Direct Measurement of Particle Inertial Migration in Rectangular Microchannels

Kaitlyn Hood,1, Soroush Kahkeshani,2 Dino Di Carlo,2 and Marcus Roper1

1

arXiv:1509.01643v1 [physics.flu-dyn] 5 Sep 2015

Department of Mathematics, UCLA, Los Angeles, CA 90095, USA

Department of Bioengineering, UCLA, Los Angeles, CA 90095, USA
(Dated: September 8, 2015)

Particles traveling at high velocities through microfluidic channels migrate across streamlines due
to inertial lift forces. There are contradictory theories predicting how the inertial lift force depends
on flow parameters, but little experimental evidence by which to validate theory. Here we directly
measure particle migration velocities and show agreement with numerical simulations and a two-term
asymptotic theory that contains no unmeasured parameters. Our data also highlight the previously
unconsidered effect of migration forces that act on particles before they enter the microchannel.
PACS numbers: 47.61.-k, 47.15.Rg, 47.55.-t

Inertial migration; the systematic movement of particles across streamlines due to finite Reynolds number
forces, is increasingly exploited in systems to separate, focus and filter particles and cells [1]. There are many theories for the magnitudes of focusing forces, but there is not
enough detailed experimental evidence, beyond information on particle equilibrium positions, to resolve the the
scaling of the theoretical inertial lift force. Indeed existing theory [25], numerical simulations [68], and indirect
experimental measurements [9] have produced contradictory scalings for the dependence of forces on particle size
and velocity. Here we directly measure inertial migration
velocities by tracking the motion of particles in a rectangular channel over Reynolds numbers ranging from 30
to 180, and find that their measured migration velocities
agree well with existing asymptotic theory [5].
Inertial migration of neutrally buoyant particles was
first reported in flows in circular pipes [10]. In circular pipes, particles are inertially focused into a ring with
radius approximately 0.6 times the pipe radius, with different sized particles focused at different rates and to
slightly different positions [1115]. However, microfluidic
channels are more readily built with a rectangular geometry, in which particles are inertially focused predominantly to either two or four stable equilibrium streamlines [6]. Focusing occurs in two phases, with apparently
well-separated natural time scales: (i) (Fast) particles
quickly focus to a two-dimensional manifold of streamlines and (ii) (Slow) particles travel within the manifold
to one of the focusing streamlines. Two stage focusing
has been directly measured [16], and is consistent with
numerical simulations of the spatial pattern of lift forces
across the channel cross-section [5, 7, 8, 17]. However, direct experimental tests of rates of particle focusing have
been wildly at variance with numerical simulations and
theory [9].
Here we present the first reconciliation of predictive
theory and direct experimental measurement of inertial
migration velocities. Our method allows accurate measurement of particle migration velocities in two dimensions, and via a velocity-based reconstruction method, of

their position and movement in the third dimension. Because it does not require holographic techniques [16], it
provides position readouts for thousands of particles, and
allows particle trajectories, rather than just positions, to
be measured, providing the first direct measurement of
inertial migration velocities. In addition to verifying the
existence of a slow-focusing manifold, our position measurements show that significant inertial focusing occurs
while particles are funneled into the channel, and that
once this contribution is accounted for, inertial migration velocities agree fully with an asymptotic theory [5].
Inertial focusing was measured in a 1.5 cm long PDMS
microchannel fabricated using standard soft lithography
methods [18]. The channel cross-section dimensions were
90m 45m, respectively, with the shortest dimension
identified as the depth (y) dimension (Fig. 1A) and the
longer as the width or lateral dimension (x). Neutrally
buoyant (density 1.05 g/cm3 ) polystyrene spherical particles, prepared at 0.004 volume fraction were dispersed
in a suspending fluid composed of deionized water and
0.002 (wt/vol) triton X-100 and pumped into the channel at controlled flow rate utilizing a syringe pump (Harvard Apparatus, Holliston MA). Four particle diameters
were separately used, d = 4.8, 10, 12, and 19m, along
with four different total flow rates Q = 160, 320, 640,
and 960L/min, corresponding to a range of channel
Reynolds numbers Re = 30 180 and particle Reynolds
numbers Rep = 0.083.2. The schematics of the channel
are displayed in Fig. 1A; in particular, particles enter the
channel through an expanded inlet region whose depth is
constant (45m) but that tapers in width from 1.5 mm to
90 m over a 2.4 mm downstream length. Particle velocities were tracked by high speed imaging (14000 frames
per second, using a Phantom V710 camera) over the first
and last 1 mm of the channel. The microchannel was
viewed from above, although the depth of field allowed
particles to be visualized over the entire channel depth.
For all diameters and velocities, particles were eventually
focused to two streamlines on the midplane x = 0 (Fig.
1B-C).
High speed videography provided only x- and z- (lat-

FIG. 1. Reconstruction of particle focusing velocities and

three dimensional positions in a rectangular channel. (A)
Schematic of the inlet of the channel. (B) Reconstructed
probability density function (PDF) of particle distributions
across the channel cross-section for 10 m particles at Re = 30
shows that within the first 1 mm of the channel particles are
initially focused to two narrow bands of streamlines (high density shown in yellow, low density shown in blue). (C) After
1.5 cm of inertial focusing, the same particles are fully focused
to two streamlines on the channel mid-line. (D) Numerically
computed downstream particle velocity as a function of x and
y positions across the channel cross-section: using this plot
and the particle velocity in the z-direction, we can compute
its yposition. (E) A hybrid PIV-particle tracking scheme
is used to track the particles, green circles show particles in
present frame, magenta circles show the particles in the next
frame. (inset) Template matching (blue circle) allows particle
center to be located with sub-pixel accuracy. (F) Representative trajectories of six particles tracked over 700 s.

