Jay A Stotsky Navier Stokes Equation 2016
Jay A Stotsky Navier Stokes Equation 2016
Jay A Stotsky Navier Stokes Equation 2016
com/science/article/pii/S0021999116300730
5a5959eaf384a8e238661dcf70160665
Abstract
The goal of this work is to develop a numerical simulation that accurately captures the biomechanical response
of bacterial biolms and their associated extracellular matrix (ECM). In this, the second of a two-part eort,
the primary focus is on formally presenting the heterogeneous rheology Immersed Boundary Method (hrIBM) and
validating our model by comparison to experimental results. With this extension of the Immersed Boundary Method
(IBM), we use the techniques originally developed in Part I ([19]) to treat biolms as viscoelastic uids possessing
variable rheological properties anchored to a set of moving locations (i.e., the bacteria locations). In particular, we
incorporate spatially continuous variable viscosity and density elds into our model. Although in [14, 15], variable
viscosity is used in an IBM context to model discrete viscosity changes across interfaces, to our knowledge this work
and Part I are the rst to apply the IBM to model a continuously variable viscosity eld.
We validate our modeling approach from Part I by comparing dynamic moduli and compliance moduli computed
from our model to data from mechanical characterization experiments on Staphylococcus epidermidis biolms. The
experimental setup is described in [26] in which biolms are grown and tested in a parallel plate rheometer. In
order to initialize the positions of bacteria in the biolm, experimentally obtained three dimensional coordinate
data was used. One of the major conclusions of this eort is that that treating the spring-like connections be-
tween bacteria as Maxwell or Zener elements provides good agreement with the mechanical characterization data.
We also found that initializing the simulations with dierent coordinate data sets only led to small changes in
the mechanical characterization results. Matlab code used to produce results in this paper will be available at
https://github.com/MathBioCU/BiolmSim.
Keywords: Navier-Stokes equation, biolm, immersed boundary method, computational uid dynamics,
viscoelastic uids
1. Introduction
The goal of this work is to show that the model and corresponding simulation method originally developed in
part I [19], can accurately capture the biomechanical response of bacterial biolms. The underlying mathematical
technique is an adaptation of the Immersed Boundary Method (IBM) that takes into account the nite volume of
bacteria, and the widely variable material parameters anchored to the positions of the bacteria in a biolm. We
call this method the heterogeneous rheology Immersed Boundary Method (hrIBM). With the hrIBM model, biolms
are resolved to a scale at which the contributions of individual bacteria are distinguishable. This allows for detailed
modeling of the highly viscous and spatially heterogeneous extracellular matrix (ECM) and, the polysaccharide
chains that link bacteria together in biolms.
A key feature of our simulations is the use of experimental data from live S. epidermidis biolms to initialize the
spatial location of each bacterium. This removes ambiguity about how to represent the biolm computationally.
Motivated by results from [17], we model the spatial heterogeneity of the ECM by treating the ECM as a polymer
brush with radially dependent viscosity and density elds centered at each bacteria location. In Figure 1, images
of the positions, connectivity, and a viscosity isosurface of bacteria in a sample biolm are shown. When using
∗ Corresponding author
Email address: dmbortz@colorado.edu (David M. Bortz)
© 2019. This manuscript version is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
(a) (b) (c)
Figure 1: a) Shows the 3D locations of bacteria from experimental biolm data. b) Each line represents a viscoelastic connection
between two bacteria. Bacteria connected if they are within 1.62µm of each other. c) A viscosity isosurface of the same biolm. The
maximum viscosity is 250µ0 where µ0 is the viscosity of water. The isosurface is the surface dened by µ(x) = 125µ0 .
experimental data to initialize the positions of bacteria, the bulk dynamic moduli and compliance modulus com-
puted through simulation are consistent with experimental results found in [26]. We also show that upon running
simulations with several dierent data sets that possess similar spatial statistics, the physical properties of the
biolm do not change signicantly. Finally, we provide quantitative results on the periodic rotation of suspended
aggregates of bacteria in shear ow.
In Section 2, we provide a brief review of the classical Immersed Boundary Method (IBM), a well known
computational technique for the simulation of coupled uid-structure interactions. Additionally, we discuss some
other IBM-based biolm models, and explain the adaptations of the IBM that lead to the hrIBM. In Section 3, we
give a description of the numerical properties of the hrIBM model and provide results from numerical tests showing
that the model is convergent. In Section 4, methods for computing relevant material properties from the hrIBM
model are discussed, and the dynamic moduli and compliance moduli computed from the model are compared to
experimental data from biolms grown in a bioreactor. In Sections 5 and 6, we conclude this work and discuss
future research directions.
To our knowledge, this work is the rst to use a model that accounts for both the heterogeneous rheological
properties (e.g., variable viscosity) and, the inter-bacterial connectivity to compute material properties of a biolm.
This work along with Part I, is also the rst to incorporate variable viscosity into an IBM-based biolm model,
and the second to use variable viscosity with an IBM model in general. Matlab code used to produce the results
obtained in this paper is be available at https://github.com/MathBioCU/BiolmSim.
2
been used in [39] to model sperm motility, and in [34] to model immersed three dimensional viscoelastic networks.
We pursue this strategy by testing our model with several viscoelastic force laws, and also build upon this class of
models with the inclusion of rheologically variable uid media.
Other IBM-based biolm models can be found in [2] and in [11]. In these models, an IBM is used directly to
couple the forces between connected bacteria with uid motion. Additionally, some validation results are performed
to show that properties such as the recovery and relaxation times of simulated biolms fall within the range of
realistic values. Although not an IBM-based model, we also note that similar results are computed and compared
to experiments in [33] where a phase-eld model is used. Other phase eld models that the reader may nd of
interest can be found in [9, 35, 36, 37].
In a recent work [32], the IBM is used along with the von Mises stress criterion to provide detailed simulations of
the detachment of a biolm in shear ow. Although we do not currently employ a means of modeling detachment, it
is possible to implement a model that allows the breakage of connections between bacteria in the hrIBM framework.
If enough connections break, detachment of one or more bacteria from the biolm can occur. A simple, strain
based criterion was used in Part I [19] to model the rupture of links connecting bacteria as deformation occured in
a biolm.
