shift could wrap one edge of the domain onto another, thus resulting
in particles on one edge colliding with particles on the other edge. This could be avoided
by making sure that they are included in separate boxes, as if on a domain of size L, with
walls at a and L a. This would result in partially lled collision boxes, however. To solve
this, one may put a set of ghost particles into that box in order to make the box have M
particles to collide. The velocities of these particles will be drawn from a Maxwell-Boltzmann
distribution, with zero mean velocity and the same temperature as the wall [11].
2.5 Equilibrium Measurements
There are a number of coecients that can both be calculated from the properties of the SRD
model and can be experimentally measured from the simulation. In order for the simulation
to be valid, both the experimental and theoretical values of these must match.
2.5.1 Momentum and Energy
The easiest parameters to measure are momentum and energy. Without the inuence of
walls or external forces, the two quantities should never change, as shown in Section 2.2.3.
As demonstrated by Figures 2.6, 2.7, and 2.8, energy and momentum are constant at their
expected values. This means that they are conserved by the SRD algorithm, as expected.
2.5.2 Velocity Distributions
While the total momentum remains constant, the distribution of velocities can vary as the
simulation runs. While the particles are initialized with uniformly distributed velocities,
their distribution should change to match the Maxwell-Boltzmann distribution,
P (v

) dv

=
_
m
2k
B
T
e

m
2k
B
T
v
2

dv

, (2.13)
where k
B
is the Boltzmann constant, T is the temperature once the system reaches equilib-
rium [21], and is a direction x, y, or z.
As shown in Figures 2.9 and 2.10, the velocity distributions quickly converge to the form
given by the Maxwell-Boltzmann distribution. Since the simulation was initialized with
a uniform velocity distribution, this shows that the SRD method correctly generates its
equilibrium distribution.
2.5.3 Pressure
In a thermodynamically accurate simulation of an ideal gas (what SRD models), the ideal gas
law PV = Nk
B
T, where P is pressure, V is volume, and N is the number of particles should
be obeyed. However, since the simulation is two dimensional, the corresponding equation is
PA = Nk
B
T , (2.14)
18
Figure 2.6: Total energy with respect to time. L = 32, a = 1, N = 4096, M = 4, k
B
T = 1,
= 1, = 90

.
Figure 2.7: Total x momentum with respect to time. L = 32, a = 1, N = 4096, M = 4,
k
B
T = 1, = 1, = 90

.
19
Figure 2.8: Total y momentum with respect to time. L = 32, a = 1, N = 4096, M = 4,
k
B
T = 1, = 1, = 90

.
Figure 2.9: Distribution of x velocities at the end of a simulation run. a = 1, = 1, k
B
T = 1,
= 90

.
20
Figure 2.10: Distribution of y velocities at the end of a simulation run. a = 1, = 1,
k
B
T = 1, = 90

.
where P is linear pressure (force per unit length) and A is the area. Pressure is determined
based on the momentum traveling through a line. Instead of measuring the force on the line,
the simulation allows one to measure the force that would have been on it, without actually
having to aect the uid.
Since all of the parameters except for P can be set for the simulation, it is easy to test
this. Figures 2.11 and 2.12 show pressure as a function of area and temperature respectively.
21
Figure 2.11: Pressure as a function of area for dierent values of temperature. N = 65, 536,
a = 1, = 1, = 90

.
Figure 2.12: Pressure as a function of temperature for dierent values of area. N = 65, 536,
a = 1, = 1, = 90

