Lecture 1: A first introduction to

Immersed Boundary Methods Circular cylinder in uniform cross-flow

• Vortex shedding from a circular cylinder in uniform cross-flow at Re=100

• Two different approaches to solve flow around a cylinder:

• Body-fitted approach: structured (curvilinear) or unstructured grid that
Wim-Paul Breugem conforms to shape of cylinder
Laboratory for Aero & Hydrodynamics, • Immersed boundary method: grid does not conform to shape of the cylinder
Delft University of Technology, The Netherlands (e.g., Cartesian grid)
Email: [email protected]

1 2

Body-fitted approach vs. IBM Additional force

• Force added to right-hand side of momentum equations to mimic boundary
conditions:
u  0
u 1 2
   u u  p   u f
t Re

body-fitted immersed boundary • IBMs differ in the way how the IB force is computed

• conformal grid: • non-conformal grid: • One of first IBMs developed by Peskin [1972]: application of IBM to simulation
structured or unstructured extended through obstacles, typically
of blood flow in a beating heart
regular (often Cartesian)
• direct imposition of b.c.’s
• no b.c.’s, but forces added to flow • Since pioneering work by Peskin many new variants of IBM have been
field to enforce b.c.’s by good
developed, see reviews by Peskin [2002], Iaccarino and Verzicco [2003] &
approximation
Mittal and Iaccarino [2005]
(Figures from: Mittal and Iaccarino [2005])

3 4

Pros of IBMs Cons of IBMs

1. Typically a regular (often Cartesian) grid is used: 1. Formulation of additional force not straightforward
a. very simple to generate compared to unstructured body-fitted grids
b. no regridding needed for moving obstacles 2. Possible influence on properties of numerical scheme such as stability,
c. simple coefficient matrix for pressure Poisson equation accuracy, conservation etc
 enables use of efficient solvers for linear equation systems
d. very simple discretization stencils 3. No good grid control for resolving boundary layers, wakes etc

2. Implementation in a programming code is usually easy

3. Robust method: this is an advantage (stable integration), but also a

disadvantage (programming bugs might not be immediately noticed)

4. Versatile method: it can be applied to many different flows and extension with
heat and mass transfer is usually straightforward

5 6

1
 IBMs: no good grid control
[Mittal and Iaccarino, 2005] Which IBM to choose?
• The requirements for an IBM are flow specific =>
there is no such thing as a universal IBM for all kind of problems, but many
variants of the IBM have been developed for specific flows

• The choice for a specific IBM depends on considerations such as:

body-fitted immersed boundary
• fixed (non-moving) obstacle vs. moving obstacle
• Consider laminar boundary layer with thickness  over body with typical size L • rigid boundary vs. flexible boundary
For laminar b.l. over flat plate:  L  Re0L.5 with Re L  U 0 L  • requirements for the accuracy, stability, conservation properties etc of the
numerical scheme
• …
• Grid dimensions in body-fitted method:  t  L (tangential),  n   (wall-normal)
Nr grid cells needed for b.l. in 3D: L t  2   n   constant in Re

• Grid dimensions in IBM with uniform Cartesian grid:    (cubical cells)

Nr grid cells need for b.l. in 3D: L  3  L   3  Re1.5
7 8

Intermezzo:
Overview of lectures today Finite-volume/pressure-correction method (I)
• Lecture 1 (this lecture) • In the Matlab exercises on the IBM today, we will make use of IBM of Fadlun
• Continuous vs. discrete forcing methods et al. [2000]. This IBM is embedded in a finite-volume/pressure-correction
• Discrete forcing method of Fadlun et al. [2000], illustrated for a fixed circular method [Ferziger and Perić, 2002] of which some details are given below.
cylinder in uniform cross flow  Matlab exercises
• Finite-volume method: Navier-Stokes equations are integrated locally over a
• Lecture 2 grid cell (= small control volume). Fluxes/stresses need to be evaluated at cell
• Stress IBM [Breugem and Boersma, 2005; Pourquie et al., 2009] for fixed faces. Global conservation of mass and momentum is guaranteed.
rectangular obstacles that can be aligned along grid
• Extension Stress IBM to scalar transport: comparison Stress IBM with IBM of Fadlun • Staggered grid [Harlow and Welsh, 1965]:
et al. [2000]
• 2nd-order central-differencing scheme
• Lecture 3 for spatial gradients:
• Discrete forcing method of Uhlmann [2005] for moving obstacles
1 p pi 1, j yz  pi , j yz pi 1, j  pi , j
xyz Vcell x u
• Modifications by Breugem [2012] dV  
xyz x
i, j

9 10

Finite-volume/pressure-correction method (II) Finite-volume/pressure-correction method (III)

• Navier-Stokes equations: • Time integration of momentum equation with 2nd-order Adams-Bashforth
scheme for advection + diffusion terms, and 2nd-order Crank-Nicolson scheme
u  0 for pressure gradient:
u 1 2 n 1
 p  r with r  u   u u u u
n
n 1  3 n 1 n 1 
t Re  p 2   r  r 
t 2 2 

• Concept of pressure-correction method for incompressible flows: with the velocity defined at time full levels and the pressure at time half levels.
a. A prediction velocity is computed from the momentum equation. This
velocity does not satisfy the continuity equation. • Pressure at time level n+1/2 is unknown  pressure decomposed into:
b. A pressure correction to this velocity is computed such that the final
velocity is divergence free. n  12 n  12
p p ~
p ~
with p the correction pressure

11 12

2
Finite-volume/pressure-correction method (IV) Finite-volume/pressure-correction method (V)
• Rewrite previous equation into following form, where all unknowns are • Correct prediction velocity for mass conservation:
grouped at left-hand side: n 1
 u  t~
*
u p
n 1  n 1 3 n 1 n 1 
 t~
p  u  t   p 2  r  r 
n
u
 2 2  n  12 n  12
• Compute pressure at n+1/2: p p ~
p
• Define left-hand side as the prediction velocity:
n 1 • Summary of pressure-correction scheme:
u u  t~
*
p
 n 1 3 n 1 n 1 
u  u  t   p 2  r  r 
* n
• Take divergence of the prediction velocity:
 2 2 
n 1 1
u  u  t2 ~ 2 ~
*
p   u
*
p
t
= 0 because of continuity equation n 1
u  u  t~
*
p
1
This yields a Poisson equation for correction pressure:  2 ~
p  u
*
n  12 n  12
t p p ~
p