eral and streamwise) coordinates for each particle, and

provided no direct measurement of the particle depth (ycoordinate). We measured the x- and z- velocities by hybridizing particle image velocimetry (PIV) and particle
tracking, similar to an algorithm previously developed for
tracking fluorescent organelles [19]. Specifically, we first
use the PIV code MatPIV [20] to develop a vector field
representing the displacements of all particles from one
frame to the next. We then used template matching to
align a template consisting of a single 88 pixel image
of a particle with both the first frame and the next. The
template matching process gives a single correlation value
for every pixel in the image, representing how closely the
template matches the real image if centered at that pixel.
We then use cubic polynomials to interpolate the correlation data and to find, with sub-pixel precision, each
particle location. After locating particles in both frames,
we evaluate the PIV velocity field at the particle centers
in the first frame to predict their location in the next
frame. We then identify the detected particle in the next

frame that is closest to this predicted location. The particle tracking adjustment allows us to correct PIV velocity
fields to obtain sub-pixel accurate particle displacements
(Fig. 1E).
Inertial migration velocities are typically two orders of
magnitude smaller than particle downstream velocities
(3 mm/s in a typical experiment, compared to 0.6 m/s
downstream velocity); in fact the lateral displacements
of particles over a single frame are typically sub-pixel.
To accurately measure the migration velocities, we track
single particles over at least 10 consecutive frames, and
average their total lateral displacement over all of these
frames (Fig. 1F).
Downstream velocities vary across the depth of the
channel, with no slip boundary conditions on the upper and lower walls of the channel and fastest velocities
attained on the midplane of the channel. The Stokes
timescale for a spherical particle, s = 2a2 /9, gives a
measure of the time needed for a particle at any point in
the channel cross-section to accelerate until it is both
force and torque free. For the particles in our study
s = 580 s, is much less than a typical tracking time of
700 s, so particles are effectively force-free and torquefree throughout their migration. We used a finite-element
model built in Comsol Multiphysics (Comsol, Los Angeles) to compute the downstream velocities for force-free
and torque-free finite sized particles [5] located anywhere
within the channel (Fig. 1D). For each x-position there
is a two-to-one mapping of downstream velocity to particle depth, allowing particles to be assigned one of two
ycoordinates that are symmetric about the depth midplane y = 0 (Fig. 1D).
We measured the two dimensional probability density
function (PDF) for the x and y coordinates of particles at the entrance to the microchannel and after 1.5cm
of inertial focusing (Fig. 1B-C). In particular, our data
shows that particles within 1mm of the microchannel entrance are not uniformly dispersed in channel depth but
instead are focused to a thin band of y coordinates (Fig.
1B). Along the channel, particles move laterally within
this band until they are also focused close to the channel center-line, with typically 71% of particles focused
to within 4 m of the focusing streamline after traveling
1.5cm through the microchannel (Fig. 1C).
The thin band on which particles are concentrated in
the first 1 mm of the channel coincides with an asymptotic calculation for the slow manifold, described in more
detail below (Fig. 2A). Since the particles are already focused to their slow manifold, the observed lateral migration within the microchannel represents only the second
phase of particle focusing, i.e. the migration of particles along the slow manifold to their eventual focusing
streamline.
We adapt the asymptotic theory in [5] to predict the
inertial forces on a particle in a channel with the experimental geometry [21]. Since numerical experiments show

0
453015 0 15 30 45
x (m)

vm (mm/s)

y (m)

15

2
1
0
1

0.02

0.04

10

3
453015 0 15 30 45
x (m)

that viscous stresses dominate momentum flux terms

over the entire fluid filled domain, V , we can perform a
regular perturbation expansion in the particle Reynolds
a2
number Re p = H
2 Re, treating the viscous and pressure
stresses as dominant terms, and the inertial stress as a
perturbative correction. By the Lorentz reciprocal theorem the total force on a particle that is constrained from
migrating across streamlines can be written as an integral
that involves two solutions of Stokes equations [22]:
Z
(
u
u u + u
u + u u) dv
(1)
F =
V

where u is the Stokes (Re p = 0) solution for a force-free

and torque-free sphere moving through the microchannel
is the Stokes solution for a particle in quiescent
and u
fluid that is moving without rotation in the lateral di is the undisturbed flow through the
rection. Here, u
channel, and can be computed analytically [23]. To expose the role played by particle size in determining the

lift force,we expanded the two Stokes solutions u and u

a
as a two-term series in H
, the ratio of the particle radius
to the channel depth. Since the particle Stokes time is
very small, we can assume the particle is force free; correspondingly we can divide the inertial migration force
on the particle from (1) by the resistance coefficient for
the particle, also computed to two-terms, and arrive at
a two term asymptotic series for the migration velocity
vm :
a 
U 2 a3 
c5 ,
c
+
4
6H 2
H

(2)

in which parameters c4 and c5 are dimensionless constants that depend only on the location of the particle
(x0 , y0 ). The remaining parameters are defined as follows: is the fluid density, is the fluid viscosity, and U

10

10

10

10

FIG. 2. (A) The slow focusing manifold for 12m particles

and Re = 30; showing agreement between predicted manifold
location (solid red line) and the measured manifold (dashed
black line). (B) Measured migration velocity along the measured manifold (black markers) agrees quantitatively with
the asymptotic theory of [5] adapted for rectangular channels
(blue line).

