AC Electrokinetic micropumps

Laurits Højgaard Olesen

+

+

+

+

+

+

+

+

V

0

cos(ωt)

Fluid rolls over

electrode edges

Average ﬂow

is a simple

Couette ﬂow

Channel lid

Electrolyte

Electrodes

Channel substrate

J = σE

Et

q

U

Supervisor: Henrik Bruus

MIC – Department of Micro and Nanotechnology

Technical University of Denmark

31 August 2006

ii

Abstract

The goal of the present thesis has been to investigate ways of controlling

and manipulating ﬂuids and suspensions in microﬂuidic systems. In partic-

ular, we focus on a theoretical description of AC electroosmotic micropumps

with asymmetric electrode arrays, that have recently been demonstrated to

permit fast pumping (velocities ∼ mm/s) with low driving voltage of a few

volt only.

The dynamical description of electrokinetics and electrochemical trans-

port at driving voltages of just a few volt is a theoretically challenging

subject, and therefore simplifying assumptions such as the Debye–H¨ uckel

approximation or linear response for weak applied ﬁeld have often been em-

ployed in the literature.

We extend previous linear theory for AC electroosmotic ﬂow into the

“weakly nonlinear” regime by accounting for nonlinear capacitance of the

Debye screening layer, and also consider the eﬀect of Faradaic current in-

jection from electrochemical electrode reactions. This allows us to explain

why the frequency of maximum pumping is sometimes shifted down when

the driving voltage is increased, but neither the linear nor weakly nonlinear

models are able to account for the reversal of the pumping direction that

has been observed experimentally.

Therefore we also study the “strongly nonlinear” regime where classical

circuit models with uniform bulk electrolyte concentration break down. We

extend recent theoretical work in this regime, by accounting for dynamics

in the diﬀusion layer developing when an AC voltage with driving frequency

around the inverse RC time is applied, and by considering ﬂuid motion and

convection of ions. Moreover, we attempt to include existing theory for

double layers driven out of quasiequilibrium from problems of DC Faradaic

conduction at “very large” (but experimentally relevant) voltage into our

dynamical model.

We solve the coupled electrohydrodynamical problem numerically for

experimental micropump geometries and display contributions to the net

pumping velocity from the diﬀerent ﬂow sources in the model: Our results

indicate that both bulk electroconvection and electroosmotic ﬂow of the

“second kind” may contribute a signiﬁcant fraction of the overall induced

ﬂow already at driving voltages of a few volt. However, a rigorous account

iv ABSTRACT

for nonequilibrium double layers in a dynamical setting would be necessary

to justify several ad hoc assumptions in our model.

Finally, we investigate diﬀerent ways of breaking the symmetry of an

electrode array and determine in each case the optimal device geometry to

maximize the pumping velocity at a given low driving voltage as described

by the simple linear theory.

Resum´e

Form˚alet med nærværende afhandling har været at undersøge m˚ader at

kontrollere og manipulere væsker og partikler i mikroﬂuide systemer. Især

fokuserer vi p˚a en teoretisk beskrivelse af AC elektroosmotiske mikropumper

med asymmetriske sæt af elektroder, der tillader hurtig pumpning (hastig-

heder ∼ mm/s) ved lav spænding p˚a nogle f˚a volt.

Dynamisk beskrivelse af elektrokinetik og elektrokemisk transport ved

spændinger p˚a bare f˚a volt er teoretisk udfordrende, og derfor benyttes i

litteraturen ofte forudsætninger s˚asom Debye–H¨ uckel approksimationen eller

lineært respons for svage p˚atrykte felter.

Vi udvider den eksisterende lineære teori for AC elektroosmose til at

gælde i det “svagt ulineære” regime, idet vi tager højde for ulineær ka-

pacitans i Debye skærmningslaget, og undersøger ogs˚a eﬀekten af faradaisk

strøm fra elektrokemiske elektrode reaktioner. Dermed kan vi forklare hvor-

for den frekvens hvor pumpehastigheden maksimeres i nogle tilfælde falder

n˚ar spændingen øges, men hverken den lineære eller svagt ulineære teori kan

forklare hvorfor pumperetningen vender, s˚aledes som det ses eksperimentelt.

Derfor studerer vi ogs˚a det “stærkt ulineære” regime hvor de klassiske

kredsløbsmodeller bryder sammen, idet vi udvider tidligere teoretisk arbej-

de i dette regime ved at tage højde for dynamik i diﬀusionslaget der dannes

n˚ar en vekselspænding med frekvens omkring den inverse RC tid p˚atrykkes,

og ved at medregne væskestrømning og deraf følgende konvektion af ioner.

Desuden forsøger vi at inkludere eksisterende teori for elektriske dobbelt-

lag drevet ud af kvasiligevægt fra problemer med DC faradaisk strøm ved

“meget høj” spænding i vores dynamiske model.

Vi løser det koblede elektrohydrodynamiske problem numerisk for ekspe-

rimentelt relevante mikropumpe geometrier, og bestemmer bidrag til pumpe-

hastigheden fra de forskellige strømningsled i modellen: Vores resultater in-

dikerer at b˚ade elektrokonvektion og elektroosmose af den “anden slags”

kan bidrage med en betydelig del af den samlede væskestrøm allerede ved

spændinger p˚a nogle f˚a volt. Imidlertid er en mere grundig behandling

af dobbeltlag drevet dynamisk ud af kvasiligevægt nødvendig for at ret-

færdiggøre ﬂere ad hoc antagelser i vores model.

Endelig undersøger vi forskellige m˚ader at bryde symmetrien af et sæt

af elektroder, og bestemmer i hvert tilfælde den geometri der maksimerer

vi RESUME

pumpehastigheden for en given lav spænding ved en beskrivelse med den

simple lineære teori.

Preface

The present thesis is submitted in partial fulﬁlment of the requirements for

the PhD degree at the Technical University of Denmark (DTU). The work

has been carried out at MIC – Department of Micro and Nanotechnology

in the Microﬂuidic Theory and Simulation Group (MIFTS) from September

2003 to August 2006. In the period January to May 2005 I worked at the

Laboratoire de Physico-Chimie Th´eorique at Ecole Sup´erieure de Physique

et de Chimie Industrielles (ESPCI) in Paris. The project was supervised by

Prof. Henrik Bruus, and while abroad by Prof. Armand Ajdari.

I would like to thank Henrik Bruus for good spirits and dedicated eﬀort

as supervisor, and for allowing me great freedom to pursue my many wild

ideas. I would also like to thank Armand Ajdari for welcoming me at ESPCI,

and for providing me invaluable insight on electrokinetic phenomena; I am

grateful for many thoughtful comments on my work also after I left Paris.

Vincent Studer digged up old impedance measurements for me to validate

my theoretical model. I also thank Patrick Tabeling and his group at ESPCI

for welcoming me and for exposing me to some of their interesting work and

problems.

Much of the work I did in the ﬁnal part of my PhD studies depends

on a theoretical model developed by Martin Bazant and Kevin Chu, with

whom I had many helpful discussions. Also I would like to thank Nicolas

Green for inviting me to see his lab in Southampton, and for sharing his long

experience with AC electrokinetic experiments and recent observations.

My collaboration with Fridolin Okkels on topology optimization was an

exciting sidetrack on the route I otherwise pursued; Misha Gregersen never

lost her temper when I once and again came up with just another interesting

measurement to do. I thank the people at MIC and in the MIFTS cave in

particular for providing a very nice and inspiring working environment.

Finally, I thank my girlfriend Hanne for always being there and for re-

minding me that there is life beyond work.

Laurits Højgaard Olesen

MIC – Department of Micro and Nanotechnology

Technical University of Denmark

31 August 2006

viii PREFACE

Contents

List of Figures xii

List of Tables xiii

List of Symbols xv

1 Introduction 1

1.1 Microﬂuidics and the lab-on-a-chip . . . . . . . . . . . . . . . 1

1.2 AC electrokinetics . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . 5

1.4 Publications during the PhD studies . . . . . . . . . . . . . . 7

2 Electrokinetic theory 9

2.1 Classical theory . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1.1 Electrochemical transport . . . . . . . . . . . . . . . . 10

2.1.2 Electrical double layer . . . . . . . . . . . . . . . . . . 11

2.1.3 Electrode reactions . . . . . . . . . . . . . . . . . . . . 15

2.1.4 Electroosmosis . . . . . . . . . . . . . . . . . . . . . . 18

2.2 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.1 Dimensionless form . . . . . . . . . . . . . . . . . . . . 22

2.2.2 Electroneutral bulk transport . . . . . . . . . . . . . . 24

2.2.3 Diﬀusion layer . . . . . . . . . . . . . . . . . . . . . . 26

2.2.4 Quasiequilibrium double layer . . . . . . . . . . . . . . 29

2.2.5 Surface conservation laws . . . . . . . . . . . . . . . . 30

2.2.6 Linear model . . . . . . . . . . . . . . . . . . . . . . . 33

2.2.7 Weakly nonlinear model . . . . . . . . . . . . . . . . . 37

2.2.8 Strongly nonlinear model . . . . . . . . . . . . . . . . 39

2.2.9 Beyond quasiequilibrium . . . . . . . . . . . . . . . . . 42

2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3 Linear and weakly nonlinear analysis 51

3.1 Linear analysis at low voltage . . . . . . . . . . . . . . . . . . 53

3.1.1 Capacitive charging . . . . . . . . . . . . . . . . . . . 54

3.1.2 Faradaic current injection . . . . . . . . . . . . . . . . 57

x CONTENTS

3.1.3 Linear response at nonzero intrinsic zeta potential . . 59

3.2 Weakly nonlinear analysis . . . . . . . . . . . . . . . . . . . . 59

3.2.1 Nonlinear surface capacitance . . . . . . . . . . . . . . 60

3.2.2 Nonlinear Faradaic current injection . . . . . . . . . . 61

3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Strongly nonlinear analysis 65

4.1 1D model problem . . . . . . . . . . . . . . . . . . . . . . . . 65

4.1.1 Dynamical solutions . . . . . . . . . . . . . . . . . . . 67

4.1.2 Comparison to a full numerical solution of the

Poisson–Nernst–Planck equations . . . . . . . . . . . . 70

4.2 2D model problems . . . . . . . . . . . . . . . . . . . . . . . . 78

4.2.1 Symmetric electrode array . . . . . . . . . . . . . . . . 78

4.2.2 Asymmetric electrode array . . . . . . . . . . . . . . . 82

4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5 Optimization 91

5.1 Planar asymmetric electrode pairs . . . . . . . . . . . . . . . 93

5.2 Planar electrodes with nonuniform coating . . . . . . . . . . . 94

5.3 3D structured electrodes with regular shape . . . . . . . . . . 100

5.4 3D structured electrodes with arbitrary shape . . . . . . . . . 105

5.5 Thermodynamic eﬃciency . . . . . . . . . . . . . . . . . . . . 110

5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6 Conclusion and outlook 113

A FEMLAB code 117

A.1 Weakly nonlinear ﬁnite element model . . . . . . . . . . . . . 117

A.2 Strongly nonlinear ﬁnite element model . . . . . . . . . . . . 122

A.3 Periodic integro-diﬀerential equations in time . . . . . . . . . 127

B Paper published in Phys. Rev. E 131

C Paper published in Int. J. Numer. Meth. Eng. 149

List of Figures

2.1 Schematic picture of electrical double layer . . . . . . . . . . 11

2.2 Relative increase in anion concentration at positive electrode 14

2.3 Free energy landscape around transition state . . . . . . . . . 16

2.4 Schematic picture of nested boundary layers . . . . . . . . . . 21

2.5 Equivalent circuit diagram for electrochemical cell . . . . . . 35

3.1 Sketch of planar pump with asymmetric pairs of electrodes . 52

3.2 Pumping velocity vs. frequency . . . . . . . . . . . . . . . . . 54

3.3 Streamlines on planar asymmetric electrode array . . . . . . . 55

3.4 Bode plot of system impedance . . . . . . . . . . . . . . . . . 56

3.5 Pumping velocity vs. frequency and charge transfer resistance 58

3.6 Pumping velocity vs. frequency and voltage . . . . . . . . . . 60

3.7 Pumping velocity vs. frequency and voltage . . . . . . . . . . 61

3.8 Time average electric ﬁeld vs. frequency and voltage . . . . . 62

4.1 Sketch of 1D model geometry . . . . . . . . . . . . . . . . . . 66

4.2 Dynamical solutions for potential vs. time . . . . . . . . . . . 68

4.3 Surface concentration and zeta potential vs. voltage . . . . . 69

4.4 Concentration proﬁle in the diﬀusion layer . . . . . . . . . . . 71

4.5 Individual ion concentrations . . . . . . . . . . . . . . . . . . 72

4.6 Individual ion concentrations on loglog scale . . . . . . . . . . 74

4.7 Electric ﬁeld on loglog scale . . . . . . . . . . . . . . . . . . . 75

4.8 Thickness of space charge layer vs. time . . . . . . . . . . . . 77

4.9 Sketch of symmetric electrode array . . . . . . . . . . . . . . 78

4.10 Slip velocity distribution on electrode . . . . . . . . . . . . . 79

4.11 Salt concentration, ﬂux, and bulk electroconvection . . . . . . 79

4.12 Salt concentration, ﬂux, and bulk electroconvection . . . . . . 80

4.13 Slip velocity distribution on electrode . . . . . . . . . . . . . 81

4.14 Total angular momentum vs. voltage . . . . . . . . . . . . . . 82

4.15 Development of salt concentration proﬁle . . . . . . . . . . . 83

4.16 Pumping velocity vs. voltage . . . . . . . . . . . . . . . . . . 84

4.17 Pumping velocity vs. frequency and voltage . . . . . . . . . . 85

4.18 Pumping velocity vs. frequency and voltage . . . . . . . . . . 86

xii LIST OF FIGURES

4.19 Pumping velocity vs. voltage . . . . . . . . . . . . . . . . . . 87

5.1 Optimal design of planar asymmetric electrode array . . . . . 94

5.2 Sketch of planar electrode array with asymmetric coating . . 95

5.3 Optimal nonuniform coating . . . . . . . . . . . . . . . . . . . 98

5.4 Piecewise uniform coating . . . . . . . . . . . . . . . . . . . . 99

5.5 Sketch 3D geometry with stepped electrodes . . . . . . . . . . 101

5.6 Flowrate vs. overall channel height . . . . . . . . . . . . . . . 102

5.7 Pumping velocity vs. electrode dimensions . . . . . . . . . . . 103

5.8 Pumping velocity vs. electrode dimensions . . . . . . . . . . . 104

5.9 3D geometries with stepped and grooved electrodes . . . . . . 105

5.10 Optimized arbitrary 3D electrode shapes . . . . . . . . . . . . 108

5.11 Optimized arbitrary 3D electrode shapes . . . . . . . . . . . . 109

A.1 Output from weakly nonlinear ﬁnite element model . . . . . . 121

List of Tables

2.1 Typical experimental values and numbers for a 1 mM KCl

solution in a device with 10 µm characteristic length . . . . . 24

xiv LIST OF TABLES

List of Symbols

To reduce the notational overload we denote dimensionless quantities by the

same symbol as their dimensional counterparts; it should be obvious from

the context whether a given quantity is dimensionless or not.

Symbol Description Unit

c (Molar) concentration (mol) m

−3

D Diﬀusivity m

2

s

−1

D

0

Average diﬀusivity m

2

s

−1

D

a

Ambipolar diﬀusion constant m

2

s

−1

∆D Diﬀusivity diﬀerence m

2

s

−1

e Elementary charge 1.602 10

−19

C

E Electric ﬁeld V m

−1

f Frequency Hz [ ≡ s

−1

]

f

a

Excess ambipolar ﬂux (mol) s

−1

m

−2

F Flux density (mol) s

−1

m

−2

F

a

Ambipolar ﬂux (mol) s

−1

m

−2

j

0

Exchange current A m

−2

j

ext

Faradaic current injection A m

−2

j Surface excess current density A m

−1

J Current density A m

−2

k

B

Boltzmann constant 1.381 10

−23

J K

−1

ℓ

0

Geometrical length scale m

n Surface outward normal –

O Oxidized species –

p Pressure Pa [ ≡ N m

−2

]

q Surface charge density C m

−2

Q Volume ﬂow rate m

3

s

−1

r Position vector m

R Reduced species –

R

ct

Charge transfer resistance Ω m

2

xvi LIST OF SYMBOLS

Symbol Description Unit

R

mt

Warburg constant Ω m

2

s

−1/2

t Time s

T Period in time s

T Temperature K

u Fluid velocity m s

−1

U Pumping velocity m s

−1

V

0

Driving voltage V

V

ext

External potential on electrode V

w Surface excess concentration (mol) m

−2

z Ion valence –

γ Excess concentration (mol) m

−3

Γ

i

Surface excess ﬂux density (mol) s

−1

m

−1

δ Surface capacitance ratio –

ǫ Permittivity F m

−1

ǫ

0

Permittivity of vacuum 8.85410

−12

F m

−1

ε Dimensionless Debye length –

ζ Zeta potential V

ζ

eq

Intrinsic zeta potential V

ς Diﬀusion layer potential drop V

η Dynamic viscosity Pa s

ϑ Overpotential V

λ

D

Debye length m

µ Electrochemical potential J

ν Kinematic viscosity m

2

s

−1

ρ Charge density C m

−3

ρ

m

Mass density kg m

−3

σ Conductivity S m

−1

σ Stress tensor Pa

τ

0

RC time s

τ

D

Debye time s

φ Electrostatic potential V

Φ Space-charge layer potential drop V

ψ Excess potential V

ω Angular frequency rad s

−1

Ω Computational domain –

LIST OF SYMBOLS xvii

Accent Description

a

∗

Equilibrium bulk reference value

¯ a Bulk variable in asymptotic expansion

˜ a Debye layer variable in asymptotic expansion

ˇ a Diﬀusion layer variable in asymptotic expansion

˘ a Space-charge layer variable in asymptotic expansion

ˆ a Complex variable

ˆ a

∗

Complex conjugate

xviii LIST OF SYMBOLS

Chapter 1

Introduction

1.1 Microﬂuidics and the lab-on-a-chip

Microﬂuidics refers to hydrodynamics and ﬂow manipulation in systems and

devices with characteristic dimensions less than a millimeter. Of course,

such systems have long been studied in biology and colloid science, but

recently there has been an enormous increase of attention on the subject, for

several reasons: First, the fabrication techniques originally developed by the

microelectronics industry now allow accurate design of microﬂuidic circuits,

and is being used to construct microelectromechanical systems (MEMS)

and microﬂuidic devices, e.g., inkjet printer heads. Secondly, advances in

biology and biotechnology require manipulation of particles or living cells

and chemical detection of very small quantities. Also in more conventional

analytical chemistry, miniaturization has the advantage of low sample and

reagent consumption. The ultimate goal is development of “lab-on-a-chip”

systems with integrated pumps, reagent dispensers, mixers, separators, and

detection units, that would automatize the complete analysis from sample to

electrical read-out [1, 2]. Fabrication from polymer materials would lower

the production cost and allow for cheap and portable devices to perform

simple analyses. Equally important, miniaturization opens for new types of

analysis that would not be possible in a macrosystem, because mass diﬀusion

and heat transfer is very rapid on the microscale and the relative importance

of surface to volume forces increases as the system is scaled down [3, 4].

An important question when designing such microﬂuidic lab-on-a-chip

systems is how to get the sample into the chip and manipulate it around.

For a simple disposable device an elegant approach is the use of capillary

forces, but for more complex systems it is necessary to use active pumping

to control the ﬂow. Depending on the application there can be very diﬀerent

requirements for such a pump: In a laboratory test setup it is often suﬃcient

to use external syringe pumps, and multilayer soft lithography provides a

very ﬂexible platform for microﬂuidic plumbing with valves and peristaltic

2 Introduction

pumps powered by external pressure transducers [5]. Devices for drug deliv-

ery need to be extremely fail-safe, and especially if designed for implantation

they should have very low power consumption. On the other hand, for dis-

posable devices the production costs and ease of integration should be much

more important. An integrated pump design that has been subject to inten-

sive research over the years is the (piezo-actuated) membrane pump, where

a membrane is displaced to create pulsating ﬂow that is rectiﬁed by valves

at the inlet and outlet to the pumping chamber [6]. However, moving parts

makes the fabrication delicate, and many valve designs are prone to failure

if there are particles or bubbles in the liquid. Micropumps based on elec-

troosmosis have no moving parts which makes them easy to fabricate, but

their performance depends strongly on the pH and chemical composition

of the pumping liquid. The basic design where an electric ﬁeld is simply

applied down a section of the microchannel can in principle pump the liquid

in a plug ﬂow with velocities of a few mm/s in free ﬂow conditions, but

only generate very small pressures so the thermodynamic (energy) eﬃciency

is low [7]. This can be overcome by introducing a micro- or nanoporous

medium, e.g., a sintered glass frit, into the channel which increases the hy-

draulic resistance dramatically but does not aﬀect the electroosmotic ﬂow;

in this way both a decent ﬂow rate and pressure can be generated at a ther-

modynamic eﬃciency of a few percent [7, 8, 9, 10, 11, 12]. Another problem

is that in order to maintain an electric ﬁeld down the pumping channel, a

DC Faradaic current has to be drawn between electrodes submerged in the

pumping liquid, which in many cases generates bubbles due to electrolysis.

This can be avoided by separating the electrodes from the pumping channel

by means of ion exchange membranes [11, 12] or a gel salt bridge [13] at the

expense of increased fabrication complexity.

An interesting alternative to standard DC electroosmotic pumps is the

use of AC electroosmosis with an asymmetric array of microelectrodes to

generate pumping [14]. This requires driving the system with an AC voltage

in the kHz range, but allows pumping velocities of the order of mm/s to

be attained with just a few volt applied between the electrodes [15]. The

fabrication is simple and only requires an array of interdigitated electrodes

to be deposited on the channel substrate, which can be achieved by a single

photolithography step followed by metal deposition and lift-oﬀ. Similar to

the basic DC electroosmotic pump design, the AC electroosmotic micropump

is very poor at generating pressure, but it allows local control of the ﬂow

and could be used, e.g., to circulate sample between two chambers on a chip.

Moreover, the pumping velocity is inversely proportional to the size of the

electrodes, i.e., the eﬃciency increases with miniaturization.

1.2 AC electrokinetics 3

1.2 AC electrokinetics

In electrochemistry impedance spectroscopy, where a small harmonic modu-

lation is added to a (slowly varying) DC voltage and the current response of

the electrochemical cell is measured, is a standard and very powerful tool for

determining reaction rate constants and also for probing the surface prop-

erties at the electrode [16]. Usually one stirs the solution to ensure uniform

bulk reagent concentrations and simplify the experiment.

In colloid science the focus is most often on colloidal particles with a

certain intrinsic surface charge and zeta potential, and one studies, e.g., the

electrophoretic velocity of the particles in an applied DC ﬁeld, or investigate

their inﬂuence on the dielectric response when an AC ﬁeld is applied [17, 18].

An AC ﬁeld also gives rise to electrokinetic ﬂow around the particles but no

net particle motion (if the ﬁeld is uniform and the particle is spherical). The

ﬂow does, however, induce interaction between particles and inﬂuence their

tendency to agglomerate [19, 20]. Interestingly, Squires and Bazant predict

that a nonspherical egg-shaped metal colloid will align with the applied ﬁeld

and “swim” due to induced-charge electroosmosis [21].

Ramos and co-workers ﬁrst observed electroosmotic ﬂow over pairs of

microelectrodes when driven with an AC voltage in the kHz range, around

the inverse RC time form the system [22, 23, 24]. Ajdari predicted that

the same mechanism would give rise to pumping on an asymmetric elec-

trode array [14], which was soon after demonstrated experimentally by

Brown et al. [25] and later by several other groups [15, 26, 27, 28, 29, 30,

31, 32]. While the general symmetry arguments of Ajdari predicts pumping

in some direction for any locally asymmetric geometry, Ramos et al. did a

detailed theoretical study with realistic boundary conditions for the partic-

ular planar geometry used by Brown et al., and obtained the same direction

of pumping as observed experimentally and similar streamline patterns [33].

They were also able to match quantitatively the experimental pumping ve-

locity and frequency of maximal pumping by assuming a compact Stern

layer on the electrodes with roughly the same capacitance as the diﬀuse

Debye layer. In some sense this good agreement is surprising because their

theory is based on the Debye–H¨ uckel approximation and linearization which

is strictly valid only at low driving voltages V

0

< k

B

T/e ≃ 25 mV, while

experiments are typically performed with at least a few volt.

Later, Studer et al. observed reversal of the pumping direction when

driving the system at around 50 kHz and above 5 V

rms

, well above the fre-

quency for maximal pumping in the “forward” direction around 1 kHz [15].

Applying the same voltage at 1 kHz, bubble formation and electrode degra-

dation was observed because they used bare platinum electrodes that sup-

port electrochemical electrode reactions and Faradaic current. Gregersen

found reversal of the pumping direction at 2.5 kHz and below 1.2 V

rms

,

changing to forward pumping above 1.2 V

rms

[32]; this was found for a dilute

4 Introduction

KCl working electrolyte on bare platinum electrodes with the same device

geometry as used by Studer. Ramos and co-workers found reversal of the

pumping direction on a travelling wave device at 1 kHz and above 2 V [34],

and on an asymmetric electrode array at 10 kHz and above 5 V

rms

[31]. They

used titanium coated gold electrodes that do not support Faradaic current

due to the formation of a thin natural titanium oxide layer. We note that

while electrode reactions leading to bubble formation is a major challenge

for DC electroosmosis, the same needs not be the case in AC because the

reactions could be run reversibly [35]. However, our guess is that many

biochemists would be uncomfortable with uncontrolled electrode reactions

taking place in same the channel where a delicate sample is to be pumped

through.

Most previous work on AC or induced-charge electroosmosis has been

based on the Debye–H¨ uckel approximation with linearization of the electrical

problem [19, 23, 33, 36], which we denote “linear theory” even though the

induced time-average electroosmotic ﬂow scales as the square of the driving

voltage. However, this theory is unable to account for the reversal of the

pumping direction observed experimentally. Therefore it has been a major

focus of our work to extend the existing theory to be able to describe these

phenomena, and also to be able to compare the theory more quantitatively

with experiments as a function of both driving voltage and frequency.

Bazant, Thornton, and Ajdari recently studied diﬀuse-charge dynamics

in a simple 1D model system analytically by means of matched asymptotic

expansions [37]. In particular they went beyond the linear theory to an-

alyze the “weakly nonlinear” regime, where the capacitance of the Debye

screening layer at the electrodes becomes nonlinear, but the bulk electrolyte

concentration remains uniform and constant at leading order, with pertur-

bations showing up only as higher order corrections. They also derived a

dimensionless criterion for the zeta potential at which the weakly nonlinear

asymptotics break down, and argued that for weakly nonlinear dynamics

to hold, the driving voltage V

0

cannot greatly exceed the thermal voltage

k

B

T/ze. Further, they wrote the governing equations for the “strongly

nonlinear” regime, where the assumption of uniform bulk electrolyte con-

centration breaks down but bulk charge neutrality remains valid; however,

because that problem is not analytically tractable they did not attempt to

solve it. Chu extended this work to three dimensions by studying a spher-

ical metal particle, and investigated both the weakly nonlinear dynamical

response to a suddenly applied DC ﬁeld and the steady-state solution to the

strongly nonlinear problem arising at larger applied DC ﬁeld [38, 39].

We have taken this fruitful approach further by actually solving the

strongly nonlinear dynamical model numerically in 2D, and by account-

ing also for electroosmotic ﬂuid motion and convection. We agree that for

weakly nonlinear dynamics to hold, the total zeta potential indeed cannot

greatly exceed the thermal voltage, but because the induced zeta potential

1.3 Outline of the thesis 5

is related to the driving voltage in a nonlinear way we also ﬁnd that the

latter can often be an order of magnitude larger than the thermal voltage

at the point where the weakly nonlinear dynamics break down.

As the driving voltage is further increased, even the strongly nonlin-

ear asymptotics break down due to complete salt depletion in the diﬀusion

layers close to the electrodes. In steady-state problems of Faradaic con-

duction it is well known that beyond the “diﬀusion-limited” current, the

electrical double layer changes from quasiequilibrium to a nonequilibrium

structure with an extended space charge layer, as described ﬁrst by Ru-

binstein and Shtilman [40]. Further, when tangential ﬁelds act upon this

space charge layer, nonlinear electrokinetic ﬂow is induced. Dukhin and co-

workers introduced the term “electroosmosis of the second kind” to describe

this phenomenon, and studied it extensively in the context of nonlinear elec-

trophoresis of conductive particles made from ion exchanger material [41].

More recently, Rubinstein and Zaltzman theoretically predicted that second

kind electroosmosis renders linearly instable the diﬀusion layer at a planar

permselective (ion exchange) membrane and provides an eﬃcient mixing

mechanism, which can account for overlimiting conductance [42, 43].

Because even our strongly nonlinear model does not seem to be able

to account for reversal of the pumping velocity, and because the driving

voltages used experimentally are often large enough to make the strongly

nonlinear model break down, we try to incorporate the basic features of the

nonequilibrium double layer and electroosmosis of the second kind into our

dynamical model. This is a challenging problem, and to be able to handle it

numerically we make some fairly crude approximations. Our results indicate

that electroosmosis of the second gives rise to net pumping in the same

(“forward”) direction as the usual electroosmosis of the “ﬁrst kind”, while

bulk electroconvection contributes in the “reverse” direction.

1.3 Outline of the thesis

Chapter 2: Electrokinetic theory

We ﬁrst give a brief overview of some of the classical electrokinetic

theory, before we proceed to construct a mathematical model based

on matched asymptotic expansions. The latter is a perturbation tech-

nique where one identiﬁes the ratio of the Debye screening length to the

characteristic geometrical length scale as a small parameter. Indeed

the assumption of thin Debye layers underlies most classical analysis

in electrochemistry, so the perturbation technique is often just a sys-

tematic way of obtaining already known results, but it also provides

a ﬁrm starting point for going beyond the classical theory. In this we

follow the work of Bazant, Thornton, and Ajdari [37] and Chu [38] on

diﬀuse charge dynamics, but where they primarily focus on transient

6 Introduction

solutions for a suddenly applied DC voltage, we study AC driving and

steady-state periodic solutions.

Chapter 3: Linear and weakly nonlinear analysis

We study the pumping from an array of asymmetric pairs of micro-

electrodes, extending the existing linear theory for AC electroosmo-

sis by accounting for Faradaic current injection both in a linearized

scheme and using the full nonlinear Butler–Volmer reaction kinetics.

We also include the nonlinear surface capacitance of the Debye layer

as described by Gouy–Chapman–Stern theory. Most of this work was

published in our paper ‘AC electrokinetic micropumps: the eﬀect of ge-

ometrical conﬁnement, Faradaic current injection, and nonlinear sur-

face capacitance’ [44], see Appendix B, but here we also take into

account the eﬀects of mass transfer on the Faradaic electrode reac-

tion.

Chapter 4: Strongly nonlinear analysis

In this chapter we ﬁrst analyse the strongly nonlinear dynamics for

a simple 1D geometry with parallel-plate blocking electrodes, both

to gain a better understanding of the dynamical solutions, but also

to validate the asymptotic model by comparison to a full numerical

solution of the Poisson–Nernst–Planck equations. Next we study a

2D model of the planar AC electroosmotic pump and determine the

contributions to the net pumping from both electroosmotic ﬂow of the

ﬁrst and second kind, and from electroconvection due to concentration

gradients in the electroneutral diﬀusion layers and bulk electrolyte.

Chapter 5: Optimization

There are many ways of breaking the left-right symmetry of an elec-

trode array to induce net pumping. In this chapter we consider both

the “standard” pump with asymmetric pairs of planar microelectrodes,

a design with asymmetric coating of the individual electrodes, and a 3D

structured electrode design recently suggested by Bazant and Ben [45].

In all cases the performance increases with miniaturization, so we in-

vestigate the optimal electrode design for a ﬁxed minimal linewidth or

feature size in the geometry. This work is related to our work done on

topology optimization earlier during the PhD studies, which has been

published in the paper ‘A high-level programming-language implemen-

tation of topology optimization applied to steady-state Navier–Stokes

ﬂow’ [46]. We have chosen not to present those results in this thesis

but have included the paper in Appendix C.

Chapter 6: Conclusion and outlook

We present concluding remarks on our work on AC electrokinetic

pumping and give some directions for future research.

1.4 Publications during the PhD studies 7

Appendix A: FEMLAB code

While the standard linear model for induced-charge electroosmosis has

been studied numerically by several authors, not many attempts have

been made in the weakly nonlinear regime. Part of the reason may

be the increased complexity involved with solving a time evolution

problem with special variables deﬁned on the boundary. However,

using commercial ﬁnite element software like Femlab [47] this is ac-

tually straightforward to implement, and to demonstrate this we have

included a code example in Appendix A. We also include the imple-

mentation of our strongly nonlinear model, and discuss how to deal

with the nonlocal coupling in time introduced by our semi-analytical

solution for the dynamics in the diﬀusion layer.

1.4 Publications during the PhD studies

Papers in peer reviewed journals

1. Electro-hydrodynamics of binary electrolytes driven by modulated

surface potentials,

N. A. Mortensen, L. H. Olesen, L. Belmon, and H. Bruus,

Phys. Rev. E 71, 056306 (2005).

2. A high-level programming-language implementation of topology

optimization applied to steady-state Navier–Stokes ﬂow,

L. H. Olesen, F. Okkels, and H. Bruus,

Int. J. Num. Meth. Eng. 65, 975 (2006).

3. Transport coeﬃcients for electrolytes in arbitrarily shaped nano

and micro-ﬂuidic channels,

N. A. Mortensen, L. H. Olesen, and H. Bruus,

New J. Phys. 8, 37 (2006).

4. AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement,

Faradaic current injection, and nonlinear surface capacitance,

L. H. Olesen, H. Bruus, and A. Ajdari,

Phys. Rev. E 73, 056313 (2006).

8 Introduction

Conference proceedings

1. Topology optimization of Navier–Stokes ﬂow in microﬂuidics,

L. H. Olesen, F. Okkels, and H. Bruus,

ECCOMAS 2004, Jyv¨askyl¨a, Finland, July 2004, proc. vol. 2, p. 224.

2. Applications of topology optimization in the design of micro- and

nanoﬂuidic devices,

F. Okkels, L. H. Olesen, and H. Bruus,

Nanotech 2005, Anaheim, USA, May 2005, proc. vol. 1, p. 575–578.

3. Applications of topology optimization in microﬂuidics: ﬂuid ﬂows,

thermal transport, and chemical reactions,

F. Okkels, L. H. Olesen, and H. Bruus,

6th EPS Liquid Matter Conf., Utrecth, the Netherlands, July 2005.

4. AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement,

Faradaic currents, and nonlinear surface capacitance,

L. H. Olesen, H. Bruus, and A. Ajdari

MicroTAS 2005, Boston, USA, October 2005, proc. vol. 1, p. 560–562.

5. Nonlinear studies of AC electrokinetic micropumps,

H. Bruus, L. H. Olesen, and A. Ajdari

APS MAR06 Meeting, Baltimore, USA, March 2006.

6. Mass and charge transport in micro and nano-ﬂuidic channels,

N. A. Mortensen, L. H. Olesen, and H. Bruus,

2nd Int. Conf. Transport Phenomena in Micro and Nanodevices,

Barga, Italy, June 2006.

7. Universality in microﬂuidic phenomena inside microchannels with

arbitrarily shaped cross-sections,

H. Bruus, N. Mortensen, F. Okkels, and L. H. Olesen,

Euromech Fluid Mechanics Conf. 6, Stockholm, Sweden, June 2006.

Chapter 2

Electrokinetic theory

In this chapter we start by giving a brief overview of the classical theory of

electrochemical transport and electrode reaction processes. This is a widely

studied subject due to its importance both for industrial applications such

as batteries and fuel cells, electroplating, and desalination by electrodialy-

sis, and in natural processes such as corrosion and ion transport through

ion channels in biological cells. A common feature underlying most of the

classical analysis is the assumption that the system can be divided into an

electroneutral bulk region and charged boundary layers that are very thin

compared to the characteristic geometrical length scale in the system. This

simple picture can be formalized by perturbation theory, and in Sec. 2.2

we go through a systematic derivation of a mathematical model based on

matched asymptotic expansions.

2.1 Classical theory

A fundamental approximation in the classical theory of electrochemical

transport is the continuum approximation where the diﬀerent chemical species

are represented by their concentration ﬁelds. For a concentrated solution

of 1 M the mean separation between ions is around 1 nm, whereas for a

dilute 1 mM solution the mean separation is around 10 nm. Therefore, in

most micro- and macroscopic applications the continuum approximation is

well justiﬁed, but in true nanoﬂuidics it may not be. However, even in a

macroscopic electrochemical experiment, the system spontaneously devel-

ops microscopic structure in the form of screening layers at electrodes or

any other charged surfaces and interfaces in the electrolyte. The thickness

of the screening layer is given by the Debye length

λ

D

=

ǫk

B

T

¸

i

c

∗

i

z

2

i

e

2

, (2.1)

10 Electrokinetic theory

where ǫ is the electrical permittivity of the solvent, k

B

is Boltzmanns con-

stant, T the absolute temperature, e the elementary charge, c

∗

i

is the bulk

concentration (number density) of the ith ionic species, and z

i

the valence

number. For a typical system λ

D

is of the order of 1–10 nm, that is, on

average there are no more than a few ions across it. This means that the

instantaneous “concentration proﬁle” over a particular point on the surface

will be highly nonuniform and rapidly ﬂuctuating, but upon proper averag-

ing in time or in the transverse direction to the screening layer one should

still be able to use continuum theory on the averaged ﬁelds.

2.1.1 Electrochemical transport

Within a continuum model the transport of the individual chemical species

is governed by conservation laws of the form

∂

t

c

i

= −∇ F

i

, (2.2)

where F

i

is the ﬂux of the ith species and we disregard any bulk chemical

reactions producing or consuming that species. In general the ﬂux is a

function of the electrochemical potential µ

j

of all the species present in

solution [53, 54]

F

i

= −c

i

¸

j

L

ij

∇µ

j

+c

i

u, (2.3)

where L

ij

are generalized mobilities that account for interactions between

diﬀerent species, and c

i

u represents transport by convection with the ﬂuid

velocity u.