.
22
As predicted, pressure and area have an inverse relation, while pressure and temperature
are directly proportional. This means that the simulation is properly modeling an ideal gas.
2.5.4 Diusive Transport
Another parameter that can be measured is the diusion coecient D. This is measured as
change in the mean squared displacement of the particles as a function of time, i.e.
x
2
= 2Dt , (2.15)
where x is the coordinate of a particle. Analytical calculation shows that the diusion
coecient should work out to [21]
D =
k
B
T
2M
_
2M
(1 cos ) (M 1 + e
M
)
1
_
, (2.16)
where is the timestep size, M is the number of particles per cell (Equation (2.1)), and
is the angle rotated through during a collision (Equation (2.6)). This coecient has no
dependence on direction, its value should be the same for both dimensions. One would expect
that in a nite simulation domain, diusion could only be measured for a short period of time
until the particles under consideration run into the edges of the simulation. This happens
quite quickly (on the order of tens to hundreds of timesteps depending on parameters). In
order to properly measure diusion for long periods of time, periodic boundary conditions
are used (Section 2.4). Due to the wrapping explained earlier, the simulation domain appears
innite. This means that distances far larger than the available space can be observed, and
the diusion coecients can be calculated over a longer time (and thus more accurately).
The larger diusion area allowed is shown in Figure 2.13.
The diusion coecient was measured for a sample of particles whose locations at the
start of the simulation were recorded, and dierences averaged at each timestep (Figure
2.14). The expected line from Equation 2.16 is plotted along with both the x and y diusion
results. The diusion coecients are the slopes of the lines. The agreement is excellent.
Before shifting was implemented, the simulation had Galilean invariance problems. When
the particles are at a low enough temperature or high enough density, and the system has
a ow, the diusion coecient in the direction of motion was greater than the diusion
coecient in the other direction (Figure 2.15). This was solved by implementing shifting as
described in Section 2.2.4. The improved version behaves as expected regardless of ow.
2.5.5 Viscous Transport
Another parameter that can be obtained from the simulations is the viscosity. This can
easily be measured by applying thermal walls to one dimension of the simulation, emulating
an innite channel. By including an acceleration term, a ow can be induced in the uid.
This is usually called Poiseuille ow, and has a parabolic velocity distribution given by
v(x) =
g
2
_
_
(x
L
2
_
2

_
L
2
_
2
_
(2.17)
23
Figure 2.13: Traced path of one particle over 20,000 timesteps. L = 32, a = 1, T = 1, = 1,
= 90

.
Figure 2.14: Mean squared displacement as a function of time for x and y components. Solid
line is the theory from Equation (2.16).
24
Figure 2.15: Diusion when a ow is imposed in the x direction. The diusion coecients
dier by 20% between the x and y directions when shifting is not enabled. M = 16, T =
0.015625, = 1, = 90

.
25
where = mM = m
N
L
2
is the density and g is the acceleration of each particle.
This can be shown by starting with the denition of viscosity, which is
F =
v
x
, (2.18)
where F is the force, v is the velocity, and x is the position. Taking a 2D channel from
L/2 to L/2, the system will be symmetric across x = 0. One can thus treat the channel
as a smaller channel from 0 to L/2, with the side at x = 0 frictionless. In equilibrium, the
forces will balance, and so a dierential segment of uid at x will support all of the uid
from x = 0 to x. Equivalently, it will support
F = gx =
v
x
, (2.19)
which leads to
v
x
=
x

. (2.20)
Integrating from 0 to x and shifting so that v(L/2) = 0 gives
v(x) =
g

1
2
_
_
L
2
_
2
x
2
_
.
A transformation from x = 0 to x = L/2 gives Equation (2.17).
This means that the experimental result can be matched against the expected prole to
extract the viscosity . The viscosity can also be calculated using
=
k
B
T
2m
_
M
(M 1 + e
M
) sin
2

1
_
+
a
2

1
6dM
_
M 1 + e
M
_
(1 cos ) (2.21)
where a is the width of each box [21].
With the Poiseuille ow, the velocity distribution, shown in Figure 2.16, was parabolic
as expected from Equation (2.13). The temperature distribution across the channel was
approximately constant across the middle of the channel, also as expected (Figure 2.17).
26
Figure 2.16: Velocity distribution across a channel. a = 1, k
B
T = 0.01, L = 64, = 1,
m = 1, M = 4, g = 0.0001, = 90

.
Figure 2.17: Temperature distribution across a channel. a = 1, k
B
T = 0.01, L = 64, = 1,
m = 1, M = 4, g = 0.0001, = 90