vm

10

10

2
0

10

< vm > (mm/s)

< vm > (mm/s)

10
a/H

10

10

10

10
Re

FIG. 3. (A) Over the range of measured particle sizes there

is no simple power law for the dependence of migration velocity upon particle size, a. Here we fixed Re = 30 and varied particle diameter (dashed green line: a3 scaling law, blue
line: Equation (2), black circles: measured average migration velocity s.e., red stars: numerical prediction of average
migration velocity). Previous indirect measurements (yellow
squares, from [9]) show a similar trend, but are an order of
magnitude smaller. (B) Average migration velocities scale
like U 2 . Here we fixed particle diameter at d = 12m and
varied the flow rate (blue line: Equation (2), black circles measured average migration velocity s.e.).

is the maximum undisturbed velocity. In particular the

asymptotic result supports that vm U 2 , just as was
found in previous numerical simulations [6]. The asymptotic theory also shows that migration velocity has no
clear power law dependence on particle size. This asymptotic theory is most accurate for small particle sizes and
moderate Reynolds numbers, in practice requiring that
a
H < 0.2, and that channel Reynolds number, Re . 80.
Lateral velocities along the manifold quantitatively
agree with the asymptotic theory. We filtered the measured velocities to select particles that were within a distance 2.25m of the slow manifold. We then binned these
particles into 3 m xintervals, and averaged migration
velocities for particles within the same bin. Experimental
measurements of migration velocity along the slow manifold agree almost exactly with the asymptotic migration
velocity along the theoretical manifold (Fig. 2B).
We performed similar analysis of migration velocities
for particles of different sizes and for different flow velocities. To compute an average migration velocity hvm i, we
computed the mean migration velocity vm for each 3 m
width bin, and then averaged sign(x)vm over all bins;
the sign(x) factor prevents left and right sides of the
channel from canceling, since for x < 0, vm is positive,
while for x > 0, vm is negative.
Average migration velocity hvm i does not have a power
law dependence upon particle size a, but agrees quantitatively with (2). For very small particles, migration velocities increase with a3 scaling law, as predicted asymptotically in [3, 4], but this power law breaks down even
at small particle sizes. Incorporating an extra term in
a
= 0.16
the series expansion produces good fit up to H
in our data. To clarify that there is no conflict between

0.3

15
0.2

y (m)

A
particle pdf

numerical data and experimental data we computed the

migration forces on a particle using the same finite element simulation that was used to extract the downstream velocity of the particle over a range of particle
a
= .04, .08, .17, and .23) that covered the entire
sizes ( H
experimental range. Numerical migration velocities averaged over the slow manifold agreed with experimental
measurements and, over their range of validity, with the
asymptotic series also (Fig. 3A).
Migration velocities scale like U 2 . If particle size is
fixed and the flow rate through the microchannel is var vary in proportion to U ,
ied then since in (1) both u and u
the total migration force, and total migration velocity vm
will scale like U 2 [35]. Our experimental measurements
confirm this scaling (Fig. 3B). Experiments at much
higher Reynolds numbers have shown that additional focusing positions appear in channel corners [24, 25], but
we find no evidence of alternate focusing positions over
the range Re = 30 180.
Our direct measurements of particle migration show
that asymptotic theory adapted for rectangular microchannels can quantitatively predict inertial lift forces on
particles, including their dependence on particle size and
channel velocity. Why have previous indirect measurements of migration velocities, such as [9] contradicted
theory? First we note that our inertial migrational velocities are an order of magnitude larger than previous experiments (Fig. 3B), likely because indirect focusing measurements do not equally weight trajectories across the
entire slow manifold, but rather only the limiting trajectories of particles close to the focusing streamlines. Additionally, previous experiments such as [9] have assumed
that particles are uniformly spread across the microchannel cross-section at the entry. Indeed, when we measured
only particle x-coordinates, we found that particles appeared to be uniformly dispersed (Fig. S1A) at the inlet. However, our reconstruction of particle depth showed
that particles entered the microchannel already focused
in their y-coordinate (Fig. 1B and S1B). Thus, our inchannel measurements showed only the second phase of
inertial migration along a single slow manifold. In particular, pre-focusing makes it impossible to separate fast
and slow phases of focusing in the manner attempted by
[9].
Pre-focusing is due to inertial lift forces acting in the
channel inlet. We can use asymptotic theory to predict these lift forces, which act primarily in the narrowest (y-) dimension where velocity shear is largest.
We model the inlet region as a linear expansion in the
xdirection, with maximum width Wi at z = Li and
minimum width W0 at the opening of the channel at
z = 0, and constant depth H. Conserving flux across
each surface z =constant gives that the characteristic velocity in the inlet region is U (z) = U0 W0 /W (z) where
z
W (z) = W0 Li
(Wi W0 ). Taylor expanding the the
migration velocity around the equilibrium position yeq

0.1

0
15

0
453015 0 15 30 45
x (m)

0.2 0.4 0.6

particle pdf

FIG. 4. Particles enter the microchannel prefocused to a thin

band of y coordinates, so only slow focusing dynamics can
be measured. (A) The particle x-position PDF is nearly uniform at channel entry (blue line) becoming focused after traveling 1.5cm through the channel (red line). (B) However,
the particle y-position PDF is strongly focused both at entry
(blue), and after particles have reached their focusing streamline. (particle diameter d = 10m, channel Reynolds number
Re = 30).