As our biolm model is based on the IBM, interested readers can nd detailed explanations of the IBM in
[27, 28, 38, 39], and additional IBM-based biolm models can be found in [1, 11, 19].
∇·u=0 (2)
ˆ
U(s, t) = u(x, t) δ̂(X(s, t) − x, ω) dx s = 1, 2, ...N (3)
Ω
∂X(s, t)
= U(s, t) (4)
∂t
F(s, t) = F(X(s, t), P) (5)
N
1 X
f (x, t) = F(s, t) δ̂(X(s, t) − x, ω) (6)
d30 s=1
N
X
ρ(x, t) = ρ0 + (2ω)3 (ρb − ρ0 ) δ̂(X(s, t) − x, ω) (7)
s=1
These equations are similar to those used in Part I [19] except for some small changes that were convenient for
implementing the simulations described in Section 4.
In most IBM literature, the Equations 6 and 7 are written as integrals over Lagrangian coordinates. In this
case, there is a xed, nite number of bacteria in the domain regardless of renement level, so we compute these as
summations instead. In these equations, we use a scaled approximation of the Dirac δ -function, denoted δ̂ , which
has a region of support that depends on a radial parameter ω , dened such that 2ω is the hydrodynamic radius of a
typical bacteria. Since only nitely many bacteria are considered, we also dene a Lagrangian label s as a number
associated with each bacteria.
The scale factor of (2ω)3 in Equations 7 and 8 is chosen to enforce that at the locations, X(s, t), the density and
viscosity are equal to ρb and µb respectively. Finally, we dene F as a force instead of a force density to allow for
standard units to be associated with the constitutive parameters (i.e. spring constants) that will be used to model
the viscoelastic links between neighboring bacteria. Complete listings of the quantities appearing in the model
equations can be found in Tables 1, 2, and 3.
Since individual bacteria are not assumed to have innitessimal volume at the scale of our simulations, the
Lagrangian quantities; X, U, µb , ρb , and F correspond to measurements taken at the center of mass of each
bacterium. As described in Section 2.3, the second argument, ω , of the smoothed Dirac δ function, δ̂(·, ω) determines
a region of support for the smoothed δ function. The choice of δ̂(·, ω) govern how the mass density, viscosity, and
force density vary around each bacterium.
3
Eulerian Variables
Symbol Denition
x Eulerian position
t time
u Eulerian velocity
P pressure
f (x, t) Eulerian force density
µ(x, t) Eulerian viscosity
ρ(x, t) Eulerian density
Table 1: Quantities assigned a value at each grid point and timestep in the discretized computational domain. Together they provide
an Eulerian description of the uid motion in the biolm.
Table 2: Quantities assigned to each bacteria in the domain at each time step. They can be thought of as Lagrangian variables associated
with each bacteria center of mass.
Model Parameters
Symbol Denition
µ0 viscosity of water
ρ0 density of water
F(X, P) constitutive relation between bacteria conguration and force
P parameters in the viscoelastic law used (i.e. spring coecients etc.)
N number of bacteria in computational domain
d30 volume of average bacterium
ω hydrodynamic radius of a bacterium
rc cuto radius for establishing which bacteria are linked
δ̂(·, ω) discrete Dirac delta function with support related to parameter ω
Table 3: These quantities are parameters to the model. They are either based on physical data (i.e. density and viscosity of water) or
are tuned through simulations to obtain realistic values.
4
The viscosity is computed using a dierent form than the density and force and is shown in Equation (8). This
formula for viscosity is not intuitive, and an explanation for this choice of viscosity can be found in Part I. However,
nding the most accurate form for µ(x) is still an area of active research. In particular, we experimented with
´
approximations to µ(x) = µ0 + Ω (µb − µ0 ) δ̂(X(t) − x, ω)dX. We made our ultimate choice for µ(x) so that the
viscosity does not overshoot µb , the force and viscosity drop o at the same rate, and to maintain consistency with
Part I.
The last part of the model is a force law, F needed to specify F based on the conguration of bacteria, X.
We use Hooke's law, Maxwell's law, and the Zener model in our simulations, and initialize the connectivity of the
biolm by using a cuto radius, rc . At the start of the simulation, any two bacteria that are separated by a distance
less than the cuto radius will be linked by a spring whose initial, resting length is their initial separation.
By using the IBM as a basis for our biolm model, we avoid treating the biolm as a two phase uid with
a distinct bulk uid region and a distinct biolm region. Instead, the use of variable rheological properties over
the entire domain couples the biolm and bulk uid motions as a single viscoelastic material. Because the uid
permeates through the entire domain, even when the simplest force law, Hooke's Law, is used to determine forces
between bacteria, the model still behaves viscoelastically, not elastically.
Further discussion of this choice for δ̂(x, ω) can be found in [19]. If ω = h, the standard discrete δ function seen
in IBM literature is obtained. For this work, we assume that the bacteria are spherical and thus ω is understood as
a hydrodynamic radius. We also note that extensions to this formalism will allow for the treatment of nonspherical
bacteria or unevenly sized bacteria. Thus, ω may be though of more generally as a shape parameter.
Lagrangian Variables
to determine U F
Eulerian to Lagrangian
Lagrangian to Eulerian
Immersed Boundary Solution Cycle
Solve Navier-
Stokes equations
δ̂(·, ω)-convolution
to get f , µ, and ρ
for u and P
Eulerian Variables
Figure 2: The coupling between the Eulerian and Lagrangian variables in the hrIBM is shown here. The Eulerian and Lagrangian
variables are coupled by the computation of U from u, and the computation of f , µ, and ρ from F and X. The IBM is a widely
applicable method in part because it allows for a great variety of uid solvers and solid structural models to be coupled through δ
function transfer identities.
In the simulations, the number of bacteria, N , is xed and independent of the mesh spacing h, thus a summation
is used instead of an integral for the density computation.