13 14

Classification of IBMs:
continuous vs. discrete forcing (I)
Finite-volume/pressure-correction method (VI) [Fadlun et al., 2000; Mittal and Iaccarino, 2005]
1
• Pressure Poisson equation: 2 ~
p  u
*

t u 1 2
   u u  p   u f
• What boundary conditions for correction pressure? t Re
n 1
For solid wall with no-slip/no-penetration b.c. put: u  u  0 . From • Continuous forcing:
*

n 1
u  u  t~ p we then get the homogeneous Neumann b.c.: 1. Force added to momentum equation prior to discretization.
*

~0 Example: original IBM of Peskin [1972], applied to simulation of blood flow in
p
a beating heart  IB represented by set of elastic fibers with fiber stress
• The pressure Poisson equation can be written in following discrete form: given by Hooke’s law  fiber stress transmitted from IB to grid by means of

A ~ 
p     u t
*
 regularized Dirac delta function.
2. Force formulation does not depend on used numerical scheme.
For Matlab exercise of flow over cylinder (2D domain, Cartesian grid), A is 3. Effect on numerical accuracy, stability, conservation etc.
symmetric and positive definite, is sparse (only 5 non-zero diagonals) and is 4. Widely used for elastic boundaries, but stiffness problems for rigid
time-independent --> linear equation system can be efficiently solved boundaries.
(for example with FFT-based solvers: A transformed into tri-diagonal matrix).

15 16

Stiffness problem in continuous forcing IBMs (I) Stiffness problem in continuous forcing IBMs (II)
• Method of Goldstein et al. [1993] to enforce no-slip/no-penetration condition • Consider 1D case with:
on rigid solid boundary: dU
t

u 1 2
   U ( X , t ' )dt '  U( X , t )
   u u  p   u f dt 0
t Re
t Solution describes a damped harmonic oscillator of form:
on Lagrangian grid
on boundary: F  X     U ( X , t ' )dt '  U ( X , t )    2  4
0 U ~ e t with 
force transmitted 2
to underlying f x ijk    F l d ( x ijk  X l )Vl damping ratio:  

Eulerian grid: l
2 
• Regularized Dirac delta function d => smoothing of sharp interface => undamped angular frequency: 0  
affects the numerical accuracy
• Coefficients  ,  must have large negative values: • If forcing terms are integrated explicitly, then severe constraint on
• choice of values for  ,  is subjective and may influence results time step when  ,  have large negative values (stiff problem 
• severe consequences for numerical stability when time integration explicit large eigenvalues!).
17 18

3
 Classification of IBMs:
Discrete forcing method of
continuous vs. discrete forcing (II)
[Fadlun et al., 2000; Mittal and Iaccarino, 2005] Fadlun et al. [2000] (I)
• Forcing applied to nearest velocity nodes just outside obstacle.
u 1 2 • Velocity on these velocity nodes are determined from linear interpolation from
   u u  p   u f
t Re second-nearest velocity nodes to obstacle boundary.
second v-node first v-node
• Discrete (or direct) forcing [Mohd-Yosuf, 1997]:
1. Force added to momentum equation after discretization. v=0 desired
on boundary
2. Force formulation depends on used numerical scheme.
3. Better control over numerical accuracy, stability, conservation etc
4. Widely used for rigid boundaries.

prior to forcing after forcing: velocity

corrected on nearest node
• Momentum equation is effectively not solved for nearest velocity nodes
• Linear velocity profile assumed up to 2nd nearest node
• Interpolation direction is subjective -> bi-linear interpolation [Kim et al., 2001]
19 20

Discrete forcing method of Discrete forcing method of

Fadlun et al. [2000] (II) Fadlun et al. [2000] (III)
• IBM is embedded in finite-volume/pressure-correction method • Forcing is applied to the second prediction velocity instead of the true velocity
1. Calculate prediction velocity in absence of obstacle: at time level n+1 -> Forcing corrected for mass conservation.
 n 1 3 n 1 n 1  Difference between the 2 velocities must therefore be kept small. For present
u  u  t   p 2  r  r 
* n

 2 2  scheme this is satisfied for sufficiently small time step:

2. Add force to nearest velocity nodes, based on linear interpolation: n  12
 p 
n  12 u lin.int.  u* n  12
u u
** n 1
 t~
n 1
 
p  u  t 2    O(t 3 )
f 
t
 u**  u*  t f  u lin.int.  t  n 1
u u
**
For stationary flows: correction pressure is expected to vanish 
with u **  u n1  t~
p the second prediction velocity, now corrected for
presence of obstacles • Be careful when using implicit time integration of diffusion terms!
3. Pressure Poisson equation:  2 ~ 1
p   u
**

t u 1 2
 p   u A with A    uu
4. Correction on prediction velocity: u
n 1
 u  t~
**
p t Re
5. Update pressure: n 1 n 1
p 2 p 2~ p
21 22

Implicit integration of diffusion term (I) Implicit integration of diffusion terms (II)
• Time integration of momentum equation: Crank-Nicolson for pressure gradient • Rewrite previous equation as follows:
and diffusion terms, Adams-Bashforth for advection terms:
t 2 **  n 1 1 2 n 3 n 1 n1  n 1 t 2 n
u   u  u  t   p 2   u  A  A   t f 2   u
** n

 
 
n 1
u 2 Re Re 2 2 2 Re
n
u n 1 1 n 1 1 2 n 3 n 1 n1 n 1
 p 2  2 u  u   u  A  A f 2
n

t 2 Re Re 2 2 • Define first prediction velocity as:

 n 1 1 2 n 3 n 1 n 1 
u  u  t   p 2   u  A  A 
* n
• Decompose pressure as follows:
 Re 2 2 
n  12 t 2 ~
n  12
p p ~ p  p with ~ p the correction pressure • So:
2 Re t 2 ** t 2 n
•
n 1
Define second prediction velocity as usual: u  u  t~
** n 1
p  u  O t 2   u 
**

2 Re
* n 1
 u  u  t f 2 
2 Re
 u
• Substitution in momentum equation yields:
• Force formulation at velocity nodes where forcing is applied:
u u
 
