Applied Mathematical Modelling 25 (2001) 803±814

www.elsevier.com/locate/apm

Calculating particle-to-wall distances in unstructured

computational ¯uid dynamic models
a,b a,b,*
Mauro Tambasco , David A. Steinman
a
Imaging Research Labs, The John P. Robarts Research Institute, 100 Perth Dr., P.O. Box 5015,
London, Ont., Canada N6A 5K8
b
Department of Medical Biophysics, University of Western Ontario, Medical Sciences Building,
London, Canada N6A 5C1
Received 27 March 2000; received in revised form 31 October 2000; accepted 4 December 2000

Abstract

Knowledge of particle deposition is relevant in biomedical engineering situations such as computational modeling of
aerosols in the lungs and blood particles in diseased arteries. To determine particle deposition distributions, one must
track particles through the ¯ow ®eld, and compute each particle's distance to the wall as it approaches the geometric
surface. For complex geometries, unstructured tetrahedral grids are a powerful tool for discretizing the model, but they
complicate the particle-to-wall distance calculation, especially when non-linear mesh elements are used. In this paper, a
general algorithm for ®nding minimum particle-to-wall distances in complex geometries constructed from unstructured
tetrahedral grids will be presented. The algorithm is validated with a three-dimensional 90° bend geometry, and a
comparison in accuracy is made between the use of linear and quadratic tetrahedral elements to calculate the minimum
particle-to-wall distance. Ó 2001 Elsevier Science Inc. All rights reserved.

Keywords: Particle deposition; Lagrangian particle tracking; Computation ¯uid dynamics; Finite element methods;
Unstructured mesh generation; Boundary layer

1. Introduction

Predictions of particle deposition are relevant in biomedical engineering situations involving

laminar and turbulent ¯ow. Examples include deposition of aerosols in lung bifurcations [1,2],
and deposition of blood particles in stenosed and/or branching arteries [3,4]. In lung bifurcations,
localized enhanced aerosol deposition has been related to clinically observed incidences of
bronchogenic carcinomas [5]. In stenosed arteries, locally enhanced platelet or monocyte depo-
sition on the artery walls may lead to thrombosis and/or plaque development [6±8]. To quantify
the distribution of particle deposition using computational models, each point particle (with
``virtual radius'' rp ) must be tracked through the computational ¯ow ®eld, and deposition is as-
sumed to occur if the particle's trajectory comes within a distance rp to the wall. One approach
that may be used to determine whether a particle has intercepted the wall is to reduce the wall

*
Corresponding author. Tel.: +1-519-685-8300; fax: +1-519-663-3900.
E-mail address: steinman@irus.rri.on.ca (D.A. Steinman).

0307-904X/01/$ - see front matter Ó 2001 Elsevier Science Inc. All rights reserved.
PII: S 0 3 0 7 - 9 0 4 X ( 0 1 ) 0 0 0 1 4 - 2
804 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814