In this model, we indirectly take into account the uid volume displacement caused by the presence of the
bacteria. We treat the localized high viscosity around each bacteria as an eective viscosity that accounts for both
the displaced uid volume and the increased viscosity near the bacteria surface [17]. Extensions based on changing
our choice for δ̂(x, ω) and µ(x) could possibly allow for a more precise computation of the volume displacement
6
caused by the biolm into the model. In Figure 3, observe that lower strain rates are found near the bacteria cells
where high viscosity is present. This is indicative of the bacteria appearing as relatively solid objects compared to
the bulk uid.
We make the assumption that the bacteria have a uniform diameter of 1.0µm based on [18] and [31] where
bacteria diameters are measured and determined to have a tightly centered distributions around 0.64 µm 0.70 µm
respectively. We choose a larger radius since we want to account for the inuence of the polymeric ECM which
extends into the uid outside the cell. We also note another source that states that S. epidermidis typically have
a diameter in the range of 0.5 − 1.0µm [16]. We also assume that the bacteria are uniformly spherical in shape.
Although we have not been able to nd data to quantify this assumption, examination of high-resolution images of
S. epidermidis bacteria found in [30] and [25] appears to conrm the validity of this assumption.
3. Numerical Methods
The numerical methods we use are based on those originally discussed in [19]. We summarize them here
for convenience and also provide convergence results. To approximate solutions to equations (1)-(8), we use an
approximate projection method similar to that used in [38]. The solution scheme uses an implicit Euler solver to
update an intermediate velocity prole at each time step and is expected to be O(∆t) convergent in the velocity
eld. To discretize the domain, we use a uniform nite dierence discretization with equal spacings in the x, y , and
z directions. The spatial derivatives are approximated with 2nd order, centered nite dierences.
1 Numerically, products of the form, Dh a(n) Gh are discretized using standard nite dierence stencils rather than as a product of
the gradient and divergence operators.
7
Quantity Value Explanation
µ0 1 · 10−3 Pa·s viscosity of water at 20◦ C
ρ0 998 kg/m3 density of water at 20◦ C
L 10 µm characteristic length scale
Re 0.0014 − 0.0175 Reynold's number. Varies based on the simulation
St 0.0075 − 0.0720 Strouhal number
t0 1s characteristic time
u0 1.40 · 10−4 − 0.0017m/s characteristic velocity, chosen to match experimental strain conditions
average volume of each bacterium,
d0 1.59 · 10−6 m
found by dividing the total volume by the number of bacteria
ρb 0.12ρ0 added density contribution from ECM
µb 450µ0 added viscosity from ECM
Fs1 1.64 · 10−11 N Hooke's law force constant
rc 1.62 · 10−6 m maximum distance by which bacteria can be connected t = 0
ω 2.5 · 10−7 m scaling constant, equal to 1/2 the hydrodynamic radius of a bacteria
Fd1 5.6 · 10−12 N/s damping coecient in Maxwell and Zener models
Fs2 1.64 · 10−11 N second force constant used in Zener model
Table 4: Values of Physical Parameters and Nondimensional constants used in simulations. The values of Fs1 , Fs2 , and Fd1 vary
depending on whether Hooke's Law, the Maxwell Model, or the Zener Model are used, so the values shown here are for the Zener model
which appeared to perform most favorably compared to experimental results. The coecients for the other two models are discussed in
Section 4.2.
In steps 1 and 2, full multigrid solvers and multigrid preconditioned conjugate gradient solvers are used to nd u∗
and P (n) 2 . After obtaining the updated velocity and pressure, the Lagrangian velocity and position updates follow,
X
U(n+1) = u(n+1) δ̂(xh − X(n) , h) h3
h∈Gh
∆t (n)
X(n) = X(n−1) + U .
St
Next the Lagrangian force density is computed based on the new positions, X(n+1) as F(n+1) = F(X(n+1) ) (see
section 4.2 for more details). Finally, the Eulerian elds,f and ρ are computed using discrete δ function interpolation
to the Eulerian grid through equations of the form:
N
X
f (n+1) = F(n+1) (s)/d30 δ̂(xh − X(n+1) (s), ω)d30
s=1
N
X
ρ(n+1) = ρ0 + (2ω)3 (ρb − ρ0 ) δ̂(xh − X(n+1) (s), ω),
s=1
In the simulations we conduct, the primary direction of uid ow is in the z direction. The height is governed
by the y coordinate and width by the x coordinate. In Figure 5, the motion of the bacteria in a simulation with
these conditions is depicted.
2 Strictly speaking, the pressure should be interpreted as being dened at the half-integer time steps as discussed in [6]. We use the
notation P (n) for simplicity
8
(a) (b)
Figure 3: a) A viscosity isosurface is shown for a small section of a biolm used in simulation. The inner isosurface is µ = 125µ0 and
the outer transparent isosurface is at µ = 50µ0 . Slices of the ||u|| velocity eld are shown as well. b) The ˙yz component of the strain
rate is plotted on the µ = 125µ0 viscosity isosurface. Additionally, the strain rate and contours of viscosity are shown in slice planes.
For a single phase uid, Newton's viscosity law is σ = µ˙. Although biolms are not Newtonian uids, we still see that in areas of low
viscosity, higher strain rates are found and in areas of higher viscosity lower strain rates occur.
Table 5: Average convergence factors of the Fluid Solver. Spatial convergence tests were carried out with grid spacings, h, set to 1/32,
1/64, and 1/128 and a time step of ∆t = 1/2500/ν . Temporal convergence tests were done with ν∆t set to π/125, π/250, and π/500
and h = 1/32. Error is computed at t = (0.5π/ν) s for temporal convergence tests and after 1250 time stepsfor the spatial convergence
tests. Convergence factors are computed as ρ(∆t) = log2 ||u(∆t/2)−u in time and by ρ(h) = log2 ||u(h/2)−u in
||u(∆t)−uexact ||2 ||u(h)−uexact (h)||2
exact ||2 exact (h/2)||2
space and uexact (h) is the exact solution on a grid with spacing h. For the spatial convergence tests, we found it necessary to increase
the Reynold's number by a factor of 1000 to ensure that the error in the iterative solver and the temporal discretization was much
smaller than the spatial discretization error. Before this change the discretization error was only O(10−9 ).