Nernst–Planck equation

For an ideal dilute system the interaction between diﬀerent species can be

neglected and the electrochemical potential can be decomposed into separate

entropic and electrical terms as

µ

i

= k

B

T log c

i

+z

i

e φ (2.4)

where is φ the mean-ﬁeld electrostatic potential. Then Eq. (2.3) for the ﬂux

reduces to the Nernst–Planck equation

F

i

= −D

i

∇c

i

+

z

i

e

k

B

T

c

i

∇φ

+c

i

u, (2.5)

where we used the Einstein relation D

i

= k

B

TL

ii

to express the mobility in

terms of the diﬀusivity D

i

. The electrostatic potential φ is determined by

the charge distribution ρ through the Poisson equation

−ǫ∇

2

φ = ρ = e

¸

i

z

i

c

i

, (2.6)

assuming constant permittivity ǫ of the solvent.

2.1 Classical theory 11

IHP OHP

Metal

Diﬀuse layer Bulk electrolyte

Solvated cation

Speciﬁcally adsorbed anion

Solvent molecule

−λ

S

0 λ

D

y

V

ext

ζ

φ(y)

Figure 2.1: Schematic picture of the diﬀuse screening layer in the electrolyte next to

a negatively biased metal electrode. Cations are attracted to the electrode and anions

are repelled so the interface appears charge neutral when seen from the bulk. The outer

Helmholtz plane (OHP) deﬁnes the closest distance that the solvated cations can move

towards the electrode. Also shown is φ(y), the mean-ﬁeld potential variation from the

electrode to the charge neutral bulk.

2.1.2 Electrical double layer

The common continuum model view of the screening layer is based on di-

vision into a compact and a diﬀuse part. This is shown schematically in

Fig. 2.1 for the screening layer next to a negatively biased metal electrode.

Cations in the electrolyte move towards the electrode to screen the negative

charge, and anions are repelled. The ﬁnite size of the ions can be taken into

account by deﬁning a smallest distance that the solvated cations can move

towards the electrode, and this point of closest approach is called the outer

Helmholtz plane (OHP). The region between the OHP and the electrode is

called the “compact layer”, and in Fig. 2.1 it is roughly two solvent molecules

thick because the approach of the cations is obstructed both by their own

solvation shell and by a monolayer of solvent molecules held at the surface

by van der Waals forces. The ions in the screening layer outside the OHP are

not static but undergo thermal motion, and this region is therefore called

the “diﬀuse layer”. Some ions in the electrolyte may interact more strongly

with the electrode material and adsorb directly on the electrode, and the

locus of their electrical centers is denoted the inner Helmholtz plane (IHP).

Together the compact and diﬀuse layers are called the “double layer”, and

the model is known as the Gouy–Chapman–Stern model [16].

In order to avoid confusion with the diﬀusion layer of width ℓ

i

(ω) ∼

12 Electrokinetic theory

D

i

/ω, to and from which the ions have time to diﬀuse during one cycle of

the driving voltage at angular frequency ω, we prefer to denote the diﬀuse

part of the double layer as the “Debye layer”. We also tend to denote the

compact part as the “Stern layer”, and mention that this can also be used

to model an oxide layer or insulating coating on the electrode.

Gouy–Chapman–Stern model

In thermal equilibrium the electrochemical potential is constant thoughout

the system and the ions in the Debye layer are Boltzmann distributed, i.e.,

c

i

= c

∗

i

exp

¸

−

z

i

e(φ −φ

∗

)

k

B

T

, (2.7)

where c

∗

i

is the reference concentration attained in the charge neutral bulk

electrolyte where φ = φ

∗

. Substituting into Eq. (2.6) we get the Poisson–

Boltzmann equation

−∇

2

ψ =

e

ǫ

¸

i

c

∗

i

z

i

exp

−

z

i

eψ

k

B

T

, (2.8)

where we introduced ψ = φ−φ

∗

, and for a symmetric binary z:z electrolyte

this reduces to the more familiar form

∇

2

ψ =

1

λ

2

D

k

B

T

ze

sinh

zeψ

k

B

T

. (2.9)

The Poisson–Boltzmann equation constitutes a single equation to solve for

the equilibrium electrostatic potential; the ion distributions then follow di-

rectly from Eq. (2.7). For one-dimensional systems it has an analytical

solution, namely the Gouy–Chapman solution

1

ψ(y) =

4k

B

T

ze

tanh

−1

¸

tanh

zeζ

4k

B

T

e

−y/λ

D

, (2.11)

where y is the distance perpendicular to the surface and ζ is the potential

attained at the outer Helmholtz plane where y = 0, cf. Fig. 2.1. The total

amount of charge accumulated in the Debye layer per unit area is

q =

∞

0

ρ dy = ǫ∂

y

ψ

y=0

(2.12)

= −sgn(ζ)

2ǫk

B

T

¸

i

c

∗

i

(e

−z

i

eζ/k

B

T

−1) (2.13)

= −

ǫ

λ

D

2k

B

T

ze

sinh

zeζ

2k

B

T

, (2.14)

1

Gouy also showed that Eq. (2.8) can be integrated for an asymmetric binary electrolyte

with z+ = −2z− = 2z, where

ψ(y) =

kBT

ze

log

¸

1 −

3

1 + cosh

y/λD + 2 tanh

−1

(1 + 2e

zeζ/k

B

T

)/3

. (2.10)

2.1 Classical theory 13

where Eq. (2.13) reduces to Eq. (2.14) for a symmetric z:z electrolyte only.

Neglecting speciﬁcally adsorbed ions, the region −λ

S

≤ y ≤ 0 behind the

outer Helmholtz plane is free of charge and the potential varies linearly, i.e.,

the Stern layer ideally behaves as a capacitor of capacitance C

S

= ǫ

S

/λ

S

.

The overall potential drop across the double layer is then

V

ext

−φ

∗

= ζ −

q

C

S

, (2.15)

and the diﬀerential capacitance per unit area becomes

C

dl

= −

dq

d(V

ext

−φ

∗

)

=

¸

1

C

S

+

λ

D

ǫ cosh(zeζ/2k

B

T)

−1

. (2.16)

In the Debye–H¨ uckel limit ζ ≪ k

B

T/e where the potential variations are

small compared to the thermal voltage, the double layer capacitance reduces

to

C

dl

=

¸

1

C

S

+

λ

D

ǫ

−1

=

ǫ

λ

D

(1 +δ)

, (2.17)

where we introduced the parameter δ = ǫλ

S

/ǫ

S

λ

D

for the capacitance ratio

between the Debye and Stern layers. At larger potential the Debye layer

capacitance becomes very large and the overall double layer capacitance is

dominated by the Stern layer.

The relevant parameter range for δ is large: If we compare the Debye

layer capacitance for a dilute 1 mM aqueous electrolyte with λ

D

= 10 nm

and ǫ = 80ǫ

0

to the molecular Stern layer deﬁned by the distance closest

approach λ

S

∼ 1

˚

A of an ion to a bare metal electrode then we get a very

small capacitance ratio δ ∼ 0.01, whereas for a titanium electrode with

a natural TiO

2

oxide of thickness λ

S

∼ 4 nm and ǫ

S

= 110ǫ

0

we ﬁnd

δ ∼ 0.3, and for even thin polymer coatings we easily obtain δ ≫1. In real

electrochemical systems the Stern layer on bare metal electrodes often does

not behave exactly as an ideal capacitor but shows some voltage dependence

in C

S

, and the speciﬁc adsorbtion and desorption of ions give rise to an

additional component in Eq. (2.16). However, because we are not concerned

with any particular material system but wish to study electrokinetics in

more general, we stick to the simplest form of the Gouy–Chapman–Stern

model and always assume a constant Stern layer capacitance and neglect

speciﬁc adsorption.

Modiﬁed Poisson–Boltzmann equation

The Gouy–Chapman–Stern model is widely used because it is simple and

yet captures the generic features observed in typical electrochemical experi-

ments. The exponential increase in ion concentration with potential implied

in Eq. (2.7) eﬀectively limits the range of zeta potentials that makes sense

14 Electrokinetic theory

10

0

10

1

10

2

10

3

10

0

10

2

10

4

10

6

10

8

ze (V

ext

−φ

∗

)/k

B

T

e

x

p

z

e

ζ

/

k

B

T

−

1

δ = 0.1

δ = 1.0

δ = 10.

Figure 2.2: Double logarithmic plot of the relative increase in anion concentration at

the outer Helmholtz plane as a function of the overall potential drop Vext −φ

∗

across the

double layer for diﬀerent capacitance ratios δ in the Gouy–Chapman–Stern model.

physically: Substituting ζ = 500 mV for a bulk dilute 1 mM KCl solu-

tion, the concentration predicted at the outer Helmholtz plane exceeds 10

32

ions per m

3

or 100 ions per

˚

A

3

which is ridiculously large. However, the

“purpose” of the Stern layer is exactly to supress this unphysical behaviour.

At large voltage when the double layer capacitance is dominated by the

Stern layer, the accumulated charge depends linearly on the applied voltage

q ≃ −C

S

(V

ext

− φ

∗

), whereas the zeta potential grows only logarithmically

with q, cf. Eq. (2.13). Therefore the increase in ion concentrations in the

Debye layer also becomes less dramatic, as shown in Fig. 2.2 where the rel-

ative increase in the anion concentration at the outer Helmholtz plane is

plotted as a function V

ext

−φ

∗

. The ﬁgure shows that the ion concentration

grows as (V

ext

− φ

∗

)

2

/δ

2

at large voltage, but also that this large voltage

regime is not reached until V

ext

− φ

∗

exceeds ≈ 1 V. If we consider again

a dilute 1 mM KCl solution then the bulk concentration is 6 10

23

ions

per m

3

. The ions have radii of a few ˚angstrøm, with 1.33 (2.32)

˚

A for a

(hydrated) K

+

ion and 2.32

˚

A for Cl

−

, so the steric limit of a close packed

phase of ions corresponds to around 10

28

ions per m

3

. The voltage at which

this limit is reached depends on the Stern layer capacitance: Assuming a

low capacitance ratio of δ = 0.1 it occurs already at 500 mV while for δ = 10

the limit is reached around 10 V.

One problem with this reasoning, where almost the entire potential drop

across the double layer is placed on the Stern layer, is that most insulators

suﬀer dielectric breakdown around 1–100 MV/m, which is far exceeded if we

put 1 V across a 1

˚

A molecular Stern layer, or even a 5 nm titanium oxide

layer.

2.1 Classical theory 15

Further, the neglect of correlations between diﬀerent ions inherent in the

use of Eq. (2.4) for the electrochemical potential obviously becomes ques-

tionable long before the limit of steric exclusion sets in, but the violation of

the steric limit is perhaps the most oﬀending feature in the Gouy–Chapman–

Stern model. Borukhov et al. have suggested a simple modiﬁcation of the

Poisson–Boltzmann equation to account for the ﬁnite size of ions [55]. They

include the entropy of the solvent molecules in the expression for the free

energy F, and obtain for a symmetric binary z:z electrolyte with ions of

characteristic size a the following expression for the electrochemical poten-

tial

µ

±

=

δF

δc

±

= k

B

T log(c

±

a

3

) +z

±

eφ −k

B

T log(1 −c

+

a

3

−c

−

a

3

), (2.18)

where a

3

is the volume occupied by a single ion. The ion concentrations

c

±

cannot then exceed the steric limit 1/a

3

, essentially because the entropy

associated with excluding the last solvent molecule from a condensed phase

of ions is inﬁnite. Their modiﬁed Poisson–Boltzmann equation becomes

∇

2

ψ =

1

λ

2

D

k

B

T

ze

sinh

zeψ

k

B

T

1 −2c

∗

a

3

+ 2c

∗

a

3

cosh

zeψ

k

B

T

. (2.19)

Kilic, Bazant, and Ajdari have recently extended this approach to the non-

equilibrium case by deriving a modiﬁed Nernst–Planck equation based on

Eq. (2.18) [56, 57]. This is very a interesting direction which may allow to

explain, e.g., the vanishing of the AC electroosmotic pumping at electrolyte

concentrations above 10 mM [15]. However, we also like stress that the

ability of the double layer to absorb ions and act as a capacitor persist long

after the steric limit is reached. In this regime the Gouy–Chapman–Stern

model systematically underestimates the zeta potential, but may still give

a good approximation to the overall capacitance provided the capacitance

ratio δ is not too low and the nominal bulk concentration is not too high.

2.1.3 Electrode reactions

Electrode reactions are chemical redox reactions that involve transfer of one

or more electrons from an electrode to the reactants. Familiar examples are

Zn

2+

(aq)

+ 2 e

−

⇋ Zn

(s)

E

◦ −

= −0.762 V, (2.20)

2 H

+

(aq)

+ 2 e

−

⇋ H

2(g)

E

◦ −

= 0.000 V, (2.21)

O

2(aq)

+ 2 H

2

O

(ℓ)

+ 4 e

−

⇋ 4 OH

−

(aq)

E

◦ −

= 0.444 V, (2.22)

where E

◦ −

is the standard reduction potential. The reaction in Eq. (2.20)

involves transfer of two electrons from the electrode to the Zn

2+

ion. Such

16 Electrokinetic theory

Reaction coordinate

F

r

e

e

e

n

e

r

g

y

State I

State II

State ‡

∆G

‡

f

∆G

‡

b

Figure 2.3: Free energy landscape along reaction path from state I (reactant) to state II

(product). The activated complex (or transition state) ‡ is the conﬁguration of maximal

free energy.

a tunneling process is very unlikely, and Eq. (2.20) must therefore repre-

sent a multistep reaction where the Zn

2+

is ﬁrst reduced to Zn

+

and then

Zn. Likewise, the reaction in Eq. (2.21) must also be a multistep reaction

consisting of one step where H

+

is reduced to form a single hydrogen atom

adsorbed on the electrode, followed by a second step where two hydrogen

atoms join and desorb from the electrode as H

2

. Often one step in such a

multistep reaction is much slower than the rest. To determine the reaction

kinetics it is then suﬃcient to focus on this rate-limiting step whereas the

other steps are considered as reaction equilibria. In the present study we

therefore assume a generic one-step, single electron process of the form

O + e

−

⇋R, (2.23)

where z

O

= z

R

+ 1 to ensure charge conservation.

Nernst equation

In thermal equilibrium the change in electrochemical Gibbs free energy as-

sociated with the reaction is zero

∆G =

G

◦ −

R

+k

B

T log c

∗

R

+z

R

eφ

∗

−

G

◦ −

O

+k

B

T log c

∗

O

+z

O

eφ

∗

−eV

ext

= 0, (2.24)

from which we obtain the Nernst equation for the equilibrium potential drop

from the electrode to the electrolyte

(V

ext

−φ

∗

)

eq

= E

◦ −

+

k

B

T

e

log

c

∗

O

c

∗

R

. (2.25)

Here ∆G

◦ −

= G

◦ −

R

− G

◦ −

O

= −eE

◦ −

is the increase in standard free energy

associated with the reaction.

2.1 Classical theory 17

Butler–Volmer kinetics

Out of equilibrium the forward and backward reaction rates no longer bal-

ance and a net reaction takes place. According to transition state theory

the rate limiting factor is associated with getting across a bump in the free

energy landscape, cf. Fig. 2.3. The state ‡ of maximal free energy along

the reaction path in conﬁguration space is denoted the transition state or

activated complex. The forward reaction rate ν

f

is proportional to the prob-

ability of getting from the reactant state into the transition state, and this

probability drops exponentially with the diﬀerence in electrochemical free

energy between the two states, i.e.,

ν

f

∝ exp

−

∆G

‡

f

k

B

T

, (2.26)

and similarly for the backward reaction ν

b

ν

b

∝ exp

−

∆G

‡

b

k

B

T

. (2.27)

Further it is assumed that the electrochemical free energy at the transition

state can be decomposed into a chemical term that is essentially independent

of the applied potential, and an electrical term (G

‡

)

e

that is a weighted

average of the electrostatic energy of the reactant and product, i.e.,

(G

‡

)

e

= (1 −α)(z

O

eφ

s

−eV

ext

) +αz

R

eφ

s

. (2.28)

Here the transfer coeﬃcient α is simply a measure of how close the transition

state is to either the reactant or product state in terms of electrostatic

energy, and φ

s

is the potential in the electrolyte at the electrode surface (or

outer Helmholtz plane). Substituting into Eqs. (2.26) and (2.27) the net

reaction rate becomes

ν

net

= ν

f

−ν

b

= k

◦ −

f

c

s

O

e

−αe(Vext−φ

s

)/k

B

T

−k

◦ −

b

c

s

R

e

(1−α)e(Vext−φ

s

)/k

B

T

, (2.29)

where k

◦ −

f

and k

◦ −

b

are the forward and backward standard rate constants,

and c

s

O

and c

s

R

are the concentrations of the oxidized and reduced species at

the surface. This result and its implications is known as the Butler–Volmer

formulation of reaction kinetics [16]. The Faradaic current injected from the

electrode into the electrolyte by the reaction is

j

ext

= −eν

net

. (2.30)

If we assume quasiequilibrium in the double layer we can relate the con-

centrations at the electrode surface to those immediately outside the Debye

18 Electrokinetic theory

layer by a Boltzmann factor, c

s

i

= c

i

e

−z

i

eζ/k

B

T

. Then Eqs. (2.29) and (2.30)

can be re-expressed as the current-overpotential equation

j

ext

= j

0

e

−(z

O

−α)eζ/k

B

T

c

R

c

∗

R

e

(1−α)eϑ/k

B

T

−

c

O

c

∗

O

e

−αeϑ/k

B

T

, (2.31)

where j

0

= ek

◦ −

(c

∗

R

)

α

(c

∗

O

)

1−α

is the exchange current, k

◦ −

is the standard

rate constant, c

O

and c

R

are the concentrations immediately outside the

Debye layer, c

∗

O

and c

∗

R

are the corresponding bulk values at equilibrium,

and the overpotential ϑ = (V

ext

−φ) −(V

ext

−φ)

eq

is the diﬀerence between

the actual potential drop across the double layer and the thermal equilibrium

for the given redox process and bulk concentration. Linearizing for small

overpotentials we get

j

ext

=

j

0

e

k

B

T

exp

¸

−

(z

O

−α)eζ

eq

k

B

T

ϑ =

1

R

ct

ϑ, (2.32)

deﬁning the charge transfer resistance R

ct

.

Reversibility of the electrode reactions

In our model calculations we always assume that the electrode reaction is

reversible. This seems natural because our primary focus is on a micropump

based on an array of electrode pairs that is driven with an AC voltage: We

model this as an inﬁnite periodic system in space and look for steady-state

periodic solutions in time after all transients have died out. If the device

is pumping the electrolyte in a closed loop and for a long time then this

should be a good approximation. However, if the device is pumping from an

inlet with “fresh” electrolyte to an outlet that is sent to “waste”, then the

chemical composition in the electrolyte might change upon passage through

the pump, e.g., by a change in pH, by absorbtion of a small amount of

electrolytic gasses, or by gradually dissolving the electrode material and

ﬂushing it out to waste. While it may or may not be feasible or desirable to

operate a pump in this way we do not account for it in our model.

2.1.4 Electroosmosis

The ﬂuid motion in the electrolyte is governed by the Navier–Stokes equation

ρ

m

[∂

t

u + (u ∇)u] = −∇p +η∇

2

u +ρE, (2.33)

where u is the ﬂuid velocity, p is the pressure, ρ

m

is the mass density, η

the dynamic viscosity, and ρE is the electrical body force acting the ﬂuid.

In the bulk electrolyte the charge density ρ is small, but when a tangential

ﬁeld acts on the charge in the Debye layers it can induce a ﬂow which in

turn drags the bulk ﬂuid along. This phenomenon is called electroosmosis.

2.2 Mathematical model 19

If we consider a uniformly charged ﬂat surface at y = 0 in contact with

an electrolyte and with a tangential ﬁeld E

x

acting on the charge in the

screening layer then the tangenial component of the Navier–Stokes equation

reduces to a balance between electrical and viscous forces

0 = η∂

2

y

u

x

−ǫ∂

2

y

ψ E

x

. (2.34)

Integrating twice and using that u

x

= 0 and ψ = ζ at the surface while

∂

y

u

x

= 0 and ∂

y

ψ = 0 in the bulk we obtain

u

x

= −

ǫ(ζ −ψ)

η

E

x

. (2.35)

The ﬂuid velocity increases from no-slip at the surface to a constant value

u

x

= −

ǫζ

η

E

x

(2.36)

outside the Debye layer, which eﬀectively acts as a slip velocity for the bulk

ﬂuid motion. This classical result is known as the Helmholtz–Smoluchowski

formula [17, 53].

In deriving Eq. (2.36) we associated the zeta potential with the no-

slip plane, whereas in Sec. 2.1.2 we deﬁned ζ as the potential at the outer

Helmholtz plane. Apriori there is no guarantee that the no-slip and the outer

Helmholtz planes should coincide, but experiments for some particularly

smooth, impermeable, and well-deﬁned surfaces indicate that they do [17,

58]. On the other hand, for rough, porous, or inhomogeneous surfaces we

cannot expect this, e.g., if the charged surface has a thin adsorption layer

of macromolecules or polymer gel then the electroosmotic ﬂow would be

greatly suppressed due to the large hydrodynamic resistance of the polymer

fragments [58].

2.2 Mathematical model

The solution of the full Poisson–Nernst–Planck equations is a challenging

task; even numerically the solution is made diﬃcult because the intrinsic

length scale λ

D

in the electrolyte is often several orders of magnitude smaller

than the relevant geometrical length scale ℓ

0

. Yet our physical intuition is

that the bulk electrolyte is simply charge neutral, and that only the screening

layers at the boundary display structure on the Debye length scale. This

simple picture can be formalized by perturbation theory, rewriting the full

Poisson–Nernst–Planck problem as a perturbation expansion in the small

parameter

ε =

λ

D

ℓ

0

. (2.37)

20 Electrokinetic theory

The Poisson equation turns out to be singularly perturbed in the sense that

the small parameter ε multiplies the highest (second) order derivative, and

mathematically this singular perturbation can be seen as the origin of the

disparate length scales appearing in the solution.

The method of matched asymptotic expansions is a generalization of

Prandtl’s viscous boundary layer theory, and it is a method for treating

singular perturbations [59]. Essentially the method works by computing

separate “outer” and “inner” approximations to the solution in the bulk

and boundary regions, respectively, where the inner approximation in the

boundary layer is obtained by scaling the spatial variables to remove the

singular perturbation before solving the problem. The two solutions are

matched by requiring that the outer limit of the inner approximation ˜ c

±

should be asymptotically equal to the inner limit of the outer approximation

¯ c

±

, i.e.,

lim

˜ y→∞

˜ c

±

(x, ˜ y, t) = lim

y→0

¯ c

±

(x, y, t), (2.38)

where ˜ y = y/ε is the rescaled spatial variable normal to the boundary in the

inner region. The standard procedure begins by seeking regular expansions

in the form of power series

¯ c

±

= ¯ c

(0)

±

+ε ¯ c

(1)

±

+ε

2

¯ c

(2)

±

+. . . , (2.39)

substituting into the governing equations, and collecting like powers of ε.

The primary interest is on the leading order term ¯ c

(0)

±

, but the higher order

corrections ¯ c

(i)

±

can also contain important information. The expansions need

not always take the form Eq. (2.39) but can also contain fractional powers

of ε, and logarithmic terms like ε log ε sometimes “pop up” in the process

of deriving higher order terms [59]. The approximations to the solution are

“asymptotic” in the sense that they converge to the true solution in the limit

ε → 0 with all other parameters in the problem held ﬁxed. However, for

any ﬁxed ε > 0 there could be ε dependent restrictions on, e.g., the driving

voltage V

0

for a truncation of the series to produce accurate results [37], and

strictly there is no guarantee that the series will converge at all, even if an

inﬁnite number of terms are included [59].

We denote “weakly nonlinear” the regime where the leading order term

in the standard asymptotic expansion remains a good approximation to the

solution. However, experimentally there are several interesting and intrigu-

ing phenomena that show up at larger voltage, so we would like to extend

the validity of the model as far up in driving voltage as possible. It turns

out that for V

0

∼ k

B

T/ze

√

ε the weakly nonlinear solution breaks down,

but by reordering the terms in the expansion to reﬂect the dominant bal-

ance we can develop another “strongly nonlinear” model valid in the joint

limit ε →0 and V

0

→∞.

Because we focus on the periodic solution for an AC driving voltage

after all transients have died out, there are actually three diﬀerent charac-

2.2 Mathematical model 21

The asymptotic model in a nutshell

Substrate

Electrode

Electrolyte

ε

√

ε

1

y

Electrode

Debye layer

Diﬀusion layer

Bulk electrolyte

Nested boundary layers Variables

¯ c,

¯

φ, ¯ u, ¯ p

ˇ c,

ˇ

φ, ˇ u, ˇ p

˜ c

±

,

˜

φ, ˜ u, ˜ p

ˇ y = y/

√

ε

˜ y = y/ε

ˇ γ, ˇ ς, ˇ u

s

˜ w

±

,

˜

ζ, ˜ u

s

Figure 2.4: Schematic picture of the nested boundary layers in the matched asymptotic

expansion: The “outer” bulk region is connected via the “middle” diﬀusion layer to the

“inner” Debye layer. Variables associated with the diﬀerent regions are denoted by bar,

check, and tilde accents, respectively. The bulk is electroneutral with a salt concentration

¯ c that is constant in time but not necessarily in space. The diﬀusion layer is also elec-

troneutral but has a dynamically varying salt concentration; the excess salt concentration

ˇ γ = ˇ c − ¯ c is determined by a simple 1D diﬀusion problem. The excess potential drop

ˇ ς across the diﬀusion layer induces an eﬀective slip velocity ˇ us. The Debye layer is in

quasiequilibrium at the RC time scale with a structure that is completely determined by

the excess amount of each ion ˜ w± =

∞

0

(˜ c±−ˇ c) d˜ y accumulated in the layer together with

the salt concentration ˇ c seen from the Debye layer. The potential drop

˜

ζ across the Debye

layer determines the electroosmotic contribution to the slip velocity ˜ us.

teristic length scales in the system that we can identify, namely, the De-

bye screening length λ

D

, the geometrical length scale ℓ

0

, and the diﬀusion

length ℓ(ω) =

D

0

/ω which is the typical distance that an ion can diﬀuse

during one period of the driving signal. Likewise we can identify three dif-

ferent characteristic time scales, namely, the charge relaxation (Debye) time

τ

D

= λ

2

D

/D

0

, the bulk diﬀusion time τ

ℓ

= ℓ

2

0

/D

0

, and the cell relaxation

(RC) time τ

0

= λ

D

ℓ

0

/D

0

. AC electroosmosis occurs primarily for driving

frequencies around the inverse RC time ω ∼ 1/τ

0

. Here the diﬀusion length

becomes ℓ(ω = 1/τ

0

) =

√

λ

D

ℓ

0

, which means that in the asymptotic analysis

we need to account for three nested regions: The “inner” screening layer of

(dimensionless) width ε, the “middle” diﬀusion layer of width

√

ε, and the

“outer” bulk region at the unit characteristic length scale [38].

Applying the method of matched asymptotic expansions to study prob-

lems in electrochemistry and electrokinetics is not a new idea: Much work

has been done on steady-state solutions for electrochemical cells operated at

22 Electrokinetic theory

DC conditions, e.g., [40, 60], and the low voltage regime, where the Debye–

H¨ uckel approximation ζ ≪k

B

T/ze can be employed or the driving voltage

or applied ﬁeld is weak enough that the problem can be linearized, has been

studied extensively, e.g., [23, 58, 61, 62]. Dynamical solution have been

looked at recently by Bazant and co-workers [37, 38, 39]. However, they pri-

marily focus on transient solutions for a suddenly applied DC voltage. We

extend their work by taking into account dynamics in the diﬀusion layers

when the system is driven into the strongly nonlinear regime by a large AC

voltage around the inverse RC time. Also we include electroosmotic ﬂuid

motion into our model and account for bulk ion transport by convection.

Below we ﬁrst cast the problem into dimensionless form and identify the

relevant dimensionless groups. Then we proceed determine the leading order

solution in the bulk, diﬀusion layer, and Debye layer, respectively, and state

the surface conservation laws governing the dynamics of the solution in the

boundary layers. The structure is sketched in Fig. 2.4; we focus exclusively

on the leading order solution and therefore drop the superscript

(0)

to sim-

plify the notation. In the derivation we retain all terms that are important

in the strongly nonlinear regime, which makes it easy in subsequent sections

to extract the linear (low voltage), weakly nonlinear, and strongly nonlinear

dynamical models. Finally we discuss the conditions for validity or break-

down of both the weakly and strongly nonlinear models, and propose an ad

hoc modiﬁcation to the strongly nonlinear model when the double layer is

driven out of quasiequilibrium at “very large” voltage.

2.2.1 Dimensionless form

For simplicity we assume a symmetric binary electrolyte with z

+

= −z

−

= z,

but we allow for asymmetry in the diﬀusivity D

+

= D

−

. The problem is

cast into dimensionless form, denoting dimensionless variables by an acute

accent,

´r =

r

ℓ

0

,

´

t =

t

τ

0

, ´ c

±

=

c

±

c

0

,

´

φ =

φ

φ

0

, ´ u =

u

u

0

, (2.40)

where ℓ

0

is the characteristic geometrical length scale,

τ

0

=

λ

D

ℓ

0

D

0

(2.41)

is the RC time for the system,

D

0

=

D

+

+D

−

2

(2.42)

is the average diﬀusivity, c

0

is the nominal bulk salt concentration,

φ

0

=

k

B

T

ze

(2.43)

2.2 Mathematical model 23

is the thermal voltage, and

u

0

=

ǫφ

2

0

ηℓ

0

(2.44)

is the characteristic electroosmotic velocity. In dimensionless form the gov-

erning equations become

1

ε

∂

´

t

´ c

±

= −

´

∇

´

F

±

, (2.45)

´

F

±

= −

´

D

±

´

∇´ c

±

± ´ c

±

´

∇

´

φ

+ Pe ´ u´ c

±

, (2.46)

−ε

2

´

∇

2

´

φ = ´ ρ =

1

2

(´ c

+

− ´ c

−

), (2.47)

1

Sc

¸

1

ε

∂

´

t

´ u + Pe(´ u

´

∇)´ u

= −

´

∇´ p +

´

∇

2

´ u −

1

ε

2

´ ρ

´

∇

´

φ (2.48)

0 =

´

∇ ´ u. (2.49)

Here ε = λ

D

/ℓ

0

is the dimensionless thickness of the Debye layer,

Pe =

u

0

ℓ

0

D

0

=

ǫφ

2

0

ηD

0

(2.50)

is the P´eclet number deﬁned as the ratio of the convective to the diﬀusive

ion ﬂux, and

Sc =

η

ρ

m

D

0

=

ν

D

0

(2.51)

is the Schmidt number deﬁned as the ratio of momentum diﬀusion to ionic

diﬀusion in the electrolyte. The factors of 1/ε in front of the time deriva-

tives in Eqs. (2.45) and (2.48) arise because we have chosen to focus on the

dynamics at the RC time

´

t and not the bulk diﬀusion time

¯

t = ε

´

t. We also

introduced dimensionless diﬀusion constants

´

D

+

=

D

+

D

0

=

2D

r

1 +D

r

and

´

D

−

=

D

−

D

0

=

2

1 +D

r

, (2.52)

where D

r

= D

+

/D

−

is the ratio of the cation to the anion diﬀusivity. By

construction they add up to

´

D

+

+

´

D

−

= 2, whereas (half) their diﬀerence

becomes

∆

´

D =

1

2

(

´

D

+

−

´

D

−

) =

D

r

−1

D

r

+ 1

, (2.53)

and is always in the range −1 < ∆

´

D < 1. It is convenient to operate also

with the average ion concentration

´ c =

1

2

(´ c

+

+ ´ c

−

), (2.54)

which we often denote as the “salt” concentration. The characteristic dif-

fusion constant for this quantity turns out to be the harmonic mean of the

individual diﬀusivities

´

D

a

=

´

D

+

´

D

−

=

4D

r

(1 +D

r

)

2

. (2.55)

24 Electrokinetic theory

Symbol Expression Value Unit Scaling

ν 1 10

−6

m

2

s

−1

D

0

2 10

−9

m

2

s

−1

φ

0

k

B

T/ze 25 mV

c

0

6 10

23

m

−3

ℓ

0

10 µm

λ

D

ǫk

B

T/2c

0

z

2

e

2

10 nm ∝ 1/

√

c

0

τ

0

λ

D

ℓ

0

/D

0

50 µs ∝ ℓ

0

/

√

c

0

u

0

ǫφ

2

0

/ηℓ

0

50 µms

−1

∝ 1/ℓ

0

ε λ

D

/ℓ

0

0.001 ∝ 1/ℓ

0

√

c

0

δ ǫλ

S

/ǫ

S

λ

D

1 ∝

√

c

0

Pe ǫφ

2

0

/ηD

0

0.25

Sc ν/D

0

5 10

2

Re ǫφ

2

0

/ην 5 10

−4

≡ Pe/Sc

Table 2.1: Typical numbers with and without dimensions, corresponding to a dilute

1 mM KCl solution at room temperature in a device with 10 µm characteristic length.

This is denoted the ambipolar diﬀusion constant and it is always in the range

0 <

´

D

a

≤ 1.

Table 2.1 shows typical values for the diﬀerent numbers with or without

dimension in the problem, corresponding to a dilute 1 mM KCl solution at

room temperature in a device with characteristic geometrical length scale

ℓ

0

= 10 µm. Notice that the Reynolds number Re = Pe/Sc = u

0

ℓ

0

/ν is

small, and also independent of ℓ

0

because the characteristic velocity scales

as 1/ℓ

0

. The nonlinear convection term in the Navier–Stokes equation can

be neglected if [Re ´ u[ ≪ 1 which is usually the case, even at large driving

voltage; in the numerical example above this corresponds to [u[ ≪0.1 m/s.

The relative diﬀusivity is typically in the range 0.1 < D

r

< 10, but the

common working electrolyte KCl is actually very close to symmetric with

D

r

≈ 0.97.

For the remainder of this thesis we drop the acute accent on dimension-

less variables to reduce the notational overload; it should be obvious from

the context whether a given quantity is dimensionless or not.

2.2.2 Electroneutral bulk transport

The standard procedure of matched asymptotic expansions begins by seeking

regular expansions (denoted here by bar accents) in the bulk “outer” region,

cf. Eq. (2.39) and similarly for

¯

φ, ¯ u, and ¯ p. Substituting ﬁrst into the Poisson

equation Eq. (2.47) and equating like powers in ε we ﬁnd that the bulk

2.2 Mathematical model 25

charge density ¯ ρ vanishes to both zeroth and ﬁrst order in ε. For a binary

electrolyte this implies that the leading order ion concentrations are equal,

i.e.,

¯ c

+

= ¯ c

−

= ¯ c. (2.56)

From the Nernst–Planck equation Eq. (2.45) we ﬁnd at leading order that

∂

t

¯ c = 0, (2.57)

i.e., the leading order bulk concentration ¯ c = ¯ c(r) is stationary in time,

with the dynamics showing up only as higher order corrections. Bulk charge

neutrality together with the Nernst–Planck equation also imply that the

leading order electric current density

¯

J =

1

2

(

¯

F

+

−

¯

F

−

) is divergence free

∇

¯

J = 0, (2.58)

where

¯

J = −¯ c∇

¯

φ −∆D∇¯ c. (2.59)

In the case of equal diﬀusivities where ∆D = 0 this reduces to an Ohmic

current with spatially varying conductivity ¯ c, but more generally the second

term related to the diﬀusion potential must also be taken into account. With

this, the leading order charge density (that is to second order in ε) can be

determined as

¯ ρ = −ε

2

∇

2

¯

φ = ε

2

∇¯ c ∇

¯

φ + ∆D∇

2

¯ c

¯ c

, (2.60)

using Eqs. (2.58) and (2.59) to eliminate ∇

2

¯

φ.

We focus on the periodic solution reached after all transients in the

system have died out. In order for a periodic ﬁrst order correction to exist,

it is necessary that the leading order time average ﬂuxes are divergence free,

i.e., that ∇ '

¯

F

±

` = 0. Writing the mean ion ﬂux as

¯

F =

1

2

(

¯

F

+

+

¯

F

−

) =

¯

F

a

+ ∆D

¯

J, (2.61)

where

¯

F

a

=

1

2

(D

−

¯

F

+

+D

+

¯

F

−

) = −D

a

∇¯ c + Pe ¯ u¯ c, (2.62)

we then obtain

2

−∇ '

¯

F

a

` = ∇

D

a

∇¯ c −Pe '¯ u`¯ c

= 0. (2.63)

At leading order in the Navier–Stokes equation for the ﬂuid motion we obtain

∂

t

¯ u = 0, (2.64)

2

It is an interesting feature of binary electrolytes that the evolution of the leading order

bulk concentration proﬁle – at the bulk diﬀusion time scale

¯

t = εt – is determined by a

simple convection-diﬀusion equation with no migration term, and with a diﬀusivity that

is simply the harmonic mean of the individual ion diﬀusivities, namely, the ambipolar

diﬀusion constant Da [54].

26 Electrokinetic theory

i.e., the leading order ﬂuid velocity ¯ u = ¯ u(r) is stationary and determined

by a time average force balance

0 = −∇¯ p +∇

2

¯ u +'∇

2

¯

φ∇

¯

φ`, (2.65)

with the pressure ¯ p determined by the incompressibility constraint

∇ ¯ u = 0. (2.66)

We denote as “bulk electroconvection” that component of the overall ﬂuid

motion that is induced by the bulk electrical body force

¯

f =

∇

2

¯

φ∇

¯

φ

**to distinguish it from, e.g., electroosmotic ﬂow induced in the Debye layer.
**

Because momentum diﬀusion is typically much faster than ionic diﬀusion

corresponding to a large Schmidt number Sc ≫ 1, cf. Table 2.1, the ﬁrst

order correction of O(εSc) may be not-so-small for the ε of interest to a

microsystem. It may therefore be more appropriate to state the leading

order problem as

1

εSc

∂

t

¯ u = −∇¯ p +∇

2

¯ u +∇

2

¯

φ∇

¯

φ, (2.67)

for εSc ∼ O(1). In any case the time average velocity entering Eq. (2.63)

remains the same.