distances by a particle diameter and check to see if the particle remains within the diameter-re-
duced model. This approach was taken in work quantifying aerosol particle deposition in
bronchial bifurcations [1±3,9,10]. In these studies, the bronchial bifurcation geometries were
constructed from a parent branch and two daughter branches that were approximated mathe-
matically by three straight-tube sections joined by a central zone. Although this method works
well in geometries de®ned mathematically by a set of parameters, it would be computationally
intensive to carry out in complex geometries that cannot be divided into a small set of para-
metrically de®ned regions. Another drawback of this method is that the particle-to-wall distances
are not explicitly calculated. As a result, additional coordinate transformation computations are
necessary if one wants to extend the simulation to more complex situations involving, for ex-
ample, particles rolling along the model boundary.
Another approach to determine whether a particle has intercepted the wall is to compute the
distance of the particle to the wall as it approaches the surface geometry. However, due to
computational discretization of the surface, one must address the problem of ®nding the surface
face having a point on its surface that is closest to the particle. For a simple geometry represented
by a structured mesh of hexahedral elements, this problem is trivial since the surface face that is
closest to the particle will simply be a wall face of the element in which the particle resides.
Moreover, if the structured mesh is composed of linear elements the surface geometry will be
approximated by ¯at faces, and in this case a simple analytic formula for computing the minimum
distance between the particle and the ¯at surface may be used [11]. However, for a complex ge-
ometry consisting of sharp bends and/or high curvatures, it may be necessary to compute the
particle's distance to the wall in an unstructured mesh composed of (possibly superlinear) tet-
rahedral elements.
With advances in digital medical image processing and analysis techniques, computational
models constructed for anatomically realistic geometries are becoming more common [12]. These
``realistic'' models usually have very complex geometries that cannot be represented by a set of
mathematical parameters and are dicult to discretize with a structured mesh. In these situations,
an unstructured tetrahedral element mesh permits an alternative approach that easily accom-
modates the complex geometric domains. However, in using an unstructured ± and, in some cases,
nonlinear ± grid [13] to subdivide the geometry, the particle-to-wall distance calculation is no
longer trivial.
In this paper, a method is presented for calculating a particle's distance to complex surfaces
represented by linear and quadratic surface faces derived, respectively, from unstructured meshes
composed of linear and quadratic tetrahedral ®nite elements. A general algorithm for determining
the ®nite element surface face that contains the point closest to the particle position is also
presented. To verify the algorithms and compare linear and quadratic surface approximations for
unstructured meshes, particles are tracked through an analytical velocity ®eld de®ned on a three-
dimensional 90° bend geometry.

2. Minimum particle-to-wall distance

For a given continuous parametric surface,

n* o
S ˆ x …u; v† ˆ …x…u; v†; y…u; v†; z…u; v††

de®ned on an open set D of R2 , the minimum distance between the surface and a point particle at
*
position p not lying on the surface may be found by minimizing
 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814 805

* *
Fig. 1. The set of surface faces, M ˆ fS1 ; S2 ; S3 ; S4 g, form a patch on surface S that ``covers'' point p . In this example, q 1
*
is a point on the surface that is closest to p .

* * 2
d…u; v† ˆ x …u; v† p …1†

with respect to the parameters u and v. However, if the parametric surface is approximated by
many unstructured piecewise continuous surface faces the problem of ®nding the minimum dis-
*
tance of p to the surface is no longer as straightforward as simply evaluating (1). For this discrete
case, one must:
1. Find the set of surface faces that form a surface patch that contains a point that is closest to the
point particle. Such a surface patch will be thought of as ``covering'' the point particle. (Fig. 1).
2. Calculate (using (1)) the minimum distance between the point and each surface face that covers
the point.
3. Determine the smallest of all these distances. This value will be the closest distance between the
point and the surface.
The algorithm presented below for step (1) is a general method that may be applied to any 2-D
surface face type. The minimization method presented for step (2) is also general, however the
detailed calculations of the minimization depend on the nature of the surface faces used to dis-
cretize the surface. An unstructured 3-D mesh is most commonly composed of tetrahedral ele-
ments, and a subdivision of the geometry with these elements creates a surface consisting of
contiguous triangular surface faces. Hence, as an example of the above calculation the case of
computing the distance of a point to a surface that has been subdivided into contiguous triangular
surface faces that bound an arbitrary 3-D tetrahedral mesh geometry will be considered.

2.1. Patch ®nding algorithm

Let Si denote a surface face on the wall of the mesh and let the set M ˆ fS1 ; S2 ; . . . ; Sm g denote
*
the patch of surface faces that cover the point particle p i ˆ …xp ; yp ; zp † within the mesh. In addi-
* *
tion, let q i ˆ …xqi ; yqi ; zqi † denote the point on surface face Si that is closest to p i . As mentioned
above, the ®rst step to calculating the minimum distance of a point to a surface is to ®nd the patch
M. If the point particle's ``virtual radius'' rp is much smaller than the smallest distance between
two connected nodes in the mesh, then the particle-to-wall distance calculation is only performed
if the particle lies within a wall element (i.e., a volume element that has at least one node attached
to the wall). This is usually the case for particle sizes of interest in the bronchial or arterial models.
Hence, for a ®nite element mesh composed of tetrahedral elements and triangular surface faces, M
will be determined by applying the following algorithm:
806 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814