structure, this is a test of the uid solver alone, and not the IBM method. For this test, the domain, Ω is chosen
to be a rectangular solid that is periodic in the x and z directions. From [8], the following boundary conditions for
y = 0 and yL ,
∂P
=0 u|0 = (0, 0, 0) u|yL = (0, 0, sin νt) (11)
∂y
provide us with an analytic solution,
1/2
sinh k y(1 + i) sinh k y(1 + i) νρ
uz (y, t) =
sin ν t + arg k= . (12)
sinh k yL (1 + i) sinh k yL (1 + i) 2µ
The values of P , ux and uy are exactly zero in this case. The values of ρ and µ are set to 998 kg/m3 and 1 mPa · s
respectively and are homogenous across the domain since no analytic solutions with variable density and viscosity
and the boundary conditions given above are known to the authors. Convergence tests were conducted with
frequencies ν = 4.991 Hz,49.91 Hz with the biolm, and at 49.91 Hz and 499.1 Hz without the biolm. In Table 5
the absolute error, temporal convergence factors, and spatial convergence factors are listed. For a full explanation
and justication of these tables, see Part I [19].
Additionally, with the same boundary conditions as above, we tested the convergence rates for simulations with
a biolm that possesses variable density and viscosity. Temporal and spatial convergence factors are shown in Table
6. As is often seen with immersed boundary methods, discretization error induced by the presence of discrete δ
functions, in the transferrence of density, viscosity, and forces between Eulerian and Lagrangian coordinates leads
to a reduction to rst order convergence [28, 20]. In some cases, second order accuracy can be recovered, however
9
Velocity, ||u|| Position, ||X||
Frequency of Boundary Oscillation
Time Space Time Space
49.91 Hz 0.983 1.105 1.022 0.952
4.991 Hz 0.991 0.910 1.007 1.054
Table 6: Convergence factors of hrIBM with biolm. For spatial convergence, h was set to 1/32, 1/64, and 1/128 with a time step of ν∆t =
1/500. To measure the temporal convergence factors, ν∆t was set to 1/250, 1/500, and 1/1000. In both cases, the boundary conditions
described in Section 3 were used. Note, that ν refers to oscillation frequency, not dynamic viscosity. Temporal convergence factors were
h
||u(∆x)−Ih/2 u(∆x/2)||
computed as ρ(∆t) = log ||u(∆t/2)−u(∆t/4)||
||u(∆t)−u(∆t/2)||
and spatial convergence factors were computed as ρ(∆x) = log h/2
||u(∆x/2)−Ih/4 u(∆x/4)||
where Ih2h is an interpolation operator from a grid of spacing h to a grid of spacing 2h. For spatial renement, we also note that the
number of Lagrangian nodes (bacteria) remains constant as we are not approximating a surface, but instead treating each Lagrangian
node as a single bacteria which should have a constant volume independent of the grid renement.
this generally requires predictor-corrector methods that are more computationally expensive than what we use
here. In addition, variable density tends to lead to rst order in time methods even when the velocity is discretized
by a formally second order accurate scheme [14, 15]. More detailed numerical convergence results for this model
with dierent boundary conditions are shown in [19]. In Table 6, temporal convergence factors for the same uid
conditions and domain as the analytical solution are listed. The pressure convergence is not shown here since
we use a pressure-free projection method in our simulations (see [6] for more details), and because the eect of
pressure on the velocity eld is several orders of magnitude less than the viscous and elastic eects on the scale of
our simulations.
10
Small Amplitude Rheology Shear Compliance
φ/ω
σyz (t) σyz (t)
σ0
Stress
Stress
G00 (ω) = (σ0 /0 ) sin φ
G0 (ω) = (σ0 /0 ) cos φ
σ0
1
(t)
Shear Rotation
J(t; σ0 ) =
tσ t t
yz (t) yz (t)
0
Strain
Strain
t t t
(a) (b)
Figure 4: a) A depiction of the experimental rheometer setup. Reproduced from [26] with permission from the Royal Society of
Chemistry. b) An image highlighting the shear rotation of the rheometer. The typical stress and strain proles expected from SAR and
Compliance experiments are shown on the right.
If we temporarily approximate the biolm as a homogenous, isotropic material, then for low amplitude rotational
shear oscillations, the spatial variation in the shear stress and strain will be proportional to the radial distance from
the center of the rheometer [10]. If we were to simulate a region of biolm with dimensions 30 × 30 × 30 µm3
positioned approximately half way between the center and the outer boundary of the rheometer, the ratio of shear
strain and shear stress exerted at the inner and outer boundaries of the simulated section of biolm is greater
than 0.99. Thus the deviation from linear shear is minimal in the computational domain. Of course, biolms are
heterogeneous and may be anisotropic however, the previous rough calculation indicates that rectangular geometry
and linear shear should provide an accurate approximation of the motion of the biolm.
functions, G1 (t) and G2 (t) correspond respectively to shear stress and dilatational stress. Analogous expressions
exist for the compliance tensor.
Although the viscoelastic moduli are spatially heterogeneous, we believe that more meaningful results are ob-
tained in the mean eld, or spatially averaged, time (and frequency) dependent values for , σ , G, and J . These
quantities depend less on the exact conguration of bacteria in a biolm and behave more like bulk material parame-
ters that can be measured experimentally. Although the interconnected links used to model the connections between
11
adjacent bacteria each individually introduce anisotropy into the model, under the conditions of our simulations,
the overall behavior of the biolm is not highly anisotropic.
As can be seen in Figure 5, uid ow is primarily in the z direction and the y coordinate indicates height in the
simulations. Thus the yz and σyz components of the stress and strain must be computed at the top plate in order
to determine the shear moduli.
In the biolm simulations, tracer particles with positions denoted by S(x, y, z), are initialized at heights yL − γ ,
yL − γ − h, and yL − γ − 2h, near the top of the biolm at t = 0. At each time step, the positions of the tracers
are updated using the same δ function interpolation used to update the bacteria positions. With these tracers, the
deformation of the biolm can be tracked throughout the simulation.