2.2.3 Diﬀusion layer

The Debye layer at the electrodes periodically absorbs and expels an excess

amount of ions. However, because the bulk transport of neutral salt is

eﬀectively a diﬀusion process with diﬀusion constant D

a

and we consider

driving frequencies around the inverse RC time, the disturbances do not have

time to propagate into the bulk region, and the leading order dynamics only

occur in a diﬀusion layer of width O(

√

ε) around the electrode. Assuming

the electrode is smooth on the diﬀusion length scale we model it locally as

a ﬂat surface and deﬁne x as the tangential and y as the normal direction.

Further, we introduce a scaled spatial variable ˇ y = y/

√

ε and seek regular

asymptotic expansions (denoted by check accents) in the “middle” diﬀusion

layer.

Substituting into the Poisson equation (2.47) we ﬁnd immediately that

the charge density vanishes to zeroth order in ε, so that again the leading

order ion concentrations are equal

ˇ c

+

= ˇ c

−

= ˇ c. (2.68)

To determine the dynamical solution we introduce the excess concentration

ˇ γ = ˇ c − ¯ c in the diﬀusion layer relative to the bulk, and form weighted

2.2 Mathematical model 27

averages of the ionic ﬂuxes similar to Eq. (2.62) to get

∂

t

ˇ γ = −∇ (

ˇ

F

a

−

¯

F

a

)

= D

a

∂

2

ˇ y

ˇ γ +εD

a

∂

2

x

ˇ γ −ε Pe

ˇ u

x

∂

x

ˇ c +

1

√

ε

ˇ u

y

∂

ˇ y

ˇ c − ¯ u

x

∂

x

¯ c − ¯ u

y

∂

y

¯ c

= D

a

∂

2

ˇ y

ˇ γ +O(ε). (2.69)

At leading order the excess concentration in the diﬀusion layer is therefore

determined by a simple 1D diﬀusion problem for ˇ y ∈ [0, ∞[. We argue

below that the normal velocity component u

y

vanishes at leading order,

which means that convection is indeed negligible provided [Pe u

x

∂

x

c[ ≪ε

−1

.

Matching to the bulk solution requires that ˇ γ = 0 at ˇ y = ∞ whereas the

boundary condition at ˇ y = 0 is determined by the excess ﬂux fed into the

diﬀusion layer from the Debye layer

−

1

√

ε

D

a

∂

ˇ y

ˇ γ =

ˇ

F

y,a

−

¯

F

y,a

ˇ y=0

≡

ˇ

f

a

. (2.70)

The solution can be expressed as a Fourier series

ˇ γ(ˇ y, t) =

ε

ωD

a

∞

¸

n=1

ˆ

f

a,n

√

in

e

inωt−

√

inω/Daˇ y

+ c.c. (2.71)

where

ˆ

f

a,n

is the nth Fourier component of

ˇ

f

a

(t)

ˆ

f

a,n

=

T

0

ˇ

f

a

(t) e

−inωt

dt. (2.72)

Equivalently, the solution can be written as a convolution integral

ˇ γ =

ε

ωD

a

1

T

T

0

((ˇ y, t −t

′

)

ˇ

f

a

(t

′

) dt

′

(2.73)

where ( the periodic 1D diﬀusion kernel (Green’s function)

((ˇ y, t −t

′

) =

∞

¸

n=1

1

√

in

e

inω(t−t

′

)−

√

inω/Daˇ y

+ c.c. (2.74)

In the weakly nonlinear regime where the excess ﬂux

ˇ

f

a

is O(1) the excess

concentration ˇ γ only exist as an O(

√

ε) small perturbation to the bulk con-

centration; the strongly nonlinear regime at driving voltages V

0

∼ 1/

√

ε is

then essentially deﬁned as the point where the perturbation reaches O(1).

Since the diﬀusion layer is charge neutral, the current must be constant

across it at leading order, i.e.,

ˇ

J

y

= −

1

√

ε

ˇ c∂

ˇ y

ˇ

φ + ∆D∂

ˇ y

ˇ c

=

¯

J

y

= −

¯ c∂

y

¯

φ + ∆D∂

y

¯ c

. (2.75)

28 Electrokinetic theory

Introducing the excess potential

ˇ

ψ =

ˇ

φ −

¯

φ and rearranging a bit we obtain

∂

ˇ y

ψ = −

√

ε

¯

J

y

1

ˇ c

−

1

¯ c

−∆D∂

ˇ y

log

ˇ c/¯ c

, (2.76)

where we used ∂

ˇ y

¯

φ =

√

ε∂

y

¯

φ and similarly for ¯ c. Integrating once we get

ˇ

ψ(ˇ y) =

√

ε

¯

J

y

∞

ˇ y

1

ˇ c(ˇ y

′

)

−

1

¯ c

dˇ y

′

−∆Dlog

ˇ c(ˇ y)/¯ c

. (2.77)

The interpretation is straightforward: The last term is the diﬀusion po-

tential due to the diﬀerence in concentration across the layer, whereas the

ﬁrst term is just the excess potential drop required to drive the bulk cur-

rent through the diﬀusion layer when the local resistivity 1/ˇ c diﬀers from

the bulk value 1/¯ c. In the weakly nonlinear regime where ˇ c is only a small

perturbation to the bulk concentration ¯ c, the magnitude of the diﬀusion

potential −∆Dlog(ˇ c/¯ c) is only O(

√

ε), whereas the excess Ohmic poten-

tial drop is O(ε). In the strongly nonlinear regime both terms reach O(1)

because log(ˇ c/¯ c) ∼ O(1) and

¯

J

y

∼ 1/

√

ε.

3

Concerning the ﬂuid motion induced in the diﬀusion layer, the incom-

pressibility constraint immediately yields ˇ u

y

= 0. In the normal direcion

the Navier–Stokes equation reduces to a force balance between the excess

electrical body force and the pressure at leading order

0 = −∂

ˇ y

ˇ p +

1

ε

∂

2

ˇ y

ˇ

ψ

∂

ˇ y

ˇ

ψ = ∂

ˇ y

− ˇ p +

1

2ε

∂

ˇ y

ˇ

ψ

2

, (2.79)

from which

ˇ p =

1

2ε

∂

ˇ y

ˇ

ψ

2

. (2.80)

In the tangential direction we obtain a balance between the viscous and

electrical forces and the tangential pressure gradient

0 = −ε∂

x

ˇ p +∂

2

ˇ y

ˇ u

x

+∂

2

ˇ y

ˇ

ψ∂

x

ˇ

φ. (2.81)

Neglecting terms proportional to ∂

x

ˇ

ψ because these are much smaller than

∂

x

¯

φ we arrive at a Helmholtz–Smoluchowski type tangential velocity com-

ponent

ˇ u

x

= (ˇ ς −

ˇ

ψ)∂

x

¯

φ, (2.82)

where ˇ ς =

ˇ

ψ(0) is the total excess potential drop across the diﬀusion layer.

3

To make a rough estimate of the integral factor we can approximate ˇ c by a piecewise

linear proﬁle ˇ c ≈ ˇ co + (¯ c − ˇ co)

ω/Da ˇ y for ˇ y ≤

Da/ω and ˇ c ≈ ¯ c for ˇ y ≥

Da/ω to get

∞

0

1

ˇ c

−

1

¯ c

dˇ y ≈ −

Da

ω

¸

log(ˇ co/¯ c)

¯ c − ˇ co

+

1

¯ c

∼ O(1). (2.78)

2.2 Mathematical model 29

2.2.4 Quasiequilibrium double layer

The singular perturbation in the Poisson equation gives rise to a boundary

layer of width ε where the charge density is nonzero to zeroth order in

ε. Introducing a scaled spatial variable ˜ y = y/ε to remove the singular

perturbation, we can seek regular asymptotic expansions (denoted by tilde

accents) in the Debye “inner” layer. Substituting into the Nernst–Planck

equations Eqs. (2.45) and (2.46) and using ∂

˜ y

= ε∂

y

we ﬁnd at leading order

0 = ∂

˜ y

∂

˜ y

˜ c

±

± ˜ c

±

∂

˜ y

˜

φ

, (2.83)

that is, the Debye layer is in quasiequilibrium and the ions are Boltzmann

distributed

˜ c

±

= ˇ ce

∓

˜

ψ

, (2.84)

where

˜

ψ =

˜

φ −

ˇ

φ is the excess potential in the Debye layer relative to the

bulk, and ˇ c is the limiting value of the concentration in the diﬀusion layer

seen from the Debye layer at ˇ y = 0. Likewise for the charge density and

average ion concentration in the Debye layer we have

˜ ρ = −ˇ c sinh(

˜

ψ) and ˜ c = ˇ c cosh(

˜

ψ). (2.85)

The quasiequilibrium arises because the Debye layer reaches local equilib-

rium on the Debye time scale

˜

t = t/ε which is much faster than the RC

time of the system. Strictly, Eq. (2.83) does not imply that the leading

order ﬂuxes

˜

F

y,±

= −

1

ε

D

±

∂

˜ y

˜ c

±

± ˜ c

±

∂

˜ y

˜

φ

(2.86)

vanish but only that they are constant; it is by matching with bulk ﬂuxes

≪ O(1/ε) that we arrive at the Boltzmann proﬁles. Substituting into the

Poisson equation, the leading order problem is the dimensionless Poisson–

Boltzmann equation

∂

2

˜ y

˜

ψ = ˇ c sinh(

˜

ψ), (2.87)

which has the well-known Gouy–Chapman solution

˜

ψ = 4 tanh

−1

tanh

˜

ζ/4

e

−

√

ˇ c ˜ y

. (2.88)

The integration constant

˜

ζ is simply the leading order zeta potential, and

1/

√

ˇ c is the local eﬀective Debye length.

As in the diﬀusion layer, the incompressibility constraint immediately

yields ˜ u

y

= 0 at leading order, and the balance between the electrical body

force and the pressure in the normal direction that gives rise to an (excess)

osmotic pressure of

˜ p =

1

2ε

2

∂

˜ y

˜

ψ

2

. (2.89)

30 Electrokinetic theory

Substituting into the Navier–Stokes equation in the tangential direction we

obtain at leading order

0 = −ε

2

∂

x

˜ p +∂

2

˜ y

˜ u

x

− ˜ ρ ∂

x

˜

φ. (2.90)

Writing ∂

x

˜

φ = ∂

x

ˇ

φ+∂

x

˜

ψ and using again ˜ ρ = −∂

2

˜ y

˜

ψ we arrive at the general

expression for electroosmotic ﬂow of the ﬁrst kind

4

˜ u

x

= (

˜

ζ −

˜

ψ)∂

x

ˇ

φ + 4 log

¸

cosh(

˜

ψ/4)

cosh(

˜

ζ/4)

∂

x

log ˇ c. (2.91)

The ﬁrst term is the familiar Helmholtz–Smoluchowski velocity due to the

action of the tangential ﬁeld from the diﬀusion layer, and the second term is

the diﬀusioosmotic ﬂow due to tangential gradients in the osmotic pressure

in the Debye layer.

2.2.5 Surface conservation laws

The redistribution of ions across the Debye layer in the normal direction is

instantaneous on the (RC) time scale that we consider here, but the total

amount of ions accumulated locally can change only by ﬂux in or out of

the layer or by tangenial ﬂuxes within the layer. Following Bazant and

Chu [37, 39] we quantify this by deﬁning ε ˜ w

±

(x, t) as the excess amount of

each ionic species accumulated locally in the Debye layer, such that

˜ w

±

=

1

ε

Debye layer

˜ c

±

− ˇ c

±

dy =

∞

0

(˜ c

±

− ˇ c

±

) d˜ y. (2.92)

Likewise we deﬁne the surface charge and average excess concentration by

˜ q =

∞

0

(˜ ρ − ˇ ρ) d˜ y and ˜ w =

∞

0

(˜ c − ˇ c) d˜ y, (2.93)

respectively, where by construction ˜ w

±

= ˜ w ± ˜ q. At leading order the

integrals can be evaluated using Eqs. (2.84) and (2.88) to get

˜ w

±

= 2

√

ˇ c

e

∓

˜

ζ/2

−1

, (2.94)

from which also

˜ q = −2

√

ˇ c sinh(

˜

ζ/2) and ˜ w = 4

√

ˇ c sinh

2

(

˜

ζ/4). (2.95)

4

The notion “electroosmosis of the ﬁrst kind” was introduced by Dukhin and co-workers

to distinguish it from their “electroosmosis of the second kind” in double layers driven out

of quasiequilibrium [41, 42].

2.2 Mathematical model 31

The time evolution of ˜ w

±

is determined by

∂

t

˜ w

±

=

∞

0

∂

t

˜ c

±

− ˇ c

±

d˜ y (2.96)

= −

∞

0

ε∂

x

˜

F

x,±

−

ˇ

F

x,±

+∂

˜ y

˜

F

y,±

−

ˇ

F

y,±

d˜ y (2.97)

= −ε∂

x

∞

0

˜

F

x,±

−

ˇ

F

x,±

d˜ y −

˜

F

y,±

−

ˇ

F

y,±

∞

˜ y=0

(2.98)

= −ε∂

x

˜

Γ

x,±

−

ˇ

F

y,±

+F

±,ext

. (2.99)

Here we used the continuity equation Eq. (2.45) to eliminate the time deriva-

tive. Next we partially integrated the normal ﬂux, used the ﬂux matching

condition

lim

˜ y→∞

˜

F

y,±

∼ lim

ˇ y→0

ˇ

F

y,±

, (2.100)

and wrote

˜

F

y,±

[

˜ y=0

= F

±,ext

for the ﬂux injected into the Debye layer at the

electrode surface, e.g., by electrochemical reaction. Finally, we introduced

the surface excess ﬂux

˜

Γ

x,±

given by

˜

Γ

x,±

=

∞

0

˜

F

±,x

−

ˇ

F

±,x

d˜ y. (2.101)

At leading order we can evaluate Γ

x,±

by rewriting the ﬂux in terms of the

electrochemical potential µ

±

= log c

±

±φ to get F

±

= −D

±

c

±

∇µ

±

+Pe uc

±

.

Then

˜

Γ

±,x

= −D

±

∞

0

(˜ c

±

∂

x

˜ µ

±

− ˇ c

±

∂

x

ˇ µ

±

) d˜ y + Pe

∞

0

(˜ c

±

˜ u

x

− ˇ c

±

ˇ u

x

) d˜ y

= −D

±

˜ w

±

∂

x

ˇ µ

±

+ Pe

∞

0

(˜ c

±

˜ u

x

− ˇ c

±

ˇ u

x

) d˜ y, (2.102)

where we used that µ

±

is constant across the Debye layer because it is in

quasiequilibrium at leading order. This is a powerful result because it is

independent of the particular model for the double layer [39], i.e., it remains

valid when the dilute approximation breaks down at large voltage, and may

even be generalized to account for interactions between the diﬀerent species

by replacing the diﬀusivities D

±

by generalized mobilities L

ij

, cf. Eq. (2.3).

However, in order to handle the convection term we do need a model, and

using Eq. (2.91) we get

∞

0

(˜ c

±

˜ u

x

− ˇ c ˇ u

x

) d˜ y =−2

± ˜ w

±

+

√

ˇ c

J

‡

˜

ζ

+

˜

ζ

¸

∂

x

ˇ

φ (2.103)

−2

˜ w

±

+

√

¯ c

J

†

˜

ζ

±

˜

ζ

¸

∂

x

log ˇ c.

Here J

†

and J

‡

are shorthands for the expressions

J

†

(ζ) = −8 log[cosh(ζ/4)] −

1

2

Li

2

[tanh

2

(ζ/4)] ≈ −

9

32

ζ

2

+

7

2048

ζ

4

(2.104)

32 Electrokinetic theory

and

J

‡

(ζ) = Li

2

[tanh(ζ/4)] −Li

2

[−tanh(ζ/4)] ≈

1

2

ζ −

1

144

ζ

3

, (2.105)

and Li

2

(z) =

¸

∞

n=1

z

n

/n

2

is the dilogarithm.

Surface conservation laws of this form were used already by Dukhin,

Deryagin, and Shilov in their studies of surface conductance and the polar-

ization of thin but highly charged double layers around spherical particles in

weak applied ﬁelds [18, 63, 64, 65]. Recently Chu and Bazant revisited the

problem using matched asymptotic expansions which allowed them to lift the

requirements for weak applied ﬁelds and small deviations from equilibrium

in the bulk [38, 39]. What we have contributed is to extend their work to

account also for surface convection and evaluate the integral in Eq. (2.103).

Dukhin and co-workers also accounted for electroosmotic convection and our

result is fully equivalent to their work except for two features: First, in the

English translation it appears that there is a sign error on their term for ﬂux

by electroosmotic convection proportional to ∂

x

log ˇ c, which in our notation

reads 2

¸

± ˜ w

±

+

˜

ζ − 8 log[cosh(

˜

ζ/4)]

¸

∂

x

log ˇ c. For a symmetric electrolyte

a change of sign of

ˇ

φ and

˜

φ while retaining ˇ c unchanged must result in a

symmetric interchange of the cation and anion ﬂuxes. Our Eq. (2.103) sat-

isﬁes this property, but the results in Refs. [18, 65] do not. Secondly, the

dilogarithm terms are absent in their work.

The presence of the dilogarithm terms qualitatively changes the form

of the convective ﬂux at low zeta potential: The slope of the coeﬃcient to

∂

x

ˇ

φ becomes nonzero at

˜

ζ = 0, and coeﬃcient to ∂

x

log ˇ c becomes a non-

monotonic function of

˜

ζ. However, these subtle changes will be very hard

to detect experimentally due to the factor of ε multiplying the divergence

of the surface ﬂuxes in Eq. (2.99). Only at high zeta potential is the excess

amount of ions accumulated in the Debye layer large enough to make the

surface ﬂux term signiﬁcant, which can occur either for colloidal particles

with a large intrinsic zeta potential determined by their surface chemistry,

or for electrodes or polarizable objects subjected to large external voltages

or applied ﬁeld. In this regime the surface excess conductance is dominated

by the factors ˜ w

±

that grow exponentially with

˜

ζ and the expression for the

surface excess ﬂux simpliﬁes to

˜

Γ

±,x

= −(D

±

+ 2Pe) ˜ w

±

∂

x

ˇ µ

±

. (2.106)

Using this we can write the surface excess current ˜

x

=

1

2

(

˜

Γ

+,x

−

˜

Γ

−,x

) and

neutral salt ﬂux

˜

Γ =

1

2

(

˜

Γ

+,x

+

˜

Γ

−,x

) as

˜

x

= −

(1 + 2Pe) ˜ w + ∆D˜ q

∂

x

ˇ

φ −

(1 + 2Pe)˜ q + ∆D ˜ w

∂

x

log ˇ c, (2.107)

and

˜

Γ

x

= −

(1 + 2Pe)˜ q + ∆D ˜ w

∂

x

ˇ

φ −

(1 + 2Pe) ˜ w + ∆D˜ q

∂

x

log ˇ c, (2.108)

2.2 Mathematical model 33

where we used that ˜ w

+

+ ˜ w

−

= 2 ˜ w, ˜ w

+

− ˜ w

−

= 2˜ q, D

+

+ D

−

= 2, and

D

+

−D

−

= 2∆D.

To close the coupled electrical problem for

¯

φ and ¯ c in the bulk, ˜ q and

˜ w in the Debye layer, and ˇ c in the diﬀusion layer, we need another two

boundary conditions. The ﬁrst is the potential drop across the boundary

layers obtained from the Gouy–Chapman–Stern model

V

ext

−

¯

φ = ˇ ς +

˜

ζ − ˜ q δ, (2.109)

where ˇ ς,

˜

ζ, and −˜ q δ are the potential drops over the diﬀusion, Debye, and

Stern layers, respectively, and δ is the surface capacitance ratio. This acts

as a Dirichlet condition for the bulk potential

¯

φ. The second condition is a

requirement of consistency in Eq. (2.95) as

˜ w = 4

√

ˇ c sinh

2

(

˜

ζ/4) =

˜ q

2

+ 4ˇ c −

√

4ˇ c, (2.110)

where we used

˜

ζ = −2 sinh

−1

˜ q/2

√

ˇ c

**to express ˜ w directly in terms of the
**

dynamical variables ˜ q and ˇ c.

With this we are ready to formulate in the following sections the leading

order dynamical models corresponding to the linear, weakly nonlinear, and

strongly nonlinear regimes.

2.2.6 Linear model

In the Debye–H¨ uckel limit

˜

ζ ≪ 1 where the potential is small compared to

the thermal voltage the electrokinetic problem can be linearized. This sim-

pliﬁes the analysis signiﬁcantly and this regime has therefore been studied

extensively in the literature. Since we drive the system with a harmonic

signal, at leading order all variables can be expressed as an equilibrum value

with a small deviation that varies harmonically in time

φ(r, t) = Re

ˆ

φ(r)e

iωt

=

1

2

ˆ

φ(r)e

iωt

+ c.c. (2.111)

and likewise for the ionic concentrations

c

i

(r, t) = c

∗

i

+ Re

ˆ c

i

(r)e

iωt

. (2.112)

Because we focus on driving frequencies around the inverse RC time, the

perturbations do not have time to propagate into the bulk, so the bulk

concentration is uniform and Eq. (2.58) reduces to a Laplace problem for

ˆ

φ

−∇

2

ˆ

φ = 0. (2.113)

The surface conservation law for charge accumulation in the Debye layer

becomes

iωˆ q = ˆ

ext

+n ∇

ˆ

φ, (2.114)

34 Electrokinetic theory

where ˆ

ext

is the Faradaic current injection at the electrode and n ∇

ˆ

φ is the

Ohmic current running into the Debye layer from the bulk. The capacitance

of the double layer at low voltage is C

dl

= 1/(1 + δ) so the accumulated

charge is related to the potential drop over the double layer by

ˆ q = −

ˆ

V

ext

−

ˆ

φ

1 +δ

. (2.115)

In the absense of Faradaic electrode reactions the boundary condition for

the potential therefore becomes [23]

−n ∇

ˆ

φ =

iω

1 +δ

ˆ

V

ext

−

ˆ

φ

. (2.116)

The slip velocity condition for the time average bulk ﬂuid motion reduces

to [33]

˜ u

s

=

˜

ζ∇

s

¯

φ

=

1

2

Re

ˆ

ζ

∗

∇

s

ˆ

φ

=

1

2(1 +δ)

Re

(

ˆ

V

ext

−

ˆ

φ)

∗

∇

s

ˆ

φ

= −

1

4(1 +δ)

∇

s

ˆ

V

ext

−

ˆ

φ

2

, (2.117)

where we used that

˜

ζ = (V

ext

−

¯

φ)/(1+δ) at low voltage and that ∇

s

V

ext

= 0

because the metal electrode is an equipotential surface. Since the bulk con-

centration is uniform there is no body force at leading order and the Stokes

ﬂow problem for the bulk ﬂuid motion is driven entirely by the electroos-

motic slip.

Faradaic current injection

The Faradaic current is determined from Eq. (2.32) as

ˆ

ext

=

1

R

ct

(

ˆ

V

ext

−

ˆ

φ) +

ˆ c

R

c

∗

R

−

ˆ c

O

c

∗

O

, (2.118)

where R

ct

= σk

B

T/j

0

eℓ

0

is the dimensionless charge transfer resistance,

and ˆ c

O

and ˆ c

R

are the concentrations of the oxidized and reduced species

immediately outside the Debye layer. Assuming we can neglect migration

of O and R in the diﬀusion layer we employ the result from Sec. 2.2.3 to get

ˆ c

O

[

ˇ y=0

=

ε

iωD

O

ˆ

f

O

(2.119)

and similarly for ˆ c

R

. Approximating the ﬂux

ˆ

f

O

into the diﬀusion layer

by the external current ˆ

ext

, which is equivalent to neglecting accumulation

of O in the Debye layer, and similarly approximating

ˆ

f

R

by −ˆ

ext

we then

arrive at

ˆ

ext

=

1

R

ct

(

ˆ

V

ext

−

ˆ

φ) −

1

c

∗

R

ε

iωD

R

ˆ

ext

−

1

c

∗

O

ε

iωD

O

ˆ

ext

. (2.120)

2.2 Mathematical model 35

R

ct

C

dl

Z

W

R

b

V

ext

Figure 2.5: Equivalent circuit diagram for the linear model: The double layer capaci-

tance C

dl

is connected in parallel with the charge transfer resistance Rct for the Faradaic

electrode reaction and the Warburg impedance

ˆ

ZW(ω) describing the eﬀects of mass trans-

fer. All are connected in series to the bulk Ohmic resistance R

b

.

This can be rearranged to

ˆ

j

ext

=

ˆ

V

ext

−

ˆ

φ

R

ct

+

ˆ

Z

W

(ω)

, (2.121)

where

ˆ

Z

W

(ω) =

√

ε

c

∗

O

D

O

+

√

ε

c

∗

R

D

R

1

√

iω

(2.122)

is known as the Warburg impedance [16]. The expression in the parenthesis

is called the Warburg constant and we denote it by R

mt

5

R

mt

=

√

ε

c

∗

O

D

O

+

√

ε

c

∗

R

D

R

. (2.123)

Finally, substituting into the surface charge conservation law we obtain a

closed relation between the complex bulk current running into the double

layer and the potential drop across it

−n ∇

ˆ

φ =

¸

iω

1 +δ

+

1

R

ct

+

ˆ

Z

W

(ω)

ˆ

V

ext

−

ˆ

φ

. (2.124)

The equivalent circuit diagram for the entire system is shown in Fig. 2.5

with the double layer capacitance C

dl

= 1/(1 + δ) connected in parallel

to the charge transfer resistance R

ct

that is in series with

ˆ

Z

W

, while all

are connected in series to the bulk Ohmic resistance R

b

= 1. This way

of representing an electrochemical cell is a standard tool in electrochemi-

cal impedance spectroscopy which allows experimental parameters such as

R

ct

to be determined with high accuracy [16]. In such experiments it is

common practice to add an excess amount of supporting electrolyte that is

5

The Warburg constant is often denoted σ in electrochemistry but we prefer the non-

standard symbol Rmt to avoid confusion with the electrolyte conductivity.

36 Electrokinetic theory

electrochemically inert but serves to minimize the bulk Ohmic resistance.

This also justiﬁes the neglect of migration of O and R in the diﬀusion layer

and of accumulation of those species in the Debye layer. In our formulation

this corresponds to the limit c

∗

O

≪ 1 and c

∗

R

≪ 1. However, because AC

electroosmosis is strongly supressed at high electrolyte concentration it is

relevant to consider the case when, e.g., the species O also plays the role

of cation in the supporting electrolyte, i.e., z

O

= z

R

+ 1 = 1 and c

∗

O

= 1.

Performing exactly the same analysis as above but taking into account mi-

gration of O in the diﬀusion layer we get

ˆ c

O

=

ε

iωD

a

ˆ

f

a

, (2.125)

where D

a

is the ambipolar diﬀusion constant and the ﬂux

ˆ

f

a

is given by

ˆ

f

a

=

1

2

(D

−

ˆ

F

+

+D

+

ˆ

F

−

) =

1

2

D

−

ˆ

ext

−iω∆D ˆ q. (2.126)

Using this we ﬁnd a relation similar to Eq. (2.124) between the bulk current

and potential drop across the double layer

−n ∇

ˆ

φ =

¸

iω

1 +δ

+

1 −

√

iωε

∆D

(1+δ)

√

Da

R

ct

+

ˆ

Z

W

(ω)

(

ˆ

V

ext

−

ˆ

φ), (2.127)

where now

ˆ

Z

W

is given by

ˆ

Z

W

(ω) =

√

ε

c

∗

R

D

R

+

√

ε

2

√

D

r

1

√

iω

(2.128)

and D

r

= D

+

/D

−

. The term proportional to ∆D in the nominator in

Eq. (2.127) is related to the accumulation of O in the Debye layer, and it is

negligible at leading order due to the factor of

√

ε. Therefore the case when

O plays the role of cation in the supporing electrolyte (or R plays the role

of anion) is formally equivalent to the that of excess supporting electrolyte,

with the only diﬀerence being the deﬁnition of R

mt

.

The Warburg impedance reaches O(1) only at driving frequencies below

ω ∼ R

2

mt

∼ ε/ min¦c

∗

O

, c

∗

R

¦

2

. (2.129)

However, we must keep in mind that our model with thin diﬀusion layers

is valid only when ω is well above the inverse bulk diﬀusion time, i.e., for

ω ≫ε. The eﬀect of mass transfer is therefore signiﬁcant in our model only if

one of the electrochemically active species is present at a concentration much

lower than the supporting electrolyte, min¦c

∗

O

, c

∗

R

¦ ≪1. On the other hand,

at very low concentration min¦c

∗

O

, c

∗

R

¦ ≪

√

ε the mass transfer resistance is

so large that the Faradaic electrode reaction is negligible alltogether.

2.2 Mathematical model 37

2.2.7 Weakly nonlinear model

Our notion of the “weakly nonlinear” regime is characterized by two features

in the solution, namely,

• The bulk electrolyte concentration is uniform and given by ¯ c = 1.

• The diﬀusion layer around the electrodes exists only as a small per-

turbation to the bulk solution.

When these conditions are satisﬁed the solution again simpliﬁes signiﬁcantly.

The leading order bulk problem is linear and given by

¯ c = 1, (2.130)

0 = −∇

2

¯

φ, (2.131)

0 = −∇¯ p +∇

2

¯ u, (2.132)

0 = ∇ ¯ u, (2.133)

which is matched with the following set of dynamical boundary conditions

in the Debye layer

∂

t

˜ q = n ∇

¯

φ +j

ext

, (2.134)

V

ext

−

¯

φ =

˜

ζ − ˜ qδ, (2.135)

¯ u = ˜ u

s

=

˜

ζ∇

s

¯

φ

, (2.136)

where the zeta potential is related to the charge accumulated in the Debye

layer by

˜

ζ = −2 sinh

−1

(˜ q/2). (2.137)

As we discuss below in Sec. 2.2.9 both the surface ﬂuxes in the Debye layer

and the perturbations in the local salt concentration in the diﬀusion layer

reach O(1) at suﬃciently large driving voltage V

0

∼

δ/ε. Our weakly

nonlinear model is therefore only valid at relatively low voltage

V

0

≪

δ/ε. (2.138)

For blocking electrodes where the Faradaic current j

ext

vanishes, the weakly

nonlinear model is described in terms of four dimensionless parameters,

namely the driving voltage V

0

and frequency ω, the capacitance ratio δ,

and the equilibrium zeta potential

˜

ζ

eq

on the electrodes. The electrical part

of the problem is nonlinear, but local in time so it is straightforward to in-

tegrate using, e.g., Femlab’s standard time-stepping algorithm. The ﬂow

problem is linear and can be computed from the time average slip velocity

once the electrical problem has been solved.

38 Electrokinetic theory

Faradaic current injection

When electrode reactions do occur the leading order Faradaic current j

ext

is determined from the current-overpotential equation Eq. (2.31)

j

ext

=

1

R

ct

e

−

˜

ζ/2

¸

ˇ c

R

c

∗

R

e

˜

ϑ/2

−

ˇ c

O

c

∗

O

e

−

˜

ϑ/2

, (2.139)

where R

ct

= σk

B

T/j

0

eℓ

0

is the dimensionless charge transfer resistance,

ϑ = (V

ext

−

¯

φ) − (V

ext

−

¯

φ)

eq

is the overpotential, ˇ c

O

and ˇ c

R

are the con-

centrations of the oxidized and reduced species immediately outside the

Debye layer, and we assume a symmetric transfer coeﬃcient α =

1

2

and take

z

O

= z

R

+1 = 1. To further simplify the solution we also assume that there

is excess supporting electrolyte such that c

∗

O

≪ 1. If this is not satisﬁed

the Faradaic reaction would invalidate our assumption of small perturba-

tions to the bulk concentration in the diﬀusion layer and make the problem

“strongly nonlinear” already at relatively low voltage V

0

∼ O(log ε) due to

the exponential voltage dependence in Eq. (2.139). For c

∗

O

≪ 1 we can ne-

glect migration of O in the diﬀusion layer and employ the 1D solution from

Sec. 2.2.3 to get

ˇ c

O

= c

∗

O

+

ε

ωD

O

1

T

T

0

ˇ

((t −t

′

)

ˇ

f

O

(t

′

) dt

′

, (2.140)

assuming the bulk concentration ¯ c

O

is at the equilibrium value c

∗

O

, and

similarly for ˇ c

R

. Using

ˇ

f

O

= j

ext

and

ˇ

f

R

= −j

ext

and substituting into

Eq. (2.139) we obtain a single integral equation for j

ext

(t), and after some

rearranging we arrive at

1

T

T

0

ˇ

((t −t

′

)j

ext

(t

′

) dt

′

=

√

ω

2 sinh

˜

ϑ/2

−R

ct

e

˜

ζ/2

j

ext

(t)

R

R

mt

e

˜

ϑ/2

+R

O

mt

e

−

˜

ϑ/2

. (2.141)

Here the constants R

O

mt

and R

R

mt

are the contributions from O and R to the

Warburg constant, cf. the linear analysis in Sec. 2.2.6

R

O

mt

=

1

c

∗

O

ε

D

O

and R

R

mt

=

1

c

∗

R

ε

D

R

. (2.142)

At large positive overpotential the Faradaic current j

ext

saturates at the

limiting value j

R

lim

(ω) = D

R

c

∗

R

/ℓ

R

(ω) = 1/R

R

mt

√

ω, where ℓ

R

(ω) =

εD

R

/ω

is the diﬀusion length for species R, and similarly for O at large negatively

overpotential. The numerical solution of the problem is more complicated

than for the simple case with blocking electrodes because the solution is

nonlocal in time. In Appendix A.3 we discuss how it can be handled by

replacing the standard time evolution with a relaxation method to determine

the steady-state periodic solution in time while employing the commercial

ﬁnite element software Femlab to solve the spatial problem [47].

2.2 Mathematical model 39

2.2.8 Strongly nonlinear model

Our notion of the “strongly nonlinear” regime is characterized by the fol-

lowing features in the solution

• The surface excess ﬂux

˜

Γ in the Debye layer induces O(1) bulk elec-

trolyte concentration gradients, although the leading order bulk con-

centration is constant in time ¯ c = ¯ c(r).

• In the diﬀusion layer the perturbations to the local electrolyte concen-

tration caused by the periodic massive uptake and release of ions in

the Debye layer reach O(1).

• The Debye layer remains in quasiequilibrium but the solution depends

on the local electrolyte concentration in the diﬀusion layer immediately

outside the Debye layer.

• The ﬂuid motion has components from both electroosmotic slip from

the Debye layer, an additional eﬀective slip from space charge in the

diﬀusion layer, and bulk electroconvection.

• We disregard Faradaic electrode reactions because the problem is com-

plicated enough without.

In the bulk we have four variables, namely, the potential

¯

φ(r, t), concentra-

tion ¯ c(r), and ﬂuid velocity and pressure ¯ u(r) and ¯ p(r), that are governed

by

0 = ∇

¯

J, (2.143)

0 = ∇

¯

F

a

, (2.144)

0 = −∇¯ p +∇

2

¯ u +'∇

2

¯

φ∇

¯

φ`, (2.145)

0 = ∇ ¯ u, (2.146)

where the electric current

¯

J and ambipolar ﬂux

¯

F

a

are given by

¯

J = −¯ c∇

¯

φ −∆D∇¯ c, (2.147)

¯

F

a

= −D

a

∇¯ c + Pe ¯ u¯ c. (2.148)

In the Debye layer we also have four dynamical variables, namely, the ac-

cumulated charge ˜ q and excess neutral salt ˜ w, together with the normal

current

¯

J

n

and salt ﬂux

ˇ

F

n

that act as Lagrange multipliers in the model

∂

t

˜ w = −

ˇ

F

n

−ε∇

s

˜

Γ, (2.149)

∂

t

˜ q = −

¯

J

n

−ε∇

s

˜, (2.150)

V

ext

−

¯

φ =

˜

ζ − ˜ q δ, (2.151)

˜ w =

˜ q

2

+ 4ˇ c −

√

4ˇ c, (2.152)

40 Electrokinetic theory

where

˜

ζ = −2 sinh

−1

˜ q/2

√

ˇ c

, (2.153)

and we neglected the potential drop ˇ ς across the diﬀusion layer in the con-

dition Eq. (2.151) for the overall potential drop over the boundary layers

because ˇ ς is small compared to −˜ q δ. In the expressions for the surface ex-

cess current ˜ and neutral salt ﬂux

˜

Γ we drop the diﬀusion term proportional

to ∇

s

log ˇ c because this is negligible compared to ∇

s

ˇ

φ at large voltage, and

moreover approximate the latter by ∇

s

¯

φ because ˇ ς is small; hence

˜ = −

(1 + 2Pe) ˜ w + ∆D˜ q

∇

s

¯

φ, (2.154)

˜

Γ = −

(1 + 2Pe)˜ q + ∆D ˜ w

∇

s

¯

φ. (2.155)

The Lagrange multipliers

¯

J

n

and

ˇ

F

n

determine the boundary conditions for

the bulk current and ambipolar ﬂux running into the Debye layer

n

¯

J =

¯

J

n

, (2.156)

n

¯

F

a

= '

ˇ

F

n

` −∆D'

¯

J

n

`, (2.157)

and the Debye layer also gives rise to an electroosmotic slip velocity of

˜ u

s

=

˜

ζ∇

s

¯

φ

, (2.158)

where again we dropped the term proportional to ∇

s

log ˇ c and approximated

the tangential ﬁeld by −∇

s

¯

φ.