and making the change of variables Y = y yeq results

in the ODE:
dY
Y =
U (z) (z)Y ,
dz

(3)

with (z) = a3 ReU0 W02 0 /6H 4 W (z)2 , where 0 is a dimensionless constant that can be computed from asymptotic theories [3, 5] and for two-dimensional parabolic
flow is approximately 0 = 120 [21]. Integrating equation (S24) yields
Y0
=
Yi

W0
Wi

Li

a3 Re0 W0
.
6H(Wi W0 )

(4)

Using the channel dimensions from this experiment, with

Re = 30 and d = 10m, we find that particles are within
1.5m of the equilibrium position yeq by the end of the inlet region, z = 0, consistent with our measurements (Fig.
4B). However, little focusing occurs in the xdirection,
so that if particle x positions only are measured, as in
[9] particles appear to be uniformly dispersed across the
channel (Fig. 4A).
Can a microchannel inlet be designed to measure fastfocusing dynamics? Equation (4) shows that shorter inlet
regions (smaller values of Li ) lead to less particle prefocusing. To enforce that focusing produces a less than
10% disturbance of particle depths during their passage
through the inlet, i.e. that YY0i > 0.9, we invert (4) and
find that if the particle radius a is measured in microns,
then the maximum inlet length, also in microns, is given
by Li = 2100/a3. In particular for a particle of diameter
10m, the maximum channel inlet is only Li = 17 m.
However, to see fast-focusing dynamics there must also
be fully developed Poiseuille flow at the channel inlet.
The inlet must therefore be longer than the development
length, Ld , required for viscous boundary layers to diffuse from the channel floor and ceiling and to fill the

5
1
entire channel. [24] give Ld = 30
ReH = 45 m at the
lowest Reynolds numbers used in our experiments, exceeding the minimum Li . These competing constraints
make it impossible to design a microchannel inlet to measure fast focusing dynamics. Fast focusing dynamics can
nevertheless be observed in glass capillaries [16] where
inlet regions can be removed, however glass microfluidic
capillaries can not be machined into de novo geometries.
Our direct observations show that there is no conflict
between asymptotic theory and the measured inertial migration velocities of particles in microchannels. However, a theory capable of quantitatively describing these
forces does not produce a simple power law dependence
of migration velocities upon particle size, contributing
to previous contradictions between experiments, numerical data and theory. Additionally, we show that in soft
lithography microchannels, fast focusing dynamics occur
in the channel inlet, causing pre-focusing of particles before they enter the microchannel limiting the control that
can be exerted over particle focusing trajectories.
This work was partly supported by the National Science Foundation through grants DGE-1144087 (to K.
Hood) and DMS-1312543 (to M. Roper). This work was
partly supported by the Office of Naval Research Young
Investigator Program Grant # N000141210847 (to D. Di
Carlo).

kaitlyn.t.hood@gmail.com
[1] H. Amini,
W. Lee,
and D. Di Carlo,
Lab Chip 14, 2739 (2014).
[2] P. G. Saffman, J. Fluid Mech. 22, 385 (1965).
[3] B. P. Ho and L. G. Leal, J. Fluid Mech. 65, 365 (1974).
[4] J.
A.
Schonberg
and
E.
J.
Hinch,
J. Fluid Mech. 203, 517 (1989).
[5] K.
Hood,
S.
Lee,
and
M.
Roper,
J. Fluid Mech. 765, 452 (2015).
[6] D. Di Carlo, J. F. Edd, K. J. Humphry, H. A. Stone, and
M. Toner, Phys. Rev. Lett. 102, 094503 (2009).
[7] C. Prohm and H. Stark, Lab Chip 14, 2115 (2014).
[8] C. Liu, G. Hu, X. Jiang, and J. Sun, Lab Chip 15, 1168
(2015).
[9] J. Zhou and I. Papautsky, Lab Chip 13, 1121 (2013).
[10] G. Segre and A. Silberberg, Nature 189, 209 (1961).
[11] R.
C.
Jeffrey
and
J.
R.
A.
Pearson,
J. Fluid Mech. 22, 721 (1965).
[12] A. Karnis, H. L. Goldsmith,
and S. G. Mason,
Can. J. Chem. Eng. 44, 181 (1966).
[13] M. Tachibana, Rheol. Acta 12, 58 (1973).
[14] W. S. Uijttewaal, E.-J. Nijhof, and R. M. Heethaar,
J. Biomech. 27, 35 (1994).
[15] Y.-S.
Choi
and
S.-J.
Lee,
Microfluid. Nanofluid. 9, 819 (2010).
[16] Y.-S. Choi, K.-W. Seo, and S.-J. Lee, Lab chip 11, 460
(2011).
[17] D. R. Gossett, H. T. K. Tse, J. S. Dudani, K. Goda,
T. A. Woods, S. W. Graves,
and D. Di Carlo,
Small 8, 2757 (2012).