In the simulations, the yz component of strain is needed at the upper boundary of the domain. Therefore,
the tracers are initialized near the top of the domain in three vertically aligned layers. This is done to make the
numerical approximation of derivatives of the form ∂dz /∂y simpler to compute. With the initial arrangement of
tracers in vertically aligned layers, the centered nite dierence approximation
1 ∂dz ∂dy 1 ∂dz 1 dz (S(y + h) − dz (S(y − h))
yz (S, t) ≈ + ≈ ≈ (16)
2 ∂y ∂z 2 ∂y 2 2h
can be used to approximate the strain. Since the entire upper plate moves at a single velocity at any given time,
∂dy /∂z is negligible in this case, whereas in general, this term is required to compute the shear strain. For our
simulations, the strain was also approximated in terms of material coordinates.
3 Boundary layers can also be mitigated through adjustments to the boundary conditions for u∗ in conjunction with a Crank-Nicholson
time stepping scheme.
12
The outward unit normal, n, is (0, 1, 0) in this case since the top plate is parallel to the xz plane. The force, Fb is
found by integrating the Eulerian force density eld that would be generated by the biolm nodes adhered to the
Fb
top plate. Additionally, since we are interested in the applied shear stress, σzy
b
, this can be found as - Az . Note that
γ was chosen arbitrarily, however we observed that with γ = 0.7 µm the results were not signicantly dierent.
with
||Xi (t) − Xj (t)|| − ||Xi (0) − Xj (0)||
Λij (X, X0 ) = . (23)
||Xi (t) − Xj (t)||
Following [19], we choose each kij to be a force constant, Fs1 divided by the initial separation of bacteria i and
j . With Maxwell's law and the Zener model, we use an integral form of the standard viscoelastic relation and,
integrate by parts to obtain a formula for F. This is done since integration by parts obviates the need to compute
any time derivatives of stress or strain. We then approximate the integral at each time step using the trapezoid
method. Since the immersed boundary method requires a Lagrangian force density, we divide the values of Fij
13
Figure 5: Starting at t = 0 on the left, the images show how the biolm is moved as the top plate oscillates. Dots are bacteria locations
and lines indicate viscoelastic connections. In the simulations the domain is periodic in the x and z directions. The periodicity is not
shown here since it makes it dicult to visualize the eect of deformation on the biolm.
14
(a) (b)
(c)
Figure 6: Comparison between experimentally measured results for G0 and G00 and simulated results for three viscoelastic models
are depicted. For these results, the force constant was Fs1 = 2.30 · 10−11 N , the connection distance between bacteria was 1.62µm,
µb = 350µ0 , and ρb = 1.12ρ0 . The dashed lines indicate the experimental error ranges. Subgure (a) uses Hooke's Law, Subgure (b) uses
Maxwell's law, and Subgure (c) uses the Zener model. For the Maxwell model, a damping coecient Fd1 = 1.03 · 10−11 N/s was used,
and for the Zener model, the spring coecient was set to Fs1 = 1.64 · 10−11 N , the damping coecient was set to Fd1 = 0.051 · 10−11 N/s
and, a second spring coecient of Fs2 = 1.64 · 10−11 N was used.
determined from the viscoelastic force law by the Lagrangian volume element, d30 to obtain an average force density
associated with each bacteria.
In previous works [11, 28], identical constitutive relations to (22) are derived from the starting point of energy
functionals from which the force density is found by taking a Fréchet derivative of the energy functional. Addition-
ally, in [34] and [12], similar viscoelastic force laws are studied in an immersed boundary method setting and in [34]
corresponding dynamic shear moduli are computed.
In Figure 6 we depict the frequency dependence of G0 and G00 as determined by the hrIBM model, and compare
with experimental data. From these results, it is clear that our model ts experimental data on G0 quite well. For
G00 the t is not as strong, although we still do see that for the Maxwell and Zener models, most of the results from
simulation are within the range of experimental error. We also observe that even with Hooke's Law which is an
elastic constitutive relation, the biolm still behaves viscoelastically (i.e. G00 6= 0) due to its immersion in a viscous
uid.
15
Figure 7: The time dependence of J(t, σ0 ) is shown above for σ0 = 0.1 P a and σ0 = 0.2 P a. The Zener model as used in these
simulations. For these levels of applied stress, the change in the compliance modulus is only slight indicating a nearly linear regime.
In (26), ρ0 and µ0 are the density and viscosity of water, and L is the characteristic length (in this case 10 µm). In
numerical experiments,
rather
than immediately impose a step in stress at time 0, we multiply the applied stress σ0
by a mollier, 1−e−αt − 1 to mitigate any possible numerical instabilities associated with a discontinuous jump
2
in the wall stress. In this case, α = 155 is chosen to be large so that the applied stress approaches its equilibrium
value within 0.1 seconds. We believe this is reasonable because the very short time scale compliance behavior is not
experimentally measurable, and also because a true step change in the stress is not physically realizable. Results
from simulations using this boundary condition and two values of σ0 are shown in Figure 7.
4.4. Similarity of Material Properties Between Dierent Bacteria Position Data Sets
In [13], the spatial statistics of bacteria in a biolm are studied. We show here that from data sets that have
similar spatial distributions of nearest neighbor connections between bacteria, similar bulk property measurements
are obtained. In our simulations, we take blocks of experimental data that are 18 µm wide, 9 µm long, and 27 µm
high and compute the dynamic moduli of each block at a xed frequency . The graphs in Figure 8 show the stress
and strain of four dierent biolms over one period of oscillation.
From Figure 8 and Table 7, we see that in three of the four biolm data sets, similar results are obtained. We
also note that the mean standard error in the experimental results for this particular test was 3.9791 for G0 and
16
(a) (b)
Figure 8: These graphs show the stress vs. time and strain vs. time for 4 dierent biolm samples. The samples are all 18µm × 27µm ×
9µm size and contain approximately 2000 bacteria positions.
Table 7: Results for G0 (ν) and G00 (ν) are shown at two frequencies for 3 dierent biolm coordinate data sets with ν = 49.91 rad/s.
These results show that the physical properties measured here do not depend solely on the exact microstructure of the biolm, but on
some type of more large scale organization of the bacteria positions in space.
0.8836 for G00 . We also suspect that if larger data sets were used, even better agreement would be seen in the
computed values of G0 and G00 .
The importance of this section is in verifying that the properties we are validating can be considered as bulk
properties. Since three out of the four data sets provided results within the experimental error deviation, we believe
this to be strong evidence the properties we measure are bulk properties.