** n
n 1 1 1 2 n 3 n 1 n1 n 1
 p 2  2 u  u   u  A  A f 2
 
** n
u lin.int.  u t 2 **
 
*
n  12
t f   u  u  t f   u  u  u lin.int.  O t 2
** * n
2 Re Re 2 2
t 2 Re

23 24

4
 Implicit integration of diffusion terms (III) Implicit integration of diffusion terms (IV)
• So, at location where forcing is applied, we get: • The following definition of first prediction velocity may result in a significant
error:
p  u lin.int.  Ot 2 
n 1
 u  t~  3 n 1 n 1 
**
u n 1 1
u  u  t   p 2  2 u  A  A 
* n n

which is as desired.  2 Re 2 2 
• Final scheme:  n 1 1 2 n 3 n 1 n 1 
u  u  t   p 2   u  A  A 
* n
[Uhlmann, 2005]
 n 1 
• Then: t 2 ** n 1
u   u  u  t f 2
** *
t 2 ** t Re2 n 2 2
u   u  u  t f   u
** * 2
2 Re
2 Re 2 Re
1
2 ~
p  u
*

t • So at velocity nodes where forcing is applied:

n 1
 u  t~
**
u p u lin.int.  u
*
t 2 **
 u  u lin.int.  Ot 
n  12
t 2 ~ f   u  u  t f 
** *
n  12 n  12
p p
 p ~
p t 2 Re
2 Re
• Important message: the first prediction velocity should be defined such that and the error scales with Ot  , while the previous scheme yields
in the absence of forcing it is a good estimate of the final velocity at time  
an error of O t 2 and vanishes for stationary flows
level n+1.

25 26

Circular cylinder in uniform cross-flow (I)

(simulated with IBM of Fadlun et al. [2000]) Circular cylinder in uniform cross-flow (II)
• Matlab exercise: circular cylinder in uniform cross-flow • Location of IBM forcing points:
• filled blue: U
• Computational domain and boundary conditions • filled red: V

• Interpolation direction (subjective):

• horizontal for U
• vertical for V

• Open symbols indicate 2nd nearest

velocity nodes used for interpolation.

DUin D
• Parameters: Re   40..100,  16, l x  30 D, l y  18D
 x

27 28

Circular cylinder in uniform cross-flow (III) Circular cylinder in uniform cross-flow (IV)
• Close-up of velocity field around cylinder • Calculation of lift coefficient
• Integrate v-momentum equation over volume of cylinder

D   1 v 
Dt  Vc
 vdV      pI 2 j   n j dA  n y   f y dV
 V  Re x j
 c   Vc

• Neglect inertia of flow within cylinder and replace integral by sum over grid cells
where forcing is applied:
 1 v 
   pI   n j dA  n y    f y V
Re x j
2j
Vc  
• For 2D flow simulation, we get:
Re=40: stationary flow, no vortex Re=100: instationary flow, periodic
shedding vortex shedding Fl
Cl   2 f y xy
1
2 U in2 D

• Drag coefficient calculated in similar vane: Cd  2 f x xy

29 30

5
 Circular cylinder in uniform cross-flow (V) Circular cylinder in uniform cross-flow (VI)
• Time series of drag and lift coefficients: • Maximum error in the enforced velocities at the IBM forcing points, in
percentage of the uniform inflow velocity

Re  40 Re  100
Re  40 : Cd  1.58, Cl  0 Re  100 : C d  1.34, Clp  0.22, St  0.17
Stationary flow  error in Instationary flow  error in
Park et al. [1998]: Cd  1.51, Cl  0 Park et al. [1998]:
enforced velocity at IBM forcing enforced velocity at IBM forcing
(body-fitted
C d  1.33, Clp  0.33, St  0.165 points goes to zero points starts oscillating when
approach) vortex shedding sets in
31 32

Circular cylinder in uniform cross-flow (VII) Circular cylinder in uniform cross-flow (VIII)
• Re=100 again, but now with forcing of interior flow within cylinder ( u **  0 ) • Influence of location of cylinder at Re  40, D x  16

C d  1.35, Clp  0.29, St  0.17

Location of IBM forcing points, including Original cylinder position: identical Cylinder shifted to right by D/33 (less
interior of cylinder (plus symbols) Error in enforced velocity at IBM distribution when rotating over 90˚ than 0.5Δx): distribution different
forcing points much larger…? when rotating over 90˚
33 34

Circular cylinder in uniform cross-flow (IX) Beating heart of chicken embryo (I)
• Grid convergence test for Cd at Re=40 • Goal: experimental and numerical determination of wall shear stress
distribution in the heart of a chicken embryo

Notice: no monotonic
grid convergence

• Results are sensitive to position cylinder wrt underlying grid: formulation of

IBM force is not translation invariant [Peskin, 2002]  influence of grid size, Experiments at TU Delft [Vennemann, 2008]
“grid locking effect” [Breugem, 2012]

35 36

6
 Beating heart of chicken embryo (II) References (I)
• Numerical simulations by Pourquie et al. [2005] based on IBM of Fadlun et al. 1. W.-P. Breugem and B.J. Boersma, Direct Numerical Simulations of turbulent flow over a permeable wall using a direct

[2000]. Heart beat prescribed by deformation of tube. and a continuum approach. Phys. Fluids, 17(2), 2005.
2. W.-P. Breugem, A second-order accurate Immersed Boundary Method for fully resolved simulations of particle-laden

flows. J. Comp. Phys. 13(1):4469-4498, 2012.

3. E.A. Fadlun, R. Verzicco, P. Orlandi and J. Mohd-Yusof, Combined immersed-boundary finite-difference methods for

three-dimensional complex flow simulations. J. Comp. Phys. 161:35-60, 2000.

4. J.H. Ferziger and M. Peric, Computational methods for fluid dynamics. Springer-Verlag, Berlin, 2002.

5. F.H. Harlow and J.E. Welch, Numerical calculation of time-dependent viscous incompressible flow of fluid with free

surface. Phys. Fluids 8:2182-2189, 1965.

6. D. Goldstein, R. Handler and L. Sirovich, Modeling a no-slip flow boundary with an external force field. J. Comp.

Phys. 105(2):354-366,1993.

7. G. Iaccarino and R. Verzicco, Immersed boundary technique for turbulent flow simulations. Appl. Mech. Rev.

56(3):331-347, 2003.
triangulation of bias in wall shear stress since simulated wall
tube surface IBM assumes linear velocity shear stress 8. J. Kim, D. Kim and H. Choi, An immersed-boundary finite-volume method for simulations of flow in complex
profile up to 2nd velocity node distribution geometries. J. Comp. Phys. 171:132-150, 2001.