In the diﬀusion layer we have just one dynamical variable, namely, the

concentration ˇ c, that is determined by

ˇ c(ˇ y, t) = ¯ c +

ε

ωD

a

1

T

T

0

((ˇ y, t −t

′

)

ˇ

f

a

(t

′

) dt

′

, (2.159)

where the ambipolar ﬂux entering

ˇ

f

a

is expressed in terms of

¯

J

n

and

ˇ

F

n

as

ˇ

f

a

=

ˇ

F

n

−∆D

¯

J

n

−'

ˇ

F

n

−∆D

¯

J

n

`. (2.160)

With this, the excess potential drop across the diﬀusion layer can be evalu-

ated

ˇ ς =

√

ε

¯

J

n

∞

0

¸

1

ˇ c

−

1

¯ c

**dˇ y −∆Dlog(ˇ c/¯ c), (2.161)
**

in terms of which the eﬀective slip velocity induced in the diﬀusion layer

becomes

ˇ u

s

=

ˇ ς∇

s

¯

φ

. (2.162)

Now, we already argued that ˇ ς is small so one might ask if ˇ u

s

is not neg-

ligible compared to the electroosmotic slip ˜ u

s

from the Debye layer. The

2.2 Mathematical model 41

point is that

˜

ζ scales only as

˜

ζ ∼ 2 log ˜ q ∼ 2 log(V

0

/δ) at large voltage, cf.

Eqs. (2.151) and (2.153), which is also a relatively weak dependence. More-

over, when the Gouy–Chapman solution in the Debye layer breaks down

due to steric exclusion there may be viscoelectric eﬀects suppressing the

electroosmotic ﬂow. While a description of that is beyond the scope of this

thesis, it is interesting to take into account other sources of electrokinetic

ﬂow that could start to dominate in that situation.

In order to simplify the numerical solution we do, however, assume that

the component of the bulk ﬂow induced by ˇ u

s

can be neglected in the

convection of neutral salt in Eq. (2.148). Upon this approximation, only

the concentration ˇ c at ˇ y = 0 enters the coupled dynamical problem for

¸

¯

φ, ¯ c, ¯ u, ¯ p, ˜ q, ˜ w,

¯

J

n

,

ˇ

F

n

, ˇ c

¸

, which is convenient. Once this problem is solved,

we can compute the full proﬁle for ˇ c as a function of ˇ y and evaluate ˇ ς along

the electrode surface, evaluating the integral over ˇ y by numerical quadra-

ture, and determine the slip velocity ˇ u

s

and the corresponding bulk ﬂow

component as a postprocessing step.

Finally, because the bulk concentration proﬁle is controlled only by ﬂux

(Neumann) boundary conditions we need an additional constraint to ﬁx the

total number of ions in the system, namely,

Ω

¯ c dr +ε

∂Ω

' ˜ w` ds =

Ω

c

∗

dr = [Ω[, (2.163)

because the nominal bulk concentration is exactly c

∗

= 1; the excess concen-

tration from the diﬀusion layer does not enter since it has no time average

by construction. Such a constraint would be redundant in an ordinary time

evolution problem, where the total amount of salt is ﬁxed by the initial con-

dition, but it is necessary in our case because we compute the steady-state

periodic solution directly using a relaxation method. Alternatively, if the

system is not closed but open and connected to a large reservoir of elec-

trolyte at unit concentration, the total amount of ions is of course not ﬁxed,

but the salt concentration level is determined by a Dirichlet condition ¯ c = 1

at the reservoir boundary. Likewise we need to impose a constraint on the

total amount of charge accumulated in the Debye layers

∂Ω

˜ q ds =

∂Ω

˜ q

eq

ds, (2.164)

where ˜ q

eq

= −2 sinh(

˜

ζ

eq

/2) is the equilibrium intrinsic surface charge. In the

special case where the ionic diﬀusivities are equal, ∆D = 0, and the intrinsic

zeta potential vanishes,

˜

ζ

eq

= 0, the problem has additional symmetry in

time: The bulk potential

¯

φ and ˜ q and

¯

J

n

in the Debye layer have odd

symmetry across one half period in time, e.g.,

¯

φ(r, t + T/2) = −

¯

φ(r, t),

whereas the rest of the solution components have even symmetry, e.g., ¯ c(r, t+

T/2) = ¯ c(r, t).

42 Electrokinetic theory

2.2.9 Beyond quasiequilibrium

In this section we ﬁrst develop an estimate for the driving voltage at which

the weakly nonlinear model breaks down, namely, V

0

∼

δ/ε. Around this

voltage both the bulk concentration gradients and the perturbations to the

bulk concentration in the diﬀusion layer reach O(1). As the driving voltage

is increased the electrolyte concentration in the diﬀusion layer approaches

zero, which makes the assumption of local charge neutrality break down and

with that also the leading order strongly nonlinear solution. The voltage

where this occurs is also around V

0

∼

**δ/ε, so it would seem that the
**

strongly nonlinear model applies only to a relatively narrow voltage range,

although certainly an experimentally relevant one. At even larger voltage

the double layer is driven out of quasiequilibrium and expands to form an

extended space charge layer: We give a brief summary of the existing theory

for the structure of this layer in the case of DC Faradaic conduction, and

discuss how the key features in this solution could be incorporated into our

strongly nonlinear model to extend its validity even beyond quasieuilibrium.

The common way of expressing the condition for validity of the weakly

nonlinear description of an electrokinetic system is in terms of the Dukhin

number Du = σ

s

/σℓ

0

∼

ε˜

/

¯

J

**, deﬁned as the ratio of the double layer
**

surface excess conductance σ

s

to the bulk conductivity σ per geometrical

length ℓ

0

[58]: Surface conductance and distortion of the double layer can

be neglected only when Du ≪1. For a symmetric binary electrolyte it can

be expressed as, cf. Eq. (2.154)

Du = ε(1 + 2Pe) ˜ w = ε(1 + 2Pe)4 sinh

2

˜

ζ/4

. (2.165)

Bazant et al. showed that the same dimensionless group governs local salt

depletion in the diﬀusion layer in the dynamical problem of a suddenly

applied voltage, and they used the low voltage estimate

˜

ζ ≃ V

0

/(1 + δ) to

argue that for weakly nonlinear dynamics to hold the applied voltage cannot

greatly exceed the thermal voltage [37].

In colloid science it is natural to express the Dukhin number in terms of

the zeta potential because for a colloidal particle the electrokinetic proper-

ties are often determined by the intrinsic zeta potential due to bound surface

charges when the particle reacts chemically with the suspending electrolyte.

However, for electrokinetic problems involving microelectrodes that are bi-

ased by an external voltage, and also for initially uncharged polarizable

particles subjected to an external DC or AC electric ﬁeld, the induced zeta

potential depends on the driving voltage in a nonlinear way and it is prefer-

able to express the conditions for validity directly in terms of V

0

rather than

˜

ζ. To this end we observe that at large voltage the surface capacitance is

dominated by the Stern layer so the accumulated charge can be estimated

as

˜ q ∼

V

0

δ

, (2.166)

2.2 Mathematical model 43

provided that the driving frequency is at or below the inverse RC time

such the screening is not negligible, and that the intrinsic surface charge

˜ q

eq

= −2 sinh

˜

ζ

eq

/2

**can be disregarded. Eq. (2.166) is an upper bound on
**

˜ q where a more accutate estimate would be

˜ q ∼

V

0

δ + sech(

˜

ζ/2)

, (2.167)

with

˜

ζ understood to be a function of ˜ q. If steric eﬀects are taken into

account the Debye layer capacitance becomes a non-monotonic function of

the accumulated charge, but ˜ q V

0

/δ remains valid also in that case. The

excess accumulated salt is bounded by ˜ w ≤ [˜ q[ by construction and at large

voltage where the absorption is dominated by uptake of coions rather than

expulsion of counterions we have ˜ w ≃ [˜ q[. The Dukhin number can therefore

be estimated as

Du ∼ ε(1 + 2Pe) ˜ w ∼

εV

0

δ

, (2.168)

assuming the P´eclet number is O(1). This shows that the surface excess

current should become signiﬁcant only at very large voltage V

0

∼ δ/ε. How-

ever, because the bulk salt ﬂux

¯

F is only O(1), the surface excess salt ﬂux

˜

Γ becomes signﬁcant when the ratio

ε

˜

Γ

¯

F

∼

ε(1 + 2Pe)˜ q∇

s

¯

φ

∼

εV

2

0

δ

(2.169)

reaches O(1), which occurs already for V

0

∼

δ/ε. Likewise, if an amount

of ions ε ˜ w (per unit area) is taken up by the Debye layer from a diﬀusion

layer of width

**ε/ω, the perturbation to the concentration in the diﬀusion
**

layer can be estimated as

ˇ γ ∼

ε ˜ w

ε/ω

∼

√

εω V

0

δ

, (2.170)

reaching O(1) for V

0

∼ δ/

√

εω , or, for V

0

∼

**δ/ε assuming the system is
**

driven around the inverse RC time with ω ∼ δ. If ε is small it is clear that

the weakly nonlinear dynamics do not break down before the V

0

signiﬁcantly

exceeds unity, i.e., not before the driving voltage is well above the thermal

voltage. This argument is consistent with our numerical results for pairs and

arrays of microelectrodes, and also with those of Chu et al. for an uncharged

metal particle subjected to a DC electric ﬁeld where O(1) perturbations to

the concentration were reached at a dimensionless ﬁeld strength of

¯

E ≃ 5 for

dimensionless Debye layer thickness ε = 0.01 and unit surface capacitance

ratio δ = 1 [38, 39].

The strongly nonlinear model both in the bulk and the diﬀusion layer

is based fundamentally on local electroneutrality. The leading order charge

44 Electrokinetic theory

density in the bulk is nominally to second order in ε and for a symmetric

electrolyte it can be expressed simply as, cf. Eq. (2.60)

¯ ρ = −ε

2

∇

2

¯

φ = −ε

2

∇¯ c

¯

J

¯ c

2

. (2.171)

Obviously, as ¯ c approaches zero the charge density diverges which is incon-

sistent with the underlying assumption of electroneutrality; that assumption

breaks down when the charge density becomes comparable to the average

concentration, i.e., for ¯ ρ ∼ ¯ c, which occurs when

¯ c ∼ ε

2

∇¯ c

¯

J

¯ c

2

or ¯ c ∼ ε

2/3

∇¯ c

¯

J

1/3

. (2.172)

In an electrochemical cell the “diﬀusion-limited” DC Faradaic current is

O(1) which leads to the conclusion that the electroneutral bulk solution

breaks down for ¯ c ∼ ε

2/3

[60, 54]. However, in our case of strong AC forcing

applied to blocking electrodes the Ohmic currents involved and the average

ion ﬂux fed into the diﬀusion layer at the electrode surface are much larger:

Taking

¯

J ∼ V

0

and using −∂

y

ˇ c =

ˇ

F

y

∼ ω ˜ w we obtain

[∂

y

ˇ c

¯

J

y

[ ∼

ωV

2

0

δ

. (2.173)

In eﬀect, the leading order strongly nonlinear solution based on electroneu-

trality breaks down already for

ˇ c ∼ (ωε)

1/3

, (2.174)

taking V

0

∼

δ/ε.

Nonequilibrium double layer structure

When the electroneutral solution breaks down, the structure of the double

layer changes from quasiequilibrium to a nonequilibrium structure and ex-

pands in width from O(ε) to O(ε

2/3

); this was ﬁrst described by Smyrl and

Newman [60]. Rubinstein and Shtilman later showed that a region com-

pletely depleted of counterions can develop at an electrode or ion exchange

membrane when driven above the diﬀusion-limited current [40]. In this

regime one can identify three sublayers within the nonequilibrium double

layer, namely [42]

• A quasiequilibrium “Debye” layer of width O(ε) at the electrode sur-

face.

• An extended “space charge” layer of width y

o

> O(ε

2/3

) that is com-

pletely depleted of counterions.

2.2 Mathematical model 45

• A “Smyrl–Newman” transition layer of width O(ε

2/3

) around y = y

o

connecting the space charge layer to the quasi-electroneutral diﬀusion

layer.

To our knowledge this fundamental structure of the nonequilibrium double

layer has only been studied for electrochemical systems in steady state with a

DC Faradaic current, although it was argued by Bazant et al. that a similar

solution with a transient space charge layer should form also in systems

driven dynamically at very large voltage [37]. Our numerical solution to the

full Poisson–nernst–Planck equations in Chap. 4 show that such a layer does

indeed form.

In the existing theory for DC Faradaic conduction, the Poisson–Nernst–

Planck equations can be reduced to a single master equation for the electric

ﬁeld whose solution can be expressed in terms of Painlev´e trancendents [42,

60, 66]. In the space charge layer migration dominates and the solution for

the electric ﬁeld reduces to (denoting variables by a breve accent)

∂

y

˘

φ = ±

1

ε

∓2

¯

J

n

(y

o

−y)/D

±

, (2.175)

where the positive sign applies to a space charge layer completely depleted

of anions (˘ c

+

= 0 and ˘ c

−

= 0) and the negative sign to depletion of cations,

and the thichkess y

o

of the layer arises simply as an integration constant.

The potential drop across the space charge layer is found by integrating

Eq. (2.175) to get

˘

Φ =

˘

φ(0) −

˘

φ(y

o

) = ∓

2

3ε

2[

¯

J

n

[

D

±

y

2/3

o

, (2.176)

and the (small) charge density from the coions by diﬀerentiation

˘ ρ = −ε

2

∂

2

y

˘

φ = ±

ε

2

2[

¯

J

n

[

D

±

(y

o

−y)

. (2.177)

The ions in the inner O(ε) quasiequilibrium Debye layer are Boltzmann

distributed and the solution for the potential takes the form [66]

˜

φ = ±log

sinh

2

(a˜ y +b)

+ const. (2.178)

where a is determined by matching to the ﬁeld in the space charge layer

a =

[

¯

J

n

[/2D

±

, (2.179)

and b is related to the total amount of charge ˜ q accumulated in both the

Debye and space charge layers by

˜ q = ±2a coth(b). (2.180)

46 Electrokinetic theory

The excess potential drop across the O(ε) Debye layer can be determined as

˜

ζ = ∓2 log

1

2

+

[˜ q[

4a

. (2.181)

Finally, in the Smyrl–Newman transition layer [y − y

o

[ ≤ O(ε

2/3

) all terms

in the master equation play in, but at least the fundamental form of the

solution is an invariant function of the rescaled spatial variable [42]

` y = (y −y

o

)

[

¯

J

n

/D

±

[

1/3

ε

2/3

. (2.182)

When we try to apply this solution structure in a dynamical rather

that steady-state setting there are several things we need to worry about,

e.g., in the space charge layer the migration term must dominate over ∂

t

˘ c

±

in the Nernst–Planck equation, and the detailed structure of the Smyrl–

Newman transition layer might change due to the dynamic nature of y

o

(t).

However, the most important complication seems to be that the solution

structure is only valid during the charging of the Debye layer, but not during

decharging. This is indicated in Eq. (2.175) where the expression under the

square root is positive only when

¯

J

n

has the opposite sign of the coions,

i.e., when

¯

J

n

< 0 for a space charge layer completely depleted of anions

with ˘ c

+

= 0 and ˘ c

−

= 0, developing during charging of the negatively

biased electrode, or vice versa at the positively biased electrode. During

decharging the interpretation of the integration constant y

o

as the thickness

of the space charge layer does not apply and matching to the solution in

the diﬀusion layer seems impossible. Moreover, at the end of the charging

period when

¯

J

n

approaches zero before it changes sign, the thickness of the

O(ε

2/3

/[

¯

J

n

[

1/3

) Smyrl–Newman transition layer diverges. It is not clear to

us how the dynamics at this point should be properly described, but from

our numerical solutions in Chap. 4 it appears that the space charge layer

collapses rapidly when the current is reversed.

Ad hoc modiﬁcation to strongly nonlinear model

In order to incorporate the key features from the nonequilibrium double layer

into our strongly nonlinear model we attempt the following modiﬁcations:

• The breakdown of the quasiequilibrium solution is identiﬁed with the

point where the concentration ˇ c in the diﬀusion layer drops to zero at

the electrode surface. Beyond this point we continue to solve a simple

diﬀusion problem for the concentration proﬁle in the diﬀusion layer,

i.e., we maintain Eq. (2.159) even if that results in unphysical negative

concentrations close to the electrode surface.

• Based on the solution for ˇ c we determine y

o

(t) as the width of the

region where ˇ c is negative during charging of the Debye layer. However,

2.2 Mathematical model 47

we always set y

o

= 0 during decharging to model a rapid collapse of

the space charge layer when the current is reversed.

• The nonequilibrium zeta potential is modelled as

˜

ζ = −2 sinh

−1

¸

˜ q

2

ˇ c + 8[

¯

J

n

/D

±

[y

o

¸

, (2.183)

where ˇ c is a strictly positive model for the concentration at the surface

ˇ c = max¦ˇ c(0, t), ε

1/3

¦. (2.184)

Here ε

1/3

is a small number introduced to avoid division by zero at the

point where ˇ c(0, t) and y

o

(t) vanish simultaneously, and the exponent

is chosen in accordance with Eq. (2.174). This approximation for

˜

ζ is

identical to the quasiequilibrium solution for ˇ c(0, t) ≥ ε

1/3

and asymp-

totically equal to Eq. (2.181) for [

¯

J

n

/D

±

[y

o

≫ε

1/3

.

• When computing the potential drop ˇ ς across the diﬀusion layer in

Eq. (2.161) we only consider the interval where ˇ c ≥ ˇ c.

• We include the potential drop

˘

Φ across the space charge layer as given

by Eq. (2.176) into the boundary condition for the bulk potential

V

ext

−

¯

φ =

˘

Φ +

˜

ζ − ˜ q δ, (2.185)

but continue to exclude ˇ ς.

• The relation between the surface excess accumulated charge and neu-

tral salt remains

˜ w =

˜ q

2

+ 4 max¦ˇ c, 0¦ −

4 max¦ˇ c, 0¦, (2.186)

i.e., we simply impose ˜ w = [˜ q[ in the nonequilibrium double layer

completely depleted of counterions.

• The surface conservation laws can still be applied to the nonequi-

librium double layer provided the thickness is small enough, y

o

≪ 1.

The (excess) surface ﬂux is dominated by the O(ε) Debye layer because

the concentration in the space charge layer is small in comparison.

Since the Debye layer is in quasiequilibrium the electrochemical po-

tential of the coions is constant, which allows the diﬀusion and migra-

tion part of the surface ﬂux to be evaluated exactly as in Sec. 2.2.5 to

get −˜ w

±

∇

s

˘ µ

±

. However, the electrochemical potential is not constant

across the space charge layer, so ˘ µ

±

= ¯ µ

±

. What we do assume is that

˘ µ

±

is dominated by the electrical term so that we can write ˘ µ

±

≈ ±

˘

φ.

The convection part of the surface ﬂux is more diﬃcult to evaluate,

but for simplicity we continue to use

˜

Γ

±

= ∓˜ w

±

(1 + 2Pe)∇

s

˘

φ, which

is consistent with our assumption that the Helmholtz–Smoluchowski

term dominates the electroosmotic ﬂow of the ﬁrst kind.

48 Electrokinetic theory

• For the electrosmotic slip induced in the nonequilibrium double layer

we have both the usual component of the “ﬁrst kind” from the O(ε)

quasiequilibrium Debye layer

˜ u

s

=

˜

ζ∇

s

˘

φ, (2.187)

but also an additional component of the “second kind” from the space

charge layer, for which we use the result of Rubinstein et al. [43]

˘ u

s

= −

y

3

o

9ε

3

D

±

∇

s

[

¯

J

n

[ +

˘

Φ∇

s

¯

φ. (2.188)

Here the ﬁrst term is expected to dominate at very large voltage, but

the second Helmholtz–Smoluchowski type term is signiﬁcant at more

realistic ﬁnite voltages.

Clearly, there are several crude approximations in this approach. Especially

our model for the transport in the diﬀusion layer is oﬀending because of the

region of negative concentration, which corresponds to ions absorbed by the

double layer that were really not there, but should have been taken up from

the region y ≥ y

o

. This eﬀectively means that we systematically underes-

timate y

o

, and therefore also

˘

Φ and ˘ u

s

. A more correct approach would

be to solve a diﬀusion problem for ˇ γ in the dynamically changing domain

ˇ y

o

(t) ≤ ˇ y < ∞ with the ﬂux boundary condition −D

a

∂

ˇ y

ˇ γ =

√

ε

ˇ

f

a

applied

at ˇ y

o

=

√

εy

o

. However, this problem is signiﬁcantly more challenging to

implement numerically than the simple form Eq. (2.159) so we have chosen

not to pursue it. On the contrary, to avoid the complexity involved with

determining y

o

from the spatial proﬁle of ˇ c dynamically and as a function

of the position along the electrode surface, we make a simple estimate for it

based on a quadratic extrapolation of ˇ c from the value at the surface, i.e.,

ˇ c(ˇ y, t) ≈ ˇ c(0, t) +∂

ˇ y

ˇ c ˇ y +

1

2

∂

2

ˇ y

ˇ c ˇ y

2

, (2.189)

to get an algorithm of the form

if ˇ c ≥ 0 or sign

˜ q

¯

J

n

≥ 0

ˇ y

o

≈ 0,

elseif ∂

2

ˇ y

ˇ c < 0

ˇ y

o

≈ −ˇ c/∂

ˇ y

ˇ c, (2.190)

else

ˇ y

0

≈ −

∂

ˇ y

ˇ c

∂

2

ˇ y

ˇ c

+

∂

ˇ y

ˇ c

∂

2

ˇ y

ˇ c

2

−2

ˇ c

∂

2

ˇ y

ˇ c

.

end

2.3 Summary 49

Because both ˇ c(0, t), ∂

ˇ y

ˇ c = −

√

ε

ˇ

f

a

/D

a

, and ∂

2

ˇ y

ˇ c = ∂

t

ˇ c/D

a

are readily avail-

able as solution variables at the surface, this estimate for ˇ y

o

is straight-

forward and cheap to implement numerically. In Chap. 4 we compare our

simple model to a full solution of the Poisson–Nernst–Planck equations in

a simple 1D geometry, and ﬁnd that they agree remarkably well: Even at

V

0

= 200 for ε = 10

−3

where the space charge layer expands a signiﬁcant

depth y

o

≈ 0.025 into the diﬀusion layer of characteristic length

√

ε ≈ 0.03,

the rough estimate from Eq. (2.190) comes within 10% of the true thickness

of the space charge layer as observed from the full numerical solution, except

just around the collapse of the space charge layer.

Another issue is that the bulk concentration variations due to surface

excess salt ﬂux in the Debye layer can become so large that ¯ c approaches

zero. This occurs ﬁrst close to the electrode edges or other sharp features

where the electric ﬁeld is particularly strong. We take this as an indication

that space charge layers should develop that are steady, i.e., not transient,

in space but with a charge density that changes with the polarity of the

driving voltage. However, it is not clear to us how the dynamics in such a

situation should be described. For the driving voltages that we consider in

Chap. 4 the region where ¯ c approaches zero is small, so we simply choose

neglect this problem except for another ad hoc modiﬁcation to our model:

• We replace ¯ c with the strictly positive ¯ c = max¦¯ c, ε

1/3

¦ in Eqs. (2.59)

and (2.60) to avoid unphysically large predictions of bulk electrocon-

vection and low conductivity.

2.3 Summary

In this chapter we ﬁrst gave a brief overview of some of the classical theory

for electrochemical transport, electrode reactions, and the standard Gouy–

Chapman–Stern model for the electrical double layer. Next we constructed

a mathematical model based on matched asymptotic expansions in the three

nested regions, namely, the “inner” Debye layer, “middle” diﬀusion layer,

and “outer” bulk region. In this we build on recent work by Bazant et al.

and Chu on dynamical solutions in electrochemical systems, where our con-

tribution has been to write an explicit semi-analytical solution (in terms of

a Fourier series) for the salt concentration in the diﬀusion layer developing

when the system is driven with an AC voltage around the inverse RC time,

and also to account for electroosmotic ﬂuid motion and convection of salt

in the bulk region. We formulated separate model equations corresponding

to the linear, weakly nonlinear, and strongly nonlinear regimes, which form

the basis for our numerical results in Chap. 3 and 4. Finally, we derived

an estimate for the driving voltage where the weakly and strongly nonlinear

models break down, and also discussed how existing theory for nonequilib-

rium double layers in electrochemical cells driven above the diﬀusion-limited

50 Electrokinetic theory

current could be incorporated into our dynamical model. However, this at-

tempt to extend the validity of our strongly nonlinear model beyond the

breakdown of local electroneutrality in the diﬀusion layers has character of

an ad hoc modiﬁcation and lacks the mathematical rigor otherwise pursued

in this chapter.

Chapter 3

Linear and weakly nonlinear

analysis

In this chapter we study the standard planar AC electroosmotic pump design

with an asymmetric array of interdigitated thin microelectrodes on a planar

substrate, as sketched in Fig. 3.1. This design was ﬁrst used by Brown

et al. [25] and later by several other groups [15, 26, 27, 29, 30, 31, 32].

It was on a device like this that Studer et al. observed pumping in

the “reverse” direction [15], where the “forward” direction is deﬁned by the

experimental observations of Brown et al. and the predictions from the

linear theory of Ramos et al. [33]. This theory predicts a single maximum

in the pumping velocity at a driving frequency around the inverse RC time,

with rapid and monotonical decrease in pumping at both lower and higher

frequencies. The only free adjustable parameter is the surface capacitance

ratio between the Debye and Stern layers, which allows the inverse RC time

and to some extent also the magnitude of the pumping velocity to be ﬁtted

to experiments. However, the theory is completely unable to account for

reversal of the pumping velocity at higher frequency.

This discrepancy between experimental observations and predictions from

existing theory has motivated us to extend the latter by including nonlinear

capacitance in the Debye layer, and Faradaic electrode reactions in both a

linearized scheme and using the full nonlinear Butler–Volmer reaction ki-

netics. Most of our results have been published in the paper Ref. [44], see

Appendix B, but in the present chapter we take into account also the ef-

fect of mass transfer limitations on the Faradaic electrode reaction. Several

groups have observed bubble formation and electrode degradation at driving

frequencies below 1 kHz and voltages of a few volt [26, 28, 29, 15, 32, 31],

which is experimental evidence that electrode reactions do actually occur.

Another motivation for including electrode reactions into the theory was

that the original results of Ajdari indicated reverse pumping in this case,

although in the low, not high frequency limit [14].

52 Linear and weakly nonlinear analysis

W

1

G

1

W

2

G

2

L

H

x

1

x

2

x

3

x

4

x

1

+L x

y 2V

0

cos(ωt)

U

Channel lid

Substrate

Electrolyte

Electrodes

Fluid rolls over

electrode edges

Average ﬂow is a

simple Couette ﬂow

Figure 3.1: Sketch of the device geometry for the planar pump design with asymmetric

pairs of interdigitated electrodes. Each pair has a narrow electrode of width W1 and a

wide electrode of width W2 separated by a narrow and wide gap of width G1 and G2.

The ﬂuid ﬂow pattern is complex with vortices above the electrode edges, but the spatial

average in the x direction is a simple Couette ﬂow proﬁle.

Idealized geometry model

As sketched in Fig. 3.1 the device consist of an array of interdigitated elec-

trodes, where each pair has a narrow electrode of width W

1

and a wide

electrode of width W

2

separated by a narrow and a wide gap of width G

1

and G

2

, respectively, and the full period is L = W

1

+ G

1

+ W

2

+ G

2

. No-

tice that in order to break the left-right symmetry it is necessary that both

W

1

= W

2

and G

1

= G

2

. On top of the array a microﬂuidic channel of height

H is placed, and we assume that the transverse width of the channel is large

enough that a 2D description in the xy plane is appropriate. In order to

reduce the number of geometrical parameters we assume an idealized geom-

etry with inﬁnitely thin and ﬂat electrodes, and a channel height H ≫ L

large enough to not inﬂuence the solution. Brown et al. originally used

W

1

= 4.2 µm, G

1

= 4.5 µm, W

2

= 25.7 µm, G

2

= 15.6 µm, and L = 50 µm,

but in this chapter we study a design with (dimensionless) electrode widths

and gaps W

1

= 1.5, G

1

= 1, W

2

= 7, G

2

= 5, and L = 14.5. As discussed in

Chap. 5 this choice is close to optimal in the sense of maximizing the pump-

ing velocity for a given (low) driving voltage, but the qualitative features

in the solution are not very sensitive to the particular choice of electrode

geometry.

In order to determine the net pumping we observe that because the

system is periodic in the x direction, the ﬂuid velocity ¯ u can be expressed

as a Fourier series [33]

¯ u(x, y) =

∞

¸

n=0

ˆ u

n

(y) e

i2πnx/L

+ c.c. (3.1)

3.1 Linear analysis at low voltage 53

where ˆ u

n

is the nth Fourier component. The net ﬂow rate is given by

Q =

H

0

¯ u

x

dy, (3.2)

and due to incompressibility Q must be independent of x, i.e., only the

zeroth order Fourier component ˆ u

0

can contribute to the net ﬂow rate. The

solution for ˆ u

0

is simply a Couette proﬁle, ˆ u

0

= y U/H e

x

, where U is the

spatial average slip velocity

U =

1

L

electrodes

˜ u

s

dx, (3.3)

and the net ﬂow rate becomes Q = HU/2. When the channel height is

suﬃciently large H ≫ L to not aﬀect the electric ﬁeld line distribution

around the electrodes, the slip velocity and hence U becomes independent

of H, whereas Q grows linearly with H. To quantify the pump performance

it is therefore convenient to report the “pumping velocity” U instead of the

net ﬂow rate Q.

In Ref. [33] Ramos et al. carried out a complete Fourier analysis in the

limit H → ∞ and determined all components ˆ u

n

(y) in terms of the slip

velocity distribution ˜ u

s

(x). They also expressed the streamfunction Ψ for

the ﬂow in terms of the Green’s function H(x − x

′

, y) for an array of ﬂuid

velocity line sources

Ψ(x, y) =

1

L

L

0

H(x −x

′

, y) ˜ u

s

(x

′

) dx

′

, (3.4)

where the Green’s function is given by

H(x −x

′

, y) =

y sinh

2πy/L

cosh

2πy/L

−cos

2π(x −x

′

)/L

. (3.5)

We adopt this for calculating the streamfunction and visualizing the ﬂow

pattern around the electrodes also at ﬁnite H ≫L.

Importantly, this Fourier analysis applies not only to the low voltage

limit, but also in the weakly nonlinear regime, or more generally to any slip

velocity driven ﬂow above a planar surface. Moreover, the linearity of the

Stokes ﬂow problem allows us to include (back) pressure driven ﬂow simply

by superposition of a parabolic Poiseuille ﬂow proﬁle [25, 31].

3.1 Linear analysis at low voltage

In the Debye–H¨ uckel limit ζ ≪ k

B

T/ze the electrical problem can be lin-

earized, and the dimensionless boundary condition for the Ohmic current

running into the Debye layer becomes, cf. Eq. (2.124)

−n ∇

ˆ

φ =

iω

1 +δ

+

1

R

ct

+R

mt

/

√

iω

(

ˆ

V

ext

−

ˆ

φ), (3.6)

54 Linear and weakly nonlinear analysis

10

−2

10

−1

10

0

10

1

10

2

10

−6

10

−5

10

−4

10

−3

10

−2

ω

U

Figure 3.2: Pumping velocity U as a function of frequency ω. The pumping is maximized

around the inverse RC time ω ∼ 1 when the screening of the electrodes is partial. Dashed

line: Analytical result at low frequency where U ∝ ω

2

.

whereas the time average electroosmotic slip velocity reduces to

˜ u

s

= '

˜

ζ∇

s

¯

φ` = −

1

1 +δ

∇

s

ˆ

V

ext

−

ˆ

φ

2

. (3.7)

Here

ˆ

V

ext

−

ˆ

φ is the potential drop across the double layer; for a given

electrode geometry there are only ﬁve dimensionless parameters describing

the problem, namely, the driving frequency ω and voltage V

0

, the surface

capacitance ratio δ, charge transfer resistance R

ct

, and Warburg constant

R

mt

. The dependence on the driving voltage in the model can be scaled out

by rescaling the potential with V

0

instead of the thermal voltage k

B

T/ze,

and the ﬂuid velocity by V

2

0

. Also the inﬂuence of the capacitance ratio

δ is almost trivial as it simply shifts the inverse RC time from ω ∼ 1 to

ω ∼ 1 + δ, and reduces the electroosmotic slip by a factor of (1 + δ)

−1

. In

our paper Ref. [44] we included this into the rescaling of the problem, but

here we simply set V

0

= 1 and δ = 0 throughout the linear analysis, with

the understanding that the eﬀect of nonzero δ and other V

0

1 is trivial.

3.1.1 Capacitive charging

In the absense of Faradaic electrode reactions the only dimensionless pa-

rameter left in the problem is the driving frequency ω. At low frequency

ω ≪1 the Debye layers are fully charged and the screening is complete at all

times, so there is no tangential electric ﬁeld to drive the electroosmotic ﬂow.

Conversely, at high frequency ω ≫ 1 the screening is negligible, so there is

no charge in the Debye layers and again no ﬂow. Only around the inverse

RC time ω ∼ 1 when there is partial screening do we have both charge and

3.1 Linear analysis at low voltage 55

Figure 3.3: Streamlines for the time average ﬂuid ﬂow at ω = 0.8 where the pumping

velocity is maximized. The streamlines are drawn as equidistant contours of the stream-

function Ψ, cf. Eq. (3.4), following Ramos et al. [33].

tangential ﬁeld in the Debye layers. Fig. 3.2 shows the pumping velocity

U as a function of ω. There is a unique maximum around ω ≃ 0.8 where

U = 0.0035, and the pumping falls oﬀ rapidly as ω

2

for ω ≪1 or as ω

−2

for

ω ≫1.

1

This scaling of the pumping velocity at low and high frequency was

predicted already in the original paper by Ajdari [14] for a simple generic

model of an electrode array as a harmonic potential modulation at the bot-

tom of the channel, whereas the particular direction and magnitude was

obtained by Ramos et al. using more realistic boundary conditions [33].

The streamline pattern for the ﬂow at ω = 0.8 is shown in Fig. 3.3:

There is a uniform ﬂow from the left to the right far above the electrodes and

more tortuous streamlines close to the electrodes, including two recirculating

vortices at the right edge of both electrodes. Notice that the stagnation

point where the slip velocity changes sign is close to the center of the narrow

electrode but is shifted more to the right on the wide electrode.

Electrical impedance

The impedance of the system can be determined as

ˆ

Z = ∆

ˆ

V /

ˆ

I, where

∆V = 2 is the magnitude of the voltage diﬀerence applied between the

electrodes, and

ˆ

I is the total current running onto the electrodes through

the power supply, i.e.,

I(t) =

electrode #1

j

ext

−∂

t

(˜ q +ε n ∇

¯

φ)

dx. (3.8)

1

In Fig. 3.2 it appears that U falls oﬀ faster than ω

−2

at ω ≫1, but this is an artefact

due to our idealized geometry model with inﬁnitely thin electrodes. For a more realistic

geometry with electrodes of ﬁnite thickness and radius of curvature R at the edges, the

pumping U matches ω

−2

for ω ≫R

−1

.

56 Linear and weakly nonlinear analysis

−π/2

−π/4

0

10

−2

10

−1

10

0

10

1

10

2

0

0.5

1

1.5

2

ω

l

o

g

1

0

[

ˆ

Z

[

θ

(

ˆ

Z

)

Figure 3.4: Bode plot of the dimensionless impedance

ˆ

Z as a function of frequency ω

at zero capacitance ratio δ = 0: Magnitude (solid line) and phase angle (dashed line).

For comparison also the impedance

ˆ

Z0 = R∞ +

1

iωC

0

for a simple RC circuit with linear

circuit elements is shown (dotted lines). The phase angle passes through −π/4 when the

frequency is at the inverse RC time at ω ≈ 0.8.

Here j

ext

is the Faradaic current passed from the electrode into the elec-

trolyte, −˜ q is the charge on the electrode required to screen the accumulated

charge in the Debye layer, and −ε n∇

¯

φ is the dielectric polarization charge.

For ω ≪ε

−1

we can neglect the last term and obtain

ˆ

I =

iω

1 +δ

+

1

R

ct

+R

mt

/

√

iω

electrode #1

ˆ

V

ext

−

ˆ

φ

dx. (3.9)

Fig. 3.4 shows a Bode plot of the dimensionless impedance of the system in

the absense of Faradaic electrode reactions and for zero capacitance ratio δ.

At low frequency the impedance is dominated by the capacitance C

0

of the

Debye layers on the two electrodes that are connected in series, i.e.,

1

C

0

=

1

W

1

+

1

W

2

=

1

1.24

, (3.10)

whereas at high frequency the bulk Ohmic resistance R

∞

= 0.88 dominates.

2

For ω ∼ 1 we observe a slight frequency dispersion in

ˆ

Z as compared to a

constant circuit element model

ˆ

Z

0

= R

∞

+1/iωC

0

, which is due to a change

2

At high frequency the resistance R∞ can be computed with high accuracy from the

power dissipation P = ∆V

2

/R∞ in the electrolyte

1

R∞

=

1

∆V

2

Ω

∇φ ∇φdr = 1.125, (3.11)

where φ is a solution to the Laplace equation with a Dirichlet boundary condition φ = Vext

on the electrodes. Likewise, the capacitance due to dielectric polarization at high frequency

3.1 Linear analysis at low voltage 57

in the electric ﬁeld line pattern between low and high frequency. The inverse

RC time can be read oﬀ as the point when the phase angle of

ˆ

Z passes

through −π/4, which coincides almost exactly with the frequency where the

pumping velocity is maximized, namely at ω ≈ 0.8. This “coincidense” is

not particular to our speciﬁc electrode geometry from Fig. 3.1, but holds for

other choices of the electrode widths and gaps as well. Which is perhaps

not surprising, given that both the impedance

ˆ

Z and the pumping velocity

U are integral quantities over the electrodes.