.
27
Chapter 3
Modeling Polymers in an SRD Solvent
The coarse-grained nature of the SRD simulation method allows it to be used as a solvent
for other methods. Taking advantage of this, a molecular dynamics module was created and
used to model polymers. This chapter explains the methods used for modeling a polymer.
Once modeled, applications range from viscous relaxation to swimming organisms.
3.1 Modeling of Polymers
Since polymers are long repeating chemical chains, they can be modeled easily as a chain
of points connected together. These connections can be modeled as potentials. A cartoon
description of a polymer is given in Figure 3.1.
Figure 3.1: An example of a continuous polymer.
28
3.1.1 Continuum Model
In order to describe the potentials that describe a polymer, one may rst consider a contin-
uum model. Using r(s) as the position of each point on the polymer, and

t(s) =
r
s
(3.1)
as the tangent vector to r is a good way to start. If

t/s = 0,

t(s) is constant, and thus the
polymer is a straight rod. Thus, bending energy must be a quadratic of

t/s. By denition
(

t is a unit tangent vector),

t
s
= 0 (3.2)
and so the next available term is
_

t
s
_
2
. This gives
U
bend
=
1
2
EI
_
L

0
ds
_

t
s
_
2
(3.3)
where L

is the length, and EI = L

p
k
B
T is controlled by L
p
, the persistence length. From
statistical mechanics, the distribution of angles will follow a canonical distribution
(U) e

U
k
B
T
e

IE
2k
B
T

0
ds

t
s

2
e

Lp
2

0
ds

t
s

2
From here, one can show that

t(s)

t(0)
_
2
=
2s
L
p
, (3.4)
which is similar in form to the diusion equation (2.15). Thus, L
1
p
acts like a diusion
coecient for the polymer.
End-to-end distance R
ee
can also be calculated, i.e.,
R
2
ee
=
_
L

0
ds
_
L

0
ds

t(s)

t(s

) (3.5)
using

t(s)

t(0) = e

s
Lp
gives
R
2
ee
= 2L

L
p
2L
2
p
_
1 e

Lp
_
(3.6)
As expected, this gives the correct behavior in its limits. With
L

Lp
1, R
2
ee
= 2L

L
p
. For
L

Lp
1, R
2
ee
= (L

)
2
.
29
Figure 3.2: An example of a discretized polymer with four segments.
3.1.2 Discrete Treatment
In order to simulate the polymer, however, it needs to be broken down into discrete lengths,
as shown in Figure 3.1.2. Making a discrete treatment of the polymer involves using r
i
, and

t
i
, instead of r(s) and

t(s), respectively. The bending energy can immediately be converted
to
U
bend
=
1
2
EI
n1

i=1
_

t
i+1

t
i
_
2
(r
i+1
r
i
)
2
l . (3.7)
Here l is the segment length and n is the number of segments. Since

t
i
is a unit vector,
combined with subtraction of zero-point energy, the numerator of the sum reduces to 2

t
i

t
i+1
. Assuming that segments have xed length, (r
i+1
r
i
)
2
= l
2
. Thus,
U
bend
=

l
n1

i=1

t
i

t
i+1
(3.8)
where EI. This means that the polymer chains can be given arbitrary rigidities,
controlled by .
3.1.3 Stretching Energy
For the component of the energy from stretching, the discrete model can be used from the
beginning. Placing a spring between each point on the chain gives a potential of,
U = 1/2k (|r
i1
r
i
| l)
2
(3.9)
such that the forces obeyed Hookes law with spring constant k.

F = k (|r
i1
r
i
| l)

t (3.10)
This constraint makes the polymer behave as a freely jointed chain. Each individual link
is kept at its given length l, and the links can rotate freely.
30
Figure 3.3: Coupling between the uid and a polymer.
3.1.4 Resulting Forces
The total internal potential energy for the polymer is then given by
U
internal
= U
bend
+ U
stretch
. (3.11)
Thus, the behavior of the polymer is
m
d
2
x
dt
2
= U
internal
+ F
random
+ F
dissipative
(3.12)
where U
internal
are the internal potentials, and there are an additional two forces: F
random
and
F
dissipative
. The random force applies Brownian kicks to the polymer, while the dissipative
force acts as a heat bath to absorb energy from the polymers relaxation. Unless solvent
molecules are included explicitly these forces do not conserve momentum, and therefore do
not produce the correct in hydrodynamic behavior.
3.1.5 SRD Heat Bath
It is possible to have molecular dynamics support full hydrodynamics, but it requires mod-
eling the solvent as well, which is very expensive.
However, it is possible to use an inexpensive SRD solvent as the heat bath for the poly-
mer. The SRD solvent can provide the friction and random force required for the MD to
work properly. Since the forces are the product of interactions with the surrounding uid,
momentum is conserved, and is passed back to the uid.
The coupling is done by having the particles that are part of MD participate in the SRD
collisions as if they were regular solvent particles. This gives
u =

i=0
m
i
v
i
+

j=0
m
j
v
j

i=0
m
i
+

j=0
m
j
(3.13)
where the rst sum (over i) is the same as in Equation (2.8), and the second sum (over j) is
for the additional M

particle(s) from the polymer.