37 38

References (II)
9. R. Mittal and G. Iaccarino, Immersed boundary methods. Annu. Rev. Fluid Mech. 37:239-261, 2005.

10. J. Mohd-Yusof, Combined immersed boundary/B-spline methods for simulations of flows in complex geometries . CTR

Annual Research Briefs, NASA Ames/Stanford University, 1997.

11. J. Park, K. Kwon and H. Choi, Numerical solutions of flow past a circular cylinder at Reynolds numbers up to 160.

KSME Int. Journal 12(6):1200-1205, 1998.

12. C.S. Peskin, Flow patterns around heart valves: a numerical method. J. Comp. Phys. 10:252-271, 1972.

13. C.S. Peskin, The immersed boundary method. Acta Numer. 11:479-517, 2002.

14. M. Pourquie et al., Blood shear stress in a chicken embryo: experiments and numerical modeling. Proceedings TSFP-

4, Turbulence and Shear Flow Phenomena, June 27-29 2005, Williamsburg, Virginia, USA.

15. M. Pourquie, W.P. Breugem and B.J. Boersma, Some issues related to the use of immersed boundary methods to

represent square obstacles. Int. J. Multiscale Comp. Eng. 7(6):509-522, 2009.

16. M. Uhlmann, An immersed boundary method with direct forcing for simulation of particulate flows . J. Comp. Phys.

209:448-476, 2005.

17. P. Vennemann, Particle Image Velocimetry for microscale blood flow measurement. PhD thesis, Delft University of

Technology, 2008.

39

7
 Lecture 2: Flow over rectangular obstacles
a discrete-forcing IBM for rectangular obstacles (urban areas, porous media)
• IBM of Fadlun et al. [2000]: assumption of linear velocity profile up to 2 nd
Breugem and nearest velocity node to boundary
Boermsa [2005]:
turbulent channel flow
over a porous bed of
5400 cubes, simulated
with the Stress IBM

IBM of Fadlun et al. [2000]: Stress IBM:

velocity (or force) imposed stress imposed on IB
Wim-Paul Breugem
Laboratory for Aero & Hydrodynamics, • Improvement possible when obstacle is rectangular and can be aligned with
Delft University of Technology, The Netherlands Cartesian grid: assumption of linear velocity profile only up to 1st node 
Email: [email protected] the “Stress IBM” [Breugem and Boersma, 2005; Pourquie et al., 2009]
1 2

Stress IBM for rectangular obstacles (I)

IBM of Fadlun et al. [2000] [Breugem and Boersma, 2005; Pourquie et al., 2009]
• IBM is embedded in finite-volume/pressure-correction method Principle of Stress IBM:
1. Calculate prediction velocity in absence of obstacle: 1. Boundary of obstacle aligned with
grid such that boundary coincides
 n 1 3 n 1 n 1 
u  u  t   p 2  r  r 
* n
with nodes for boundary-normal
 2 2 
2. Add force to nearest velocity nodes, based on linear interpolation: velocity.
2. For nodes on boundary, put 2nd
u lin.int.  u
*
f   u  u  t f  u lin.int.
** *
prediction velocity in pressure-
t correction scheme to zero:
with u **  u n1  t~
p the second prediction velocity, now corrected for un**  0 
presence of obstacles  n 1 
~    p  2 
unn 1  t
1 p
3. Pressure Poisson equation:  ~ p   u
**  O t 2
n  t 
2
n  
t  
n 1
4. Correction on prediction velocity: u
n 1
 u  t~
**
p For stationary flow: un  0 (exact! )
5. Update pressure: n 1 n 1
p 2 p 2~ p  
For instationary flow: unn1  O t 2 (very small)
3 4

Stress IBM for rectangular obstacles (II) Stress IBM for rectangular obstacles (III)
Principle of Stress IBM (continued): • The Stress IBM can be easily combined with RANS (Reynolds-Averaged Navier-
3. At velocity nodes half grid cell away from Stokes) or LES (Large-Eddy Simulation) models where logarithmic velocity
boundary: adjust stencil for discretization profile is assumed near IB:
of wall-normal advection and diffusion term
and impose correct stress/flux at wall => u 1  y
 ln    desired stress at wall: u2 

 u ( 1 y )
2
2
“exact” no-slip condition imposed! u   y0    1 y 
2
For example, in absence of boundary we ln  

get for diffusion term at node (i,k):   2 y0 
1  2u 1  ui ,k 1  2ui ,k  ui ,k 1  IBM of Fadlun et al. [2000]: logarithmic interpolation required

Re z 2 Re  z 2 

but if boundary present then:
• Adjustment of discretization stencil for flux/stress on IB has implications for
1  2u 1  ui ,k 1  ui ,k   ui ,k  0   numerical stability; some theoretical analysis is possible
   
Re z 2 Re z  z   z 2  
1  ui ,k  ui ,k 1 
The required forcing is the difference ft    
between the two formulations: Re  z 2 
5 6

1
 Accuracy of Stress IBM:
Stress IBM for rectangular obstacles (IV) flow through a periodic array of square bars
• Von Neumann stability analysis [Pourquie et al., 2009]:

u
3
  1 2 
 Lu  0, Lu    U l   u  ft
t i 1  xl Re xl2 
f t  
l
1
Re xl2

l um  umel   l um  umel  
where  l , l  0 or 1 are forcing-indicator functions, l  1..3 , m  (i, j, k ) is grid
index and e1  (1,0,0) etc represent a shift in grid index
• Eigenvalues of discretised NS-operator L  must lie within stability domain of
used time-integration scheme. For 3D simulation with AB2 scheme and cubical
grid cells, it can be derived that:
2

t  U 
1
l   l 
t 1 l
Re x  1 • Distribution of IBM forcing points along boundary
 , l

Re x 2 12 x 3 • Staggered grid: different distribution for each velocity component
• Close to wall: first (viscous) criterion dominates => IBM no effect on stability • Forcing at both sides of boundaries in simulations

7 8

Laminar case (I) Laminar case (II)

Note: IBM of Uhlmann [2005] developed Definition of permeability:
for moving obstacles (diffuse interface)  u
K
 dpe 
 
u d  dx 
Re   0.05

(Stokes flow)
u d
Re   0.05


• Velocity field and pressure distribution

• Pressure gradient inside object put to zero in simulation
• Stationary flow: zero penetration across boundaries, velocities inside object zero =>
results identical as for standard body-fitted method!