3.1.2 Faradaic current injection

When we include electrode reactions and Faradaic current into our model,

one of the most important new features is that the complete screening at

frequencies below the inverse RC time is unbalanced, that is, provided the

electrode reaction is not limited by mass transfer. In general, we expect AC

electroosmotic ﬂow whenever there is partial screening of the electrodes. In-

specting the boundary condition Eq. (3.6) there are three regimes in param-

eter space where the surface impedance is O(1), namely, (i) when capacitive

charging dominates for ω ∼ 1 and either R

ct

1 or R

mt

1, (ii) when

the electrode reaction kinetics dominate for R

ct

∼ 1 and R

2

mt

≤ ω ≤ 1, and

(iii) when mass transfer dominates for ω ∼ R

2

mt

and R

ct

1. These diﬀerent

regimes are illustrated in Fig. 3.5, where the pumping velocity U is shown

as a function of ω and R

ct

for diﬀerent values of R

mt

. In Fig. 3.5(a) we have

R

mt

= 1, so at the baseline R

ct

≫ 1 we recover the previous results from

Fig. 3.2 for purely capacitive charging, whereas for R

ct

≤ 1 we see a minor

downshift in the optimal frequency for pumping and a factor of 3 reduction

in the magnitude of the pumping velocity. Figs. 3.5(b) to (d) show results

for R

ct

= 0.46, 0.1, and 0.01, respectively, where we see the peak around

ω ∼ R

2

ct

shifting down and even outside the ﬁgure window in Fig. 3.5(d).

What we do not see, however, is the peak in the pumping velocity for R

ct

∼ 1

and R

2

mt

ω 1. This is puzzling, in particular because there is actually

electroosmotic ﬂuid motion in this parameter range, but only no net pump-

ing. It turns out that the vanishing of the pumping is particular to the case

when the charge transfer resistance takes the same value on both electrodes.

If R

ct

is larger on the wide electrode than on the narrow electrode we do

get pumping in the usual “forward” direction, whereas if R

ct

is larger on the

narrow electrode the pumping is in the “reverse” direction; see our paper

in Appendix B for more details on this.

Although we have not gained an intuitive understanding of why the

can be determined from the electric ﬁeld energy | =

1

2

C∞∆V

2

as

C∞ =

2ε

∆V

2

Ω

1

2

∇φ ∇φdr = 1.125ε. (3.12)

58 Linear and weakly nonlinear analysis

0

.

0

0

3

0

.

0

0

2

0

.

0

0

1

R

c

t

10

3

10

2

10

1

10

0

10

-1

10

-2

0

.

0

0

3

0

.

0

0

2

0

.

0

0

1

0

.

0

0

0

5

0

.

0

0

3

0

.

0

0

2

0

.

0

0

1

0

.

0

0

0

5

R

c

t

ω

10

-3

10

-2

10

-1

10

0

10

1

10

2

10

3

10

2

10

1

10

0

10

-1

10

-2

0

.

0

0

3 0

.

0

0

2

0

.

0

0

1

ω

10

-3

10

-2

10

-1

10

0

10

1

10

2

(a) (b)

(c) (d)

Figure 3.5: Contour plots of the pumping velocity U as a function of frequency ω and

charge transfer resistance Rct for diﬀerent values of the Warburg constant Rmt: (a) Rmt =

1, (b) Rmt = 0.46, (c) Rmt = 0.1, (d) Rmt = 0.01.

Faradaic charging does not give rise to any pumping when R

ct

is uniform,

we do have a qualitative explanation for the direction of the pumping when

R

ct

takes diﬀerent values on the two electrodes: The net pumping contribu-

tion from the narrow electrode tends to be in the reverse direction, whereas

the wide electrode contributes in the forward direction. In the simple case

with purely capacitive charging, the wide electrode dominates and the net

pumping is in the forward direction. At lower frequency the Debye layer on

the electrode where R

ct

is smaller tends to be “short-circuited” by the elec-

trode reaction, which inhibits the electroosmotic ﬂow and leaves the other

electrode dominating and determining the pumping direction. This idea can

be useful in other designs as well where, e.g., one could inhibit the ﬂow from

the reverse slipping electrode by modifying the intrinsic zeta potential on

the electrodes as in Sec. 3.1.3 below, by depositing a porous layer permable

to ions but with a large hydraulic resistance, or by geometrically lowering

the reverse slipping sections from the channel surface as suggested by Bazant

and Ben [45] and investigated in detail in Sec. 5.3.

3.2 Weakly nonlinear analysis 59

3.1.3 Linear response at nonzero intrinsic zeta potential

Many electrode-electrolyte systems spontaneously form a Debye layer and

possess an intrinsic zeta potential typically of the order ζ

eq

∼ 100 mV,

which is well beyond the Debye–H¨ uckel limit ζ ≪ k

B

T/ze ≃ 25 mV. Our

previous linear analysis does not immediately apply to such systems, but

we can still attempt a linear response analysis if the driving voltage is low

enough V

0

≪(1 +δ)k

B

T/ze. In the absense of Faradaic electrode reactions

we obtain essentially the same results as in Sec. 3.1.1, except that the RC

time for the system is determined by the diﬀerential capacitance, and the

dimensionless frequency where the pumping occurs is therefore shifted down

from ω ∼ 1 +δ to

ω ∼

¸

−

d˜ q

d(V

ext

−

¯

φ)

−1

= sech(

˜

ζ

eq

/2) +δ. (3.13)

Also the characteristic ﬂuid velocity is proportional, not to (1 +δ)

−1

but to

d

˜

ζ

d(V

ext

−

¯

φ)

=

1

1 +δ cosh(

˜

ζ

eq

/2)

(3.14)

which can be come very small if

˜

ζ

eq

is large. Of course, at very large

˜

ζ

eq

where the Dukhin number reaches O(1) we would also have to include tan-

gential surface ﬂuxes into the model. When Faradaic electrode reactions do

occur we need to replace R

ct

by the diﬀerential charge transfer resistance

R

ct

(

˜

ζ

eq

) =

¸

dj

ext

d(V

ext

−

¯

φ)

−1

= R

ct

(0)e

(z

O

−α)

˜

ζ

eq

. (3.15)

However, we also need to take into account the possibility that a time average

current to second order in V

0

could run between the electrodes. This would

give rise to an additional component to the ﬂow when the corresponding time

average electric ﬁeld interacts with the constant intrinsic zeta potential, as

we discuss at the end of Sec. 3.2.2.

3.2 Weakly nonlinear analysis

The weakly nonlinear model is a ﬁrst step towards better understanding of

the dynamics at larger voltage, and yet there have been few theoretical inves-

tigations that go beyond the Debye–H¨ uckel limit (or linear response) in the

context of AC electrokinetics. One reason for this may be that the numerical

solution is slightly more complicated: Instead of a single linear problem for

the complex potential

ˆ

φ(r) one has to solve a time evolution problem for

the bulk potential

¯

φ(r, t) with a special variable ˜ q(x, t) deﬁned only on the

electrode boundary. However, using commercial ﬁnite element software like

Femlab [47] this is straightforward to implement, and to demonstrate this

we have included a code example in Appendix A.1.

60 Linear and weakly nonlinear analysis

0

.

0

0

3

0

.

0

0

2

0

.

0

0

1

V

0

ω

10

-2

10

-1

10

0

10

1

10

3

10

2

10

1

10

0

10

-1

Figure 3.6: Contour plot of the rescaled dimensionless pumping velocity U/V

2

0

as a

function of driving frequency ω and voltage V0 for a capacitance ratio of δ = 0.1 and no

Faradaic current. The dashed white line marks the frequency where the pumping velocity

is maximized as a function of V0.

3.2.1 Nonlinear surface capacitance

In the absense of Faradaic electrode reactions there are four dimensionless

parameters in the model, namely, the driving voltage V

0

and frequency ω,

the capacitance ratio δ, and the intrinsic zeta potential

˜

ζ

eq

on the electrodes.

Fig. 3.6 shows the pumping velocity U as a function of V

0

and ω for δ =

0.1 and

˜

ζ

eq

= 0, with the data rescaled to U/V

2

0

to remove the inherent

quadratic voltage dependence and ease comparison to the linear analysis.

At the baseline V

0

≪ 1 we recover the result from Fig. 3.2 with a peak

in the pumping velocity for ω ∼ 1 + δ. However, at higher voltage the

double layer capacitance is dominated by the Stern layer, and the inverse

RC time where the pumping is maximized drops to ω ∼ δ. It also appears

from Fig. 3.6 that the pumping velocity falls oﬀ at large voltage, but this

is simply because the electroosmotic slip ˜ u

s

=

˜

ζ∇

s

¯

φ no longer scales as V

2

0

at large voltage: There

˜

ζ ∼ 2 log(V

0

/δ) rather than

˜

ζ ∼ V

0

/(1 + δ), cf. the

discussion in Sec. 2.2.9, and the slip velocity grows only as V

0

logV

0

.

When the intrinsic zeta potential is nonzero, the pumping velocity follows

the linear response analysis from Sec. 3.1.3 up to the point where the induced

perturbations in

˜

ζ becomes comparable to

˜

ζ

eq

, and well above that point

the results are largely unaﬀected by

˜

ζ

eq

.

We must emphasize that the upper voltage limit in Fig. 3.6 where V

0

extends to 10

3

is far outside the range of validity of weakly nonlinear model

for a typical microsystem. Nevertheless, the conclusion that according to

the weakly nonlinear model the only eﬀect of having a nonlinear surface

3.2 Weakly nonlinear analysis 61

0

.0

0

3

0

.

0

0

2

0

.

0

0

1

0

.0

0

0

5

V

0

ω

10

-3

10

-2

10

-1

10

0

10

1

10

3

10

2

10

1

10

0

10

-1

Figure 3.7: Contour plot of the rescaled dimensionless pumping velocity U/V

2

0

as a

function of driving frequency ω and voltage V0, for capacitance ratio δ = 0.1 and Faradaic

current with charge transfer resistance Rct = 10 and R

O

mt

= R

R

mt

= 0.1. The dashed white

line marks the frequency where the pumping velocity is maximized as a function of V0,

and the dashed black line roughly marks the voltage where the limiting current is reached.

Some data points at large voltage V0 ≥ 10

2

are missing due to numerical diﬃculties with

obtaining a self-consistent solution for jext(t) in Eq. (2.141).

capacitance is to downshift the inverse characteristic RC time and to change

the scaling of U with driving voltage from V

2

0

to V

0

logV

0

is important as a

starting point for interpreting the results from the strongly nonlinear model

in Chap. 4.

3.2.2 Nonlinear Faradaic current injection

With electrode reactions included into our model, the number of dimen-

sionless parameters increases by the charge transfer resistance R

ct

, mass

transfer coeﬃcients R

O

mt

and R

R

mt

, and the intrinsic zeta potential

˜

ζ

eq

on

the electrodes and channel substrate. From the analysis in Sec. 3.1.2 we

know that if R

mt

1, the eﬀect of the Faradaic electrode reaction is always

relatively small, so in order to actually see a nonvanishing eﬀect we consider

R

O

mt

= R

R

mt

= 0.1. Further we take R

ct

= 10 and ζ

eq

= 0 on the elec-

trodes. Fig. 3.7 shows the pumping velocity U as a function of the driving

voltage V

0

and frequency ω, where again we rescaled to U/V

2

0

to remove

the inherent quadratic voltage dependence. At the baseline for V

0

≪ 1 we

recover the results from Fig. 3.5(c), except that the scaling diﬀers by a fac-

tor of 1 + δ. For V

0

∼ 5 there is a peak in the pumping velocity at low

frequency which is due to Faradaic current injection. Although R

ct

is rela-

tively large, the exponential increase in Faradaic current with overpotential

62 Linear and weakly nonlinear analysis

0.0002

0.0004

0.0006

0.001

V

0

ω

10

-3

10

-2

10

-1

10

0

10

1

10

3

10

2

10

1

10

0

10

-1

Figure 3.8: Contour plot of the rescaled average tangential ﬁeld Es/V

2

0

along the insu-

lating substrate, cf. Eq. (3.17), for the same parameters as in Fig. 3.7. Depending on the

intrinsic zeta potential

˜

ζ

eq

this can induce pumping in either the “forward” or “reverse”

direction, but above the diﬀusion-limited current Es saturates and the eﬀect becomes

negligible compared to the “ordinary” induced-charge electroosmosis.

always makes it signiﬁcant for V

0

≫1. Moreover, because the potential drop

across the double layer is generally larger on the narrow electrode than on

the wide, the eﬀective value of R

ct

is smaller, i.e., the nonlinearity induces

an asymmetry between R

ct

on the two electrodes that leads to pumping in

the forward direction, cf. the discussion in Sec. 3.1.2 and in more detail in

our paper [44] in Appendix B. The peak fades out for ω ≪ 10

−2

due to

the mass-transfer limitations on the Faradaic current, a feature which was

missing in our earlier results in that paper. Speciﬁcally, the dashed black

line in Fig. 3.7 roughly marks the voltage where the deviations in ˇ c

O

and

ˇ c

R

relative to the bulk reach O(1); beyond this voltage the diﬀusion-limited

current is reached, i.e., the Faradaic current saturates, and for V

0

≫ 1 the

system behaviour is again dominated by the capacitive charging.

The nonlinearity in the reaction kinetics also implies that a DC Faradaic

current can be running between the electrodes even when a harmonic AC

voltage diﬀerence is applied. The corresponding DC tangential electric ﬁeld

'−∂

x

¯

φ` will interact with any intrinsic zeta potential on the electrodes or

channel substrate and induce electroosmotic ﬂow. In the special case when

˜

ζ

eq

takes the same value on both electrodes and substrate, this ﬂow does not

give rise to any net pumping on a planar device because there is no global

potential gradient along the electrode array, i.e.,

L

0

˜

ζ

eq

∂

x

¯

φ

dx =

˜

ζ

eq

L

0

∂

x

¯

φ

dx =

˜

ζ

eq

¯

φ

L

x=0

= 0. (3.16)

3.3 Summary 63

However, if

˜

ζ

eq

is nonuniform there can be pumping, and Fig. 3.8 shows the

average tangential ﬁeld along the channel substrate

E

s

=

1

L

substrate

−'∂

x

¯

φ` dx, (3.17)

which gives rise to an additional component to the pumping velocity of

U

′

=

˜

ζ

eq

e

−

˜

ζ

eq

s

E

s

, where

˜

ζ

eq

e

is the constant intrinsic zeta potential on the

electrodes and

˜

ζ

eq

s

is the corresponding value on the channel substrate. As

before the data are rescaled to E

s

/V

2

0

to remove the inherent voltage depen-

dence. The ﬁgure shows that at low voltage and for ω ∼ 0.1 there is a max-

imum in the average tangential ﬁeld along the substrate with E

s

= 0.0012.

Therefore, if the (diﬀerence in) dimensionless intrinsic zeta potential is larger

that 2, the pumping induced by the time average ﬁeld becomes comparable in

magnitude to the peak pumping velocity from the “ordinary” induced-charge

electroosmotic ﬂow, while the direction of this ﬂow component depends on

the sign of ∆

˜

ζ

eq

=

˜

ζ

eq

e

−

˜

ζ

eq

s

. Beyond the voltage where the diﬀusion-limited

current is reached, E

s

saturates and therefore appears to fall oﬀ rapidly in

Fig. 3.8 due to the rescaling with V

2

0

. Since the induced-charge electroos-

motic ﬂow continues to grow as V

0

logV

0

, this means that any net pumping

from E

s

is negligible at large voltage.

Strictly, a DC Faradaic current implies a time average ﬂux of O and

R between the two electrodes. In our model we assumed uniform bulk

concentrations ¯ c

O

= c

∗

O

and ¯ c

R

= c

∗

R

of the electrochemically active species,

which means that we neglected any bulk mass transfer limitations. Taking

such into account would reduce the value of the DC current and hence E

s

,

but we do not expect it to have much impact on the results in Fig. 3.7.

3.3 Summary

In this chapter we have studied the planar AC electroosmotic pump de-

sign used experimentally by several groups, extending the existing linear

theory to account for Faradaic electrode reactions. When the electrode

reaction is fast enough to eﬀectively “short-circuit” the Debye layer, i.e.,

when the charge transfer resistance is small R

ct

≪ 1, the electroosmotic

ﬂow is suppressed because it depends fundamentally on partial screening of

the electrodes. However, the electrode reaction can run no faster than the

electrochemically active species O and R are transported to and from the

electrode by diﬀusion, and depending on whether this mass-transfer limita-

tion is weak or strong, the pumping velocity is either suppressed and shifted

down in frequency to ω ∼ R

2

mt

or virtually unaﬀected by the electrode reac-

tion.

We also extended the theory into the weakly nonlinear regime, where the

nonlinear capacitance of the Debye layer causes the characteristic inverse RC

64 Linear and weakly nonlinear analysis

time to shift down from ω ∼ 1 + δ to ω ∼ δ, and the electroosmotic slip to

scale as V

0

logV

0

with the driving voltage rather than V

2

0

. Gonz´ alez et al.

have reported a nonlinear theoretical study of a travelling wave device, and

their results are fully equivalent to ours [67]. The nonlinearity in the Butler–

Volmer electrode reaction kinetics gives rise to more dramatic results: A DC

Faradaic current scaling as V

2

0

at low voltage should interact with any intrin-

sic zeta potential on the electrodes or channel substrate, and the direction

of the induced pumping velocity depends on the sign of the (diﬀerence in)

intrinsic zeta potential. At larger voltage there is pumping at low frequency

due to an asymmetry in the eﬀective charge transfer resistance between the

narrow and wide electrode. Ultimately, when the diﬀusion-limited current

is reached and the concentration of O and R (dynamically) approaches zero

at the electrodes, the Faradaic current saturates and our weakly nonlinear

model becomes dominated by capacitive charging. We should stress that this

last conclusion is valid only because we assume an excess of inert supporting

electrolyte, while otherwise the Faradaic current would induce bulk concen-

tration gradients and strong concentration polarization which we consider

“strongly nonlinear” phenomena.

In many aspects the weakly nonlinear model does not diﬀer qualita-

tively from the linear theory, and could therefore be classiﬁed as a “circuit

model” with nonlinear circuit elements. Neither the linear or weakly nonlin-

ear models are able to account for the reversal observed experimentally at

large voltage and frequencies above the inverse RC time. Another feature

is that the bulk concentration almost completely scales out of the problem,

entering only through λ

D

in the surface capacitance ratio δ = ǫλ

S

/ǫ

S

λ

D

.

Therefore these models are also unable to account for the experimental fact

that AC electroosmosis is suppressed at concentrations above 10 mM [15].

Chapter 4

Strongly nonlinear analysis

In this chapter we present numerical results obtained using the strongly

nonlinear model derived in Chap. 2. We ﬁrst apply the model to a simple

1D geometry in Sec. 4.1. The purpose of studying this system is both that

it gives good understanding of the dynamical properties of the model, which

is necessary when we later interpret the results in higher dimensions, but

also that it allows us to test the validity by comparing with a dynamical

solution of the full Poisson–Nernst–Planck equations.

In Sec. 4.2 we study the electrokinetic ﬂow induced on both a symmet-

ric and asymmetric electrode array. Of course, the symmetric device does

not pump, but it allows us to monitor the strength of the diﬀerent ﬂow

components, namely, electroosmotic ﬂow of the ﬁrst and second kind, and

electroconvection induced when a current is passed through concentration

gradients in the electroneutral diﬀusion layer and bulk. Such a distinction

between diﬀerent ﬂow components is artiﬁcial in the sense that experimen-

tally only the overall ﬂow is observed, but it helps us gain understanding of

the system behaviour. Finally, we investigate the net pumping induced in

the asymmetric device.

4.1 1D model problem

In this section we study the dynamics of the strongly nonlinear model in a

one-dimensional geometry as shown in Fig. 4.1 with an electrolyte conﬁned

between parallel-plate blocking electrodes, i.e., we neglect Faradaic electrode

reactions, and the (dimensionless) width of the electrode gap is L = 2. The

system is identical to that studied by Bazant et al. in Ref. [37], except that

we apply an AC voltage rather than a step voltage, and study the periodic

response after all transients have died out.

66 Strongly nonlinear analysis

2V

0

sin(ωt)

−1 0 1

y

Electrolyte

Figure 4.1: Sketch of 1D model problem. The electrolyte is conﬁned between parallel-

plate blocking electrodes separated by a gap of width L = 2, and a harmonic potential of

±V0 sin(ωt) is applied to the left and right electrode, respectively, so the overall voltage

drop across the system is 2V0 sin(ωt).

In one dimension the solution to the leading order bulk problem simpli-

ﬁes a lot. By symmetry there can be no ﬂuid motion,

1

so the only dynamical

variables in the electroneutral bulk are the potential

¯

φ(y, t) and the average

concentration ¯ c(y, t). Moreover, there can be no time average ﬂux from the

boundary layers into the bulk, so the leading order bulk concentration is sim-

ply constant ¯ c(y, t) = ¯ c

o

, and using Eq. (2.163) to impose mass conservation

across the entire system we obtain

¯ c

o

= 1 −ǫ' ˜ w`. (4.1)

The bulk electric current

¯

J(t) is constant in space and the potential is given

by

¯

φ =

¯

φ

o

(t) −

¯

J(t)y

¯ c

o

, (4.2)

where

¯

φ

o

is the potential at the centre of the cell. The bulk ﬁelds are matched

to the solution in the boundary layers at y = ±1 by requiring

∂

t

˜ q = ±

¯

J, (4.3)

∂

t

˜ w = ±

ˇ

F, (4.4)

V

ext

−

¯

φ

o

±

¯

J

¯ c

o

=

˜

ζ − ˜ q δ, (4.5)

˜

ζ = −2 sinh

−1

˜ q/2

√

ˇ c

, (4.6)

˜ w =

˜ q

2

+ 4ˇ c−

√

4ˇ c, (4.7)

in the Debye layers, where ˇ c is the concentration in the diﬀusion layer seen

1

Unless, of course, some kind of electroconvective instability occurs, but we do not

consider that here.

4.1 1D model problem 67

from the electrode surface

ˇ c = ¯ c

o

∓

ε

D

a

ω

1

T

T

0

((t −t

′

)

ˇ

F(t

′

) −∆D

¯

J(t

′

)

dt

′

. (4.8)

The bulk potential must have odd symmetry upon reﬂection in space and

translation by one half period in time

¯

φ(−y, t +T/2) = −

¯

φ(y, t), (4.9)

which implies that also

¯

J and

¯

φ

o

must have odd symmetry in time. In the

special case where ∆D = 0, the central potential

¯

φ

o

vanishes and both ˜ q

and

˜

ζ have odd symmetry, whereas ˜ w,

ˇ

F, and ˇ c have even symmetry, e.g.,

˜ w(t + T/2) = ˜ w(t). We do not include the excess potential drop ˇ ς across

the diﬀusion layer into Eq. (4.5) because this vanishes at low voltage and

typically remains small compared to Ohmic potential drop across the bulk

at larger voltage.

The coupled diﬀerential, algebraic, and integral equations are solved

numerically by substituting a truncated Fourier series and using a collocation

scheme, e.g.,

¯

J(t) =

N/2

¸

m=−N/2

′

ˆ

J

m

e

imωt

=

N

¸

n=1

=

¯

J(t

n

)S

N

(t −t

n

), (4.10)

where S

N

is the periodic sinc function

S

N

(t −t

′

) =

1

N

N/2

¸

m=−N/2

′

e

imω(t−t

′

)

, (4.11)

see Appendix A.3 or Ref. [68] for details.

4.1.1 Dynamical solutions

We solve the problem for ε = 0.001, ω = 1, δ = 1, and ∆D = 0, at

a relatively large voltage V

0

= 50, and using N = 128 grid points in time.

The solution for the diﬀerent components in the overall potential drop across

(half) the system, namely, the Ohmic potential drop

¯

J/¯ c

o

across the bulk

electrolyte, the Stern layer potential drop −˜ q δ, and the zeta potential

˜

ζ

across the Debye layer, is shown in Fig. 4.2(a). At this driving voltage

the Debye layer capacitance is large so

˜

ζ is small and the potential drop

across the double layer is primarily on the Stern layer. The nonlinearity

is not very pronounced and the system behaves almost as an ideal linear

impedance

ˆ

Z = 1 − i, except when

˜

ζ changes sign and the Debye layer

capacitance shortly drops to 1, which results in a small decrease in the

current. As usual for an RC circuit driven at a frequency around the inverse

68 Strongly nonlinear analysis

−40

−20

0

20

40

P

o

t

e

n

t

i

a

l

0 1 2 3 4 5 6

0

1

2

Time t

ˇc

0 1 2 3 4 5 6

−0.5

0

0.5

(a)

−20

−10

0

10

20

P

o

t

e

n

t

i

a

l

0 10 20 30 40 50 60

0

1

2

Time t

ˇc

0 20 40 60

−0.5

0

0.5

(b)

Figure 4.2: Dynamical solution to the 1D model problem as a function of time for two

sets of parameter values: (a) V0 = 50, ω = 1, and δ = 1; (b) V0 = 25, ω = 0.1, and δ = 0.1.

The top panel shows the diﬀerent components to the overall potential drop across (half)

the system, namely, the driving voltage V0 sin(ωt) (dotted), bulk Ohmic potential

¯

J/¯ co

(solid), Stern layer potential −˜ q δ (dashed), and zeta potential

˜

ζ (dash-dot). The inset

shows the potential drop ˇ ς across the diﬀusion layer which is O(

√

ε) smaller than

¯

J/¯ co.

The bottom panel shows the concentration ˇ c in the diﬀusion layer as seen from the surface.

RC time, the phase angle of the current is around π/4 ahead of the driving

voltage, whereas the charge accumulated in the Debye layer is π/2 ahead the

current. Fig. 4.2(a) also shows the surface concentration ˇ c that oscillates at

the double frequency 2ω. Although the variations in ˇ c are O(1), the diﬀusion

layer potential drop is small, ˇ ς 0.4, as shown in the inset.

For comparison Fig. 4.2(b) shows the solution for a diﬀerent set of pa-

rameter values where now ω = 0.1, δ = 0.1, and V

0

= 25. Again, at this

voltage the Debye layer capacitance is large, but because the intrinsic capac-

itance ratio δ is small, the potential drop across the Debye and Stern layers

are comparable. Therefore the frequency ω = 0.1 is actually also a little

below the inverse RC time. The current has a signiﬁcant bump at the point

where

˜

ζ changes sign and the double layer capacitance drops from ∼ 1/δ

to 1/(1 + δ). The maximal amount of charge ˜ q and salt ˜ w accumulated in

the Debye layer is about a factor of 3 larger than in Fig. 4.2(a), but the

variations in the surface concentration ˇ c are similar because the diﬀusion

length ℓ(ω) ∝ 1/

√

ω is also increased by roughly a factor of 3.

The concentration variations in the diﬀusion layer are directly linked

to the amount of salt accumulated in the Debye layer. At low voltage

˜ w = 4

√

ˇ c sinh

2

˜

ζ/4

scales as V

2

0

, while at higher voltage our estimate

˜ w ∼ V

0

/δ from Sec. 2.2.9 applies. Therefore we also expect the pertur-

bations to the concentration in the diﬀusion layer to grow linearly with V

0

4.1 1D model problem 69

0 20 40 60 80 100

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

0

2

4

6

8

10

12

14

16

18

20

V

0

ˇc

m

i

n

ˇς

m

a

x

,

˜

ζ

m

a

x

Figure 4.3: Minimal salt concentration mintˇ c in the diﬀusion layer (dashed line), and

maximal values of the zeta potential maxt

˜

ζ (upper solid line) and diﬀusion layer potential

drop maxt ˇ ς (lower solid line), as a function of the driving voltage V0 for ω = 1 and δ = 1.

Around V0 ≈ 90 the concentration ˇ c approaches zero which results in unphysically large

values of both

˜

ζ and ˇ ς. The dotted lines show the result when this is suppressed by

replacing ˇ c with ˇ c = max¦ˇ c, ε

1/3

¦ in Eq. (4.6) and similarly when computing ˇ ς, cf. the ad

hoc modiﬁcations suggested in Sec. 2.2.9.

for V

0

≫ 1. Fig. 4.3 shows the minimal concentration min

t

ˇ c in the diﬀu-

sion layer as a function of the driving voltage V

0

at ω = 1 and δ = 1, and

indeed for V

0

10 there is a clear linear dependence. Around V

0

≈ 90

the minimal concentration approaches zero, which marks the breakdown of

local electroneutrality in the diﬀusion layer close to the electrodes. The

ﬁgure also shows the maximal values of

˜

ζ and ˇ ς. At low voltage the zeta

potential

˜

ζ ∼ V

0

/(1+δ) grows linearly, while at higher voltage the scaling is

only logarithmic with

˜

ζ ∼ 2 log(V

0

/δ). The excess potential drop across the

diﬀusion layer depends on both the current passing though the layer and

the perturbations to the concentration, and therefore scales as V

3

0

at low

voltage changing to V

2

0

at higher voltage. When ˇ c approaches zero both

˜

ζ

and ˇ ς diverge logarithmically, which in turns prevents ˇ c from going negative

in Fig. 4.3. However, this divergent behaviour of

˜

ζ and ς is unphysical: As

discussed in Sec. 2.2.9 the leading order strongly nonlinear solution based

on local electroneutrality in the diﬀusion layer breaks down already when

ˇ c ∼ ε

1/3

. The divergence is suppressed by replacing ˇ c with ˇ c = max¦ˇ c, ε

1/3

¦

in Eq. (4.6) and similarly when computing ˇ ς, as suggested also in Sec. 2.2.9,

and the result is shown with dotted lines in Fig. 4.3.

70 Strongly nonlinear analysis

4.1.2 Comparison to a full numerical solution of the

Poisson–Nernst–Planck equations

Based on the matched asymptotic expansions we can construct a uniformly

valid approximation in space by adding the solutions from the inner, middle,

and outer regions, and subtracting the overlaps, i.e.,

c(y, t) = ¯ c

o

+

¸

ˇ c

1 +y

√

ε

, t

− ¯ c

o

+

¸

˜ c

1 +y

ε

, t

− ˇ c(0, t)

+

¸

ˇ c

1 −y

√

ε

, t

− ¯ c

o

+

¸

˜ c

1 −y

ε

, t

− ˇ c(0, t)

, (4.12)

where ˜ y = (1 ±y)/ǫ and ˇ y = (1 ±y)/

√

ǫ are the rescaled spatial variables in

the inner and middle regions. In order to validate the asymptotic model we

can compare this uniformly valid approximation to a full numerical solution

of the Poisson–Nernst–Planck equations in time and space. For a symmetric

electrolyte in one dimension Eqs. (2.45) to (2.47) reduce to

−ε

2

∂

2

y

φ = ρ, (4.13)

∂

t

ρ = ε∂

y

(∂

y

ρ +c∂

y

φ), (4.14)

∂

t

c = ε∂

y

(∂

y

c +ρ∂

y

φ), (4.15)

with boundary conditions at y = ±1

V

ext

−φ = ∓εδ∂

y

φ, (4.16)

0 = −∂

y

ρ −c∂

y

φ, (4.17)

0 = −∂

y

c −ρ∂

y

φ. (4.18)

When solving this problem for ε = 0.001 it is necessary to use a very ﬁne

ﬁnite element mesh of h

min

10

−4

close to the electrodes in order to resolve

the highly compressed Debye layer structure, or even ﬁner at larger voltage.

In 1D this is straightforward to achieve with a graded mesh, but in 2D or

3D it can become a serious problem.

We compare the two models for the same parameter values as in Fig. 4.2(a),

i.e., ε = 0.001, δ = 1, ω = 1, and V

0

= 50. The results for the average con-

centration c(y, t) is shown in Fig. 4.4 as a number of snapshots over one half

period in time. The ﬁrst frame at t =

1

16

T is just about when the charge

density in the Debye layer changes sign, cf. Fig. 4.2(a), so for t =

2

16

−

4

16

T

we see the Debye layer charging up and accumulating a lot of ions, holding

them at t =

5

16

T, before it decharges and expels them back into the diﬀusion

layer for t =

6

16

−

8

16

T. The agreement between the full numerical solution

and the asymptotic expansion is very good, with a maximal relative error

less than 0.5% in the diﬀusion layer and around around 2.5% in the Debye

4.1 1D model problem 71

0.5

1

1.5

2

t =

1

16

T

0.5

1

1.5

2

t =

2

16

T

0.5

1

1.5

2

t =

3

16

T

0 0.02 0.04 0.06

0.5

1

1.5

2

t =

4

16

T

0.5

1

1.5

2

t =

5

16

T

0.5

1

1.5

2

t =

6

16

T

0.5

1

1.5

2

t =

7

16

T

0 0.02 0.04 0.06

0.5

1

1.5

2

t =

8

16

T

1 + y 1 + y

Figure 4.4: Snapshots of the average concentration c in the diﬀusion layer over one

half period in time for ε = 0.001, ω = 1, δ = 1, and V0 = 50. The ﬁgure compares a

full numerical solution of the Poisson–Nernst–Planck equations (circles) to the uniformly

valid approximation from the asymptotic model (dotted line), and the agreement between

the two is very good. Also the leading order approximation ˇ c in the diﬀusion layer alone

is shown (solid line). Notice that the vertical axis does not go to zero.

layer. The error is of the order O(

√

ε) because we neglected the excess po-

tential drop across the diﬀusion layer ˇ ς, which is only O(

√

ε) smaller that

the bulk Ohmic potential drop.

If we increase ε the quality of the leading order asymptotic approximation

of course decreases, and, e.g., for ε = 0.05 and δ = 0.1, which are the param-

eter values used throughout Ref. [37], the error for ω ∼ 0.1 at the inverse

RC time is large because the diﬀusion layer of thickness ℓ(ω) ∼

ε/ω ≃ 0.7

extends across the entire system. However, even at very small ε the lead-

ing order strongly nonlinear solution breaks down when the driving voltage is

72 Strongly nonlinear analysis

0

1

2

3

4

t =

1

16

T

0

1

2

3

4

t =

2

16

T

0

1

2

3

4

t =

3

16

T

0 0.025 0.05 0.075 0.1 0.125

0

1

2

3

4

t =

4

16

T

0

1

2

3

4

t =

5

16

T

0

1

2

3

4

t =

6

16

T

0

1

2

3

4

t =

7

16

T

0 0.025 0.05 0.075 0.1 0.125

0

1

2

3

4

t =

8

16

T

1 + y 1 + y

Figure 4.5: Individual ion concentrations for ε = 0.001, ω = 1, and δ = 1, at a driving

voltage V0 = 200 which is large enough to drive the double layer out of quasiequilibrium.

The ﬁgure shows the full numerical solution for the concentration of the cations c+ (solid

line) and anions c− (dashed line). Between t =

3

16

T and

6

16

T a layer completely depleted

of cations appear between the electrode and the electroneutral diﬀusion layer.

4.1 1D model problem 73

increased enough that the solution in the diﬀusion layer approaches zero. For

ε = 0.001 and ω = δ = 1 this occurs around V

0

≈ 90, cf. Fig. 4.3, and for

other values of ε and δ the breakdown occurs approximately at V

0

≈ 3

δ/ε;

the prefactor 3 is particular to the simple 1D geometry that we study here,

but the scaling with

**δ/ε should hold generally.
**

When the driving voltage is further increased beyond the point where the

concentration at the surface approaches zero, the Debye layer does not stop

charging but continues to take up ions from the diﬀusion layer, and a layer

almost completely depleted of ions develops between the Debye layer and the

electroneutral diﬀusion layer. This is shown in Fig. 4.5, again as a number

of snapshots over one half period for ε = 0.001, δ = 1, ω = 1, and V

0

=

200. The solid and dashed lines show the cation and anion concentrations,

respectively, from a full solution to the Poisson–Nernst–Planck equations.

Between t =

1

16

T and

2

16

T we see the charge density in the Debye layer

change sign, followed by a rapid uptake of ions for t =

3

16

T to

5

16

T. At

t =

4

16

T and

5

16

we can clearly distinguish a region completely depleted of

cations, while the anion concentration c

−

∼ 0.1 is also low but ﬁnite. This

is the extended space charge layer discussed in Sec. 2.2.9, which was ﬁrst

described by Rubinstein and Shtilman for a permselective membrane driven

above the DC “diﬀusion-limited” current [40]. Here we see the same basic

structure of the nonequilibrium double layer developing and dynamically

propagating into the diﬀusion layer. At t =

6

16

T, which is immediately

before the current reverses, we see the space charge layer retreating, and for

t =

7

16

T and

8

16

T the diﬀusion layer is quickly reﬁlled by an equal amount of

anions expelled from the Debye layer and cations transported by migration

from the opposite electrode.

Fig. 4.6 shows the same data as Fig. 4.5 but on double logarithmic axis

which better allow the diﬀerent boundary layers to be distinguished. The

ﬁgure also shows the results from our strongly nonlinear asymptotic model

when the ad hoc modiﬁcations suggested in Sec. 2.2.9 are employed, namely,

we continue to solve a simple 1D diﬀusion problem for ˇ c even beyond the

breakdown of local electroneutrality, and identify the unphysical region of

negative “salt” concentration with the width y

o

of the space charge layer.

In the electroneutral diﬀusion layer for y ≥ y

o

(t) we have c

+

= c

−

= ˇ c

at leading order, while in the space charge layer c

+

= 0 and c

−

= [2˘ ρ[,

where the charge density ˘ ρ is given by Eq. (2.177), and vice versa during the

second half period when a space charge layer completely depleted of anions

develops. Finally, in the O(ε) quasiequilibrium “Debye” layer the leading

order anion concentration is

c

−

= [2˜ ρ[ = 2[∂

2

˜ y

˜

φ[ =

4a

2

sinh

2

(a˜ y +b)

, (4.19)

cf. Eq. (2.178), where a =

[

¯

J

n

/2D

−

and b = coth

−1

([˜ q[/2a). Fig. 4.6

shows that the asymptotic model captures the overall dynamical behaviour

74 Strongly nonlinear analysis

10

−2

10

−1

10

0

10

1

10

2

t =

1

16

T

10

−2

10

−1

10

0

10

1

10

2

t =

2

16

T

10

−2

10

−1

10

0

10

1

10

2

t =

3

16

T

10

−4

10

−3

10

−2

10

−1

10

0

10

−2

10

−1

10

0

10

1

10

2

t =

4

16

T

10

−2

10

−1

10

0

10

1

10

2

t =

5

16

T

10

−2

10

−1

10

0

10

1

10

2

t =

6

16

T

10

−2

10

−1

10

0

10

1

10

2

t =

7

16

T

10

−4

10

−3

10

−2

10

−1

10

0

10

−2

10

−1

10

0

10

1

10

2

t =

8

16

T

1 + y 1 + y

Figure 4.6: Individual ion concentrations: Same data as in Fig. 4.5 but plotted on

double logarithmic axis. Comparison of the full numerical solution for the concentration

of the cations c+ (stars) and anions c− (circles) to the asymptotic model for the cations

(solid line) and anions (dashed line).