31
These particles behave almost exactly the same way as the regular SRD particles, with
one exception. Instead of moving normally as described in Section 2.2.1, they move based
on the molecular dynamics calculations. These are done using a Velocity-Verlet algorithm.
3.1.6 Velocity-Verlet Algorithm
In order to get very rigid rods and joints, large spring constants are required. This works,
though when the forces are are too strong, the simulation does not converge with reasonably
sized timesteps. Since strong springs are essential to the algorithm working, other options
are required to ensure convergence. The simplest way to improve this is to just use a smaller
timestep. As such, the MD calculations were done with a timestep on the order of 1:100 the
size of the SRD calculations.
Even with this, convergence is a bit of a problem. In order to help them converge faster,
a velocity integration method with a lower error term is helpful. For this, velocity-corrected
Verlet integration was used:
r (t + t) = r (t) +v (t) t +
1
2
a (t) t
2
(3.14)
v (t + t) =
1
2
(a (t) +a (t + t)) t (3.15)
Since the particles in the polymer chains are using a smaller timestep, they spend their
movement time doing MD instead of the normal Eulerian streaming described previously.
3.2 Equilibrium Measurements
3.2.1 Overall Length
The simplest measure of a polymer is the end to end length. This is simply dened as
R
ee
= | r
n
r
0
| (3.16)
In the case of a polymer with = 0, the points along the polymer chain follow a random
walk. As such, the total length will follow a path similar to that a single point takes (Equation
(2.15)). This means that the total length L
t

n. However, when the polymer has a = 0,

it follows a more straight path, and as approaches innity, R
ee
L

. This is shown in
Figure 3.4, by how the = 0 curve follows

n.
3.2.2 Radius of Gyration
The radius of gyration, R
g
gives a measure of the dimensions of the chain. It is dened as
R
2
g
=
1
n
n

i=0
(r r
cm
)
2
(3.17)
32
10
n
1
10
L
t

(
a
)
=0
=10
Figure 3.4: Polymer length for = 0, 10. a = 1, l = 1, k
B
T = 1, = 1, L = 32, M = 5.
where the center of mass is
r
cm
=
1
n
n

i=0
r (3.18)
R
g
depends on both the chain length n, and the stiness, , as shown in Figure 3.5. A stier
chain will be more spread apart, and thus have a greater R
g
.
33
10
n
1
10
R
g

(
a
)
=0
=10
Figure 3.5: Radius of Gyration for = 0, 10. a = 1, l = 1, k
B
T = 1, = 1, L = 32, M = 5.
34
Chapter 4
Non-ideal and Binary Fluids
Since SRD uses point particles, the resulting equation of state is that of an ideal gas. In
order to have a non-ideal equation of state, particles need to act as if they have non-zero
diameter. This is done using a new type of collision rule, as explained in the next section.
4.1 Description of Model
The non-ideal model extends SRD by also including an inter-cell collision term. In each
timestep, the collision step has an additional part, where all of the particles in a cell are
collided against all of the particles in a neighboring cell.
4.1.1 Super-cells
In order to do the collisions, cell pairs are required. This is done by splitting the domain
into super-cells of size 2a, where a is the size of a SRD collision cell (Figure 4.1). In each
super-cell, during each timestep, the particles can either collide horizontally, vertically, or
diagonally. The probability of collisions is equal for each option. However, since there are
two dierent ways of having horizontal collisions, it has twice the probability of the other
two. Thus, horizontal and vertical collisions each have a 25% probability, while diagonal
collisions have a 50% probability.
The super-cells used for non-ideal collisions also have shifting. The shifting used is similar
to the shifting used in normal SRD (Section (2.2.4)), but it uses a range of [a, a]. This
is because the super-cells are of size 2a, and spreading the shifting over that requires the
greater range.
4.1.2 Modication of SRD Rules
All of the particles in a super-cell collide against all of the particles in the paired super-cell.
The collision is simple, where the parallel components of the velocity interchange while the
perpendicular components are unchanged. This gives a set of rules as described in Table
4.1.2.
35
Figure 4.1: A super-cell on an SRD domain.
(a) Horizontal: = x (b) Vertical: = y (c) Diagonal:
=
x+ y