9 10

Turbulent case (I) Turbulent case (II)

u d
• Re   625

• 192x96x96 grid points
penetration velocity
across walls:

• corresponding correction
unn 1  t
~
p
n
 
 O t 2

pressure and velocities inside

objects (multiplied by 104)
• largest penetration velocity
1000 times smaller than bulk
velocity:
t 2 (at fixed spatial resolution)
 
• Max. penetration velocity scales with
~
unn 1  t
p
 O t 2
n
11 12

2
 Extension of Stress IBM to triangular obstacles
Turbulent flow over a porous bed possible for some specific cases
Breugem and
Boermsa [2005]:
turbulent channel flow
over a porous bed of
5400 cubes, simulated
with an IBM

• Porous bed of 30x20x9 = 5400 cubes angle of wall determines

• 600x400x400 = 96∙106 grid cells (203 for each unit cell of porous bed) aspect ratio of grid cells

Stokes flow around triangle (periodic b.c.’s)

13 14

Extension of Stress IBM to scalar transport (I)

[Pourquie et al.,2009] Extension of Stress IBM to scalar transport (II)
• Stress IBM can be easily extended with a scalar • In presence of boundary with zero wall-normal
transport equation: concentration flux:
c
 rc with rc    uc  D 2c  2c D  ci, k 1  ci, k 
   0 
t D
z  z
z 2 
• Discretize and adjust stencil for wall-normal
 c c 
(diffusive) flux at wall. This is equivalent to Required source term at (i,k): S  D i, k i, k 1 
 
adding a source term at right-hand side. With  z 2 
AB2 time integration scheme we get: • In presence of boundary with prescribed
concentration:
c n 1  c n 3
 rc  S n  rc  S n 1
1
 2c D  ci, k 1  ci, k ci, k  Cs 
t 2 2 D    
z 2 z  z z 2 
• Wall-normal diffusion term at (i,k) in absence of Required source term at (i,k):
boundary:
 ci, k  2Cs  ci, k 1 
 2c  ci, k 1  2ci, k  ci, k 1  S   D 
D  D  
  z 2 
z 2  z 2 

15 16

Scalar dispersion in street canyon

[Pourquie et al., 2009] Case A: zero flux across walls (I)
• DNS of low-Re turbulent flow over street 3
C wall L VB stress
3
C wall R VB stress
C wall L VB verz C wall R VB verz

canyon
C wall L standard C wall R standard
2.5 2.5

2 2
HEIGHT

HEIGHT

• Inflow, outflow boundary conditions 1.5 1.5

1 1

• At bottom of cavity a scalar is released

0.5 0.5

0 0
-1 0 1 2 3 4 5 6 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
C C

• Comparison of IBM of Fadlun et al. mean concentration close mean concentration close
[2000] and Stress IBM with standard to left wall of cavity to right wall of cavity
body-fitted method
• Mean and rms concentration close to left
• Legend: black = standard body-fitted method ; red = Stress IBM ; blue =
and right wall of cavity considered IBM of Fadlun et al. [2000]
(dashed lines)
• 2 different cases considered: a) zero flux • Results of both IBMs in good agreement with standard body-fitted method
and b) prescribed concentration at walls

17 18

3
 Case B: prescribed concentration at walls
Case A: zero flux across walls (II) (near-wall rms concentration scales with concentration flux across wall)
3 3
RMS C wall L VB stress RMS C wall R VB stress 3 3
RMS C wall L VB verz RMS C wall R VB verz C wall R VB stress RMS C wall R VB stress
RMS C wall L standard RMS C wall R standard C wall R VB verz RMS C wall R VB verz
2.5 2.5 C wall R standard RMS C wall R standard
2.5 2.5

2 2
2 2

HEIGHT
HEIGHT

HEIGHT
HEIGHT
1.5 1.5
1.5 1.5

1 1 1
1

0.5 0.5 0.5

0.5

0 0 0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
RMS C RMS C C RMS C

rms concentration close to rms concentration close to mean concentration close rms concentration close to
left wall of cavity right wall of cavity to right wall of cavity right wall of cavity

• Legend: black = standard body-fitted method ; red = Stress IBM ; blue =

• Legend: black = standard body-fitted method ; red = Stress IBM ; blue = IBM IBM of Fadlun et al. [2000]
of Fadlun et al. [2000]
• Conclusion: Stress IBM in good agreement with standard method, while IBM
• IBM of Fadlun et al. [2000] less accurate at right wall in particular. Reason: grid of Fadlun et al. [2000] is accurate only if grid is sufficiently fine near wall. The
too course near right wall for assuming linear profile up to 2 nd concentration interpolation error appears more important in this case than error in no-
(pressure) node (y+ ~ 6 sometimes). penetration condition at wall (since Stress IBM appears much more accurate).
19 20

References
1. W.-P. Breugem and B.J. Boersma, Direct Numerical Simulations of turbulent flow over a permeable wall using a direct

and a continuum approach. Phys. Fluids, 17(2), 2005

2. E.A. Fadlun, R. Verzicco, P. Orlandi and J. Mohd-Yusof, Combined immersed-boundary finite-difference methods for

three-dimensional complex flow simulations. J. Comp. Phys. 161:35-60, 2000.

3. M. Pourquie, W.P. Breugem and B.J. Boersma, Some issues related to the use of immersed boundary methods to

represent square obstacles. Int. J. Multiscale Comp. Eng. 7(6):509-522, 2009.

4. M. Uhlmann, An immersed boundary method with direct forcing for simulation of particulate flows . J. Comp. Phys.

209:448-476, 2005.