*
Case 1: The wall element within which p resides has only one node g1 on the wall (Fig. 2(a)).
Carry out steps 1±5 below.
*
1. Calculate the point q i on each of the surface faces Si attached to the wall node.
* *
2. Check to see if each point q i lies within its corresponding surface face Si that is, q i 2 Si . The set
*
of surface faces M ˆ fS1 ; S2 ; . . . ; Sm j q i 2 Si g forms a patch covering the point.
*
3. If M is an empty set, that is, if for each surface face Si ; q i 62 Si , then search the mesh for the
*
surface corner wall node gc closest to q i .
4. If gc 6ˆ g1 , then repeat steps 2 and 3.
*
5. If gc ˆ g1 , or one still has that for each surface face Si , q i 62 Si then one assumes the point par-
*
ticle p must lie at the base of a surface enclosing a concave volume (Fig. 3). In this case the

*
Fig. 2. Two-dimensional examples: (a) The wall element ei within which p resides has only one node g1 on the wall.
*
(b) The wall element ej within which p resides has several nodes on the wall that de®ne the surface face g1 g2 g3 .

* *
Fig. 3. The point particle p is closest to surface point q i which lies on the curve at the base of a surface that encloses a
concave volume.
 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814 807

*
point on the surface that gives the minimum distance to p will lie on the curve at the base of the
surface of the concave volume, and this curve will form the patch covering the point. Proceed
with steps 6 and 7.
6. Find nodes attached to gc that de®ne edges on the surface faces Si containing gc .
*
7. Determine the closest distance of p to each edge, and keep the smallest distance.
*
Case 2: The wall element within which p resides has several nodes on the wall that de®ne a surface
face Se (Fig. 2(b)).
*
1. Find the corner node of surface face Se that is closest to the point p .
*
2. Calculate the point q i on each of the surface faces Si attached to this corner node.
3. Repeat steps 3±5 in Case 1.

3. Minimum particle-to-wall distance calculations and surface face checks

3.1. Determining whether a particle is within an element

*
The global coordinates x ˆ …x; y; z† de®ne a computational mesh that subdivides a geometry
*
and forms the physical or global space. To ®nd the location of a point particle p that is being
tracked through this global space one must ®rst locate the element that contains the particle. To
do this, the particle's global coordinates and the nodes of the element in question are mapped to a
natural or local (non-dimensional) coordinate system …a; b; c† (Fig. 4).
Four nodal coordinates are used to de®ne a linear tetrahedral element, and the mapping from
local to global coordinates is given by:
* * * * * * * *
x …a; b; c† ˆ x 1 ‡ … x 2 x 1 †a ‡ …x3 x 1 †b ‡ …x4 x 1 †c; …2†
*
where x i ˆ …xi ; yi ; zi †; i ˆ 1; 2; 3; 4, represent the nodal coordinates. Eq. (2) can be inverted ana-
lytically because it does not have any non-linear terms. In the case of quadratic tetrahedral ele-
ments, 10 nodal coordinates are used to specify an element, and the mapping from local to global
coordinates is given by:
* * * * * * * * * * *
x …a; b; c† ˆ c 1 ‡ c 2 a ‡ c 3 b ‡ c 4 c ‡ c 5 ab ‡ c 6 ac ‡ c 7 bc ‡ c 8 a2 ‡ c 9 b2 ‡ c 10 c2 ; …3†

*
Fig. 4. Linear tetrahedral element mapping from local to global coordinates, x : f…a; b; c† 2 Lg 7! f…x; y; z† 2 Gg, and the
* 1
inverse mapping from global to local coordinates, x : f…x; y; z† 2 Gg 7! f…a; b; c† 2 Lg.
808 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814