[18] D. C. Duffy, J. C. McDonald, O. J. A. Schueller, and

G. M. Whitesides, Anal. Chem. 70, 4974 (1998).
[19] M. Roper, A. Simonin, P. C. Hickey, A. Leeder, and N. L.
Glass, Proc. Nat. Acad. Sci. U.S.A. 110, 12875 (2013).
[20] J. K. Sveen, Mech. Appl. Math. (2004).
[21] See Supplemental Material at url for detailed derivations
of the migration velocity and the inlet length calculation.
[22] L. G. Leal, Annu. Rev. Fluid Mech. 12, 435 (1980).
[23] T. C. Papanastasiou, G. C. Georgiou, and A. N. Alexandrou, Viscous Fluid Flow (CRC Press, 1999).
[24] A. T. Ciftlik, M. Ettori,
and M. A. M. Gijs,
Small 9, 2764 (2013).
[25] K. Miura, T. Itano,
and M. Sugihara-Seki,
J. Fluid Mech. 749, 320 (2014).
[26] J. Happel and H. Brenner, Low Reynolds number hydrodynamics: with special applications to particulate media, Mechanics of Fluids and Transport Processes, Vol. 1
(Springer, 1982).
[27] H. Lamb, Hydrodynamics (Dover Publications, 1945).
[28] S.
Kim
and
S.
Karrila,
Microhydrodynamics: Principles and Selected Applications,
Butterworth - Heinemann series in chemical engineering
(Dover Publications, 2005).

SUPPLEMENTAL MATERIAL FOR

Direct Measurement of Particle Inertial Migration in Rectangular Microchannels
ASYMPTOTIC CALCULATION OF MIGRATION VELOCITY

Here we provide more detail for the asymptotic calculation of the particles migration velocity (Equation (2) in
the main text). We consider a single spherical particle of radius a suspended in a rectangular channel with aspect
ratio two. The origin is located at the center of the particle, and the particle is allowed to translate downstream with
velocity Up = Up ez and to rotate with angular velocity p . Up and p are chosen so that the particle is totally
torque free and force free in the downstream direction. It will in general experience forces in the x and y direction.
From these forces we can compute the migration velocity for a particle that is totally force and torque free.
First we define the three-dimensional undisturbed flow, u
, which is rectangular channel Poiseuille flow with centerline
velocity U , width W , and height H [S23], and takes the form u
=u
(x, y)ez , where ez is a unit vector pointing in the
downstream direction. The velocity u
and pressure p solve the Stokes equations with boundary condition u
= 0 on
the channel walls. We will also need the Taylor series expansion for u
around the center of the particle:
u(x, y) = + x x + y y + xx x2 + xy xy + yy y 2 + O(x3 , y 3 , xy 2 , x2 y)

(S1)

where we define our origin of coordinates to coincide with the center of the sphere, and with the x-axis aligned with
the width dimension, and y-axis with the depth dimension.
Within the microchannel, the fluid velocity u and pressure p are governed by the dimensionless steady-state 3D
Navier-Stokes Equations (NSE) in the reference frame of the moving particle:
2 u p = 2 Re(
u u + u
u + u u) ,
u = 0,

u = Up + p r u

(S2)
on |r| = 1 ,

u = 0 on the channel walls, and as z .

The dimensionless equations are obtained by scaling lengths by the particle radius a, velocities by the velocity U a/H,
and pressures are scaled by U/H where is the dynamic viscosity.
The dynamics of the flow around a particle in a channel is characterized by two dimensionless parameters: Re =
U H/ gives the ratio of inertial and viscous effects, and = a/H the ratio of particle radius to characteristic channel
width. The parameter ranges of interest for our experiments are 1 Re 200 and 0 < < .5.
The calculation of the inertial lift force FL , and consequently the migration velocity vm is outlined as follows.
First we make a regular perturbation expansion in the particle Reynolds number Re p and use the Lorentz reciprocal
theorem to represent the lift force FL in terms of the perturbation expansion. Then we further expand the terms
in the reciprocal theorem integral as a series expansion in the relative particle size , assumed to be asymptotically
small. As 0, the reciprocal theorem integral must be calculated by dividing it into two subdomains, in which
different terms dominate within the integrand, we call the contributions from these two regions the inner and outer
integrals. We must combine the inner and outer integrals to find the inertial lift force FL . Finally, we show how to
compute particle trajectories and derive the location of the slow focusing manifold as well as the migration velocity
along this manifold.

Perturbation Expansion

For small particle sizes, the particle Reynolds number Re p = 2 Re is a small parameter. While a priori estimates
suggest that inertial stresses will become co-dominant with viscous and pressure forces sufficiently far from the particle
[S4], numerical examination of the terms of (S2) shows that the inertia is subdominant throughout the channel; because
of this, we can treat inertial stresses as a small perturbation to the solution produced by balancing viscous and pressure
stresses across the entire channel cross-section, i.e. perform a regular perturbation expansion in Re p [S5]. We then
expand further in the small parameter , following for this second part, the method proposed by [S3], but using
numerical PDE methods to compute boundary corrections that arise in the solution, and extending the solution to
include higher order terms in , to capture the fact that the particle migration velocity has no simple power law
dependence on particle size.

2
We expand the fluid velocity u, pressure p, particle velocity Up , and particle rotation p in the small parameter
Re p ,
u = u(0) + Re p u(1) + . . . ,

p = p(0) + Re p p(1) + . . . ,

etc.,

(S3)

and substitute into Equation S2 and collect like terms in Re p . The first order velocity and pressure solve the
homogeneous Stokes problem:
2 u(0) p(0) = 0,

u(0) = 0,

on r = 1,
u(0) = Up (0) + p (0) r u
u

(0)

(S4)

= 0 on channel walls and as z ,

while the second order velocity and pressure solve the inhomogeneous Stokes problem:
2 u(1) p(1) = (
u u(0) + u(0)
u + u(0) u(0) ),

u(1) = 0,

u(1) = Up (1) + p (1) r on r = 1,

u

(1)

(S5)

= 0 on channel walls and as z .