2
,
x y

2
Figure 4.2: Possible collision pairs.
36
Collision
x
v
x
(t + ) = 2u
x
v
x
(t)
v
y
(t + ) = v
y
(t)
y
v
x
(t + ) = v
x
(t)
v
y
(t + ) = 2u
y
v
y
(t)
x+ y

2
v
x
(t + ) = u
x
+ u
y
v
y
(t)
v
y
(t + ) = u
x
+ u
y
v
x
(t)
x y

2
v
x
(t + ) = u
x
u
y
+ v
y
(t)
v
y
(t + ) = u
y
u
x
+ v
x
(t)
Table 4.1: Non-ideal collision equations.
In order to have the proper thermodynamics, collisions happen with a probability P,
given by
P = tanh(AuM
a
M
b
) , (4.1)
where A is a parameter (like inverse temperature), u is the dierence in velocities between
box a and b, and M
a
and M
b
are the number of particles in each of the two participating
boxes.
The hyperbolic tangent function is used because it provides a nice curve for mapping
the domain [0, ) onto [0, 1]. It is linear with slope d tanh(x)/dx = 1 at x = 0, but
asymptotically approaches 1 as x approaches innity. This helps keep the probability in the
appropriate domain, though A should be chosen such that the probability does not saturate.
If A is too high, the probability looses its dependence on M
A
, M
B
, and u, which is a
reection of k
B
T.
All of the dierent steps in a timestep should happen simultaneously in order to be
invariant under time-reversal. For SRD this is not a problem, because the collision step does
not aect position, and the streaming step does not aect velocity. For SRD and non-ideal
collisions in the same system, however, this is important, and needs to be addressed. The
simplest way of solving the problem is to randomly average it out, similar to the spacial
random averaging used for shifting. By randomly picking which collision type to do rst, it
means that on average the collisions happen at the same time, and time-reversal invariance
is restored.
4.2 Test of Model
4.2.1 Conservation Laws
The primary test of the new collision rules is conservation of momentum and energy. In
theory both are conserved, since the collisions are elastic. Simulation results show that they
are indeed conserved, as expected (data not shown).
37
0 8 16 24 32
X
0
8
16
24
32
Y
Figure 4.3: A frozen non-ideal uid. L = 32, a = 1, M = 5, k
B
T = 0.001, = 0.1, = 90