where
* *
c 1 ˆ x 10 ;
* * * *
c2 ˆ 4x7 x1 3 x 10 ;
* * * *
c3 ˆ 4x8 x3 3 x 10 ;
* * * *
c4 ˆ 4x9 x5 3 x 10 ;
* * * * *
c5 ˆ 4x2 4x7 4 x 8 ‡ 4 x 10 ;
* * * * *
…4†
c6 ˆ 4x6 4x9 4 x 7 ‡ 4 x 10 ;
* * * * *
c7 ˆ 4x4 4x8 4 x 9 ‡ 4 x 10 ;
* * * *
c8 ˆ 2x1 4 x 7 ‡ 2 x 10 ;
* * * *
c9 ˆ 2x3 4 x 8 ‡ 2 x 10 ;
* * * *
c 10 ˆ 2 x 5 4 x 9 ‡ 2 x 10 ;
*
and x i ˆ …xi ; yi ; zi †; i ˆ 1; . . . ; 10, are the nodal coordinates located on the corners and mid-sides
of the elements (Fig. 5). Eq. (3) is non-linear, and hence must be inverted numerically. Once
Eq. (2) or (3) are inverted and the local coordinates …a; b; c† are evaluated, one may proceed to
determine whether a particle lies within an element.
*
In order for particle p to lie within an element, its local coordinates must satisfy a set of
conditions (which depend on the de®ned global to local coordinate mapping). For a tetrahedral
*
mesh whose mapping is de®ned by Eq. (2) or (3), the following four conditions must be valid for p
to lie within the tetrahedron:
06a61 : 06b61 : 06c61 : 06k ˆ 1 a b c 6 1: …5†
Condition (5) may be used to determine whether a particle has landed in a wall element.

3.2. Minimum distance of a particle to a wall face

Let
n* o
Se ˆ x e …u; v† ˆ …xe …u; v†; ye …u; v†; ze …u; v†† j 0 6 1 u v61 …6†

Fig. 5. (a) A quadratic tetrahedral element shown in the local coordinate system. The corner and mid-side nodes de®ne
the four surface faces as follows: Face (1): nodes 1, 2, 3, 4, 5, 6; Face (2): nodes 1, 7, 10, 8, 3, 2; Face (3): nodes 1, 6, 5, 9,
10, 7; Face (4): nodes 3, 8, 10, 9, 5, 4.
 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814 809

represent a surface face parameterized by local coordinates …u; v†, then the minimum distance
*
between the surface face and a point particle p may be found by minimizing

* * 2
de …u; v† ˆ x e …u; v† p …7†

with respect to …u; v†. That is, the minimum of de occurs when rde ˆ 0, where r ˆ …o=ou; o=ov† is
the gradient operator. In the case of tetrahedral elements, each face may be parameterized in
terms of two of the three local coordinates …a; b; c† (Fig. 5).
To illustrate this minimization procedure let us consider the case of quadratic tetrahedral el-
ements. The parametric equation for a surface of a quadratic tetrahedral element may be obtained
by collapsing the mapping described by Eq. (3) to one of the four faces. For example, for face 3 in
Fig. 5, b ˆ 0, so Eq. (3) reduces
* * * * * * * *
x …a; 0; c† ˆ x e …a; c† ˆ c 1 ‡ c 2 a ‡ c 4 c ‡ c 6 ac ‡ c 8 a2 ‡ c 10 c2 : …8†
*
To ®nd the minimum distance of p to this face, Eq. (8) is substituted into (7) with …u; v† ˆ …a; c†,
and de is minimized with respect to …a; c†.
In general, for a given tetrahedral face that lies on the surface of the geometry, Eq. (3) reduces
to an equation of the form
* * * * * * *
x e …u; v† ˆ c a ‡ c b u ‡ c c v ‡ c d uv ‡ c e u2 ‡ c f v2 ; …9†

where …u; v† represent two of the three local coordinates …a; b; c† that parametrize one of the four
*
tetrahedral faces that may lie on the surface, and the c 's represent the corresponding coecients
from Eq. (4). Next, Eq. (9) is substituted into (7), to get

* * * * * * * 2
de …u; v† ˆ c a ‡ c b u ‡ c c v ‡ c d uv ‡ c e u2 ‡ c f v2 p : …10†

Taking the gradient of Eq. (10) to ®nd the minimum distance results in two coupled third-order
polynomials of the form