This is a regular perturbation expansion: the right hand side of (S5) is the inertial stress associated with the solution
of (S4).
Since only the force on the particle is required, and not the complete velocity field u(1) , we can use the Lorentz
Reciprocal Theorem [S22], to express the inertial lift force FL as an integral containing only a solution of (S4) u(0) :
e FL =



u
u
u(0) + u(0)
u + u(0) u(0) dv.

(S6)

Here to calculate the lift force acting on the particle in the direction e we must integrate the inertial stresses against
, for the same particle moving at unit velocity in the the direction e in a quiescent
the Stokes (Re = 0) solution, u
and an associated pressure p solve the homogenous Stokes problem:
fluid. In other words u

= 0,
2 u
p = 0,
u
= e on r = 1,
u
= 0 on channel walls and as z .
u

(S7)

If the particle size is known this method reduces the problem of calculating the focusing force from solving a nonlinear
. However, this problem
Navier-Stokes problem for u to solving two linear homogenous Stokes problems for u(0) and u
is still numerically time-intensive since the values of Up and p which make the particle drag free and torque free must
be found through optimization, and the dependence of force upon particle size is not made explicit in the solution.

Series Expansion in

We need an asymptotic theory for how the lift force (and thus migration velocity) depend on the size of the particle
as power series in using the method
and its position within the channel. We further expand the velocities u(0) and u
of reflections.
Specifically, we follow [S3, S26] and expand each velocity field as a sum of corrections:
(0)

(0)

(0)

(0)

u(0) = u1 + u2 + u3 + u4 + . . . ,
(0)

(S8)

with similar expansions for p, u

, and p. Here, u1 is the Stokes solution for a particle in unbounded flow (ignoring the
(0)
(0)
(0)
channel wallks), u2 is the Stokes solution with boundary condition u2 = u1 applied on the channel walls (but
(0)
(0)
(0)
ignoring the particle boundaries), and u3 is the unbounded Stokes solution with boundary condition u3 = u2
on the particle surface, etc. Odd terms impose the boundary conditions on the particle, whereas even terms impose
the boundary conditions on the channel walls.

3
(0)

The first term in the series, u1 , is the solution for a particle in unbounded flow, can be found analytically using
the Lambs solution [S27, S28]. Note that we have corrected an error from [S5] in the series below:


2 xx 5
xz
zr
x2 zr 1
y  1
x2
5xr  x
+
e
+
10
e
+
5

35

3
u=
z
x
x
y
z
2r2 r
r
r2
8
3
r2
r2
r2
r4
r3
2 xy  xy
yz
xz
xyzr  1
3 2 ez + 5 2 ex + 5 2 ey 35 4
+
8  r
r
r
r
r3
2
2
2
yz
y zr 1
x  z
y
x
xzr  1
yy 5

ez 3 2 ez + 10 2 ey 35 4
e
+
e

5
+
x
z
8
3
r
r
r
r3
2
r
r
r3 r4
(S9)




2
2
x
y z
xx
xz
zr
x2 zr 1
y
yzr 1
ez 5 2 ez 10 2 ex 5 2 + 35 4

+
ey + ez 5 3
2
r
r
r
r4
8
r
r
r
r
r5


2
yz
xy
xz
xy
xyzr 1
5 2 ex 5 2 ey 5 2 ez + 35 4
+
5
8  r
r
r
r
r
2
2
2
y
yz
zr
y zr 1
yy
ez 5 2 ez 10 2 ey 5 2 + 35 4
.
+
8
r
r
r
r
r5
Likewise, u
1 can be calculated explicitly. Assuming that e = ey , then:


3
3yr 1
yr  1 1
u
1 =
ey 2
.
+
ey + 2
4
r
r
4
r
r3

(S10)

The remaining odd order terms can be found similarly. The even terms in the series expansions of u(0) and u
are
found numerically using a Finite Element Model implemented in Comsol Multiphysics (Comsol, Los Angeles).
Evaluation of the reciprocal theorem integral

Given the Stokes velocities u(0) and u

we can compute the inertial lift force FL up to terms of O(Re p ) using the
reciprocal theorem (S6). As in [S3], it is advantageous to divide the fluid filled domain V into two subdomains, V1
and V2 , where:
V1 = {r V : r }

and V2 = {r V : r }.

(S11)

The intermediate radius is any parameter satisfying 1 1 . Call the corresponding integrals the inner integral
and the outer integral, and identify their contributions to the lift force as FL1 and FL2 , respectively (FL = FL1 +FL2 ).
The division of the integral into inner and outer regions allows one to incorporate varying length scales (a for the
inner region and for the outer region) into our model. Note that, distinct from [S4], inertia remains subdominant
even in the outer region V2 . We will separately consider the contributions from the inner and outer integrals.
Inner Integral

Since the odd terms in the method-of-reflections expansions for u(0) and u
are prescribed on the boundary of
the particle, each gives rise to several terms that contribute to the inner integral FL1 . By contrast, the outer terms
influence Up and p , but do not contribute to the inner integrals directly. Since the odd terms are derived analytically
from the Lambs solution, it follows that FL1 can also be computed analytically. We continue to scale lengths by a,
so that 1 r 1 . The inner integral can be expressed as the following expansion in .:
FL1 = U 2 a2 (h4 2 + h5 3 + . . . ) .