21

4
 Lecture 3: Application of the Immersed Boundary Method to
simulation of particle-laden flows with finite-size particles Governing equations for particle-laden flows
externally imposed, spatially
• Navier-Stokes equations for the fluid phase: uniform pressure gradient

gravity imposed
pressure
gradient • Newton-Euler equations for (spherical) particles:

moment of inertia
for solid sphere:
Pressure-driven vertical plane 2
channel flow laden with 96 Ip   pV p R 2
5 collision terms
particles
velocity of particle segment:
Wim-Paul Breugem
Laboratory for Aero & Hydrodynamics, • Boundary conditions:
Delft University of Technology, The Netherlands
Email: [email protected]
• Body-fitted methods: direct imposition of boundary conditions -> expensive
1 2

An explicit, 2nd-order accurate pressure-correction

Concept of Immersed Boundary Method (IBM) scheme for integrating N-S equations
• No direct imposition of boundary conditions, but instead additional force: • Like standard, but with 2nd prediction velocity that includes IB forcing:
 n 1 3 
u*  u n  t   pe  p 2  r n  r n 1  with r 
1 1 2
1st prediction velocity:  u  u u
no account of  2 2  Re
solid obstacles
IB force that accounts for
u**  u*  t f n  2
1
obstacle
2nd prediction velocity:
presence of solid obstacles
correction for presence
• Pros: (instead of using b.c.’s) of solid obstacles
• no regridding needed for moving obstacles fluid phase
u**  u n 1  t~
p  2 ~
p
1
  u**    u n 1
1
• simple grids allow use of efficient solvers t t
=0
n 1
• “easy” implementation in programming code  u  t~
**
u p correction pressure kept small by
• robust and versatile method updating pressure =>
• Cons: p
n  12
p
n  12
~
p  
u n 1  u**  O t 2
• imposition of IB forces not straightforward
• (possible) effects on accuracy, stability, conservation properties etc of used • Compute IB force from requirement that at solid boundary:
numerical scheme u**  u d ( = desired velocity at solid boundary)
• no good grid control (boundary layers, wakes etc)

3 4

Determination IB force: 2 different approaches IBM of Uhlmann [2005] for particle-laden flows
1. Determine IB force directly at Eulerian grid: • IBM of Uhlmann [2005] for moving particles makes use of 2 grids:
• Eulerian: fixed, continuous and 3D Cartesian grid for fluid phase
• Lagrangian: quasi-2D, uniform grid attached to (and moving with) solid boundaries

Requires interpolation of desired particle velocity to Eulerian grid:

[e.g., Fadlun et al., 2000]. Special case: aligned cubes [Breugem & Boersma, 2005].

2. Determine IB force at Lagrangian grid:

• Coupling between grids through transfer kernel (= regularized delta function)

Requires interpolation of 1st prediction velocity to Lagrangian grid, ,
and subsequent spreading of IB force to Eulerian grid, . • Uhlmann’s IBM combines concept of discrete forcing method [Fadlun et al.,
Advantage: force distribution on Lagrangian grid needed anyway for N-E eqs. 2000] with smooth interface representation of Peskin’s IBM [Peskin, 2002]
5 6

1
 Transfer kernel for interpolation/spreading operations Calculation scheme for IB force
Interpolation :
1. (no account of solid obstacles)

Spreading : 2. (interpolation)

Roma et al. [1999] 3. (calculation force on Lagrangian grid)

4. (spreading)

5. (correction for
presence solid
obstacles)

7 8

Useful properties of regularized delta function Evaluation of force on particle (I)

• For regularized delta function of Roma et al. [1999], it holds that: • Momentum equation with additional force:

 x d 
3
ijk  X l x  1
f
Du
Dt
 pe       f f with    p I   f u
ijk

 x  xijk d 
3
ijk  X l x  X l
• Compute drag force on (fluid) particle:
 
ijk  
  
D
  ndA    f f dV   f udV   V ppe
• Consequently, conservation of force and torque is automatically satisfied: Dt  
V  
Vp V p 
 
  
f xijk x3 

F Xl    
 d xijk  X l x3  Vl  
F X l Vl    • Substitute in Newton’s law:
ijk l  ijk  l 



   pV p
d uc
 f
 f dV  
D
  
udV    p   f V p g
   
f
Dt  
        X   
  dt
x  x c  f xijk x3  x  x c  d xijk  X l x3   F X l Vl  V p 
l  x c  F X l Vl  
Vp

l  ijk 
ijk l
IB force corrected for inertia of interior flow

9 10

Evaluation of force on particle (II) Evaluation of torque on particle (I)

• Approximation of inertia term, assuming rigid-body motion on surface of • Momentum equation with additional force:
sphere [Uhlmann, 2005]:   Du
  f  pe       f f with    p I   f u


D d uc Dt
f udV    f V p
Dt   dt
V p  • Compute torque from drag on (fluid) particle:
 
• Then:  

d uc 1  
    
  f dV  g
 
D
dt  p  f 1 Vp r    n dA    f r  f dV   f r  udV  with r  x  x c
Vp Dt  
V  
• Discretize and replace integral by discrete sum:
Vp V p 

u cn 1  u cn • Substitute in Euler equation:

t
 
 p
1
 f 1 Vp  F n  12
l Vl  g
d c




 r  f dV   
D
l Ip  f f r  udV 
dt Dt  
• Compute new particle position: V p 
Vp  
x cn 1  x cn
t
1
2

 u cn 1  u cn  torque from IB force corrected for inertia of interior flow

11 12

2
 Evaluation of torque on particle (II) Explicit versus implicit coupling
• Approximation of inertia term, assuming rigid-body motion inside sphere • IBM of Uhlmann [2005]: explicit coupling of Navier-Stokes and Newton-
[Uhlmann, 2005]:   Euler equations
 
I d c

D
f r  udV    f c • Explicit coupling of N-S and N-E eqs:
Dt    p dt
V p  • Given particle velocity and positions at time level n, the N-S eqs are
 
• Then: d c f integrated to time level n+1
dt
 
I p 1  f  p r  f dV
 • Next, given the IBM force distribution at time level n+1/2, the N-E eqs are
Vp integrated from n to n+1
• Discretize and replace integral by discrete sum:
 cn 1   cn f • Implicit coupling: integrate N-S and N-E eqs simultaneously forward
t
 
1   f 
p Ip  X
l
l  xc F l 
n  12
Vl with IB force based on particle velocity at time level n+1 => iterative
scheme required, which is computationally expensive
• Compute new angular position:
 cn 1   cn
  cn 1   cn
1
t 2
  not needed for spheres because of
rotational symmetry
• Implicit coupling is stable, whereas explicit coupling becomes unstable
for low particle-to-fluid mass density ratio
13 14

Explicit coupling unstable for low mass density ratio (I) Explicit coupling unstable for low mass density ratio (II)
[Hu et al., 2001] [Hu et al., 2001]
• Newton’s law:  
F g   p   f Vp • Substitute in Newton’s law:
n  12 is net buoyancy force n  12 n  12 F n  32 
n  12   d uc Fg ma d u c Fg  g ma d u c
ma 
       
 
d uc
   f f n  2 dV   f 
D
udV 
1
mp  Fg dt mp m p dt mp  mp
m p m p dt 
dt  Dt   
    2 n 1  n 1
Vp  Vp  Fg  ma   ma   m    ma  d u c 2
n  12   1          ..    a 
  
m p   pV p  F d  drag force on particle, mp   m p   m p   m p 
   m p  dt
computed from particle velocity and position  
is mass of particle
at (old) time level n