.
Though not thermodynamically consistent, A is chosen extremely large to emphasize the
eect.
4.2.2 Freezing
Since the new collisions model a hard-sphere uid with a non-ideal equation of state, it is
capable of freezing at low enough temperatures. The square structure of the SRD model
means that the frozen uid has a square crystal structure. When it freezes, sets of particles
clump together. These particles behave as normal SRD, since they are close enough that
the non-ideal collision rarely aects them (only when shifting causes a boundary to cross the
group). The groups then collide with each other via the non-ideal term, since they are too
far away for SRD to aect them. This keeps them stable.
4.3 Binary Fluid
Using the non-ideal uid collisions, a binary uid can also be simulated. Instead of having
just a single species of particle in the simulation, one can have two of them. They interact
with each other dierently (although symmetrically), resulting in an immiscible binary uid,
as explained in the following sections.
4.3.1 Binary Collisions
While the non-ideal collisions normally happen with all the particles in two adjacent boxes,
in the case of binary collisions, it is only between dierent particle types. With two particle
species A and B, A particles in the rst box collide with B particles in the second box, while
B particles in the rst box collide with A particles in the second box. This means that
38
(a) Spontaneously cooled homogeneous
mixture.
(b) After some time, the uid has par-
tially separated into its components.
Figure 4.4: Binary mixture separating.
Figure 4.5: A binary droplet.
a homogeneous (only type A) uid will behave as normal SRD. However, at an interface
between A and B, the particles will collide with each other and be mutually repelled.
4.3.2 Phase Separation and Droplet Formation
The repulsion between the two species is eective enough to let the two uids phase separate
from a homogeneous mixture into areas of their components. Figure 4.3.2 shows 50%-50%
mixture phase separating after starting from a completely random distribution. Figure 4.3.2
shows an example of a droplet of radius 12, which is made with a with a 56%-44% mixture.
39
Chapter 5
Conclusions and Future Directions
We wrote an implementation of the SRD uid simulation technique. This implementation
was analytically proven to conserve momentum an energy, and demonstrated to do so in
practice as well. Previous work has analytically determined the expected diusion coecient
and kinematic viscosity for an SRD uid, and simulation results match these theoretical
expectations. Additionally, pressure was measured, and the ideal gas properties of SRD
were demonstrated.
This method has a number of potential applications. The rst of these is combining
molecular dynamics and multiple-component simulations to model surfactant systems. This
would allow us to study capillary waves in the surface of a droplet, and look at how surface
tension changes with the addition of a surfactant. Extending the molecular dynamics part
would allow us to model a Purcell swimmer. While the Purcell swimmer has been modeled
before in a continuum system, the Brownian kicks of the SRD solvent would provide a novel
(and more realistic) environment. Depending on the implementation, it could be possible to
have the rods connecting particles in molecular dynamics be solid boundaries. This would
also enable the simulation of impermeable and semi-permeable membranes, commonly found
in biological systems.
40
Appendix A
Code Samples
As our implementation of the SRD simulation method was written in C, the example pieces
of code provided here are also.
A.1 Data Types
There are structs for particles and for boxes:
typedef struct {
double sX; //position
double sY;
double vX; //velocity
double vY;
int rX; //wrap count
int rY;
int mass;
} particle;
typedef struct {
double mX; //total momentum
double mY;
double count; //particles in the box
int rotDir; //direction of rotation
} box;
A.2 SRD Collisions
The core of the SRD collision:
void collide(particle particles[], box boxes[], double boxShiftX, double boxShiftY) {
int i;
double nX;
double nY;
int boxnum;
for(i=0;i<PART_COUNT;i++) {
41
boxnum = CALC_BOX_NUM; //a macro for determining the ID of the box the particle is in.
particles[i].vX -= boxes[boxnum].mX; //subtract average velocity
particles[i].vY -= boxes[boxnum].mY;
nX = particles[i].vX*Wxx+boxes[boxnum].rotDir*particles[i].vY*Wxy; //do rotation
nY = boxes[boxnum].rotDir*particles[i].vX*Wyx+particles[i].vY*Wyy;