4.1 1D model problem 75

10

0

10

1

10

2

10

3

10

4

t =

1

16

T

10

0

10

1

10

2

10

3

10

4

t =

2

16

T

10

0

10

1

10

2

10

3

10

4

t =

3

16

T

10

−4

10

−3

10

−2

10

−1

10

0

10

0

10

1

10

2

10

3

10

4

t =

4

16

T

10

0

10

1

10

2

10

3

10

4

t =

5

16

T

10

0

10

1

10

2

10

3

10

4

t =

6

16

T

10

0

10

1

10

2

10

3

10

4

t =

7

16

T

10

−4

10

−3

10

−2

10

−1

10

0

10

0

10

1

10

2

10

3

10

4

t =

8

16

T

1 + y 1 + y

Figure 4.7: Magnitude of the electric ﬁeld E = −∂yφ for the same parameters as in

Fig. 4.5. Comparison of the full numerical solution for the ﬁeld (stars or circles depending

in the sign of E) to the asymptotic model (solid or dashed lines). The agreement is good

except for t =

6

16

T and

7

16

T just before and after the space charge layer collapses.

76 Strongly nonlinear analysis

of the full numerical solution well, both in the electroneutral bulk and dif-

fusion layer, and in the quasiequilibrium or nonequilibrium double layer.

At t =

2

16

T, just after the charge density in the Debye layer has changed

sign, there is a visible diﬀerence between the solutions inside the O(ε) Debye

layer, but once the space charge layer has formed the asymptotic model is

spot on the full numerical solution in the Debye layer. For t =

6

16

T and

7

16

T,

immediately before and after the collapse of the space charge layer, there is

a signiﬁcant discrepancy in the diﬀusion layer around 1 + y ∼

√

ε ≃ 0.03.

And the quality of the asymptotic model within the space charge layer is

generally not as good as in the other regions, part of the reason for which is

simply that we have not included the O(ε

2/3

/[

¯

J

n

[

1/3

) Smyrl–Newman tran-

sition layer, and therefore the divergent tail of ˘ ρ close to y

o

is visible in

the plots. We already argued in Sec. 2.2.9 that the nonequilibrium solution

breaks down close to the point where the current reverses, because migra-

tion then no longer dominates in the space charge layer, and because the

thickness of the Smyrl–Newman transition layer diverges. Also, much of the

discrepancy is due to our crude model for salt diﬀusion, allowing for nega-

tive values of ˇ c. A solution for ˇ c with the ambipolar ﬂux boundary condition

applied on a dynamically changing domain ˇ y

o

≤ ˇ y < ∞ would certainly im-

prove the quality of the model. However, having said this, we still ﬁnd that

our simple model captures the qualitative features of the full numerical so-

lution well. To further support this, Fig. 4.7 compares the magnitude of

the electric ﬁeld E = −∂

y

φ from the asymptotic model to the full numerical

solution. Again, the agreement is good, except in an O(ε

2/3

) region around

y

o

where the ﬁeld from the diﬀusion layer −∂

y

ˇ

φ =

¯

J

n

/ˇ c diverges, and except

for t =

6

16

T and

7

16

T immediately before and after the current reverses and

the space charge layer collapses.

Finally, Fig. 4.8 compares the actual thickness of the space charge layer

from the full numerical solution, deﬁned (arbitrarily) as the width of the

region where the cation or anion concentration drops below ε

2/3

, to the width

of the region of negative concentration from the asymptotic model, and to

the approximate expression for y

o

based on Eq. (2.190). The ﬁgure shows

that the width of negative concentration tends to overestimate y

o

, whereas

the approximate expression tends to underestimate. Both are within roughly

10 % of the full numerical solution, except close to the point where the space

charge layer collapses. This accuracy is good enough that we dare employ

the simple ad hoc modiﬁcations to the strongly nonlinear model and use it

below to predict the direction and magnitude of electroosmotic ﬂow of the

second kind.

Returning to Fig. 4.6 we notice that the Debye layer is highly compressed

at this large voltage. The concentration exceeds 10

2

at 1 + y = 10

−4

and

increases to almost 10

4

for 1 + y ∼ 10

−5

. For a dilute electrolyte with

nominal bulk concentration of 1 mM this corresponds to a concentration of

about 10

28

ions per m

3

or 10 ions per nm

3

at the surface which is approx-

4.1 1D model problem 77

0 1 2 3 4 5 6

0

0.01

0.02

0.03

Time t

T

h

i

c

k

n

e

s

s

y

o

Figure 4.8: Thickness yo of the space charge layer as a function of time: Comparison of

the space charge layer thickness from the full numerical solution (solid line) to the width

of the region of negative concentration in the asymptotic model (dashed line) and to the

approximation for yo based on Eq. (2.190) (dash-dot).

imately at the steric limit, cf. the discussion in Sec. 2.1.2, and for more

concentrated solutions this limit is reached already at lower voltages. At

the same time, even though the Debye length is nominally λ

D

= 10 nm for

a 1 mM electrolyte, the highly compressed Debye layer is eﬀectively no more

than about an ˚angstrøm wide. This means that the Gouy–Chapman–Stern

model or more generally the use of dilute solution theory to describe the

structure of the O(ε) quasiequilibrium Debye layer at the electrode becomes

questionable. In particular, our model for the zeta potential

˜

ζ breaks down,

but as we argued also in Sec. 2.1.2 this does not necessarily imply that our

entire model becomes invalid: Even beyond the steric limit the Debye layer

continues to accumulate charge, so the eﬀects of salt depletion in the dif-

fusion layer, including the development of an extended space charge layer,

are still relevant. The surface conductance could also be aﬀected by strong

interactions between the ions. While this is an interesting topic, it is also

beyond the scope of the present thesis, so we continue to investigate the

dynamical behaviour in our model as derived from dilute solution theory.

78 Strongly nonlinear analysis

4.2 2D model problems

Figure 4.9: Sketch of the geometry of the symmetric electrode array. The interdigitated

electrodes all have width W = 2 and ﬁnite thickness and radius of curvature R = 0.1 at

the edges, and are separated by gaps of G = 1. Dashed lines mark the computational

domain or unit cell, and the streamlines show the usual ﬂow pattern with a vortex above

the electrode edge, driven by electroosmotic slip from the electrode edges towards the

center.

4.2.1 Symmetric electrode array

We consider a symmetric array of interdigitated electrodes as sketched in

Fig. 4.9. The electrodes have (dimensionless) width W = 2 and are sep-

arated by a gap G = 1, and to avoid introducing unphysically strong salt

depletion due to the ﬁeld singularity at the edges we do not model the

electrodes as being inﬁnitely thin but allow a ﬁnite thickness and radius of

curvature R = 0.1 at the edges. Further, we take ε = 0.001, δ = 1, ω = 1,

and assume a symmetric electrolyte with ∆D = 0. Fig. 4.9 also shows the

ﬂuid ﬂow pattern for V

0

= 50 and no convective transport, i.e., Pe = 0. The

dominant component in the ﬂow is electroosmosis, so we observe the “usual”

pattern with slip from the electrode edges towards the center, driving pairs

of ﬂuid vortices above the electrodes. The slip pattern is shown in detail

in Fig. 4.10, where the solid line shows electroosmotic slip and the dashed

line is electroconvection induced in the diﬀusion layer. The two components

have opposite sign, but the electroosmotic slip dominates because the zeta

potential

˜

ζ is much larger than the excess Ohmic potential drop ˇ ς across the

diﬀusion layer.

Fig. 4.11(a) shows the bulk concentration proﬁle as a contour plot: At the

electrode edges the concentration is lower than the equilibrium bulk value

¯ c = 1, whereas at the electrode center the concentration is higher. This

nonuniform concentration proﬁle is maintained by a continuous surface ﬂux

in the Debye layer running from the electrode edges to the center, i.e., on

average in time the Debye layer sucks in salt at the electrode edges, and

transports it towards the center where it is expelled and diﬀuses back; this

4.2 2D model problems 79

0 0.2 0.4 0.6 0.8 1

−50

−40

−30

−20

−10

0

Arc length s

u

s

Figure 4.10: Slip velocity us as a function of the distance s from the electrode center.

The electroosmotic slip (solid line) and electroconvection from the diﬀusion layer (dashed)

have opposite directions but electroosmosis dominates.

diﬀusive ﬂux is shown in Fig. 4.11(b). The feature that the surface excess

ﬂux

˜

Γ = −(1 + 2Pe)˜ q∇

s

¯

φ is in the same direction as the electroosmotic

slip ˜ u

s

=

˜

ζ∇

s

¯

φ holds more generally for a symmetric electrolyte because

˜ q and

˜

ζ have opposite sign. Fig. 4.11(c) shows the bulk electroconvection

component of the overall ﬂuid motion due to the volume electrical body

force

¯

f = ∇

2

¯

φ∇

¯

φ, arising when a current is passed through a region of

nonuniform concentration. This ﬂow takes the form of a ﬂuid vortex with

the same direction as that from electroosmotic slip, but the magnitude is

much smaller than the electroosmotic ﬂow at V

0

= 50.

Now, it turns out that neglecting bulk salt transport by convection is a

poor approximation. Although for a typical aqueous electrolyte the nominal

1.05

1

.0

2

5

1

0

.9

7

5

0.95

0.925

(a) (b) (c)

Figure 4.11: Solution at V0 = 50 and Pe = 0. (a) Contour plot of the bulk concen-

tration proﬁle. The concentration at the electrode edges is lower than the equilibrium

concentration ¯ c = 1 and higher at the electrode center. (b) Purely diﬀusive bulk salt ﬂux

¯

F = −∇¯ c. (c) Bulk electroconvection component of ﬂuid ﬂow. The ﬂow takes the form

of a ﬂuid vortex with the same direction as that from electroosmotic slip.

80 Strongly nonlinear analysis

1.05

1

.0

2

5

1

0.975

0.95

0.925

0

.9

(a) (b) (c)

Figure 4.12: Solution for V0 = 50 and Pe = 0.25. (a) The salt concentration proﬁle is

distorted and (b) the bulk salt ﬂux is dominated by convection. (c) The bulk electrocon-

vection induced with the distorted concentration proﬁle features a second counterrotating

vortex in the electrode gap.

Peclet number Pe = 0.25 is relatively small, cf. Table 2.1, the true ratio of

convective to diﬀusive ion transport

Pe

∗

=

[Pe ¯ u¯ c[

[∇¯ c[

∼ Pe[¯ u[ (4.20)

is large because the dimensionless ﬂuid velocity is much larger than unity

at V

0

= 50, cf. Fig. 4.10. Therefore in Fig. 4.12 we show the solution at

Pe = 0.25: The salt ﬂux in Fig. 4.12(b) is dominated by convection, which

distorts the concentration proﬁle in Fig. 4.12(a) as compared to Fig. 4.11(a).

Interestingly, the bulk electroconvective ﬂow in Fig. 4.12(c) induced with the

distorted concentration proﬁle has diﬀerent topology than for Pe = 0: There

is again a large ﬂuid vortex rotating in the same direction as the electro-

osmotic ﬂow, but a second smaller vortex appears in the gap between the

electrodes. However, the overall ﬂow pattern is still dominated by electro-

osmosis so this second small vortex does not show up in the overall picture.

The minimal concentration in the bulk region close to the electrode edge

is around min

r

¯ c ≈ 0.8 and min

r

¯ c ≈ 0.7 for Pe = 0 and Pe = 0.25, re-

spectively, but the concentration in the diﬀusion layer is very close to dy-

namically reaching zero at V

0

= 50. Upon a further increase in the driving

voltage, a transient space charge layer forms and gives rise to electroosmotic

ﬂow of the second kind. Fig. 4.13 shows the slip velocity distribution along

the electrode for V

0

= 100 and Pe = 0.25, and in particular we notice a sig-

niﬁcant contribution from electroosmotic ﬂow of the second kind close to the

electrode edge, that drives the ﬂuid in the same direction as electroosmosis

of the ﬁrst kind: The potential drop

˘

Φ across the space charge layer has the

same sign as

˜

ζ, i.e., opposite of ˜ q, so in general the Helmholtz–Smoluchowski

component

˘

Φ∇

s

¯

φ of the second kind electroosmotic slip should have the same

direction as that of the ﬁrst kind.

In order to quantify the relative strength of the diﬀerent components in

4.2 2D model problems 81

0 0.2 0.4 0.6 0.8 1

−150

−100

−50

0

50

Arc length s

u

s

Figure 4.13: Slip velocity distribution at V0 = 100. The electroosmotic ﬂow of the ﬁrst

kind (solid line) and second kind (dash-dot) are in the same direction, while electrocon-

vection from the diﬀusion layer (dashed) is opposite.

the overall ﬂuid ﬂow we compute the total angular momentum

L

tot

=

Ω

r ¯ u dr, (4.21)

where Ω is the computational domain, cf. Fig. 4.9. Because of symmetry the

linear momentum of the ﬂuid inside Ω vanishes, which makes L

tot

indepen-

dent of the choice of origin for r. The contribution to L

tot

from the diﬀerent

ﬂow components is shown in Fig. 4.14(a) for Pe = 0 and in Fig. 4.14(b) for

Pe = 0.25. In both cases electroosmosis of the ﬁrst and second kind and

bulk electroconvection all contribute negatively to L

tot

, i.e., they give rise

to ﬂuid motion in the clockwise direction, while electroconvection induced

in the diﬀusion layer contributes positively. In general it appears that bulk

convection of salt for Pe = 0 enhances all components in the ﬂow, except

bulk electroconvection. The mechanism behind this is as follows: The sur-

face excess salt ﬂux from the electrode edges towards the center tends to

produce a region of low salt concentration and therefore low conductivity in

the gap between the electrodes. This slows down the charging close to the

electrode edges, which means that the accumulated charge becomes more

uniformly distributed along the electrode. That in turns suppresses the tan-

gential electric ﬁeld because that is fundamentally dependent on nonuniform

charging of the Debye layer, and both electroosmotic slip of the ﬁrst and

second kind and electroconvection from the diﬀusion layer are driven by the

tangential ﬁeld. However, bulk convection of salt tends to counteract the

development of a low conductivity region between the electrodes by stir-

ring the solution and bringing down unit concentration electrolyte into the

gap. This explains why electroosmosis of the ﬁrst kind saturates already for

82 Strongly nonlinear analysis

0 25 50 75 100 125 150 175 200

−125

−100

−75

−50

−25

0

25

V

0

A

n

g

u

l

a

r

m

o

m

e

n

t

u

m

0 25 50 75 100 125 150 175 200

−125

−100

−75

−50

−25

0

25

V

0

(a) (b)

Figure 4.14: Total angular momentum in ﬂuid vortex as a function of driving voltage V0

for: (a) Pe = 0 and (b) Pe = 0.25. The solid line shows the overall Ltot, and the symbols

show contributions from electroosmosis of the ﬁrst kind (circles) and the second kind

(stars), and from electroconvection from the diﬀusion layer (squares) and bulk (triangles).

For comparison the dotted line shows the corresponding result from the weakly nonlinear

model. The two arrows mark the voltage where local electroneutrality ﬁrst breaks down

in the diﬀusion layer and bulk region, respectively.

V

0

100 in Fig. 4.14(a) but not until above V

0

175 in Fig. 4.14(b), and

why both electroosmosis of the second kind and electroconvection from the

diﬀusion layer is about twice as strong in Fig. 4.14(b) as in Fig. 4.14(a). The

development of the low conductivity zone between the electrodes is shown in

Fig. 4.15 for V

0

= 50, 100, 150, and 200, and Pe = 0 and 0.25, respectively.

The bulk electroconvection only contributes very little to L

tot

at Pe =

0.25, but this is partly because the secondary vortex from Fig. 4.12(c) grows

so large that the overall contribution from the two vortices almost integrates

to zero. The arrows in Fig. 4.14 marks the point around V

0

≈ 50 where local

electroneutrality ﬁrst breaks down in the diﬀusion layer close to the electrode

edge, as well as the point around V

0

≈ 175 where the bulk concentration

drops below ε

1/3

= 0.1 and our ad hoc modiﬁcation, replacing ¯ c by ¯ c =

max¦¯ c, ε

1/3

¦ in the bulk conductivity and ﬂuid body force, becomes active.

4.2.2 Asymmetric electrode array

We now break the left-rigth symmetry of the electrode array and consider

a planar device as that sketched in Fig. 3.1, with pairs of electrodes of

width W

1

= 1.5 and W

2

= 7, separated by gaps of G

1

= 1 and G

2

= 5,

respectively. To avoid problems associated with the ﬁeld singularity at sharp

edges we allow for electrodes of ﬁnite thickness and curvature R = 0.2 at the

edges. Further, we take ε = 0.001 and δ = 1, and consider ﬁrst a symmetric

electrolyte with ∆D = 0. Because the device is no longer completely ﬂat,

and because bulk electroconvection is not induced at the surface, we cannot

apply the simple Fourier analysis from Chap. 3 to compute the pumping

4.2 2D model problems 83

1

1

1

1

0.8 0.8

0

.9 0

.9 1

1

1

1

1

.1

1

.1

0

.8

0

.8

0

.9

0

.9

1

1

1

1

1

.1

1

.

1

1

.2

1

.2

0.8

0

.

8

0

.9

0

.

9

1

1

1

1

1

.1

1

.1

1

.2

1

.2

0.9 0.9

1

1

1

1

0

.8

0

.8

0

.

9

0

.

9

1

1

1

1

1

.1

1

.1

0

.8

0

.8

0

.

9

0

.

9

0

.9

0

.

9

1

1

1

.1

1

.1

0.8 0.8

0

.9

0

.

9

0

.9

0

.9

0

.

9

0

.9

1

1

1

.

1

1

.

1

V

0

= 50

V

0

= 100

V

0

= 150

V

0

= 200

Pe = 0 Pe = 0.25

Figure 4.15: Salt concentration proﬁle for Pe = 0 (left) and 0.25 (right) at V0 = 50,

100, 150, and 200. The low conductivity region between the electrodes locally reduces

the current into the Debye layer at the electrode edges, which makes the charging of the

Debye layer more uniform and reduces the tangential ﬁeld. Bulk convection of salt when

Pe ,= 0 counteracts this by ﬂushing “fresh” electrolyte down into the gap region.

84 Strongly nonlinear analysis

0 25 50 75 100 125 150 175 200

−5

0

5

10

15

20

25

30

V

0

0 25 50 75 100 125 150 175 200

−5

0

5

10

15

20

25

30

V

0

P

u

m

p

i

n

g

v

e

l

o

c

i

t

y

(a) (b)

Figure 4.16: Pumping velocity as a function of driving voltage V0 at ω = 1 for: (a)

Pe = 0 and (b) Pe = 0.25. The solid line shows the net pumping U, and the symbols

show contributions from electroosmosis of the ﬁrst kind (circles) and the second kind

(stars), and from electroconvection from the diﬀusion layer (squares) and bulk (triangles).

For comparison the dotted line shows the corresponding result from the weakly nonlinear

model. The two arrows mark the voltage where local electroneutrality ﬁrst breaks down

in the diﬀusion layer and bulk region, respectively.

velocity from the spatial average slip. Instead we determine the full bulk

ﬂow proﬁle, compute the net volume ﬂow rate Q by integrating over a cross

section of the channel, cf. Eq. (3.2), and deﬁne the eﬀective pumping velocity

as U = 2Q/H, where H is the height of the channel.

Fig. 4.16 shows the pumping velocity as a function of driving voltage

V

0

for ω = 1 at the inverse RC time, both with and without convective

transport of salt. In both cases the net pumping is dominated by electroos-

motic ﬂow of the ﬁrst kind, so the model does not predict any reversal of the

pumping direction for V

0

200. The contribution from electroconvection

falls oﬀ as V

4

0

at low voltage, where the ﬂow induced in the diﬀusion layer

contributes in the “forward” direction, while bulk electroconvection pumps

in the “reverse” direction with much smaller strength. Note that we might

actually have expected the opposite, since we learned in Sec. 4.2.1 that bulk

electroconvection induces vortices in the same direction as electroosmosis,

while electroconvection induced in the diﬀusion layer slips in the opposite

direction. However, for a planar asymmetric electrode array it is not only

the slip direction, i.e., towards the center or the edges of the electrodes, that

determines the net pumping, but also the balance between diﬀerent sections

with opposite slip direction at diﬀerent strength.

Around V

0

≈ 50 the concentration in the diﬀusion layer dynamically

approaches zero close to the right edge of the narrow electrode, and above

this voltage the electroosmotic ﬂow of the second kind contributes to pump-

ing in the reverse direction. When the driving voltage is further increased,

the narrow electrode experiences breakdown of local electroneutrality and

formation of transient space charge layers over the entire surface of the elec-

4.2 2D model problems 85

2

4

6

8

1

0

1

2

V

0

0

25

50

75

100

125

150

175

200

0

−

0

.

5

0

1

2

3

V

0

0

25

50

75

100

125

150

175

200

ω

10

-1

10

0

10

1

2

0

−

2

−

4

ω

10

-1

10

0

10

1

2

4

6

8

1

0

1

2

(a) (b)

(c) (d)

Figure 4.17: Contour plot of pumping velocity as a function of driving frequency ω

and voltage V0 for Pe = 0. (a) Net pumping velocity U and contributions from: (b)

Electroosmotic ﬂow of the ﬁrst kind, (c) electroosmotic ﬂow of the second kind, and (d)

bulk electroconvection. The dashed black and white lines mark the voltage where local

electroneutrality breaks down in the diﬀusion layer and bulk region, respectively.

trode, while the wide electrode experiences this only close to the left edge.

Therefore for V

0

110 the overall contribution from electroosmosis of the

second kind to the net pumping changes to be in the forward direction. The

electroosmotic ﬂow of the ﬁrst kind saturates for V

0

125 in Fig. 4.16(a)

with Pe = 0, but not in Fig. 4.16(b) where Pe = 0.25. As for the symmetric

device this occurs because the surface excess ﬂux gives rise to a region of low

conductivity in the gap between the electrodes which reduces the tangential

ﬁeld, and again bulk convection of salt counteracts the formation of such a

region for Pe = 0.

Fig. 4.17 shows the pumping velocity on a contour plot as a function of

the driving frequency ω and voltage V

0

for Pe = 0. In Fig. 4.17(a) the total

pumping velocity U is maximized around ω ∼ 1 at the inverse RC time,

but at large voltage there is a more complex frequency dependence due to

the contribution from bulk electroconvection. The dashed lines mark the

breakdown of local electroneutrality in the diﬀusion layer and bulk region,

respectively, cf. the arrows in Fig. 4.16. Fig. 4.17(b) shows the contribution

86 Strongly nonlinear analysis

2

4

6

8

1

2

1

6

2

0

V

0

0

25

50

75

100

125

150

175

200

0

−

0

.

5

0

1

2

3

V

0

0

25

50

75

100

125

150

175

200

ω

10

-1

10

0

10

1

−

0

.

5

0

−

1

−

2

ω

10

-1

10

0

10

1

2

4

6

8

1

2

1

6

2

0

(a) (b)

(c) (d)

Figure 4.18: Contour plot of pumping velocity U as a function of driving frequency ω

and voltage V0 for Pe = 0.25. (a) Net pumping velocity U and contributions from: (b)

Electroosmotic ﬂow of the ﬁrst kind, (c) electroosmotic ﬂow of the second kind, and (d)

bulk electroconvection. Again, dashed black and white lines mark the voltage where local

electroneutrality breaks down in the diﬀusion layer and bulk region, respectively.

from electroosmosis of the ﬁrst kind, featuring a single peak around ω ∼ 1+δ

at low voltage that shifts down to ω ∼ δ at larger voltage. The contribution

from electroosmosis of the second in Fig. 4.17(c) starts out negative, i.e., in

the reverse direction, but changes to positive at larger voltage. Finally, bulk

electroconvection in Fig. 4.17(d) contributes in the reverse direction around

ω ∼ 1 but in the forward direction at larger frequency. We do not show the

electroconvection induced in the diﬀusion layer because its contribution to

the net pumping is small.

Fig. 4.18 shows the results at Pe = 0.25, which are qualitatively very

similar to those at Pe = 0, but quantitatively the electroosmotic ﬂow of the

ﬁrst kind is about twice as strong because convection of salt prevents forma-

tion of a low conductivity region in the electrode gap. The electroosmotic

ﬂow of the second kind displays a twin peak at large voltage, and the same

is true for bulk electroconvection in the reverse direction. A similar feature

with a secondary peak in the pumping velocity has been observed experi-

mentally by Urbanski et al., although for a diﬀerent electrode geometry [69].

4.2 2D model problems 87

0 20 40 60 80 100

−2

0

2

4

6

8

10

V

0

0 20 40 60 80 100

−2

0

2

4

6

8

10

V

0

P

u

m

p

i

n

g

v

e

l

o

c

i

t

y

(a) (b)

Figure 4.19: Pumping velocity as a function of driving voltage V0 at ω = 1 for an

asymmetric electrolyte with ∆D = 0.33: (a) Pe = 0 and (b) Pe = 0.25. The solid line

shows the net pumping U, and the symbols show contributions from electroosmosis of

the ﬁrst kind (circles) and the second kind (stars), and from electroconvection from the

diﬀusion layer (squares) and bulk (triangles). For comparison the dotted line shows the

corresponding result from the weakly nonlinear model. Arrows mark the breakdown of

local electroneutrality in the diﬀusion layer and bulk region.

However, the convergence of the solution for large voltage and frequencies

ω 3 was relatively slow, so we are not fully conﬁdent with our numerical

results here.

Until this point we have focused on the symmetric electrolyte case with

∆D = 0 in the strongly nonlinear model which is already farily complex,

but as a ﬁnal twist let us consider an asymmetric electrolyte with cations

twice as mobile as the anions, i.e., D

+

/D

−

= 2, such that ∆D = 0.33 and

D

a

= 0.88. The results are shown in Fig. 4.19 as a function of driving voltage

at ω = 1, and they are very similar to those for a symmetric electrolyte

from Fig. 4.16. The contribution from electroconvection induced in the

diﬀusion layer scales as V

2

0

at low voltage because the diﬀusion potential

−∆Dlog(1 + ˇ γ/¯ c) is linear in the excess concentration ˇ γ, which in turns is

linear in V

0

due to the electric current component in the excess ambipolar

ﬂux, cf. Eqs. (2.159) and (2.160). Moreover, salt depletion in the diﬀusion

layer is more pronounced due to the lower eﬀective diﬀusivity D

a

, and the

concentration at the electrode approaches zero already around V

0

≈ 40.

Therefore electroosmotic ﬂow of the second kind also shows up earlier, but

with magnitude similar to the symmetric electrolyte solution.

Limitations of the numerical scheme

The temporal resolution in Fig. 4.16 is N = 32 points in time over one

period, while in Figs. 4.17, 4.18, and 4.19 it is only N = 16. The diﬀerence

between the results obtained with N = 16 and N = 32 is small, around 5%

or less, but also not completely negligible. We compute the periodic solution

88 Strongly nonlinear analysis

using a relaxation method as described in Appendix A.3, which allows us

to account for the nonlocal time dependence of the surface concentration

on the ﬂux injected from the Debye layer. However, because the computer

memory required to solve the problem grows as O(N

2

), this eﬀectively limits

the time resolution that we can obtain; for the spatial resolution in the

ﬁnite element mesh that we use, we run out of computer memory already at

N = 64. Obviously, it would be a great advantage to develop a timestepping

algorithm to compute the solution because the memory requirements for a

single timestep would be O(1) only, allowing for a much better resolution in

time and space than we have been able to achieve here. However, we have

not had the time to pursue this.

4.3 Summary

In this chapter we have studied numerical solutions to our strongly nonlinear

dynamical model, both for a simple 1D geometry and for 2D models of a

symmetric and an asymmetric electrode array.

We compared the strongly nonlinear model to a full numerical solution of

the Poisson–Nernst–Planck equations in one dimension, and found excellent

agreement at driving voltages below the point where local electroneutrality

breaks down in the diﬀusion layer. At even larger voltage we found that

the simple modiﬁcations to the model to account for formation of transient

space charge layers captures the qualitative features of the full numerical

solution well, except close to the point where the current reverses and the

space charge layer collapses, but also admit that there is room for general

improvement of the model.

For the symmetric electrode array we found that both electroosmotic

ﬂow of the ﬁrst and second kind and bulk electroconvection gives rise to ﬂuid

motion in the same direction, whereas the electroconvection induced in the

electroneutral diﬀusion layer slips in the opposite direction of electroosmosis.

We also found that electroosmotic ﬂow is limited at large voltage by the

formation of a region of low conductivity in the gap between the electrodes,

but bulk convection of salt tends to counteract this.

On the asymmetric array our model predicts that electroosmotic ﬂow

of the ﬁrst and second kind contributes to pumping in the same “forward”

direction at large voltage, whereas bulk electroconvection contributes in the

“reverse” direction, and electroconvection induced in the diﬀusion layer con-

tributes only little to the net pumping. However, the electroosmotic ﬂow

dominates over the entire voltage range that we have considered, so we are

not able to account for the reversal of the net pumping direction that has

been observed experimentally [15, 31, 34].

The particular parameter values of ε = 0.001 and δ = 1 used throughout

this chapter corresponds to, e.g., a dilute 1 mM KCl solution with λ

D

=

4.3 Summary 89

10 nm in a device with 10 µm characteristic feature size, cf. Table 2.1, with

an electrode material such as titanium that forms a natural oxide layer and

therefore has a moderate Stern layer capacitance. For electrolytes of even

lower concentration or devices of smaller feature size, the relevant value of

ε would be larger, and the strongly nonlinear phenomena would become

signiﬁcant at lower driving voltage. Likewise, for electrodes made from gold

or platinum that can have very large Stern layer capacitance, the amount

of ions accumulated in the Debye layer is larger, which again makes the

strongly nonlinear phenomena show up at lower voltage.

For electrolytes of higher concentration the use of dilute solution theory

to describe the Debye screening layer also breaks down at lower driving

voltage. The eﬀect of steric exclusion due to ﬁnite size of the ions is ﬁrst

to decrease the capacitance of the Debye layer, i.e., to increase the zeta

potential for a given amount of accumulated charge [56], which would in

turn enhance the electroosmotic ﬂow of the ﬁrst kind. However, one might

also envisage viscoelectric eﬀects from strong interactions in a dense phase

of ions that would tend to inhibit ﬂuid motion. In that case the other

sources of electrokinetic ﬂow, i.e., electroosmosis of the second kind and

electroconvection from the quasielectroneutral diﬀusion layers and bulk as

studied in this chapter, would become more signiﬁcant to the overall ﬂuid

motion.

90 Strongly nonlinear analysis

Chapter 5

Optimization

Device optimization is an important discipline in engineering. It is also a

complex task with many objectives to be pursued simultaneously, including

device performance, price, robustness, ease of manufacturing, etc. There are

many trade-oﬀ’s to be made and it is often not easy to measure what is the

“optimal” design.

In mathematics an optimization problem is formulated in a much more

clean and abstract way: The goal of the optimization is to minimize or

maximize some quantity f(x), the objective function, which is a function

of all the parameters x that can be adjusted in the optimization process.

In the present chapter we focus on optimizing the electrode geometry to

maximize the pumping velocity U as a function of driving frequency and of

the diﬀerent parameters describing the electrode shape. E.g., for a simple

planar array of asymmetric electrode pairs as studied in detail Chap. 3 we

have x = [ω, W

1

, W

2

, G

1

, G

2

]

T

. There may be constraints on the values

that the parameters can take, e.g., a geometrical length scale cannot be

negative, or more complex nonlinear constraints that depends on several of

the parameters together which are typically written as g

i

(x) ≤ 0. Hence a

large class of optimization problems can be formulated as

minimize f(x)

subject to g

i

(x) ≤ 0 for i = 1, . . . , m, (5.1)

x

min

j

≤ x

j

≤ x

max

j

for j = 1, . . . , n.

Depending on the nature of the problem there are several diﬀerent solution

algorithms to choose from: For constrained linear problems the method of

choice is the so-called simplex algorithm, whereas for unconstrained nonlin-

ear problems one can apply either a conjugate gradient or a quasi-Newton

method, provided the gradient of the objective function with respect to x can

be computed [70]. For low-dimensional nonlinear constrained optimization

a popular method is sequential quadratic programming where in each itera-

tion the objective and constraints are approximated by quadratic functions

92 Optimization

of x; implementations can be found in many software libraries, including the

Matlab Optimization Toolbox. The algorithm requires the gradient and

Hessian (second derivative) of the objective and constraints to be evalu-

ated, either exactly or by ﬁnite diﬀerencing. However, for large-dimensional

problems typical of structural optimization, in particular topology optimiza-

tion, it becomes very expensive to compute the Hessian. In such cases it

is preferable to use a simple gradient based algorithm, like steepest descent,

to determine the optimal solution. We have used the so-called method of

moving asymptotes (MMA) where the approximation for the objective and

constraints is chosen to be both convex and separable [71, 72, 73].

Topology optimization is an approach to designing mechanical structures

where one does not a priori specify or even know the topology of the optimal

geometrical structure [74, 75, 76]. It was originally developed in mechanical

engineering to optimize beams for maximal stiﬀness at a given amount of

material. More recently the method has been extended to optimize acous-

tical systems [77], photonic crystals [78], and ﬂuidic structures [79, 80]. We

have been working with the method in the ﬂuidic domain, extending the

original results of Borrvall and Petterson for Stokes ﬂow [79] to account for

inertial forces, and using it to study a small ﬂuidic network whose optimal

topology changes as a function of the Reynolds number [46]; see our paper

in Appendix 5. The basic idea is to replace the discrete transition between

solid structure and void with a smooth one by introducing an (artiﬁcal)

porous medium whose permeability to the ﬂuid depends on the local mate-

rial density. The optimization algorithm is then allowed to distribute a ﬁxed

amount of material in a continuous way, which allows changes in topology to

occur smoothly rather than discretely. Often this improves the convergence

towards the globally optimal topology, but for ﬂuidic systems we observed

problems when there are two strong compeeting local optima [46]. F. Okkels

continues our work by optimizing ﬂow in chemical enzyme reactors and cell

metabolism [51, 81].

While topology optimization is a powerfull tool, there are two major

obstacles that makes it impractical for optimizing the electrode geometry

for AC electroosmotic pumping: First, the electroosmotic slip is an inherent

surface phenomenon which is incompatible with a fuzzy transition between

solid electrode and ﬂuid electrolyte. Second, the electrode topology is not

really up for discussion: A device with free-hanging metal islands above

the base electrodes would certainly be an interesting sort of combination

of AC- and induced-charge electroosmosis but it would also be extremely

diﬃcult to fabricate. Therefore, in the present chapter we use more classical

shape optimization of the electrode structures, both with low- and high-

dimensional design spaces.

We always stick to the low voltage regime V

0

≪ k

B

T/e. The primary

reason is that optimization is an expensive task numerically and therefore we

prefer to work with the linear model which is in itself inexpensive to solve.

5.1 Planar asymmetric electrode pairs 93

Our experience from Chap. 3 is that the weakly nonlinear regime is not very

diﬀerent from the low voltage regime, so it seems reasonable that a device

geometry optimized at low voltage will also perform well at larger voltage.

In the strongly nonlinear regime we still need a better understanding of the

system behaviour before we can start thinking of systematic optimization.

5.1 Planar asymmetric electrode pairs

Most of the AC electroosmotic pump designs so far tested experimentally

[25, 26, 28, 29, 15, 30] have been planar devices with asymmetric electrode

pairs as studied in Chap. 3 and sketched in Fig. 3.1. Often the relative

size of the electrodes has been chosen as in the original paper by Brown

et al. [25], with narrow and wide electrodes of width W

1

= 4.2 µm and

W

2

= 25.7 µm, respectively, separated by gaps of G

1

= 4.5 µm and G

2

=

15.6 µm. Mpholo et al. tested devices with three diﬀerent overall periods

L = 50 µm, 30 µm, and 20 µm, but used the same relative electrode sizes in

each case; this allowed them to conﬁrm that the pumping velocity increases

with downscaling of the system.

In practice the minimal size of the microelectrodes and gaps is limited by

the lithography used to fabricate the device. However, for a given minimal

feature size one can still improve on the pumping velocity U by choosing

the relative electrode sizes optimally. This problem can be expressed as a

constrained optimization of the form of Eq. (5.1) writing

maximize U( log ω, W

1

, W

2

, G

1

, G

2

) (5.2)

subject to 1 ≤ W

1

≤ W

2

,

1 ≤ G

1

≤ G

2

,

where ω is the driving frequency.

1

Because the pumping velocity increases

with downscaling we can be sure that in the optimal design at least one

of W

1

and G

1

will attain their lower limit, whereas both W

2

and G

2

must

be strictly larger than W

1

and G

1

, respectively, because the net pumping

vanishes for W

1

= W

2

or G

1

= G

2

due to symmetry. It turns out that in

the optimal design G

1

= 1 and W

1

> G

1

, and the exact location of the opti-

mum can therefore be determined using a simple unconstrained optimization

algorithm such as the Matlab function fminunc to solve the problem

2

maximize U( log ω, W

1

, W

2

, G

1

= 1, G

2

), (5.3)

1

Optimizing for log ω eﬀectively removes any problems associated with zero or negative

frequencies; indeed we would argue this is the natural way of reporting any frequency.