Fg   a p   
 1   m m n   m n d u 2
a 
1

    m  dt
c
mp  1  m m
 a p
  p
• Consider particle accelerating from rest => main contribution to drag
force from added mass. Assume that in case of explicit coupling: • Above analysis suggests that explicit coupling becomes unstable when:
n  12 1
ma  m p   p  f  for sphere in freespace
n 1 d uc 1
F 2 d  ma with ma   f V p theadded mass of a spherein freespace 2
dt 2
• Implicit coupling only needed for light particles
15 16

Grid locking effect (I) Grid locking effect (II)

• Regularized delta function of Roma et al. [1999] suppresses force oscillations • Origin of force oscillations: the interpolation and spreading operations are
when particle moves over fixed grid [Uhlmann, 2005] translation invariant [Peskin, 2002]
=> force distribution dependent on particle position wrt to underlying fixed grid
=> wave length of force oscillations scale with grid spacing: “grid locking”
yc (t )  A sin  2 f t 

• Force oscillations can be suppressed by choosing support for regularized delta

function sufficiently wide, but for computational efficiency it is desired that the
support is compact
 the regularized delta function of Roma et al. [1999] is a compromise
vertically oscillating oscillations in Cd small oscillations in Cd large
cylinder in uniform
cross flow at • All IBMs with a sharp interface representation will most likely suffer from grid
U in D
Re D   185
f IBM of Uhlmann [2005] with IBM of Kajishima and Takiguchi [2002], locking => a diffuse interface representation is therefore desired for
regularized delta function of based on volume penalization:  u  u*  simulations of moving particles!
Roma et al. [1999]
u**   u d  1   u*  f n  2    d
  
1

 t 
with  the solid volume fraction  

17 18

3
 Some modifications possible to IBM of Uhlmann [2005]
[Breugem, 2012] Modification 1: modified scheme for IB force
1. Multi-direct forcing scheme: improves enforcement of no-slip/no-penetration 1. (no account of solid obstacles)
conditions
2. (interpolation)
2. Inward retraction of Lagrangian grid: correction of effective particle diameter

3. Direct account of fluid inertia within particle: enhanced numerical stability for 3. (calculation force)
particle-fluid mass density ratio near unity

4. (spreading)

(correction for
5. presence solid
obstacles)
Problem: diffuse force distribution results in which an
Eul. grid point often used for more than 1 Lagr. grid point!

19 20

Overlap in Eulerian force distributions from

Iterative (multi-direct forcing) scheme for computation IB force
different Lagrangian points [Luo et al., 2007]

• Repeat steps 2-5 to improve estimate of IB force:

2.

3.

4.

5.

• Same Eulerian grid point used more than once for forcing velocity at
• Nq is number of iterations until 2nd prediction velocity satisfies:
Lagrangian grid. Remedy: multi-direct forcing scheme [Luo et al., 2007] in
which required force distribution is calculated iteratively.

21 22

Modification 2: Modification 3:
inward retraction of Lagrangian grid direct account of inertia fluid inside particle
• Newton’s law:

drag force
• The drag force can be related to the IBM force distribution as follows:
 
 
  
D
  ndA    f f dV   f udV   V ppe
Dt  
V V p 
Vp  
inertia of fictitious fluid contained within particle
 
 

D d uc
• Uhlmann assumed rigid-body motion at surface particle:  f udV    f V p
Dt   dt
 
V p 
 f dV  g
d uc 1
• Interface smeared out => effective particle diameter larger than actual
diameter. Remedy: slight inward retraction of Lagrangian grid [Hoefler &

dt

 p 
 f 1 Vp
Vp
Schwarzer, 2000]. Optimal value: rd  0.3x (determined from simulations). Problem: equation becomes singular for p   f !

23 24

4
 Improved stability for mass density ratio Performance of modified IBM:
near unity [Kempe et al., 2009] Stokes flow through periodic array of fixed spheres
• Remedy: direct evaluation of fluid inertia with help of level-set method • flow driven by imposed pressure gradient,
[Kempe et al., 2009] -> Reynolds number very low (< 0.1)
• 16 grid cells over sphere diameter  323 cells
• retraction distance Lagrangian grid = 0.25 ∆x
where ijk is the solid volume fraction in grid cell with index (i,j,k).
• 756 Lagrangian grid points
Computation of ijk from signed distance (level-set) function to interface:
(determined from )

• Method is second-order accurate.

• Method is stable for  p  f  0.3 .

Instabilities at lower density ratio due to weak coupling of N-S with N-E eqs.
25 26

Accuracy as function of nr iterations IB force Accuracy as function of retraction distance

(retraction distance fixed at 0.25 ∆x) (IB iters fixed at 2)

max
mean

Max error in velocity at Lagrangian Error in normalized permeability

grid points (velocity ideally zero). (equiv. to error in bulk velocity). Darcy scales appr. linearly with retraction
distance. Slope is resolution dependent!

27 28

Approx. 2nd accuracy for retraction of 0.3 ∆x 2 spheres sedimenting in a closed rectangular
(IB iters fixed at 2) container (I)
• Closed cavity of 1x1x4 cm. Diameter spheres = 1.67 mm. Density fluid = 1000
kg/m3, density spheres = 1140 kg/m3. Dynamic viscosity fluid = 1e-3 kg/(ms).
Gravitational acceleration = -9.81 m/s2. IBM specs: 2 IB iters, rd = 0.3 ∆x.

(Error estimate based on

Richardson extrapolation.)

“drafting” stage “kissing” stage

29 30

5
 2 spheres sedimenting in a closed rectangular 2 spheres sedimenting in a closed rectangular
container (II) container (III)

Vertical position as function of time Vertical velocity as function of time

distance between spheres as function of time (lines: different resolutions) (lines: different resolutions);
(lines: different resolutions) maximum sphere Reynolds number ≈ 120.