particles[i].vX = nX + boxes[boxnum].mX; //add average velocity back
particles[i].vY = nY + boxes[boxnum].mY;
}
}
A.3 Eulerian Streaming
Simple Eulerian Streaming:
void move(particle particles[]) {
int i;
for(i=0;i<PART_COUNT;i++) {
particles[i].sX += particles[i].vX*TAU;
particles[i].sY += particles[i].vY*TAU;
}
}
A.4 Thermal Walls
Thermal Walls:
inline void thermalWallsX(particle* part) { //inlined to improve efficiency
double t;
if(part->sX < 0) {
t = part->sX/part->vX;
part->sY -= t*part->vY;
part->vX = randm()*SQRT_WALL_TEMP*SQRT_2;
part->vY = randn()*SQRT_WALL_TEMP;
part->sX = part->vX*t;
part->sY += part->vY*t;
} else if(part->sX > WIDTH) {
t = (part->sX-WIDTH)/part->vX;
part->sY -= t*part->vY;
part->vX = -1*randm()*SQRT_WALL_TEMP*SQRT_2;
part->vY = randn()*SQRT_WALL_TEMP;
part->sX = WIDTH+part->vX*t;
part->sY += part->vY*t;
}
}
42
A.5 Periodic Boundary Conditions
inline void periodicBoundaryCondX(particle* part) {
part->rX += floor(part->sX/WIDTH);
part->sX -= WIDTH*floor(part->sX/WIDTH);
}
43
Bibliography
[1] F. J Alexander and A. L Garcia. The direct simulation monte carlo method. Computers in Physics,
11(6):588, 1997. 8
[2] G. A. Bird. Molecular Dynamics and the Direct Simulation of Gas Flow. Oxford Science Publications,
Oxford, 1994. 8
[3] W. D Cornell, P. Cieplak, C. I Bayly, I. R Gould, K. M Merz, D. M Ferguson, D. C Spellmeyer, T. Fox,
J. W Caldwell, and P. A Kollman. A second generation force eld for the simulation of proteins, nucleic
acids, and organic molecules. Journal of the American Chemical Society, 117(19):51795197, 1995. 7
[4] E. R Davidson and D. Feller. Basis set selection for molecular calculations. Chemical Reviews,
86(4):681696, 1986. 7
[5] J. K.G Dhont. An introduction to dynamics of colloids. Elsevier Science, 1996. 8
[6] P. Espanol. Hydrodynamics from dissipative particle dynamics. Physical Review E, 52(2):17341742,
1995. 7
[7] P. Espanol and P. Warren. Statistical mechanics of dissipative particle dynamics. EPL (Europhysics
Letters), 30:191, 1995. 7
[8] Robert Eymard, Thierry Gallouet, Rapha`ele Herbin, P.G. Ciarlet, and J.L. Lions. Finite volume
methods. In Solution of Equation in [real]n (Part 3), Techniques of Scientic Computing (Part 3),
volume Volume 7, pages 7131018. Elsevier, 2000. 6
[9] U. Frisch, B. Hasslacher, and Y. Pomeau. Lattice-gas automata for the Navier-Stokes equation. Physical
review letters, 56(14):15051508, 1986. 6
[10] Alejandro Garcia. Numerical methods for physics. Prentice Hall, Upper Saddle River N.J., 2nd ed.
edition, 2000. 8
[11] G. Gompper, T. Ihle, D. Kroll, and R. Winkler. Multi-particle collision dynamics: A particle-based
mesoscale simulation approach to the hydrodynamics of complex uids. In Christian Holm and Kurt
Kremer, editors, Advanced Computer Simulation Approaches for Soft Matter Sciences III, volume 221
of Advances in Polymer Science, pages 187. Springer Berlin / Heidelberg, 2009. 10.1007/978-3-540-
87706-6-1. 8, 17, 18
[12] X. He, S. Chen, and G. D Doolen. A novel thermal model for the lattice boltzmann method in incom-
pressible limit* 1. Journal of Computational Physics, 146(1):282300, 1998. 6
[13] P. J. Hoogerbrugge and J. Koelman. Simulating microscopic hydrodynamic phenomena with dissipative
particle dynamics. EPL (Europhysics Letters), 19:155, 1992. 7
[14] T. Ihle and D. Kroll. Stochastic rotation dynamics:a galilean-invariant mesoscopic model for uid ow.
Physical Review E, 63(2), 2001. 8, 14
44
[15] R. G Larson. The structure and rheology of complex uids. New York: Oxford, 2001. 9
[16] Anatoly Malevanets and Raymond Kapral. Mesoscopic model for solvent dynamics. The Journal of
Chemical Physics, 110(17):8605, 1999. 8, 10
[17] S. L Mayo, B. D Olafson, and W. A Goddard. DREIDING: a generic force eld for molecular simulations.
Journal of Physical Chemistry, 94(26):88978909, 1990. 7
[18] G. R McNamara and G. Zanetti. Use of the boltzmann equation to simulate lattice-gas automata.
Physical Review Letters, 61(20):23322335, 1988. 6
[19] M. Ripoll, K. Mussawisade, R. G. Winkler, and G. Gompper. Dynamic regimes of uids simulated by
multiparticle-collision dynamics. Physical Review E, 72(1):016701, 2005. 8
[20] X. Shan and H. Chen. Lattice boltzmann model for simulating ows with multiple phases and compo-
nents. Physical Review E, 47(3):1815, 1993. 6
[21] E. T uzel, T. Ihle, and D. Kroll. Dynamic correlations in stochastic rotation dynamics. Physical Review
E, 74(5), 2006. 18, 23, 26
[22] Frank White. Viscous uid ow. McGraw-Hill Higher Education, New York NY, 3rd ed. edition, 2006.
9
45