A1 u3 ‡ A2 u2 ‡ A3 u ‡ A4 ˆ 0; …11†

and

B1 v3 ‡ B2 v2 ‡ B3 v ‡ B4 ˆ 0; …12†
* * * * * * * * * * * *
where Ai ˆ Ai …v; c a ; c b ; c c ; c d ; c e ; c f †, Bi ˆ Bi …u; c a ; c b ; c c ; c d ; c e ; c f †, and i ˆ 1; 2; 3; 4.
Since Eqs. (11) and (12) cannot be solved analytically, Newton's method for nonlinear
equations is used to ®nd the local coordinates …u; v†. The0 initial local coordinate guess is taken
* *
to be the projection of the particle's local coordinates p (mapping of p given by Eq. (3)) onto
the wall face. With this initial guess, convergence of the local coordinates occurs in one to ®ve
*
iterations and typically in three. The global coordinates of the closest point x e on the wall
*
face to the particle p are found by transforming …u; v† to the global coordinate system using
*
Eq. (9). In order for the global coordinate x e to lie on the global surface face the corre-
sponding local coordinates …u; v† must satisfy condition (5) (that is, 0 6 u 6 1, 0 6 v 6 1, and
0 6 1 u v 6 1). If this condition is satis®ed, the surface face lies above (covers) the point.
* *
Finally, the Euclidean distance from x e to p is calculated using the square root form of
Eq. (7).
810 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814

4. Numerical validation and results

To verify the wall distance calculation codes, and to compare the accuracy of approximating a
curved surface with quadratic as opposed to linear surface faces, the following 3-D geometry
(Fig. 6) was de®ned:

1 6 r 6 5;
…13†
0 6 h 6 p=2;
p
where r ˆ x2 ‡ y 2 , and h ˆ arctan…y=x†. On this 90° bend geometry an analytic velocity ®eld
given by
*

u ux ; uy ; uz ˆ …k sin h; k cos h; 0†; …14†

where k ˆ 0:003, was imposed. The above geometry (13), and velocity ®eld (14) were chosen so
that a particle seeded within the geometry would be passively advected along a constant radial
path (Fig. 7). In this way, the particle's distance from the curved walls is known to be constant
along its trajectory, and deviations in the particle wall distance calculations could be computed
easily. In addition, the above geometry (13), allows one to compare the particle-to-wall distance
calculation for two di€erent surface curvatures (high curvature at r ˆ 1:0, lower curvature at
r ˆ 5:0). The velocity ®eld was chosen to have a constant magnitude in order to yield identical
particle trajectories for the linear and quadratic meshes. In doing this, the wall distance calcu-
lations depend only on the order of surface face (quadratic or linear) used to approximate the
surface geometry, and a meaningful comparison between the two surface face types may be made.
To test the wall distance algorithm, quadratic and linear tetrahedral meshes for the 90° bend
geometry were constructed. The linear meshes were derived from a subdivision of the quadratic
meshes [14], and each linear mesh contained eight times the number of elements of the corre-
sponding quadratic mesh from which it was derived. The total number of nodes in the meshes
varied from 3972 to 113702 with corresponding nodal densities varying from 134 nodes per unit
volume to 3840 nodes per unit volume (Fig. 6).
For each mesh, 25 particles were seeded equally spaced along the z-axis
*
fz ˆ 0:15; 0:20; . . . ; 1:35g above the lower wall at x ˆ …0:0; 1:003; z†, and below the upper wall at

Fig. 6. Sample 90° bend meshes for various nodal densities ranging from the lowest to the highest.
 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814 811

Fig. 7. Cross-sectional slice of 90° bend geometry at z ˆ 0:15. The equally spaced points just above the inner radius,
r ˆ 1:0, and below the outer radius, r ˆ 5:0, illustrate the particle trajectories used in the numerical validation. The
*
velocity ®eld …j v j ˆ 0:003† is also shown.

*
x ˆ …0:0; 4:997; z†. The particle paths were calculated using the fourth-order Runge±Kutta
method to solve the di€erential equation:
*
dx *
ˆ u …r; h†; …15†
dt
*
where u is given by Eq. (14). The time interval used for a given Runge±Kutta integration step was
always chosen such that the particles traveled a ®xed distance of 3  10 3 units with each step
*
along the velocity ®eld. Also, since u is only known at the nodes of a mesh, the velocity of a
particle at each Runge±Kutta integration step was found by spatially interpolating the velocity
®eld from the known nodal velocities.
Since the velocity ®eld is de®ned such that a particle traces a constant radial path, it is known
that particles seeded at radial distances of 1.003 and 4.997 units will trace paths that are 0.003
units away from the lower and upper surfaces, respectively (Fig. 7). In our numerical experiments,
the algorithms described in this paper were used to compute the particle-to-wall distances for the
lower and upper surfaces of the 90° bend. To determine the accuracy of the wall distance cal-
culation, the average percentage wall distance error