2

In order to use the fminunc function from the Matlab Optimization Toolbox, all the

user has to do is write a function that takes ( log ω, W1, W2, G2) as arguments, sets up and

solves the electrokinetic problem, e.g., using Femlab, and returns (minus) the pumping

velocity U. However, due to the truncation error and noise introduced by remeshing the

geometry in each function call, it is important choose appropriate not-so-small values

for the termination tolerance and for the step size used when fminunc approximates the

94 Optimization

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

0

2

4

6

8

10

12

14

16

18

20

0

.

0

0

3

5

0

.

0

0

3

0

.0

0

2

5

0

.0

0

2

W

1

W

2

S

y

m

m

etry

Figure 5.1: Contour plot of the maximum pumping velocity max ω U as a function of

W1 and W2 for G1 = 1 and G2 = 5. When W1 = W2 the net pumping vanishes due to

symmetry. There is a unique optimum close to W1 = 1.5 and W2 = 7 with U = 0.00358

at ω = 0.8.

provided we start the algorithm somewhere in the interior of the feasible

region. In this way we ﬁnd that the pumping velocity is maximized for

W

1

= 1.51, W

2

= 6.55, and G

2

= 4.74, where U = 0.00359 at ω = 0.80.

The pumping velocity is actually not very sensitive to the particular

choice of planar electrode geometry. This is demonstrated in Fig. 5.1 where

the maximum pumping velocity max

ω

U is shown as a function of W

1

and

W

2

for G

1

= 1 and G

2

= 5. The maximum is close to W

1

= 1.5 and W

2

= 7

where U = 0.00358 at ω = 0.8, but the ﬁgure also shows that the range of

designs that come within the 0.0035 contour is fairly broad. It is for this

reason that we have chosen in Chap. 3 to work with a design with rounded

numbers, namely W

1

= 1.5, G

1

= 1, W

2

= 7, and G

2

= 5. This does not

correspond to any device tested experimentally (yet), but it is close to the

global optimum according to our present linear analysis.

5.2 Planar electrodes with nonuniform coating

There are many ways that one can break the left-right symmetry of an elec-

trode array. From a fabrication point of view the asymmetric pairs of planar

electrodes are particularly simple because the device can be processed with

just a single lithography step to deﬁne metal electrodes on the substrate.

Another simple approach is to use pairs of planar electrodes that are sym-

gradient and Hessian by ﬁnite diﬀerencing. A robust but also less eﬃcient alternative is

direct search algorithms like the Matlab fminsearch Nelder–Mead simplex method.

5.2 Planar electrodes with nonuniform coating 95

W G

Figure 5.2: Planar electrodes (black) with a spatially nonuniform thin oxide layer or

coating (orange). The coating thickness is not drawn to scale; streamlines correspond to

the optimal solution at driving frequency ω = 5.1.

metric in size, but with an oxide layer or coating that makes the Stern layer

capacitance left-right asymmetric on the individual electrodes, see Fig. 5.2.

This was actually one of the suggestions for breaking the symmetry in the

original paper by Ajdari [14].

We consider the general problem of an array of blocking electrode where

the Stern layer capacitance varies along the electrodes, i.e., the capacitance

ratio δ(x) is a function on the position with 0 ≤ δ(x) ≤ ∞. For our present

purpose it turns out to be more convenient to parametrize the problem in

terms of Λ = (1+δ)

−1

, which is simply the ratio of the overall capacitance of

the double layer to the Debye layer capacitance [23, 33], because this stays in

the unit interval 0 ≤ Λ(x) ≤ 1. In the low voltage regime the dimensionless

problem then takes the form

−∇

2

ˆ

φ = 0 (5.4)

in the bulk with

−n ∇

ˆ

φ = iωΛ(x)(

ˆ

φ −

ˆ

V

ext

) (5.5)

on the electrodes and n ∇

ˆ

φ = 0 on the channel walls. The time average

slip velocity can be expressed as

u

s

= Λ(x)

1

2

Re

(

ˆ

V

ext

−

ˆ

φ)∂

x

ˆ

φ

∗

= −Λ(x)

1

4

∂

x

ˆ

V

ext

−

ˆ

φ

2

. (5.6)

96 Optimization

Discretization and sensitivity analysis

We approximate both the complex potential

ˆ

φ and the design variable Λ by

expanding on a ﬁnite basis set

ˆ

φ(r) =

¸

m

ϕ

m

(r)

ˆ

φ

m

and Λ(x) =

¸

k

χ

k

(x)Λ

k

. (5.7)

Multiplying Eq. (5.4) by a test function ϕ

n

and integrating over the com-

putational domain Ω, we obtain the weak form of the problem

Ω

∇ϕ

n

∇

ˆ

φdr +iω

electrodes

ϕ

n

Λ(x)(

ˆ

φ −

ˆ

V

ext

) dx = 0 for all n, (5.8)

where we used Gauss’ law with Eq. (5.5) to eliminate n ∇

ˆ

φ. Inserting the

expansion for

ˆ

φ we obtain a linear problem for the coeﬃcients

ˆ

φ

m

¸

m

¸

Ω

∇ϕ

n

∇ϕ

m

dr +iω

electrodes

ϕ

n

Λϕ

m

dx

ˆ

φ

m

= iω

electrodes

ϕ

n

Λ

ˆ

V

ext

dx. (5.9)

This can be written more compactly in matrix form

ˆ

K

ˆ

Φ =

ˆ

F, (5.10)

where

ˆ

K

nm

=

Ω

∇ϕ

n

∇ϕ

m

dr +iω

electrodes

ϕ

n

Λ(x)ϕ

m

dx (5.11)

and

ˆ

F

n

= iω

electrodes

ϕ

n

Λ(x)

ˆ

V

ext

dx. (5.12)

Notice that

ˆ

K is symmetric but not Hermitian. We also deﬁne the residual

ˆ

L =

ˆ

F −

ˆ

K

ˆ

Φ in terms of which the electrical problem becomes

ˆ

L(

ˆ

Φ, Λ, ω) = 0. (5.13)

We want to maximize the pumping velocity U, and for an array of planar

electrodes this is simply deﬁned as as the spatial average slip velocity

U =

1

L

electrodes

u

s

dx = −

1

4L

electrodes

Λ(

ˆ

V

∗

ext

−

ˆ

φ

∗

)∂

x

(

ˆ

V

ext

−

ˆ

φ) dx + c.c. (5.14)

Since the number of parameters Λ

k

describing the distribution of the coating

material is large, we use the gradient based algorithm MMA [71, 72] to

determine the optimum. Hence we need to compute the derivatives

d

dΛ

U

ˆ

Φ(Λ, ω), Λ

and

d

dω

U

ˆ

Φ(Λ, ω), Λ

, (5.15)

5.2 Planar electrodes with nonuniform coating 97

where Φ(Λ, ω) is considered an implicit function of Λ and ω, deﬁned by

the constraint Eq. (5.13). However, because

ˆ

Φ(Λ, ω) is implicit and ∂

ˆ

Φ/∂Λ

and ∂

ˆ

Φ/∂ω are unknown, we cannot simply use the chain rule. Instead we

use the adjoint method [82] and introduce the Lagrangian L,

L(

ˆ

Φ,

ˆ

Φ

∗

,

ˆ

W,

ˆ

W

∗

, Λ, ω) = U(

ˆ

Φ,

ˆ

Φ

∗

, Λ) +

1

2

ˆ

W

T

ˆ

L +

1

2

ˆ

W

∗T

ˆ

L

∗

, (5.16)

where both the potential

ˆ

Φ and its complex conjugate

ˆ

Φ

∗

are considered

as independent variables, and

ˆ

W is a Lagrange multiplier for the constraint

ˆ

L(

ˆ

Φ, Λ, ω) = 0. The derivatives in Eq. (5.15) can be computed as the

gradient of L with respect to Λ and ω at a point in (

ˆ

Φ,

ˆ

Φ

∗

,

ˆ

W,

ˆ

W

∗

, Λ, ω)

space where L is stationary with respect to both

ˆ

Φand

ˆ

W and their complex

conjugates. That is, for a given Λ and ω we ﬁrst compute the potential

ˆ

Φ

from

∂L

∂

ˆ

W

=

ˆ

F −

ˆ

K

ˆ

Φ = 0, (5.17)

and the Lagrange multiplier

ˆ

W from

∂L

∂

ˆ

Φ

=

∂U

∂

ˆ

Φ

−

1

2

ˆ

K

T

ˆ

W = 0. (5.18)

Then the desired derivatives are given by

d

dΛ

U

ˆ

Φ(Λ, ω), Λ

=

∂L

∂Λ

=

∂U

∂Λ

+ Re

ˆ

W

T

∂

ˆ

L

∂Λ

, (5.19)

and

d

dω

U

ˆ

Φ(Λ, ω), Λ

=

∂L

∂ω

= Re

ˆ

W

T

∂

ˆ

L

∂ω

. (5.20)

Numerical implementation

We build the physical model and solve the problem Eq. (5.17) using Femlab.

Therefore the both system matrix

ˆ

K = −∂

ˆ

L/∂

ˆ

Φ and the matrix ∂

ˆ

L/∂Λ are

computed automatically as part of the solution process and can be obtained

directly as Matlab sparse matrices. Further we need to compute the partial

derivatives of U with respect to

ˆ

Φ and Λ

∂U

∂

ˆ

φ

n

=

1

4L

electrodes

Λ

(

ˆ

V

∗

ext

−

ˆ

φ

∗

)∂

x

ϕ

n

+ϕ

n

∂

x

(

ˆ

V

∗

ext

−

ˆ

φ

∗

)

dx (5.21)

∂U

∂Λk

= −

1

4L

electrodes

χ

k

∂

x

ˆ

V

ext

−

ˆ

φ

2

dx. (5.22)

98 Optimization

0 0.5 1 1.5 2

0.0

0.2

0.4

0.6

0.8

1.0

1.2

C

o

a

t

i

n

g

Λ

0 0.5 1 1.5 2

−0.2

0.0

0.2

0.4

0.6

0.8

1.0

P

o

t

e

n

t

i

a

l

ˆ

φ

0 0.5 1 1.5 2

−0.2

−0.1

0.0

0.1

0.2

x

S

l

i

p

v

e

l

o

c

i

t

y

u

s

0 0.5 1 1.5 2

−0.1

−0.05

0.0

0.05

0.1

x

S

e

n

s

i

t

i

v

i

t

y

d

U

d

Λ

(a) (b)

(c) (d)

Figure 5.3: Comparison of the solution for optimal coating (solid) at ω = 5.1 where

U = 0.011 to the solution for uncoated electrodes (dashed) at ω = 4.0 where U = 0:

(a) Design variable Λ(x). (b) Potential

ˆ

φ(x, y = 0) on electrodes; optimal Re[

ˆ

φ] (solid)

and Im[

ˆ

φ] (dotted) vs. uncoated Re[

ˆ

φ] (dashed) and Im[

ˆ

φ] (dash-dot). (c) Time-average

slip velocity us(x). (d) Sensitivity dU/dΛ(x). The position of the electrode on the x-axis

is marked in gray.

This can also be done with Femlab although on the way we need to make

an auxiliary copy of the original model; see our paper in Appendix C for

details. Finally we need to evaluate

ˆ

W

T

∂

ˆ

L

∂ω

= i

electrodes

Λ

ˆ

W (

ˆ

V

ext

−

ˆ

φ) dx (5.23)

=

ˆ

W

T

ˆ

L(

ˆ

Φ, Λ, 1) −

ˆ

L(

ˆ

Φ, Λ, 0)

, (5.24)

where the last equality follows because

ˆ

L(

ˆ

Φ, Λ, ω) is a linear function of ω.

Optimal solution

We consider a symmetric device with electrodes of a ﬁxed width W = 1 and

gap G = 1, cf. Fig. 5.2. The design variable Λ is approximated with ﬁrst

order Lagrange elements, i.e., it is piecewise linear, and we use second order

elements for the potential

ˆ

φ. The optimization algorithm is initialized with

a uniform distribution Λ(x) = 1 corresponding to uncoated electrodes, and

with a frequency of ω = 4 which is about twice the relaxation frequency

5.2 Planar electrodes with nonuniform coating 99

for the uncoated device. The initial solution is shown with dashed lines in

Fig. 5.3: The potential

ˆ

φ has left-right symmetry so the induced electroos-

motic slip velocity does not give rise to any net pumping. The sensitivity

dU/dΛ of the pumping velocity U to local changes in Λ shows that U can be

increased by decreasing Λ on the left half of the electrode or by increasing

Λ on the right half; however, the latter is not an option due to the bound

Λ ≤ 1. At ﬁrst this result may seem counterintuitive: From Eq. (5.6) we

know that at ﬁxed tangential ﬁeld and potential drop across the double layer,

the slip velocity decreases locally with decreasing Λ. However, a decrease in

Λ on the left half of the electrode also induces changes in the potential in

such way that the overall result is an increase in the net pumping U.

The optimization algorithm adjusts Λ(x) according the sensitivity and

quickly reaches a solution that is close to optimal, except for some small wig-

gles at the resolution of the ﬁnite element mesh, but these slowly fade away

as the design iterations proceed. After 71 iterations the relative increase

in pumping velocity is less than 10

−5

for two successive iterations and we

terminate the algorithm. The maximum pumping velocity is U = 0.011 at

ω = 5.1, and the solution is shown with solid lines in Fig. 5.3: Close to the

right edge of the electrode there is an uncoated section of width ∼ 0.25 where

Λ = 1, but for the remaining part it turns out to be optimal to add coating.

The potential varies almost linearly and the time-average slip velocity u

s

is

positive throughout the coated section. Notice that in the optimal solution,

dU/dΛ = 0 only where Λ is limited by the upper bound; in optimization

jargon this is known as the complementary slackness condition. Finally, the

thickness of the coating in Fig. 5.2 corresponds to the optimal capacitance

ratio δ = Λ

−1

− 1, and the streamlines correspond to the optimal solution

at ω = 5.1.

0 0.5 1 1.5 2

−0.2

0.0

0.2

0.4

0.6

0.8

1.0

x

P

o

t

e

n

t

i

a

l

ˆ

φ

0 0.5 1 1.5 2

−0.2

−0.1

0.0

0.1

0.2

x

S

l

i

p

v

e

l

o

c

i

t

y

u

s

(a) (b)

Figure 5.4: Approximating the optimal solution from Fig. 5.3 by a design with electrodes

with a uniformly coated section of width W1 = 0.75 where Λ = 0.25 and an uncoated

section of width W2 = 0.25. The maximal pumping velocity is U = 0.0105 at ω = 5.4.

(a) Potential on electrodes; Re[

ˆ

φ] (solid) and Im[

ˆ

φ] (dashed). (b) Slip velocity us.

From an engineering point of view it may be possible to fabricate devices

100 Optimization

with a nonuniformly graded coating thickness, but with standard lithog-

raphy techniques it is much easier to obtain piecewise constant coating

thickness. Therefore, inspired by the optimal solution we consider a de-

vice where the electrodes are divided into two sections of width W

1

= 0.75

and W

2

= 0.25 where the design variable is Λ

1

= 0.25 and Λ

2

= 1.0, respec-

tively. The maximum pumping velocity for this simple design is U = 0.0105

at ω = 5.4, which is merely 5% lower than the optimum from Fig. 5.3. The

solution is shown in Fig. 5.4; notice that the slip velocity is discontinuous

at the point where Λ changes abruptly from Λ

1

to Λ

2

.

One may object to the above solutions because they have features below

the (unit) characteristic length scale. If the same lithography technique

is used to deﬁne both the electrodes and the coating on them it is more

appropriate to formulate the problem as

maximize U( log ω, W

1

, W

2

, G, Λ

1

, Λ

2

= 1) (5.25)

subject to 1 ≤ W

1

,

1 ≤ W

2

,

1 ≤ G,

0 ≤ Λ

1

≤ 1.

Here the optimal design is W

1

= 1.4, W

2

= 1, G = 1, Λ

1

= 0.31, and

Λ

2

= 1, where U = 0.0096 at ω = 1.9. This is about 15% lower than the

optimum from Fig. 5.3, but better suited for fabrication at the lower limit

on resolution set by standard lithography.

Compared to the device with asymmetric pairs of electrodes, asymmetric

coating allows the pumping velocity to be increased by a almost a factor of

3, at the expense of an additional lithography step to deﬁne the coating. A

disadvantage of the use an asymmetric coating to break the left-right sym-

metry is that the design variable Λ depends on the electrolyte concentration

through the Debye layer capacitance. Therefore the coating thickness will

have to be designed for a particular concentration, while oﬀ that concentra-

tion the performance is suboptimal.

5.3 3D structured electrodes with regular shape

While most work so far have focused on planar designs with thin ﬂat elec-

trodes on a substrate, one may obtain a signiﬁcant gain in pumping eﬃciency

by stepping into the third dimension: Bazant and Ben recently suggested

the use electrodes like those shown in Fig. 5.5 [45], for which they predict

pumping velocities an order of magnitude larger than the planar design of

Brown et al. They also tested the 3D design experimentally and obtained

already a factor of 7 larger pumping velocity than with planar devices of the

same minimal feature size [83, 69].

5.3 3D structured electrodes with regular shape 101

W

1

W

2

G

H

1

Figure 5.5: Stepped 3D electrode geometry: Each electrode in the array has a raised

section of width W1 and height H1 and a base section of width W2 and the gap between

adjacent electrodes is G. Streamlines correspond to the solution with W1 = W2 = H1 =

G = 1 at driving frequency ω = 0.8; close to the electrode extra streamlines are plotted

to show the details of the ﬂow pattern.

The basic design principle that they propose is to raise those sections

of the electrodes where the slip is in the “forward” direction and to lower

the sections with “reverse” slip. This enhances the forward ﬂow because the

reverse vortex is living in a trench, where it is conﬁned by the vertical solid

electrode surface rather than struggling up against the forward ﬂow. On the

contrary: At the level of the forward slipping surface the recirculation from

the vortex enhances the forward pumping instead of inhibiting it. Eﬀectively,

the electrode array now acts as a ﬂuid “conveyor belt” where all sections

help pumping the ﬂuid in the forward direction [45].

We consider ﬁrst a geometry as shown in Fig. 5.5 with electrodes that

have both a raised section of width W

1

and height H

1

, a base section of

width W

2

, and with a gap G between adjacent electrodes. In the simple

case when all dimensions are set to unity, W

1

= W

2

= H

1

= G = 1, the net

pumping is maximized around ω = 0.8. By analogy with the planar pumps

we calculate the eﬀective pumping velocity as U ≈ 2 Q/H

tot

≈ 0.05, where

H

tot

≫ 1 is the overall channel height measured from the substrate to the

channel top wall. This is an order of magnitude larger than the optimal

result from Sec. 5.1, which clearly demonstrates the potential of 3D designs.

For planar designs the pumping velocity U, deﬁned as the spatial average

slip velocity, comes out of a simple Fourier analysis as the natural way of

quantifying the pump performance. For 3D structured electrodes it is not

immediately obvious how the pumping velocity should actually be deﬁned,

or if it is appropriate for quantifying the pump performance at all. To clarify

this we compute the net ﬂow rate Q as a function of the channel height H

tot

.

The result is shown in Fig. 5.6 where we clearly see a linear dependence for

102 Optimization

0 2 4 6 8 10

0

0.05

0.1

0.15

0.2

0.25

H

tot

F

l

o

w

r

a

t

e

Q

Figure 5.6: Maximal ﬂowrate maxω Q vs. total channel height Htot measured from the

substrate to the channel top wall in a geometry with W1 = W2 = H1 = G = 1. For

Htot ≫ 1 there is a linear dependence: Data points (circles) and linear ﬁt (dashed line).

For Htot = 1 the top wall touches the raised electrode section and the ﬂow is blocked.

H

tot

≫1. For a given electrode design we can therefore uniquely deﬁne the

pumping velocity as

U = lim

Htot→∞

2 Q(H

tot

)

H

tot

. (5.26)

Extrapolating Q(H

tot

) to zero we get a nonzero intercept at around H

∗

=

0.7, which is perhaps not so surprising, given the ambiguity in the deﬁnition

of the channel height, i.e., should it be measured from the substrate or from

the top of the electrodes? For H

tot

= H

1

= 1 the channel top wall touches

the raised electrode section and the ﬂow is blocked. In order to estimate

U from data at ﬁnite H

tot

we could either use U ≈ 2 Q/(H

tot

− H

∗

) or

U ≈ 2 ∂Q/∂H

tot

. However, because H

∗

depends on H

1

, and ∂Q/∂H

tot

requires ﬁnite diﬀerencing, we simply choose a “large” H

tot

= 40 and use

U ≈ 2 Q/H

tot

throughout this section.

In order to compute the net ﬂow rate Q we need to solve the Stokes ﬂow

problem with a slip condition on the electrodes. To avoid numerical diﬃ-

culties associated with inﬁnite slip velocity from the electric ﬁeld singularity

at sharp corners on the electrode we use a geometry model with rounded

corners of radius R. Smaller R gives stronger ﬁeld at the protruding corners

which actually tends to increase the pumping velocity U, but the limit of

sharp corners is ﬁnite. We use R = 0.02 which is not very sharp, and there-

fore the pumping velocity that we report here can be up to a few percent

lower than the sharp limit.

5.3 3D structured electrodes with regular shape 103

0 0.5 1 1.5 2

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

W

1

P

u

m

p

i

n

g

v

e

l

o

c

i

t

y

U

0 0.5 1 1.5 2 2.5 3

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

W

2

0 0.5 1 1.5 2

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

G

P

u

m

p

i

n

g

v

e

l

o

c

i

t

y

U

0 0.5 1 1.5 2

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

H

1

Figure 5.7: Maximum pumping velocity maxω U as a function of each one of the in-

dividual geometrical parameters W1, W2, G, and H1, with the other parameters ﬁxed

at unity. Dashed and solid lines show the results using 1 and min¦1, W1, W2, G¦ as the

characteristic length scale, respectively.

Optimal solution

Using standard planar lithography to process the device, it is much easier

to control the thickness of the raised section on the electrodes than the

transverse dimensions. Therefore we formulate the optimal design problem

as

maximize U( log ω, W

1

, W

2

, H

1

, G) (5.27)

subject to 1 ≤ W

1

,

1 ≤ W

2

,

1 ≤ G.

In Fig. 5.7 we investigate the dependence of the maximum pumping velocity

max

ω

U on each the individual geometrical parameters W

1

, W

2

, G, and H

1

,

when the other parameters are ﬁxed at unity. The ﬁgure shows that the

optimum for both W

1

and G is at the lower limit W

1

= G = 1 whereas for

W

2

and H

1

it is around W

2

≈ 1.7 and H

1

≈ 0.6, respectively. The global

optimum can be determined from a contour plot as in Fig. 5.8 showing the

maximum pumping velocity as a function of both W

2

and H

1

: It is close to

104 Optimization

0.055

0

.0

5

0

.0

4

5

0

.0

4

0

.0

3

5

1 1.25 1.5 1.75 2 2.25 2.5

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

H

1

W

2

Figure 5.8: Contour plot of the maximum pumping velocity maxω U as a function of

W2 and H1 for W1 = G = 1. There is a unique optimum around W2 = 1.25 and H1 = 0.7

where U = 0.0566 at ω = 0.8.

W

2

= 1.25 and H

1

= 0.7, where U = 0.0566 at ω = 0.8. However, the ﬁgure

also shows that the optimum is broad. In particular, the simple design with

H

1

= 0.7 and W

1

= W

2

= G = 1 has U = 0.0545 at ω = 0.9 which is within

2% of the global optimum, and our starting point W

1

= W

2

= H

1

= G = 1

is within 12%.

A virtue of the simple 3D structured rectangular design in Fig. 5.5 is that

it can be fabricated with just two lithography steps to deﬁne the base and

raised sections of the metal electrodes on top of a ﬂat substrate. However,

the basic design principle in [45] speaks of both raising sections of “forward”

slip and recessing sections of “reverse” slip on the electrode. Therefore

we can generalize the simple design by allowing the base section of the

electrode to be lowered by an amount H

2

relative to the substrate level as

shown in Fig. 5.9(a). The optimal solution to this problem turns out to

be W

1

= 1, W

2

= 1, G = 1, H

1

= 0.4, and H

2

= 1, where U = 0.078 at

ω = 0.65. Yet another possibility is to add thin metal electrodes on top

of a 3D structured substrate as shown in Fig. 5.9(b); D. Burch reported

that such a design would increase the pumping velocity over the solid metal

electrode design [84]. The optimal solution is, roughly, W

1

= 1, W

2

= 1.2,

G = 1.0, H

1

= 0.9, and H

2

= 0.3, where U = 0.089 at ω = 2.1. Whether or

not these more complex designs are actually superior to the simple design in

Fig. 5.5 depends on how diﬃcult it is to pattern both the metal electrodes

and the substrate: If it turns out that the minimal linewidth that one can

(routinely) produce in the more complex designs is, say, twice as large as

for the simple design, then the simple design is actually superior due to the

5.4 3D structured electrodes with arbitrary shape 105

W

1

W

2

G

H

1

H

2

W

1

W

2

G

H

1

H

2

(a) (b)

Figure 5.9: Stepped and grooved 3D electrode geometries: (a) Design with the “reverse”

slipping section of the electrode put into a groove. The optimal solution is W1 = 1,

W2 = 1, G = 1, H1 = 0.4, H2 = 1, where U = 0.078 at ω = 0.65. (b) Design with thin

metal electrodes deposited on a 3D structured substrate. The optimal solution is W1 = 1,

W2 = 1.2, G = 1, H1 = 0.9, H2 = 0.3, where U = 0.089 at ω = 2.1.

fundamental scaling of the pumping velocity with one over the characteristic

length scale.

5.4 3D structured electrodes with arbitrary shape

A natural question to ask is whether the rectangular electrode structures

we investigated in the previous section are (close to) optimal in general, or

if some other shape with sloped side walls could improve the performance

even further. In order to answer this question we now consider electrodes

with a general shape described by some continuous function ¯r(s). Because

the number of parameters to describe ¯r(s) is large (inﬁnite) we will need to

compute the derivative dU/d¯r of the pumping velocity U with respect to

local changes in the electrode shape exactly rather than using ﬁnite diﬀer-

ence approximation. One way to obtain this is by an arbitrary Lagrangian

Eulerian (ALE) type formulation, introducing a mapping from the physical

frame (x, y) to a ﬁxed reference frame (X, Y ). The idea is to continuously

deform the physical domain Ω to match the electrode shape, but keep a

ﬁxed ﬁnite element mesh in the reference domain Ω

∗

where we solve the

problem numerically. When we transform the weak form of the governing

equations from the physical frame to the reference frame we make use of the

transformation rules

Ω

f(r) dr =

Ω

∗

f

r(R)

detJdR (5.28)

106 Optimization

and

∂

∂x

i

=

∂X

j

∂x

i

∂

∂X

j

, (5.29)

where detJ is the determinant of the Jacobian matrix J

ij

= ∂x

i

/∂X

j

for

the mapping from the reference domain Ω

∗

to the physical domain Ω, and

∂X

j

/∂x

i

is an element of the inverse Jacobian invJ. The gradient dU/d¯r is

computed using the adjoint method as described in Sec. 5.2. It is straightfor-

ward in principle, but more cumbersome because there are more dependent

variables in the present problem, including the mesh deformation r(R), com-

plex potential

ˆ

φ(r), ﬂuid velocity u(r) and pressure p(r), and a number of

Lagrange multipliers that we account for explicitly to gain full control over

which constraint forces act where in the governing equations.

Discretization

Our bold statement that we consider a general continuous curve ¯r(s) for the

electrode shape is subject to a number of constraints, e.g., in order to be

a valid shape ¯r(s) cannot intersect itself. Moreover we want it to be fairly

“nice” with not too many or too sharp wiggles.

We discretize the problem by introducing a piecewise linear approxima-

tion to ¯r(s) with N line segments deﬁned by N + 1 nodes ¯r

n

where the

endpoints ¯r

0

and ¯r

N

are ﬁxed. In order to gain more control over the shape

of ¯r and avoid clustering of the nodes we force all line segments to have the

same length a. Then the nodal coordinates can be expressed as

¯ x

n

= ¯ x

0

+a

n−1

¸

m=1

cos θ

m

and ¯ y

n

= ¯ y

0

+a

n−1

¸

m=1

sin θ

m

, (5.30)

where θ

m

is the angle between the mth segment and the horizontal axis,

i.e., tan θ

m

= ∆¯ y

m

/∆¯ x

m

. The shape of ¯r is uniquely determined by the

N − 1 ﬁrst angles whereas the ﬁnal angle θ

N

and the segment length a are

chosen to match the ﬁnal node with the ﬁxed endpoint ¯r

N

. Alternatively we

can parametrize in terms of the N −1 relative angles Θ

m

= N (θ

m+1

−θ

m

)

between adjacent line segments; Θ is also a measure of the local curvature

κ = ∂

s

θ of ¯r, where s is the arc length in the physical frame, since

Θ

m

= N

θ

m+1

−θ

m

= Na

∆θ

m

a

≈ L

tot

κ

m

, (5.31)

writing L

tot

= Na for the overall arc length of ¯r. In the examples below we

control the shape of ¯r by limiting Θ

m

to a ﬁnite interval Θ

min

≤ Θ

m

≤ Θ

max

.

This eﬀectively limits the curvature of ¯r and determines how sharp features

can be on the electrode, but also sets a limit on the overall arc length, cf.

Eq. (5.31). In the design process the optimization algorithm will therefore

have to make a trade-oﬀ between curvature and arc length. This makes the

5.4 3D structured electrodes with arbitrary shape 107

interpretation of the results more complex, but since we like the electrodes

to have neither too sharp features nor too large a surface area we are still

content with our approach.

Optimal solutions

As a ﬁrst design problem, inspired by the results for rectangular electrodes

in the previous section, we consider a device with a ﬁxed width W = 2 of

the electrode and a ﬁxed gap of G = 1. To avoid diﬃculties with fabrication

we do not allow undercut or backward sloping structures, i.e., we require

−π/2 ≤ θ ≤ π/2, and formulate the optimization problem as

maximize U( log ω, Θ) (5.32)

subject to −Θ

max

≤ Θ

n

≤ Θ

max

for n = 1, . . . , N −1,

−π/2 ≤ θ

m

≤ π/2 for m = 1, . . . , N.

The optimal solution to this problem for diﬀerent values of Θ

max

with N =

100 is shown in Fig. 5.10. In panel (a) we set Θ

max

= 6.66 which is just

enough to allow for two half circles of radius R ≈ 0.5, i.e., curvature κ =

1/R ≈ 2, and a total arc length around π. The maximal pumping velocity is

U = 0.078, which is almost as good as the corresponding optimal rectangular

grooved structure from Sec. 5.3, even though the present electrode shape

completely lacks sharp corners to enhance the electric ﬁeld strength. In

panels (b)-(f) we gradually increase Θ

max

: We see ﬁrst the width of the

raised section decrease, which is in accordance with the trend in Fig. 5.7(a)

for W

1

< 1. Secondly, we see the lowered section develop into a double

groove, where the “purpose” of the extra “pillar” dividing the groove into

two seems to be to weaken the counter-rotating vortex. In the ﬁrst four

panels (a)-(d) the optimal solution is everywhere limited either by [Θ

n

[ ≤

Θ

max

or by [θ

m

[ ≤ π/2, but in panels (e) and (f) this is no longer the case:

In particular in panel (f) neither of the two “pillars” are capped with half

circles but have a width and shape that start to be “optimal” in itself rather

than limited by the particular constraints that we impose.

In our second design we relax the constraint on backward sloping struc-

tures, but maintain that all nodes should conform with the overall electrode

width W. Thus we formulate the problem as

maximize U( log ω, Θ) (5.33)

subject to −Θ

max

≤ Θ

n

≤ Θ

max

for n = 1, . . . , N −1,

¯ x

0

≤ ¯ x

m

≤ ¯ x

N

for m = 1, . . . , N −1.

The result is shown in Fig. 5.11, again with N = 100 line segments to

describe the electrode shape. The overall features in the solutions are the

same as in Fig. 5.10: The width of the raised section decreases and a double

108 Optimization

U = 0.078 U = 0.090

U = 0.097 U = 0.111

U = 0.121 U = 0.123

(a) (b)

(c) (d)

(e) (f)

Figure 5.10: Electrode shapes (black) computed as optimal solutions to the prob-

lem Eq. (5.32) for diﬀerent values of the maximum curvature Θmax: (a) Θmax = 6.66.

(b) Θmax = 10. (c) Θmax = 13.3. (d) Θmax = 20. (e) Θmax = 40. (f) Θmax = 80. The elec-

trode width is ﬁxed at W = 2 and the gap at G = 1 and the electrode shape is described

by N = 100 line segments (nodes in red).

5.4 3D structured electrodes with arbitrary shape 109

U = 0.078 U = 0.092

U = 0.101 U = 0.116

U = 0.130 U = 0.127

(a) (b)

(c) (d)

(e) (f)

Figure 5.11: Electrode shapes (black) computed as optimal solutions to the prob-

lem Eq. (5.33) for diﬀerent values of the maximum curvature Θmax: (a) Θmax = 6.66.

(b) Θmax = 10. (c) Θmax = 13.3. (d) Θmax = 20. (e) Θmax = 40. (f) Θmax = 80. The

shapes in (e) and (f) are not fully converged solutions but were terminated after 200 design

iterations.

110 Optimization

groove develops when Θ

max

is increased. At the same time the extent of

the backward sloping becomes more and more pronounced, and the two

last solutions in panels (e) and (f) look pathological in the sense that they

contain very narrow parts in the electrode shape. Actually, panels (e) and

(f) do not correspond to fully converged optimal solutions because in both

cases the design process was terminated after a total of 200 iterations.

It is diﬃcult to compare the performance of our optimized electrode

shapes from Figs. 5.10 and 5.11 with the rectangular geometries from Sec. 5.3

because strictly they are not optimal solutions to the same problem: In

Sec. 5.3 we optimized the geometry subject to a constraint on the minimal

transverse linewidth allowed, whereas in the present section we ﬁx the total

width of the electrode at W = 2 and optimize subject to a constraint on

the minimal feature size as represented by the maximal curvature. However,

looking at the results in Fig. 5.10 and 5.11 we ﬁnd that, largely, the opti-

mization algorithm seems to follow the basic design principle of raising the

forward slipping and recessing the reverse slipping sections on the electrodes,

even though it does not know of any such principle but only has information

on the local gradient dU/d¯r. It therefore seems likely that the simple rect-

angular geometries from Sec. 5.3 are indeed (close to) optimal because we

constructed their low-dimensional design spaces exactly based on this basic

principle together with our knowledge of what electrode structures can most

easily be manufactured in practice.

5.5 Thermodynamic eﬃciency

Throughout this chapter we have focused on the maximum pumping velocity

that can be obtained for a given driving voltage. Equally important might

be the maximum pumping velocity for a given electrical power consump-

tion used to drive the electrode array, or the maximum amount of “useful

work” that the pump can deliver at a given power consumption, i.e., the

thermodynamic eﬃciency [7, 9]. Provided convection can be neglected in

the electrical part of the problem, which is true for our linear and weakly

nonlinear models but not in the strongly nonlinear regime, the net ﬂow rate

Q can be decomposed into separate and independent electroosmotic and

pressure-driven components

Q = Q

eo

+Q

p

. (5.34)

The electroosmotic contribution to the net ﬂow rate per unit width of the

channel is

Q

eo

= UH/2 (5.35)

where U is the pumping velocity and H is the height of the pumping channel.

The pressure driven ﬂow is given by

Q

p

= −∆p/R

hyd

(5.36)

5.5 Thermodynamic eﬃciency 111

where ∆p is the backpressure that the pump is pumping up against, and

R

hyd

is the (dimensionless) hydraulic resistance per unit width

R

hyd

=

12L

tot

H

3

(5.37)

where L

tot

is the total length of the channel. The Q-∆p diagram for the

pump is linear, with maximal ﬂow rate Q

max

= Q

eo

obtained at zero back-

pressure, whereas the maximal backpressure ∆p

max

= Q

eo

/R

hyd

is deﬁned

as the point where the pump stalls and Q vanishes. The work done when

pumping up against an external backpressure is P

out

= Q∆p which is max-

imized at ∆p = ∆p

max

/2 where

P

out

=

1

4

R

hyd

Q

2

eo

. (5.38)

The electrical power consumption is

P

in

=

1

2

Re[

ˆ

I

∗

∆

ˆ

V ] = 2NRe

¸

1

ˆ

Z

(5.39)

where ∆

ˆ

V = 2 is the (dimensionless) voltage applied between the electrodes

when V

ext

= ±Re[e

iωt

], I is the total current passed though the device, N is

the number of electrode pairs, and

ˆ

Z is the impedance for a single pair, cf.

Sec. 3.1.1. To optimize for thermodynamic eﬃciency one should therefore

maximize the ratio

P

out

P

in

= β

LU

2

Re[1/

ˆ

Z]

, (5.40)

where L = W

1

+G

1

+W

2

+G

2

is the array period and β is a dimensionless

parameter given by

β =

3

8

ǫ

σH

ǫV

2

0

ηℓ

0

(1 +δ)

2

=

3τ

D

u

0

8H(1 +δ)

2

. (5.41)

The dimensionless function LU

2

/Re[1/

ˆ

Z] should be optimized geometrically,

but otherwise the scaling of β shows that it pays to pump an electrolyte of

low conductivity σ because this lowers the power consumption at a given

driving voltage, but does not (in theory) aﬀect the pumping velocity. It

pays to squeeze down the channel lid and lower H because this increases

the internal hydraulic resistance and enables to pump up against a more

decent backpressure. It pays scale down the device and decrease ℓ

0

as much

as possible. And ﬁnally, according to the low-voltage linear theory, it pays

to increase the driving voltage V

0

and decrease the capacitance ratio δ,

which should hold also at larger voltage although the scaling is diﬀerent.