The boundary velocities are scaled by 10−3 to provide an average shear rate of approximately 114s−1 . This in the
range of shear rates examined in [7]. To ensure that the biolm is not attached to the plates and is suciently far
from the plate to induce a rotating, or tumbling motion, these simulations only include bacteria that are greater
than 8.8 µm from either plate at the start. With the physical parameters we use, the bacteria aggregation rotates
and is deformed by the uid shear forces exerted by the uid [7]. Several snapshots from a simulation of an aggregate
tumbling are shown in Figure (9).
In [4], analytical results on the frequency at which a solid ellipsoid will rotate in shear ow are provided. For
an ellipsoid with axis aligned with the direction of uid motion, the rotational frequency is found as
2π(a21 + a22 )
T =
a1 a2 τ
where τ is the shear rate and a1 and a2 are the principle axes of the ellipse undergoing rotation. In our simulation
(see Figure 9), we show that a bacterial aggregate approximated as a hydrodynamically equivalent ellipse will rotate
at a frequency similar to the theoretically expected result. The rotational frequency of the aggregate is found by
computing the average frequency of rotation of bacteria in the yz plane about the center of mass of the aggregate.
In the simulations, we observed a frequency of approximately 0.158 seconds for an aggregate approximated by an
ellipse with major axis a1 = 2.344 µm and rst semimajor axis a2 = 1.268, and with shear rate of 114 s−1 containing
110 bacteria. The theoretical result in this case is 0.1317 seconds. For smaller aggregates, the elliptical assumption
appears to become less accurate although it still provides results that are on the same order of magnitude. These
17
Figure 9: Biolm aggregate suspended in shear ow rotate over time. Snap shots shown at 0, 0.4 and 0.8 seconds into the simulation.
The attening of the ellipse in response to the shear ow can be distinguished between the rst and third gure. Several bacteria are
marked red to help show the rotation of the aggregate. The blue lines indicates the trajectories of the marked cells relative to the center
of mass of the aggregate. Distances are in micrometers.
Major axis, a1 First minor axis, a2 Theoretical Period Observed Period Relative Error
2.344µm 1.268µm 0.1317s 0.1579s +19.9%
1.646µm 1.194µm 0.1116s 0.1542s +38.2%
results are shown in Table 8. Although we do not currently have experimental results to compare our simulations
with, rotational frequency measurement do provide a strong metric for model validation in elds such as red blood
cell modeling (see [14]).
5. Conclusions
Based on the experimental results shown above and in Part I [19], the heterogeneous rheology Immersed Boundary
Method model or hrIBM (Equations (1)-(8)), appears to be an accurate model of the biomechanical response of
bacterial biolms to uid motion. With the hrIBM model, we resolve biolms to the scale of individual bacteria.
This enables the use of various viscoelastic stress-strain relationships to account for polysaccharide links between
bacteria. The spatially heterogeneous ECM also plays a prominent role in the hrIBM model and is treated as a
variable viscosity and variable density uid. We recreate a continuous spatially varying viscosity eld, with increased
viscosity in the vicinity of the center of mass of each bacteria, by adapting the interpolation procedure found in
the immersed boundary method. Although variable viscosity was introduced into the IBM's in [14, 15], this work
is the rst to consider viscosity as a continuously varying eld, not an approximation to a step change in viscosity.
Additionally, the authors believe that this work and Part I are the rst to consider variable viscosity in an IBM
model of biolms.
In Section 4, a major eort was spent towards validating the hrIBM model by comparing material properties
computed by the hrIBM model to experimental data. In particular, we show that the model yields close agreement
with experimental results from [26] in which the bacterium S. epidermidis was grown in a bioreactor and charac-
terized using a parallel plate rheometer. Based on the simulations we have conduced it appears that viscoelastic
models that contain a Maxwell element provide results comparable to experimental data whereas models that lack a
Maxwell element become less accurate at low frequencies. Another development is in showing the similarity between
the bulk dynamic moduli over dierent experimental data sets that possess similar spatial statistics. We also show
that suspended aggregates of bacteria in shear ow rotate with a similar period as a hydrodynamically equivalent
ellipse. To our knowledge, the hrIBM model is the rst that can accurately compute bulk material properties of
biolms based on coupling the microscale connectivity of the biolm with the heterogeneous rheology of the ECM.
18
6. Future Directions
Currently the scheme is O(∆t) and O(h) for the velocity eld and bacteria positions. In future work, a predictor-
corrector method as used in [14, 15], or the Crank-Nicholson time-stepping scheme may be used since, at least in the
constant viscosity and density case, these methods can lead to O(∆t2 ) convergent Navier-Stokes solvers as shown
in [6]. However, even without heterogeneous material properties, obtaining O(∆t2 ) convergence in the overall IBM
is more complicated and also depends on properties of the discrete Dirac δ function as discussed in [21, 22].
Modelling and simulation of biolm detachment is an area of interest in biolm research. One way to model
detachment in our biolm model would be the inclusion of a stress or strain criterion for when the viscoelastic links
in the biolm should rupture. This approach is explored in [32] where the von Mises stress criterion is used. Another
direction would be to follow along the lines of [5], and include model of the reconguration of the connectivity of
the biolm over time, especially as the biolm is deformed due to uid motion. It is also possible to model the
viscoelastic properties of the biolm through a more complex constitutive model of the Lagrangian force based on
the nodal conguration of the bacteria.
Another area that could be explored is the shape of the discrete δ function used to approximate each bacteria
and its associated viscosity halo. It is possible that adjusting this function may allow for more accurate modeling
of the mass displacement induced by the bacteria bodies in the bulk uid. Adjustments to the δ function may also
allow for the inclusion of non-spherical bacteria into the model. It would be interesting to see if similar results are
obtained for bacteria that are dierent shapes.
One possible diculty in adjusting the discrete δ function is the preservation of mass in the model. In Equation
(2), there is no density dependence as is often seen with the Navier-Stokes equations for a variable density system.