Dtrue Dcomp
Derror ˆ  100% …16†
Dtrue
812 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814

was computed. In Eq. (16), Dtrue represents the true distance of a particle to the surface (0.003
units in this case), and Dcomp represents the average computed distance of the particles to one
of the surfaces. To compute Dcomp for a given surface (lower or upper), each particle's distance
to the surface was calculated at equal intervals of 0.03 units as they traveled along their tra-
jectories. Next, for each surface, the average of all the particle-to-wall distances was calculated.
To illustrate the dependence of the wall distance error on mesh resolution, Derror was plotted
as a function of increasing average nodal density (average number of nodes per unit volume)
(Fig. 8).
Due to surface discretization, one expects that the particle-to-wall distances will be overes-
timated for particles traveling above the lower surface and underestimated for particles trav-
eling below the upper surface resulting in positive and negative percentage wall distance errors,
respectively. This is what was have obtained in our results (Fig. 8). From the linear mesh
results in Fig. 8, one also observes that the percentage wall distance error is consistently larger
for particles tracked just above the more highly curved lower surface, than for particles tracked
just below the less curved upper surface. This is also expected since, for uniform element
density, the discrete surface approximation will be better for lower curvatures, than for higher
curvatures. In addition, it is observed, as expected, that in all the cases Derror approaches zero as
the average nodal density increases. Typical average nodal densities used in physiological
models, for example the carotid artery bifurcation [13], vary from about 80 nodes per unit
volume in the less re®ned regions to about 600 nodes per unit volume in the more re®ned
regions. From Fig. 8, one sees that in this range of nodal densities the absolute value of Derror is
between 1% and 2% for quadratic meshes and between 15% and 150% for linear meshes. Hence

Fig. 8. The percentage wall distance error, Derror , as a function of the average nodal density. The error bars represent
standard deviations in Derror .
 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814 813

the average particle-to-wall distance errors are signi®cantly smaller when quadratic surfaces are
used.

5. Discussion and conclusion

Most ®nite element and ®nite volume CFD codes use linear elements, either tetrahedral or
hexahedral, in which case, as mentioned above, it is trivial to compute the distance from a particle
to the wall. FIDAP, a commercial ®nite element CFD code that does support quadratic elements,
is capable only of detecting whether a particle has left the computational domain. In this case the
user may ignore the particle altogether, have the particle deposited at the wall, or have it ricochet
back into the domain [15]. In the present study, an ecient searching technique is presented for
detecting not only when a particle has left the domain, but also when that particle is within a user-
speci®c distance to the wall. This is particularly useful, for example, when tracking particles in the
near wall region referred to as the ``numerical boundary layer'' [9,16], where even small pertur-
bations to the velocity vector (i.e., below convergence level or even within ¯oating point error) can
spuriously deposit a particle travelling along a near-wall streamline. Using the techniques de-
scribed in this paper, such spurious depositions can be eliminated by convecting particles traveling
within a ®xed distance to the wall using only the component of the velocity vector instantaneously
tangent to the wall.
Admittedly, the quadratic particle-to-wall distance calculation involves more operations than
the corresponding calculation for the linear case. However, one can signi®cantly reduce the time
spent in calculating particle-to-wall distances for a given particle trajectory, by reducing the
number of particle-to-wall distance checks. This may be done by allowing a particle within a wall
element to travel ®xed distances, DSi , between successive checks, where i ˆ 1 to the total number
of particle checks for a given trajectory. The distance DSi traveled between successive checks will
depend on the particle-to-wall distance found at each check. For example, if Di is the particle's
distance from the wall as determined from the ith check and rp is the particle radius, then a
particle may safely travel a distance DSi ˆ Di rp in any direction before the next check must be
made. In addition, it is important to note that the particle-to-wall distance calculation is per-
formed for only a small percentage of the total number of particles seeded. For example, in our
experience with particle deposition studies in the carotid artery bifurcation models it has been
observed that for particle sizes of interest (0:5  10 3 to 2:5  10 3 radial units) only 3±4% of the
total number of particles seeded come within a radial distance to the artery wall. Hence, this fact
coupled with the ecient particle-to-wall distance checking scheme described above, leads to CPU
times for particle deposition runs in quadratic meshes being comparable to those using linear
meshes. This is because most of time is spent tracking the particles, and relatively little time is
spent performing the particle-to-wall distance calculation. One may therefore conclude that for
the large accuracy gained and only small computational eciency lost, it is worth retaining
quadratic surfaces to calculate particle-to-wall distances when quadratic ®nite element meshes are
employed.