(S12)

In order to calculate the terms h4 and h5 , we sort the terms of the Stokes velocities by leading order in . We refer
the interested reader to [S5] for the details of this calculation. The first order contribution evaluates to zero, h4 = 0.
The next order contribution h5 = (h5,x , h5,y ) is listed below (note that we correct an error from [S5]):
h5,x =

11y xy
19x yy
26x xx
+
+
,
9
12
18

h5,y =

26y yy
11x xy
19y xx
+
+
9
12
18

(S13)

4
Outer Integral

For the outer integral we will consider alternate dimensionless variables, by using the rescaled distance R = r.
This corresponds to using H to non-dimensionalize lengths, rather than a. We call these variables the outer variables,
and we will denote them with uppercase roman letters. In the outer region V2 , we must express our functions in
terms of R and rearrange our functions by order of magnitude in . Then the reciprocal theorem integral takes the
following dimensional form:
Z


2 2
U
U(0) + U(0) U
+ U(0) U(0) dv,
(S14)

fL2 = Um
U
VC

where we have expanded our domain of integration from V2 = {R V : R } to the entire empty channel VC . As
we did for the inner integral, we can write the outer integral as an expansion in .
FL2 = U 2 2 (k4 4 + k5 5 + . . . ) .

(S15)

Likewise, in order to calculate the terms k4 and k5 , we sort the Stokes velocities by leading order in . We refer the
interested reader to [S5] for details of this calculation. We note here that both even terms and odd terms from the
method-of-reflections expansion contribute to the outer integral. In particular, since the even terms are computed
numerically, the outer integral must also be computed numerically, rather than as a closed analytic formula.
Inertial Lift Force

The total lift force is the sum of the inner and outer integrals FL = FL1 + FL2 ; combining the results from inner
and outer expansions, we can then calculate the coefficients of the series expansion to obtain the following scaling law
for the lift force
FL =

U 2 a5
U 2 a4
c4 +
c5 + O(a6 ) .
2
H
H3

(S16)

Recall that is the fluid density, H is the channel height, U is the centerline velocity of the undisturbed flow, a is
the particle radius, and c4 and c5 are dimensionless constants including both analytical and numerically computed
components, and that depend on the location of the particle and the aspect ratio of the rectangular cross-section. A
text file with the values of c4 and c5 for different particle locations is included in the supplemental materials.
Particle Trajectories, Manifold, and Migration Velocity

The method above, which adapts the results from [S5] for a channel with aspect ratio 2, gives only the focusing
force on a particle that is not free to migrate across streamlines. The particles in our experiments are free to migrate
under inertial focusing forces. We find the migration velocity um of a force-free particle by equating the lift force
(S16) with the drag force computed for a particle translating with a general velocity um [S26]. This drag force can be
and u(0) above. To the same order of accuracy
evaluated by the method of reflections, similar to the calculation of u
as our previous calculations this gives us an equation:
6a(um + uim ) = FL .

(S17)

where um is the migration velocity and uim is the first method-of-reflections correction to the flow created by the
moving particle, evaluated at the center of the particle. If the location of the particle is known then uim is linearly
related to the lift force FL , namely there exists a matrix S such that uim S FL . To compute this matrix denote by
u
2,i the method-of-reflections correction for a point force located at the particle center, and pointing in the direction
ei . In this case S = Sij is defined as:
Sij = (
u2,i ej )

(S18)

i.e. the jth component of the solution due to a point force in the i-direction, evaluated at the particle center.
Rearranging the terms above gives for the migration velocity:
um = (1 + S)

FL
.
6a

(S19)

5
the prefactor here represents the tensorial mobility of the particle.
We are interested in how particles travel due to this migration velocity, which can be computed at any point in the
channel. Let X(t) = (X(t), Y (t)) be the location of a given particle in the channel cross-section as a function of time
t. Particles trajectories are found using a Forward Euler solution of the ODE:
dX
= um ,
dt

X(0) = (x0 , y0 ).

(S20)

The slow-focusing manifold is evaluated numerically by advecting particles according to (S20) and finding the curve
which is invariant under (S20). Note that depends on the relative particle size . On the slow-focusing manifold,
the migration velocity satisfies:
a 
U 2 a3 
(S21)
c4 + c5 .
vm
2
6H
H
where the coefficients c4 and c5 include both contributions from the lift force coefficients c4 and c5 restricted to the
focusing manifold, and from the size dependent mobility (S19).
CALCULATION OF PARTICLE FOCUSING PRIOR TO ENTERING THE CHANNEL

Here we derive Equations (3) and (4) in the main paper, which describe to what extent particles are focused in the
channel inlet. Recall that the channel inlet expands linearly in width away from the channel, with maximum width
Wi at z = Li and minimum width (W0 , called W in the main text) at the opening of the channel at z = 0; the
depth of the channel is constant and equal to H (Fig. S1a).
Assuming constant flow rate Q throughout the channel, and self-similar velocity profiles across each cross-section
of the channel inlet, the downstream characteristic velocity in the inlet region takes the form:
U (z) =

U0 W0
,
W (z)

(S22)

where W (z) is the width of the channel inlet, and takes the form
W (z) = W0

z
(Wi W0 ).
Li

(S23)

By Taylor expanding the migration velocity around the equilibrium position yeq , and making the change of variables
Y = y yeq we have the following ODE:
Y = (z)Y,

(S24)

d
vm . Let 0 = (0) be the rate of change of the migration velocity at the beginning of the channel
where (z) = dy
z = 0, then since the migration velocity scales with U 2 we have:

(z) =

W02
0 .
W (z)2

(S25)

10

cL

yeq

10
20
30

(a)

(b)

0.1

0.2 0.3
y/H

0.4

FIG. S1. (a) Diagram of inlet region (not to scale). (b) Plot of cL , the particle lift force cofficient, the slope of the tangent line
at the equilibrium focusing depth is 0 .