For the original IBM, Equation (2) is in fact the correct statement of mass preservation as described in [28]. For
the hrIBM, there is an error however, the error is expected to be small in our situation since ρ(x, t) only varies by
≈ 20% over the domain (density of bacteria is not highly variable), all simulations are at low Reynold's numbers,
and because biolms are essentially incompressible even though they may have variable density.
7. Acknowledgements
This work was supported in part by the National Science Foundation grants PHY-0940991 and DMS-1225878 to
DMB, and PHY-0941227 to JGY and MJS, and by the Department of Energy through the Computational Science
Graduate Fellowship program, DE-FG02-97ER25308, to JAS. This work utilized the Janus supercomputer, which is
supported by the National Science Foundation (award number CNS-0821794), the University of Colorado Boulder,
the University of Colorado Denver, and the National Center for Atmospheric Research. The Janus supercomputer
is operated by the University of Colorado Boulder. We would also like to thank the reviewers for several useful
suggestions that improved this paper.
[1] Alpkvist, E., Klapper, I., 2007-02-22. A Multidimensional Multispecies Continuum Model for Heterogeneous
Biolm Development. Bulletin of Mathematical Biology 69 (2), 765789.
URL http://link.springer.com/10.1007/s11538-006-9168-7
[2] Alpkvist, E., Picioreanu, C., van Loosdrecht, M. C., Heyden, A., 2006-08-05. Three-dimensional biolm model
with individual cells and continuum EPS matrix. Biotechnology and Bioengineering 94 (5), 961979.
URL http://doi.wiley.com/10.1002/bit.20917
[3] Balestrino, D., Ghigo, J.-M., Charbonnel, N., Haagensen, J. A. J., Forestier, C., 2008-03. The characterization
of functions involved in the establishment and maturation of klebsiella pneumoniae in vitro biolm reveals dual
roles for surface exopolysaccharides. Environmental Microbiology 10 (3), 685701.
URL http://doi.wiley.com/10.1111/j.1462-2920.2007.01491.x
[4] Blaser, S., 2002-02. Forces on the surface of small ellipsoidal particles immersed in a linear ow eld. Chemical
Engineering Science 57 (3), 515526.
URL http://www.sciencedirect.com/science/article/pii/S000925090100389X
[5] Bottino, D. C., 1998-11. Modeling Viscoelastic Networks and Cell Deformation in the Context of the Immersed
Boundary Method. Journal of Computational Physics 147 (1), 86113.
URL http://linkinghub.elsevier.com/retrieve/pii/S0021999198960740
[6] Brown, D. L., Cortez, R., Minion, M. L., 2001-04. Accurate projection methods for the incompressible navier-
stokes equations. Journal of Computational Physics 168 (2), 464499.
URL http://linkinghub.elsevier.com/retrieve/pii/S0021999101967154
19
[7] Byrne, E., Dzul, S., Solomon, M., Younger, J., Bortz, D. M., Apr. 2011. Postfragmentation density function
for bacterial aggregates in laminar ow. Physical Review E 83 (4), bibtex: Byrne2011.
URL http://link.aps.org/doi/10.1103/PhysRevE.83.041911
[8] Carlslaw, H. S., Jaegar, J. C., 1959. Conduction of Heat in Solids, 2nd Edition. Oxford University Press.
[9] Chen, C., Ren, M., Srinivansan, A., Wang, Q., 2011. 3-d numerical simulations of biolm ows. East Asian
Journal on Applied Mathematics.
URL http://www.csrc.ac.cn/research/publications/201111/W020111117518710826964.pdf
[10] Christensen, R. M., 1982. Theory of Viscoelasticity: An Introduction, 2nd Edition. Academic Press.
[11] Dan Vo, G., Brindle, E., Heys, J., 2010-06. An experimentally validated immersed boundary model of uid-
biolm interaction. Water Science & Technology 61 (12), 3033.
URL http://www.iwaponline.com/wst/06112/wst061123033.htm
[12] Dillon, R. H., Fauci, L. J., Omoto, C., Yang, X., 2007-02-15. Fluid dynamic models of agellar and ciliary
beating. Annals of the New York Academy of Sciences 1101 (1), 494505.
URL http://doi.wiley.com/10.1196/annals.1389.016
[13] Dzul, S. P., Thornton, M. M., Hohne, D. N., Stewart, E. J., Shah, A. A., Bortz, D. M., Solomon, M. J.,
Younger, J. G., 2011-03-01. Contribution of the klebsiella pneumoniae capsule to bacterial aggregate and
biolm microstructures. Applied and Environmental Microbiology 77 (5), 17771782.
URL http://aem.asm.org/content/77/5/1777
[14] Fai, T. G., Grith, B. E., Mori, Y., Peskin, C. S., 2013-01. Immersed Boundary Method for Variable Viscosity
and Variable Density Problems Using Fast Constant-Coecient Linear Solvers I: Numerical Method and Re-
sults. SIAM Journal on Scientic Computing 35 (5), B1132B1161.
URL http://epubs.siam.org/doi/abs/10.1137/120903038
[15] Fai, T. G., Grith, B. E., Mori, Y., Peskin, C. S., 2014-01. Immersed Boundary Method for Variable Viscosity
and Variable Density Problems Using Fast Constant-Coecient Linear Solvers II: Theory. SIAM Journal on
Scientic Computing 36 (3), B589B621.
URL http://epubs.siam.org/doi/abs/10.1137/12090304X
[16] Foster, T., 1996. Staphylococcus. In: Baron, S. (Ed.), Medical Microbiology, 4th Edition. University of Texas
Medical Branch at Galveston, Galveston (TX).
URL http://www.ncbi.nlm.nih.gov/books/NBK8448/
[17] Gaboriaud, F., Gee, M. L., Strugnell, R., Duval, J. F. L., 2008-10-07. Coupled Electrostatic, Hydrodynamic,
and Mechanical Properties of Bacterial Interfaces in Aqueous Media. Langmuir 24 (19), 1098810995.
URL http://pubs.acs.org/doi/abs/10.1021/la800258n
[18] Ganesan, M., Stewart, E. J., Szafranski, J., Satorius, A. E., Younger, J. G., Solomon, M. J., 2013-05-13. Molar
mass, entanglement, and associations of the biolm polysaccharide of staphylococcus epidermidis. Biomacro-
molecules 14 (5), 14741481.