31 32

2 spheres sedimenting in a closed rectangular Motion of neutrally buoyant sphere

container (IV) in tube Poiseuille flow (I)
• Grid convergence test at t=0.28646 s (at the end of drafting stage) for old • Segre-Silberberg or tubular pinch effect: a neutrally-buoyant sphere released in
method of Uhlmann and for improved method (with retraction). tube Poiseuille flow will move towards an equilibrium radial position of 0.6 times
the tube radius (at finite Reynolds number).
‘old’
‘old’

‘new’ ‘new’

• Round tube mimicked with IBM. For Poiseuille flow

second-order accuracy found when
retraction of Lagrangian grid equal to  0.35x .

% error in vertical position of spheres % error in vertical velocity of spheres

33 34

Motion of neutrally buoyant sphere Modified IBM for particle-laden flows: conclusions
in tube Poiseuille flow (II) [Breugem, 2012]
• Reynolds number based on centerline velocity and tube radius = 50. Diameter sphere =
0.3 * tube radius. Good agreement with FE-simulations of Pan & Glowinski [2005]. • The improved IBM shows second-order accuracy for solid spheres when
the Lagrangian grid is retracted inwards by 0.3x .

• The multidirect forcing scheme improves the approximation of the no-

slip/no-penetration conditions; 2 iterations appear optimal.

• The direct account of the inertia of the fictitious fluid within the
particles enables simulation of neutrally buoyant spheres.

Difference plot with background Radial position sphere as function of

velocity field. time (normalized with centerline z
velocity and tube radius). x
35 36

6
 Normal approach of sphere to wall: Normal approach of sphere to wall:
how well is lubrication film resolved? calculation of lubrication force

without correction with correction

• Gap width: 0.25 * radius  corresponds to 2 grid cells for D/∆x = 16 • For very small gap (< ∆x): lubrication film not well-resolved anymore
• Reynolds number fixed at 0.1 (based on diameter and velocity sphere) • Correction similar to LBM sims of Ladd & Verberg [2001]:
• Large domain (8D x 8D x 8D), periodic b.c.’s in lateral directions, p=0 & free-slip at
sw
top, no-slip & no-penetration at bottom

37 38

Extension IBM with soft-sphere collision model

Soft-sphere collision model [Breugem, 2010]

• Linear-spring / dashpot model in which force on particle A is given by: • Linear-spring / dashpot model in which force on particle A is given by:

F ab,n  kn δn n ab  ηn u ab,n F ab,n  kn δn n ab  ηn u ab,n

(spring) (dashpot) (spring) (dashpot)

39 40

Particle/wall collision: Particle/wall collision from DNS:

coefficient of restitution vs. Stokes number coefficient of restitution vs. Stokes number

D  5.5 mm

exp( 35 / St )
r
r r
 exp( 35 / St ) rdry
rdry rdry
D  4 mm
Vr
r = coefficient of restitution • Steel sphere in silicone oil RV20
Vi D  1.25 mm
• rdry  0.97
m pVi particle inertia
St  ~ • collision time: 8 time steps
6πμR 2 viscous force
St • threshold values in collision model:
St  sw  0.075 , 1  103 ,  2  102
• Experimental data assembled by Legendre et al. [2006]
results sensitive to choice
of threshold values

41 42

7
 Outlook: some remarks References (I)
• IBM has become a very popular and well accepted method for simulating 1. W.-P. Breugem and B.J. Boersma, Direct Numerical Simulations of turbulent flow over a permeable wall using a direct

particle-laden flows and flow in complex geometries in general: and a continuum approach. Phys. Fluids, 17(2), 2005.

• computationally efficient 2. W.-P. Breugem, A combined soft-sphere collision / immersed boundary method for resolved simulations of particulate

• relatively easy to implement flows. In: Proceedings of the ASME-FEDSM 2010 conference, August 1-5 2010, Montreal, Canada.

• robust 3. W.-P. Breugem, A second-order accurate Immersed Boundary Method for fully resolved simulations of particle-laden

• versatile flows. J. Comp. Phys. 13(1):4469-4498, 2012.

4. E.A. Fadlun, R. Verzicco, P. Orlandi and J. Mohd-Yusof, Combined immersed-boundary finite-difference methods for

• Simulations of flows with – O(104 105)

finite-size particles are in reach three-dimensional complex flow simulations. J. Comp. Phys. 161:35-60, 2000.

now, but require efficient parallelization of programming code 5. K. Hoefler and S. Schwarzer, Navier-Stokes simulation with constraint forces: finite-dierence method for particle-laden

ows and complex geometries. Phys. Rev. E 61:7146-7160, 2000.

• First steps towards IBMs extended with physically realistic collision models 6. H.H. Hu, N.A. Patankar and M.Y. Zhu, Direct numerical simulations of fluid-solid systems using the arbitrary

 study of dense particle-laden flow Lagrangian-Eulerian technique. J. Comp. Phys. 169:427-462, 2001.
7. T. Kajishima and S. Takiguchi, Interaction between particle clusters and particle-induced turbulence. Int. J. Heat Fluid

• First steps towards extensions with heat and mass transfer  study of Flow 23:639-646, 2002.

multi-physics problems
43 44

References (II)
9. T. Kempe, S. Schwarz and J. Froehlich, Modelling of spheroidal particles in viscous flows. In: Proceedings of the

Academy Colloquium Immersed Boundary Methods: Current Status and Future Research Directions (KNAW,

Amsterdam, The Netherlands, 15-17 June 2009).

10. A.J.C. Ladd and R. Verberg, Lattice-Boltzmann simulations of particle-fluid suspensions. J. Stat. Phys. 104:1191-1251,

2001.

11. D. Legendre, R. Zenit, C. Daniel and P. Guiraud, A note on the modelling of the bouncing of spherical droplets or

solid spheres on a wall in viscous fluid. Chem. Eng. Sci. 61:3543-3649, 2006.

12. K. Luo, Z. Wang, J. Fan and K. Cen, Full-scale solutions to particle-laden flows: Multidirect forcing and immersed

boundary method, Phys. Rev. E 76, 2007.

13. T.W. Pan and R. Glowinski, Direct simulation of the motion of neutrally buoyant balls in a three-dimensional Poiseuille

flow, Comptes Rendus Mecanique 333:884-895, 2005.

14. C.S. Peskin, The immersed boundary method . Acta Numer. 11:479-517, 2002.

15. A.M. Roma, C.S. Peskin and M.J. Berger, An adaptive version of the immersed boundary method , J. Comp. Phys.

153:509-534, 1999.

16. M. Uhlmann, An immersed boundary method with direct forcing for simulation of particulate flows . J. Comp. Phys.

209:448-476, 2005.

45