6
Upon making a change of variables, Equation (S24) yields:
dY U0 W0
W02
dY dz
0 Y
=
=
dz dt
dz W (z)
W (z)2

(S26)

Integrating and rearranging gives:

Y0
=
Yi

W0
Wi

Li

0 W0
.
U0 (Wi W0 )

(S27)

Recall that 0 is a yderivative of the migration velocity vm , and that vm has the following scaling:
vm =

U02 a3
cL ,
6H 2

(S28)

where cL is a dimensionless constant that is computed analytically and numerically. Then 0 has dimensions:


 3

a ReU0 cL 
U02 a3 cL 
=
0 =
(S29)


6H 2 y y=y0
6H 4
y y =y0


1
L
L
from the derivative c
where 0 = c
y y =y0 is a constant computed numerically. Note that 0 picks up an H
y y =y0
during the change of variables from dimensional coordinate y to dimensionless coordinate y = y/H. Now we can
rewrite the exponent as:
=

a3 Re0 W0
.
6H 4 (Wi W0 )

(S30)

All that remains is to calculate the dimensionless constant:

0 =

cL 
= 120.3.
y y =y0

(S31)

This derivative is calculated numerically using difference quotients (Fig. S1b). The coefficient cL is approximated as
the ycomponent of c4 evaluated along the line x = xeq = 0. Since the width of the channel is very large compared
to the depth in the inlet region, we can neglect the effect of the walls in the xdirection on the lift coefficient cL , i.e.
neglect variation in the flow in the x-direction. Thus we use for cL , the the function computed by [S3] for migration
in Poiseuille flow in a 2D channel.

[S1]
[S2]
[S3]
[S4]
[S5]
[S6]
[S7]
[S8]
[S9]
[S10]
[S11]
[S12]
[S13]
[S14]
[S15]
[S16]
[S17]
[S18]
[S19]
[S20]

kaitlyn.t.hood@gmail.com
H. Amini, W. Lee, and D. Di Carlo, Lab Chip 14, 2739 (2014).
P. G. Saffman, J. Fluid Mech. 22, 385 (1965).
B. P. Ho and L. G. Leal, J. Fluid Mech. 65, 365 (1974).
J. A. Schonberg and E. J. Hinch, J. Fluid Mech. 203, 517 (1989).
K. Hood, S. Lee, and M. Roper, J. Fluid Mech. 765, 452 (2015).
D. Di Carlo, J. F. Edd, K. J. Humphry, H. A. Stone, and M. Toner, Phys. Rev. Lett. 102, 094503 (2009).
C. Prohm and H. Stark, Lab Chip 14, 2115 (2014).
C. Liu, G. Hu, X. Jiang, and J. Sun, Lab Chip 15, 1168 (2015).
J. Zhou and I. Papautsky, Lab Chip 13, 1121 (2013).
G. Segre and A. Silberberg, Nature 189, 209 (1961).
R. C. Jeffrey and J. R. A. Pearson, J. Fluid Mech. 22, 721 (1965).
A. Karnis, H. L. Goldsmith, and S. G. Mason, Can. J. Chem. Eng. 44, 181 (1966).
M. Tachibana, Rheol. Acta 12, 58 (1973).
W. S. Uijttewaal, E.-J. Nijhof, and R. M. Heethaar, J. Biomech. 27, 35 (1994).
Y.-S. Choi and S.-J. Lee, Microfluid. Nanofluid. 9, 819 (2010).
Y.-S. Choi, K.-W. Seo, and S.-J. Lee, Lab chip 11, 460 (2011).
D. R. Gossett, H. T. K. Tse, J. S. Dudani, K. Goda, T. A. Woods, S. W. Graves, and D. Di Carlo, Small 8, 2757 (2012).
D. C. Duffy, J. C. McDonald, O. J. A. Schueller, and G. M. Whitesides, Anal. Chem. 70, 4974 (1998).
M. Roper, A. Simonin, P. C. Hickey, A. Leeder, and N. L. Glass, Proc. Nat. Acad. Sci. U.S.A. 110, 12875 (2013).
J. K. Sveen, Mech. Appl. Math. (2004).

7
[S21]
[S22]
[S23]
[S24]
[S25]
[S26]

See Supplemental Material at url for detailed derivations of the migration velocity and the inlet length calculation.
L. G. Leal, Annu. Rev. Fluid Mech. 12, 435 (1980).
T. C. Papanastasiou, G. C. Georgiou, and A. N. Alexandrou, Viscous Fluid Flow (CRC Press, 1999).
A. T. Ciftlik, M. Ettori, and M. A. M. Gijs, Small 9, 2764 (2013).
K. Miura, T. Itano, and M. Sugihara-Seki, J. Fluid Mech. 749, 320 (2014).
J. Happel and H. Brenner, Low Reynolds number hydrodynamics: with special applications to particulate media, Mechanics
of Fluids and Transport Processes, Vol. 1 (Springer, 1982).
[S27] H. Lamb, Hydrodynamics (Dover Publications, 1945).
[S28] S. Kim and S. Karrila, Microhydrodynamics: Principles and Selected Applications, Butterworth - Heinemann series in
chemical engineering (Dover Publications, 2005).