URL http://pubs.acs.org/doi/abs/10.1021/bm400149a
[19] Hammond, J. F., Stewart, E., Younger, J. G., Solomon, M. J., Bortz, D. M., 2014. Variable Viscosity and
Density Biolm Simulations using an Immersed Boundary Method, Part I: Numerical Scheme and Convergence
Results. Computer Modeling in Engineering and Sciences 98 (3), 295340.
URL http://techscience.com/doi/10.3970/cmes.2014.098.295.html
[20] Lai, M.-C., Peskin, C. S., May 2000. An Immersed Boundary Method with Formal Second-Order Accuracy
and Reduced Numerical Viscosity. Journal of Computational Physics 160 (2), 705719.
URL http://linkinghub.elsevier.com/retrieve/pii/S0021999100964830
[21] Liu, Y., Mori, Y., 2012-01. Properties of Discrete Delta Functions and Local Convergence of the Immersed
Boundary Method. SIAM Journal on Numerical Analysis 50 (6), 29863015.
URL http://epubs.siam.org/doi/abs/10.1137/110836699
20
[22] Liu, Y., Mori, Y., 2014-01. $Lp$ Convergence of the Immersed Boundary Method for Stationary Stokes
Problems. SIAM Journal on Numerical Analysis 52 (1), 496514.
URL http://epubs.siam.org/doi/abs/10.1137/130911329
[23] Luo, H., Mittal, R., Zheng, X., Bielamowicz, S. A., Walsh, R. J., Hahn, J. K., 2008-11. An immersed-boundary
method for ow-structure interaction in biological systems with application to phonation. Journal of Compu-
tational Physics 227 (22), 93039332.
URL http://linkinghub.elsevier.com/retrieve/pii/S0021999108002623
[24] Orszag, S. A., Israeli, M., Deville, M. O., 1986-03-01. Boundary conditions for incompressible ows. Journal of
Scientic Computing 1 (1), 75111.
URL http://link.springer.com/article/10.1007/BF01061454
[25] Pavlovsky, L., Sturtevant, R. A., Younger, J. G., Solomon, M. J., Feb. 2015. Eects of temperature on the
morphological, polymeric, and mechanical properties of Staphylococcus epidermidis bacterial biolms. Langmuir
31 (6), 20362042.
URL http://pubs.acs.org/doi/abs/10.1021/la5044156
[26] Pavlovsky, L., Younger, J. G., Solomon, M. J., 2013. In situ rheology of Staphylococcus epidermidis bacterial
biolms. Soft Matter 9 (1), 122.
URL http://xlink.rsc.org/?DOI=c2sm27005f
[27] Peskin, C. S., 1977-11. Numerical analysis of blood ow in the heart. Journal of Computational Physics 25 (3),
220252.
URL http://linkinghub.elsevier.com/retrieve/pii/0021999177901000
[28] Peskin, C. S., 2002-01. The immersed boundary method. Acta Numerica 11.
URL http://www.journals.cambridge.org/abstract_S0962492902000077
[29] Stewart, E. J., Ganesan, M., Younger, J. G., Solomon, M. J., 2015-08-14. Articial biolms establish the role
of matrix interactions in staphylococcal biolm assembly and disassembly. Scientic Reports 5, 13081.
URL http://www.nature.com/articles/srep13081
[30] Stewart, E. J., Satorius, A. E., Younger, J. G., Solomon, M. J., Jun. 2013. Role of environmental and antibiotic
stress on staphylococcus epidermidis biolm microstructure. Langmuir 29 (23), 70177024.
URL http://pubs.acs.org/doi/abs/10.1021/la401322k
[31] Stull, V. R., 1972-03. Size distribution of bacterial cells. Journal of Bacteriology 109 (3), 13011303.
[32] Sudarsan, R., Ghosh, S., Stockie, J. M., Eberl, H. J., 2015-01-28. Simulating biolm deformation and detach-
ment with the immersed boundary method. arXiv:1501.07221 [physics]ArXiv: 1501.07221.
URL http://arxiv.org/abs/1501.07221
[33] Tierra, G., Pavissich, J. P., Nerenberg, R., Xu, Z., Alber, M. S., 2015-03-25. Multicomponent model of de-
formation and detachment of a biolm under uid ow. Journal of The Royal Society Interface 12 (106),
2015004520150045.
URL http://rsif.royalsocietypublishing.org/cgi/doi/10.1098/rsif.2015.0045
[34] Wrobel, J. K., Cortez, R., Fauci, L., Nov. 2014. Modeling viscoelastic networks in Stokes ow. Physics of Fluids
26 (11), 113102.
URL http://scitation.aip.org/content/aip/journal/pof2/26/11/10.1063/1.4900941
[35] Zhang, T., Cogan, N. G., Wang, Q., 2008-01. Phase eld models for biolms. i. theory and one-dimensional
simulations. SIAM Journal on Applied Mathematics 69 (3), 641669.
URL http://epubs.siam.org/doi/abs/10.1137/070691966
[36] Zhang, T., Cogan, N. G., Wang, Q., 2008-07. Phase-eld models for biolms II. 2-d numerical simulations of
biolm-ow interaction. Communications in Computational Physics 4 (1), 72101.
URL http://www.global-sci.com/freedownload/v4_72.pdf
[37] Zhao, J., Shen, Y., Haapasalo, M., Wang, Z., Wang, Q., 2016-03. A 3d numerical study of antimicrobial
persistence in heterogeneous multi-species biolms. Journal of Theoretical Biology 392, 8398.
URL http://linkinghub.elsevier.com/retrieve/pii/S0022519315005597
21
[38] Zhu, L., Peskin, C. S., 2003. Interaction of two apping laments in a owing soap lm. Physics of Fluids
15 (7), 1954.
URL http://scitation.aip.org/content/aip/journal/pof2/15/7/10.1063/1.1582476
[39] Zhuo, J., Dillon, R., 2010-12. Using the immersed boundary method to model complex uids-structure inter-
action in sperm motility. Discrete and Continuous Dynamical Systems - Series B 15 (2), 343355.
URL http://www.aimsciences.org/journals/displayArticles.jsp?paperID=5783
22