Acknowledgements

The authors thank Jaques Milner for constructing the ®nite element meshes used in this study.
This work was supported by a Medical Research Council of Canada Group Grant (GR-14973).
The ®rst author was supported by the Ontario Graduate Scholarship in Science and Technology.
814 M. Tambasco, D.A. Steinman / Appl. Math. Modelling 25 (2001) 803±814

The second author was supported by a Research Scholarship from the Heart & Stroke Foun-
dation of Canada.

References

[1] I. Balashazy, W. Hofmann, Particle deposition in airway bifurcations ± I. Inspiratory ¯ow, J. Aerosol Sci. 24 (6)
(1993) 745±772.
[2] I. Balashazy, W. Hofmann, Particle deposition in airway bifurcations ± II. Expiratory ¯ow, J. Aerosol Sci. 24 (6)
(1993) 773±786.
[3] I. Balashazy, W. Hofmann, T. Heistracher, Computation of local enhancement factors for the quanti®cation of
particle deposition patterns in airway bifurcations, J. Aerosol Sci. 30 (2) (1999) 185±203.
[4] D. Bluestein et al., Steady ¯ow in an aneurysm model: correlation between ¯uid dynamics and blood platelet
deposition, J. Biomech. Eng. 118 (1996) 280±286.
[5] R.B. Schlesinger, M. Lippmann, Selective particle deposition and bronchogenic carcinoma, Environ. Res. 15
(1978) 424±431.
[6] D. Bluestein et al., Fluid mechanics of arterial stenosis: relationship to the development of mural thrombus, Ann.
Biomed. Eng. 5 (1997) 344±356.
[7] W.F. Pritchard et al., E€ects of wall shear stress and ¯uid recirculation on the localization of circulating
monocytes in a three-dimensional ¯ow model, J. Biomech. 28 (12) (1995) 1459±1469.
[8] R.T. Schoephoerster et al., E€ects of local geometry and ¯uid dynamics on regional platelet deposition on arti®cial
surfaces, Arteriosclerosis and Thrombosis 13 (12) (1997) 1806±1813.
[9] I. Balashazy, Simulation of particle trajectories in bifurcating tubes, J. Comput. Phys. 110 (1994) 11±22.
[10] T. Heistracher, W. Hofmann, Physiologically realistic models of bronchial airway bifurcations, J. Aerosol Sci. 26
(3) (1995) 497±509.
[11] C. Green®eld, G. Quarini, A Lagrangian simulation of particle deposition in a turbulent boundary layer in the
presence of thermophoresis, Appl. Math. Model. 22 (1998) 759±771.
[12] Q. Long et al., The combination of magnetic resonance angiography and computational ¯uid dynamics: a critical
review, Crit. Rev. Biomed. Eng. 26 (4) (1998) 227±274.
[13] J.S. Milner et al., Hemodynamics of human carotid artery bifurcations: computational studies with models
reconstructed from magnetic resonance imaging of normal subjects, J. Vascular Surg. 28 (1) (1998) 143±156.
[14] A. Liu, B. Joe, Quality local re®nement of tetrahedral meshes based on 8-subtetrahedron subdivision, Math.
Comput. 65 (1996) 1183±1200.
[15] Fluent Inc., Personal communication, 2000.
[16] K. A€eld et al., Fluid mechanics of the stagnation point ¯ow chamber and its platelet deposition, Arti®cial Organs
19 (7) (1995) 597±602.