Substituting typical numbers for a 1 mM KCl solution in a device with

ℓ

0

= 5 µm, H = 50 µm, δ = 0, and V

0

= 1 V, and using U ≃ 0.0035

112 Optimization

and L ≃ 15 as corresponding to the optimized planar device, we obtain a

thermodynamic eﬃciency of

P

out

P

in

≈ 10

−8

. (5.42)

What this tells us is that an AC electroosmotic pump may be employed to

move liquid rapidly through a microchannel, but not against any signiﬁcant

backpressure.

This should actually be no surprise to us because the same limitation

applies to ordinary inline DC electroosmotic pumps. To overcome this, sev-

eral authors have introduced a porous material or sintered glass frit into

the pumping channel, with eﬀective pore radius a from a few µm down

to 100 nm [8, 9, 10, 11, 12]. The thermodynamic eﬃciency is maximized

at a ≈ 2.5λ

D

, and experimentally eﬃciencies of around 1 % have been

obtained pumping deionized water [9] an up to 5.6 % for a buﬀer solution

with the strongly dipolar additive tris(hydroxymethyl)aminomethane hydro-

cloride [10]. However, introducing porous material in the pumping channel

does not enhance the eﬃciency of the AC electroosmotic pump. Our results

from Ref. [44] show that strong geometrical conﬁnement of the pumping

channel down to H ≪ L does not, as one might have worried, mask the

asymmetry of the electrodes and destroy the ability to pump. Hence ge-

ometrical conﬁnement could be a route for increasing the thermodynamic

eﬃciency, but there is a long way to go from 10

−8

. We believe it is more

fruitful to think of AC electroosmosis as a means for local ﬂow control in a

microﬂuidic system, and to focus on ease integration rather than (lack of)

energy eﬃciency.

5.6 Summary

In this chapter we have investigated diﬀerent strategies for breaking the left-

right asymmetry of an electrode array to create an AC electroosmotic pump,

and determined in each case the design that maximizes the pumping velocity

for a given (low) driving voltage. From a fabrication point of view the easiest

device to produce is that based on asymmetric pairs of electrodes, but the

3D structured design suggested by Bazant and Ben [45] could increase the

pumping velocity by an order of magnitude at the expense of an additional

lithography step. The thermodynamic eﬃciency is low, so one should focus

on the use of AC electroosmotic pumps for local generation of ﬂow but not

for pumping up against large external backpressure or hydraulic resistance.

Chapter 6

Conclusion and outlook

Conclusion

In the present thesis we have studied AC electrokinetics with focus on mi-

cropumps based on asymmetric electrode arrays.

We have extended the existing linear theory for AC electroosmosis into

the weakly nonlinear regime by accounting for nonlinear capacitance of the

Debye layer and Faradaic current injection from electrochemical reactions at

the electrodes. The primary eﬀect of the nonlinear capacitance is to change

the scaling of the pumping velocity with driving voltage, and to shift down

the inverse RC time to lower frequency at large driving voltage. Depending

on the application it may be feasible to operate a micropump with Faradaic

current injection, provided that the electrode reaction runs reversibly such

that electrolysis and bubble formation is avoided. We conclude that fast elec-

trode reactions act to “short-circuit” the Debye layer and thereby suppress

electroosmotic ﬂow, unless the reaction itself is limited by mass-transfer

of the reactants. If the electrochemically active species are only present

in small concentrations with an excess of inert supporting electrolyte, the

Faradaic current from the reaction levels oﬀ when the diﬀusion-limited cur-

rent is reached, and the dynamical behaviour at larger voltage is dominated

by capacitive charging. On the other hand, if there is no inert supporting

electrolyte the AC Faradaic current should give rise to strong concentration

polarization above the diﬀusion-limited current with transient space charge

layers and electroosmosis of the second kind. However, that is a strongly

nonlinear phenomenon that we have have not investigated.

Neither the linear nor weakly nonlinear models are able to account for

the reversal of the pumping velocity observed experimentally, and therefore

we have also studied the strongly nonlinear regime where the classical as-

sumption of uniform bulk electrolyte concentration breaks down. Building

on recent theoretical work for diﬀuse-charge dynamics in electrochemical

systems, we have constructed a strongly nonlinear model based on bound-

114 Conclusion and outlook

ary layer analysis and matched asymptotic expansions. In particular, we

have contributed by coupling an explicit dynamical solution in the diﬀusion

layer to the evolution of the quasiequilibrium Debye layer and to the bulk

concentration proﬁle that is nonuniform but stationary in time. We also

account for ﬂuid motion and convection of salt in the bulk.

We have obtained numerical solutions for this model in both a simple

1D geometry and in 2D for a symmetric and an asymmetric electrode array.

Moreover, because even our strongly nonlinear model based on local elec-

troneutrality breaks down at “very large” but still experimentally relevant

voltage, we have attempted to include existing theory for nonequilibrium

double layers and electroosmosis of the second kind into our model. We

predict that electroosmotic ﬂow of the ﬁrst and second kind contributes to

pumping in the forward direction, while bulk electroconvection contributes

to pumping in the reverse direction. However, for the voltage range that we

have considered, electroosmosis dominates, so we are not able to account

for the experimentally observed reversal of the pumping direction with our

strongly nonlinear model, either.

Finally, we have looked at various ways of breaking the symmetry of an

electrode array and determined in each case the design that maximizes the

pumping velocity at a given low driving voltage. The solutions are expected

to be close to optimal also at larger voltage as long as electroosmosis of

the ﬁrst kind dominates the electrokinetic ﬂow. It is important to realize

that AC electrokinetic micropumps can provide fast pumping of the order

of ∼ mm/s, but not against a signiﬁcant backpressure. Thermodynamic

eﬃciency will therefore always be low, but the technology is still attractive

due to ease of integration in a microsystem or lab-on-a-chip, and because it

oﬀers local control over the ﬂuid motion in the device.

Outlook

There remains many open problems and possible directions for future re-

search on strongly nonlinear electrokinetics. What makes this subject so

challenging, and the experimental results so diﬃcult to interpret, is that

there is a whole zoo of strongly nonlinear eﬀects showing up at roughly the

same time, i.e., salt depletion in the diﬀusion layer followed by breakdown of

local electroneutrality and development of an extended space charge layer,

large surface excess ﬂux in the Debye layer leading to nonuniform bulk con-

centration with convection dominated bulk salt ﬂux, and even breakdown

of dilute solution theory in the highly charged Debye layer. A full solution

of the Poisson–Nernst–Planck equations provides only little insight into the

solution structure, but for simple geometries it can be useful for validating

the strongly nonlinear theory.

Because the diﬀerent eﬀects interact in a nontrivial way we can envisage

115

complex dynamical behaviour in geometries such as the AC electroosmotic

pump with 3D structured electrodes. However, before this can fully be ex-

ploited to optimize the pump performance or to engineer robust “forward”

operation at large voltage, it is necessary to gain a more detailed under-

standing of the dynamical behaviour, e.g., by studying simple geometries

such as metal spheres or cylinders.

The theoretical description for DC rather than AC driving is simpliﬁed

because the diﬀusion layer and bulk region merge and the number of nested

boundary layers drop by one. It would therefore be interesting to extend

the work of Chu on a metal sphere in an externally applied DC ﬁeld to

even larger ﬁelds, where local electroneutrality breaks down in the bulk

region close to the poles of the sphere due to large surface ﬂuxes towards

the equator. The space charge layers developing in this case are immediately

describable by the existing theory for DC Faradaic conduction.

Alternatively, the existing theory for nonequilibrium double layers could

deserve a rigorous account for the solution structure when the width of the

space charge layer changes in time with the applied current, and in particular

the dynamics of the collapse of the layer when the current reverses is a

challenging problem. In relation to the simple ad hoc modiﬁcations that

we suggested, the quality of the model would certainly improve by solving

a 1D semi-inﬁnite diﬀusion problem on the interval ˇ y

o

(t) ≤ ˇ y < ∞ for the

concentration in the diﬀusion layer, and it would be a great advantage to

device a timestepping algorithm to allow for better resolution in time when

computing the periodic solution.

Finally, the dependence on the model parameters ε, δ, ∆D, and possibly

˜

ζ

eq

, deserves more attention and systematic variation. In particular, lower

values of the capacitance ratio δ are relevant to many experiments. To

this end it would also be interesting to actually determine the experimental

values of δ and

˜

ζ

eq

by ﬁtting the linear response theory to impedance spectra

recorded as a function of frequency and with a DC voltage applied relative

to a distant reference electrode.

116 Conclusion and outlook

Appendix A

FEMLAB code

A.1 Weakly nonlinear ﬁnite element model

In this section we give an example of our implementation of the weakly

nonlinear model from Sec. 2.2.7 in the absense of Faradaic electrode reaction.

The leading order electrical problem reduces to a Laplace problem in the

bulk

0 = −∇

2

¯

φ, (A.1)

with

0 = n ∇

¯

φ, (A.2)

on the insulating channel walls and a set of dynamical boundary conditions

∂

t

˜ q = n ∇

¯

φ, (A.3)

V

ext

−

¯

φ =

˜

ζ − ˜ q δ, (A.4)

on the electrodes, where the zeta potential is given by

˜

ζ = −2 sinh

−1

(˜ q/2). (A.5)

The solution to the electrical problem determines the electroosmotic slip

velocity that is boundary condition for the bulk time average Stokes ﬂow

problem for the ﬂuid velocity ¯ u and pressure ¯ p

0 = ∇ ¯ σ, (A.6)

0 = ∇ ¯ u, (A.7)

where ¯ σ is the stress tensor deﬁned by

¯ σ

ij

= −¯ p δ

ij

+∂

i

¯ u

j

+∂

j

¯ u

i

. (A.8)

118 FEMLAB code

The slip condition on the electrodes is

¯ u =

˜

ζ∇

s

¯

φ

, (A.9)

and on the channel walls we impose the no-slip condition ¯ u = 0.

The electrical problem is converted to weak form by multiplying Eq. (A.1)

by a test function ϕ for the potential and integrating over the volume Ω of

the computational domain, employing Gauss’ law, and using Eq. (A.3) to

eliminate n ∇

¯

φ, i.e.,

0 = −

Ω

ϕ∇

2

¯

φdr =

Ω

∇ϕ ∇

¯

φdr +

∂Ω

ϕn ∇

¯

φds

=

Ω

∇ϕ ∇

¯

φdr +

electrodes

ϕ∂

t

˜ q ds, (A.10)

which is supposed to be satisﬁed for all test functions ϕ.

1

The Gouy–

Chapman–Stern boundary condition for the potential drop across the double

layer is imposed by multiplying a test function ̺ for the accumulated charge

and integrating over the electrodes

0 =

electrodes

̺

V

ext

−

¯

φ −

˜

ζ + ˜ q δ

ds, (A.11)

which is to be satisﬁed for all ̺. The ﬂow problem is converted to weak form

by multiplying the force balance (A.6) by a test function u for the velocity,

and the incompressibility constraint by a test function ξ for the pressure,

and integrating over Ω, i.e.,

0 = −

Ω

u (∇ ¯ σ) +ξ ∇ ¯ udr

=

Ω

∇u : ¯ σ −ξ ∇ ¯ udr +

∂Ω

u (n ¯ σ) ds. (A.12)

In Femlab we use so-called strong or ideal constraints to impose the no-slip

condition on the channel walls, but for the slip condition on the electrodes we

need weak constraints. The reason is that with an ideal constraint Femlab

would consider Eq. (A.9) equally much as a constraint on

¯

φ as on ¯ u, whereas

we want it to be a constraint on ¯ u only, and also want the reaction forces

or Lagrange multipliers from that constraint to act in the ﬂow problem only

and not in the electrical problem. Therefore we introduce the reaction force

¯

f = n ¯ σ from the electrode surface as a separate variable, and use this to

eliminate n ¯ σ in the boundary integral to get

0 =

Ω

∇u : ¯ σ −ξ ∇ ¯ udr +

electrodes

u

¯

f +f

¯ u −

˜

ζ∇

s

¯

φ

ds, (A.13)

1

Unlike the usual notation with Gauss’ law our boundary normal n points out of the

electrodes and channel walls and into the electrolyte and Ω.

A.1 Weakly nonlinear ﬁnite element model 119

for all test functions u, ξ, f, and subject to the no-slip condition ¯ u = 0

on the channel walls. In this form the problem can be typed directly into

Femlab [47] as shown in the code example below:

Code also available on http://www.mic.dtu.dk/mifts

% AC ELECTOOSMOTIC PUMPING BY WEAKLY NONLINEAR MODEL

clear fem

% GEOMETRY

G1 = 1; % gap between electrodes = reference dimension

W1 = 1.5; % width of narrow electrode

W2 = 7; % width of wide electrode

G2 = 5; % gap between adjacent electrode pairs

H = 25; % height of channel

L = W1+W2+G1+G2; % period of electrode array

R = 0.05; % electrode corner radius

fem.geom = rect2(0,L,0,H) - fillet(rect2(G2/2,G2/2+W1,-1,R),’Radii’,R) ...

- fillet(rect2(G2/2+W1+G1,G2/2+W1+G1+W2,-1,R),’Radii’,R);

% MESH

fem.mesh = meshinit(fem,’Hmax’,2,’Hmaxedg’,[4 6; 0.05 0.2]);

% PARAMETERS

fem.const.w = 1; % driving frequency

fem.const.V0 = 10; % driving voltage

fem.const.delta = 1; % surface capacitance ratio

% VARIABLES

fem.sdim = {’x’ ’y’};

fem.dim = {’phi’ ’u’ ’v’ ’p’ ’q’ ’fx’ ’fy’};

fem.shape = [2 2 2 1 2 2 2];

% GOVERNING EQUATIONS

fem.form = ’weak’;

fem.equ.shape = {[1:4]};

fem.equ.weak = {{’phix_test*phix+phiy_test*phiy’ ’-p_test*(ux+vy)’ ...

’ux_test*(-p+2*ux)+vy_test*(-p+2*vy)+(uy_test+vx_test)*(uy+vx)’}};

% BOUNDARY CONDITIONS

% group [1 2] narrow and wide electrode

% group [3] insulating walls

% group [4] master periodic boundary

% group [5] slave periodic boundary

fem.bnd.ind = {[4 9 10] [6 11 12] [2 3 5 7] [1] [8]};

fem.bnd.shape = {[1:7] [1:7] [1:4] [1:4] [1:4]};

fem.bnd.expr = {’Vext’ {’V0*sin(w*t)’ ’-V0*sin(w*t)’ {} {} {}} ...

’zeta’ ’-2*asinh(q/2)’ ’us’ ’zeta*phiTx’ ’vs’ ’zeta*phiTy’};

fem.bnd.weak = {{’phi_test*qt’ ’q_test*(Vext-phi-zeta+q*delta)’ ...

’fx_test*(u-us)+fy_test*(v-vs)’ ’u_test*fx+v_test*fy’} ...

120 FEMLAB code

{’phi_test*qt’ ’q_test*(Vext-phi-zeta+q*delta)’ ...

’fx_test*(u-us)+fy_test*(v-vs)’ ’u_test*fx+v_test*fy’} {} {} {}};

% implement no-slip and periodic boundary conditions as ideal constraints

fem.bnd.constr = {{} {} {’u’ ’v’} {} ...

{’phi-phi_cpl’ ’u-u_cpl’ ’v-v_cpl’ ’p-p_cpl’}};

% POINT CONDITIONS

% clamp pressure at a single point to fix the pressure level

% clamp constraint force [fx fy] where the electrodes and substrate meet

% and [fx fy] are overruled by the ideal constraint Lagrange multipliers

fem.pnt.ind = {[1] [3 6 7 10]};

fem.pnt.shape = {[4] [6 7]};

fem.pnt.constr = {’p’ {’fx’ ’fy’}};

% COUPLING ELEMENT FOR PERIODIC B.C.

% This is black magic. For an explanation see: elcplextr_example.m

% from www.comsol.com/support/knowledgebase/942.php

clear cpl src dst map

% extrusion coupling variables acting in geometry one

cpl.elem = ’elcplextr’;

cpl.g = {’1’};

cpl.var = {’phi_cpl’ ’u_cpl’ ’v_cpl’ ’p_cpl’};

% source expressions and trivial map {0} *from* source *to* intermediate mesh

src.expr = {{’phi’} {’u’} {’v’} {’p’}};

src.map = {’0’ ’0’ ’0’ ’0’};

src.ind = {{’1’}};

cpl.src = {{{} src {}}};

% linear transformation *from* destination *to* intermediate mesh

map.type = ’linear’;

map.sg = ’1’;

map.sv = {’11’,’12’};

map.dg = ’1’;

map.dv = {’1’,’2’};

cpl.map = {map};

% declare {1} as map *from* destination *to* intermediate mesh

dst.map = {{’1’} {’1’} {’1’} {’1’}};

dst.ind = {{’8’}};

cpl.geomdim = {{{} dst {}}};

% finally add cpl to fem.elem and clean up

fem.elem = {cpl};

clear cpl src dst map

% BUILD EXTENDED MESH

fem.xmesh = meshextend(fem);

fem.sol = asseminit(fem);

% TRANSIENT SOLUTION OVER 2 PERIODS USING FEMTIME

% solve nonlinear electrical problem

fem.sol = femtime(fem,’Tlist’,[0 4*pi],’Solcomp’,{’phi’ ’q’});

% solve linear flow problem with constant system matrix

t = fem.sol.tlist;

fem.sol = femlin(fem,’Solcomp’,{’u’ ’v’ ’p’ ’fx’ ’fy’}, ...

A.1 Weakly nonlinear ﬁnite element model 121

’Pname’,’t’,’Plist’,t,’U’,fem.sol,’Keep’,{’K’ ’N’});

% compute volume flow rate Q(t) [large Schmidt number limit]

Q = postint(fem,’u’,’Edim’,1,’Dl’,1,’Solnum’,1:length(t));

% COMPUTE PERIODIC SOLUTION USING CUSTOM WRITTEN ALGORITHM FEMFREQ

fem0 = fem;

nt = 16; tt = 2*pi*(0:nt-1)/nt;

% solve nonlinear electrical problem

fem.sol = femfreq(fem,’Tlist’,tt,’Solcomp’,{’phi’ ’q’},’Symmetry’,’odd’);

% solve linear flow problem

fem.sol = femlin(fem,’Solcomp’,{’u’ ’v’ ’p’ ’fx’ ’fy’}, ...

’Pname’,’t’,’Plist’,tt,’U’,fem.sol,’Keep’,{’K’ ’N’});

% compute time average flow rate <Q>

Qavg = postint(fem,’u’,’Edim’,1,’Dl’,1,’U’,mean(fem.sol.u,2));

plot(t,Q,’b.-’,[0 4*pi],[Qavg Qavg],’r:’)

0 2 4 6 8 10 12

0

0.5

1

1.5

2

2.5

3

3.5

Time t

F

l

o

w

r

a

t

e

Q

Figure A.1: Flowrate Q as a function of time t from weakly nonlinear Femlab code

example. Result from Femlab’s built-in timestepping algorithm femtime (solid line) and

time average from custom written pseudospectral algorithm femfreq (dashed). Strictly,

for εSc ≪1 we ought to solve a stationary Stokes ﬂow problem driven by the time average

slip

˜

ζ∇s

¯

φ

**, while the code example computes a stationary solution at each point in time
**

corresponding to the limit εSc ≫ 1. However, because the problem is linear we get the

same result for the time average ﬂow rate 'Q) in any case.

122 FEMLAB code

A.2 Strongly nonlinear ﬁnite element model

In this section we give an example of our implementation of the strongly non-

linear model from Sec. 2.2.8. The weak form of the bulk electrical problem

∇

¯

J becomes

0 = −

Ω

ϕ∇

¯

Jdr =

Ω

∇ϕ

¯

Jdr +

∂Ω

ϕn

¯

Jds

=

Ω

∇ϕ

¯

Jdr +

electrodes

ϕ

¯

J

n

ds, (A.14)

for all test functions ϕ for the potential

¯

φ at all times, where we used the

Lagrange multiplier

¯

J

n

to eliminate the normal current n

¯

J in the boundary

integral, and the boundary condition for the potential drop across the double

layer is imposed by requiring

0 =

electrodes

.

n

(V

ext

−

¯

φ −

˜

ζ + ˜ q δ) ds, (A.15)

for all test functions .

n

for

¯

J

n

. The conservation law for the accumulated

charge ˜ q in the Debye layer takes the form

0 =

electrodes

̺

∂

t

˜ q +

¯

J

n

+ε∇

s

˜

ds =

electrodes

̺

∂

t

˜ q +

¯

J

n

) −ε∇

s

̺ ˜ ds, (A.16)

for all test functions ̺ for ˜ q, where we used partial integration to move the

derivative ∇

s

onto ̺ and assumed ˜ = 0 at the electrode edges. For the bulk

salt concentration ¯ c we write

0 = −

Ω

C ∇

¯

F

a

dr =

Ω

∇C

¯

F

a

dr +

∂Ω

C n

¯

F

a

ds

=

Ω

∇C

¯

F

a

dr +

electrodes

C

ˇ

F

n

−∆D

¯

J

n

ds, (A.17)

for all test functions C for ¯ c. Alternatively, it sometimes gives more stable

numerical results to rewrite ∇

¯

F

a

in non-conservative form

∇

¯

F

a

= ∇ (−D

a

∇¯ c + Pe ¯ u¯ c) = −D

a

∇

2

¯ c + Pe ¯ u ∇¯ c, (A.18)

using ∇ ¯ u = 0, and performing partial integration only on the ∇

2

¯ c term.

The conservation law for the surface excess accumulated salt ˜ w becomes

0 =

electrodes

̟(∂

t

˜ w +

ˇ

F

n

+ε∇

s

˜

Γds =

electrodes

̟(∂

t

˜ w +

ˇ

F

n

) −ε∇

s

̟

˜

Γds, (A.19)

for all test functions ̟ for ˜ w, which eﬀectively determines the normal ﬂux

ˇ

F

n

, whereas the excess amount of salt is determined by

0 =

electrodes

T

n

˜ w −

˜ q

2

+ 4ˇ c +

√

4ˇ c

ds, (A.20)

A.2 Strongly nonlinear ﬁnite element model 123

for all test functions T

n

for

ˇ

F

n

. Finally, the surface excess concentration

ˇ γ = ˇ c − ¯ c at ˇ y = 0 is determined by the (excess) ambipolar ﬂux in and out

of the diﬀusion layer

0 =

electrodes

Y

ˇ γ −

ε

ωD

a

1

T

T

0

((t −t

′

)

ˇ

F

n

−∆D

¯

J

n

dt

′

ds, (A.21)

for all test functions Y for ˇ γ; it is not necessary to subtract

ˇ

F

n

− ∆D

¯

J

n

**because the kernel ((t − t
**

′

) is blind to any time average component by

construction.

In the strongly nonlinear model we need to include the bulk electrical

body force

¯

f into the Stokes ﬂow problem, where

¯

f = ∇

2

¯

φ∇

¯

φ = −

∇¯ c ∇

¯

φ + ∆D∇

2

¯ c

¯ c

∇

¯

φ, (A.22)

cf. Eq. (2.60). The dependence on ∇

2

¯

φ or ∇

2

¯ c is problematic because

with standard ﬁnite element approximations the solution is only continuous

and piecewise diﬀerentiable, with the second derivative undeﬁned on ele-

ment boundaries. In the special case of equal diﬀusivities where ∆D = 0,

Eq. (A.22) expresses

¯

f in terms of the gradients ∇¯ c and ∇

¯

φ only, and the

problem does not arise. Otherwise it is necessary to rewrite the body force

as

¯

f = ∇

2

¯

φ∇

¯

φ = ∇

¯

T, (A.23)

where

¯

T is the Maxwell stress tensor

¯

T

ij

= ∂

i

¯

φ∂

j

¯

φ −δ

ij

∂

k

¯

φ∂

k

¯

φ/2. (A.24)

With this formulation we can use Gauss’ law to move the derivative from

¯

T onto the test function u for the ﬂuid velocity, and the weak form of the

Stokes ﬂow problem becomes

0 =

Ω

∇u : (¯ σ +

¯

T) −ξ ∇ ¯ udr +

electrodes

u ¯ µ +f (¯ u − ˜ u

s

) ds, (A.25)

for all test functions u, ξ, f, and subject to the no-slip condition on the

channel walls. Here ¯ µ = n (¯ σ +

¯

T) is the total reaction force from the

electrode surface, and ˜ u

s

=

˜

ζ +ˇ ς

∇

s

¯

φ

**is the eﬀective slip velocity induced
**

in the Debye and diﬀusion layers.

In the code example below we consider the simple equal diﬀusivity case

with ∆D = 0 for the symmetric geometry from Sec. 4.2.1. For simplicity

we assume an open system and impose unit concentration ¯ c = 1 and no

shear force n ¯ σ = 0 boundary conditions far above the electrode array. For

∆D = 0 both

¯

φ, ˜ q, and

¯

J

n

have odd symmetry in time, ˜ w,

ˇ

F

n

, and ˇ γ have

even symmetry, whereas ¯ c, ¯ u, ¯ p, and ¯ µ are constant in time. This allows

the computational cost of solving the full coupled problem to be reduced

124 FEMLAB code

by a factor of two as compared to the case ∆D = 0. The convection from

ﬂuid motion induced in the diﬀusion layer is neglected when solving the

coupled electrohydrodynamical problem, but in a postprocessing step we

single out the components of the net ﬂuid motion due to electroosmotic slip,

bulk electroconvection, and ﬂow induced in the diﬀusion layer. The auxil-

iary function fkernel evaluates the kernel ((t − t

′

) at ˇ y = 0, and excessr

computes the integral

ω/D

a

∞

0

1

ˇ c

−

1

¯ c

dˇ y given ˇ γ(t) and ¯ c by numerical

quadrature.

Code also available on http://www.mic.dtu.dk/mifts

% AC ELECTOKINETIC FLOW BY STRONGLY NONLINEAR MODEL

clear fem

% GEOMETRY

G = 1; % gap between electrodes

W = 2; % width of electrodes

H = 5; % height of channel

L = W+G; % period of electrode array

R = 0.1; % electrode corner radius

fem.geom = rect2(0,L/2,0,H) - fillet(rect2(-W/2,W/2,-1,R),’Radii’,R);

% MESH

fem.mesh = meshinit(fem,’Hmax’,0.75,’Hmaxedg’,[2; 0.1]);

% PARAMETERS

fem.const.omega = 1; % driving frequency

fem.const.V0 = 50; % driving voltage

fem.const.delta = 1; % surface capacitance ratio

fem.const.epsilon = 0.001; % Debye layer thickness

fem.const.Pe = 0.25; % Peclet number

% VARIABLES

fem.sdim = {’x’ ’y’};

fem.dim = {’phi’ ’c’ ’u’ ’v’ ’p’ ’q’ ’w’ ’Jn’ ’Fn’ ’gamma’ ’Zeta’ ’fx’ ’fy’};

fem.shape = [2 2 2 2 1 2 2 2 2 2 2 2 2];

% GOVERNING EQUATIONS

fem.form = ’weak’;

fem.equ.shape = {[1:5]};

fem.equ.expr = {’rho’ ’(cx*phix+cy*phiy)/c’ ...

’Jx’ ’-c*phix’ ’Jy’ ’-c*phiy’ ’Fx’ ’-cx+Pe*u*c’ ’Fy’ ’-cy+Pe*v*c’};

fem.equ.weak = {{’phix_test*Jx+phiy_test*Jy’ ...

’-cx_test*cx-cy_test*cy-c_test*Pe*(u*cx+v*cy)’ ...

’ux_test*(-p+2*ux)+vy_test*(-p+2*vy)+(uy_test+vx_test)*(uy+vx)’ ...

’-p_test*(ux+vy)’ ’u_text*rho*phix+v_test*rho*phiy’}};

fem.equ.init = {{0 1 0 0 0 0 0 0 0 0 0 0 0}};

% BOUNDARY CONDITIONS

A.2 Strongly nonlinear ﬁnite element model 125

% group [1] electrode

% group [2] substrate

% group [3] even symmetry

% group [4] odd symmetry

% group [5] "infinity"

fem.bnd.ind = {[2 6] [4] [1] [5] [3]};

fem.bnd.shape = {[1:13] [1:5] [1:5] [1:5] [1:5]};

fem.bnd.expr = {’Vext’ {’V0*sin(omega*t)’ {} {} {} {}} ...

’zeta’ ’-2*asinh(q/(2*sqrt(c+gamma)))’ ...

’Gammax’ ’-(1+2*Pe)*q*phiTx’ ’Gammay’ ’-(1+2*Pe)*q*phiTy’ ...

’jx’ ’-(1+2*Pe)*w*phiTx’ ’jy’ ’-(1+2*Pe)*w*phiTy’ ...

’us’ ’(zeta+Zeta)*phiTx’ ’vs’ ’(zeta+Zeta)*phiTy’};

fem.bnd.weak = {{’phi_test*Jn’ ’Jn_test*(Vext-phi-zeta+q*delta)’ ...

’c_test*Fn’ ’Fn_test*(w-sqrt(q^2+4*(c+gamma))+sqrt(4*(c+gamma)))’ ...

’q_test*(qt+Jn)-epsilon*(qTx_test*jx+qTy_test*jy)’ ...

’w_test*(wt+Fn)-epsilon*(wTx_test*Gammax+wTy_test*Gammay)’ ...

’fx_test*(u-us)+fy_test*(v-vs)+u_test*fx+v_test*fy’ ...

’gamma_test*(gamma-Fntt*sqrt(epsilon/omega))’} {} {} {} {}};

fem.bnd.constr = {{} {’u’ ’v’} {’u’} {’u’ ’phi’} {’c-1’}};

fem.bnd.init = {{0 1 0 0 0 0 eps 0 0 0 0 0 0}};

% POINT CONSTRAINTS

% clamp weak constraint force where overruled by ideal constraints

fem.pnt.ind = {[1] [4]};

fem.pnt.shape = {[1:13] [1:13]};

fem.pnt.constr = {{’fx’} {’fx’ ’fy’}};

% BUILD EXTENDED MESH

fem.xmesh = meshextend(fem);

% COMPUTE PERIODIC SOLUTION USING FEMFREQ

nt = 32; tlist = 2*pi*(0:nt-1)/nt;

fem.sol = femfreq(fem,’Tlist’,tlist, ...

’Solcomp’,{’phi’ ’c’ ’q’ ’w’ ’Jn’ ’Fn’ ’gamma’ ’u’ ’v’ ’p’ ’fx’ ’fy’}, ...

’Symmetry’,{’odd’ ’const’ ’odd’ ’even’ ’odd’ ’even’ ’even’ ...

’const’ ’const’ ’const’ ’const’ ’const’}, ...

’Rule’,’leapfrog’,’Kernel’,fkernel(0,nt,4));

% GET INDICES OF [Zeta gamma Jn c] IN GLOBAL SOLUTION VECTOR

iZeta = find(asseminit(fem,’Init’,{’Zeta’ 1},’Out’,’U’));

ndofs = flngdof(fem);

igamma = asseminit(fem,’Init’,{’Zeta’ ’gamma’},’U’,(1:ndofs)’,’Out’,’U’);

iJn = asseminit(fem,’Init’,{’Zeta’ ’Jn’},’U’,(1:ndofs)’,’Out’,’U’);

ic = asseminit(fem,’Init’,{’Zeta’ ’c’},’U’,(1:ndofs)’,’Out’,’U’);

igamma = igamma(iZeta); iJn = iJn(iZeta); ic = ic(iZeta);

% COMPUTE EXCESS POTENTIAL DROP ACROSS THE DIFFUSION LAYER

gamma = fem.sol.u(igamma,:); Jn = fem.sol.u(iJn,:); c = fem.sol.u(ic,1);

R = excessr(gamma,c);

Zeta = sqrt(fem.const.epsilon/fem.const.omega)*Jn.*R;

u0 = fem.sol.u; u0(iZeta,:) = Zeta;

fem.sol = femsol(u0,’tlist’,tlist);

126 FEMLAB code

% COMPUTE INDIVIDUAL FLOW COMPONENTS

% flow due to slip induced in the Debye layer

ueo = femlin(fem,’Solcomp’,{’u’ ’v’ ’p’ ’fx’ ’fy’}, ...

’Const’,{’Zeta’ 0 ’rho’ 0},’U’,fem.sol,’Out’,’U’, ...

’Pname’,’t’,’Plist’,tlist,’Keep’,{’K’ ’N’});

% flow due to slip induced in the diffusion layer

uec = femlin(fem,’Solcomp’,{’u’ ’v’ ’p’ ’fx’ ’fy’}, ...

’Const’,{’zeta’ 0 ’rho’ 0},’U’,fem.sol,’Out’,’U’, ...

’Pname’,’t’,’Plist’,tlist,’Keep’,{’K’ ’N’});

% flow due to bulk electroconvection

ub = femlin(fem,’Solcomp’,{’u’ ’v’ ’p’ ’fx’ ’fy’}, ...

’Const’,{’zeta’ 0 ’Zeta’ 0},’U’,fem.sol,’Out’,’U’, ...

’Pname’,’t’,’Plist’,tlist,’Keep’,{’K’ ’N’});

% COMPUTE ANGULAR MOMENTUM FROM FLUID VORTEX

Leo = postint(fem,’x*v-y*u’,’U’,mean(ueo,2));

Lec = postint(fem,’x*v-y*u’,’U’,mean(uec,2));

Lb = postint(fem,’x*v-y*u’,’U’,mean(ub,2));

Ltot = Leo + Lec + Lb;

A.3 Periodic integro-diﬀerential equations in time 127

A.3 Periodic integro-diﬀerential equations in time

We wish to solve an integro-diﬀerential equation in time of the form

D∂

t

U+

1

T

T

0

G(t, t

′

)U(t

′

) dt

′

= L −N

T

Λ, (A.26)

subject to the constraint

0 = M, (A.27)

where the solution U(t) is periodic in time, i.e., U(t) = U(t + T). Here

D(U, t) is the mass matrix, G(t, t

′

) the kernel matrix in the convolution,

L(U, t) the residual vector, M(U, t) the constraint vector, Λ(t) a Lagrange

multiplier for that constraint, and N = −∂M/∂U the matrix that couples

the Lagrange multiplier into the problem. Apart from the convolution oper-

ator, this is identical to the generic form of a time-dependent problem after

discretization in Femlab. However, we wish to use a relaxation method

to compute the periodic solution instead of the standard time-steppig algo-

rithm. Linearizing around a given trial solution U

0

(t) we get

D∂

t

δU+

1

T

T

0

G(t, t

′

) δU(t

′

) dt

′

= L

0

−KδU−N

T

Λ, (A.28)

subject to

0 = M

0

−NδU. (A.29)

The derivative and integral terms of the trial solution are absorbed into the

residual vector

L

0

= L(U

0

, t) −D(U

0

, t) ∂

t

U

0

−

1

T

T

0

G(t, t

′

)U

0

(t

′

) dt

′

, (A.30)

and further M

0

= M

U

0

, t

**and K = −∂L/∂U. We neglect the dependence
**

of D and N on U when linearizing; this may cause problems in the Newton

itertions if those are strongly nonlinear functions.

If we represent the temporal solution for U by N equispaced points in

time t

n

= nT/N for n = 1, 2, . . . , N then it can be written as a discrete

Fourier transform

U(t

n

) =

N

¸

m=1

ˆ

U

m

e

imω tn

, (A.31)

where the coeﬃcients

ˆ

U

m

=

1

N

N

¸

n=1

U(t

n

)e

−imω tn

(A.32)

128 FEMLAB code

can be computed by a fast Fourier transform. By construction

ˆ

U

m

=

ˆ

U

m+N

,

and since U is real we also have

ˆ

U

m

=

ˆ

U

∗

−m

=

ˆ

U

∗

N−m

. The appropriate

way of interpreting U(t) for t = t

n

is

U(t) =

1

N

N/2

¸

m=−N/2

′

ˆ

U

m

e

imωt

=

N

¸

n=1

U(t

n

)

1

N

N/2

¸

m=−N/2

′

e

imω(t−tn)

, (A.33)

where the prime indicates that the terms m = ±N/2 are counted half only.

For an explanation why this is “appropriate”, and for a nice introduction to

spectral methods in general, see Ref. [68]. The function in the curly braces

is the periodic sinc function S

N

(t −t

′

)

S

N

(t −t

′

) =

1

N

N/2

¸

m=−N/2

′

e

imω(t−t

′

)

=

1

N

sin[Nω(t −t

′

)/2]

tan[ω(t −t

′

)/2]

, (A.34)

which allows the time detivative of U(t) to be evaluated as

∂

t

U(t

n

) =

N

¸

m=1

S

′

N

(t

n

−t

m

)U(t

m

). (A.35)

The convolution integral can be approximated as

1

T

T

0

G(t

n

, t

′

)U(t

′

) dt

′

≈

1

N

N

¸

m=1

G(t

n

, t

m

)U(t

m

). (A.36)

Substituting into the linearized problem we get

N

¸

m=1

D

n

S

′

N

(t

n

−t

m

) +

1

N

G

nm

δU

m

+K

n

δU

n

+N

T

n

Λ

n

= L

0,n

, (A.37)

N

n

δU

n

= M

0,n

, (A.38)

for n = 1, 2, . . . , N, i.e., a simultaneous matrix problem for all the solution

components δU

m

.

Unless N is small this problem will make most linear solvers choke; the

point is that for harmonically driven weakly nonlinear systems the num-

ber of Fourier components required to represent the solution is often fairly

small. If the problem cannot be solved accurately with just a few Fourier

components, we could replace the spectral derivative with a low order ﬁnite-

diﬀerence scheme, e.g., the leapfrog scheme [∂

t

U]

n

≈ [U

n+1

− U

n−1

]/2∆t,

which would reduce the computational cost for a diﬀerential equation from

O(N

2

) to O(N) at the expense of the spectral accuracy. For comparison,

the cost of one step forward in time by a standard timestepping algoritm

is O(1) which is much more memory eﬃcient but sometimes requires inte-

grating for a long time before a periodic solution is reached. If the problem

A.3 Periodic integro-diﬀerential equations in time 129

involves convolution there is no obvious way to work around the nonlocal

coupling in time; a pseudo-timestepping algorithm designed speciﬁcally for

rapid convergence towards the periodic solution might provide a fairly in-

expensive solution to the problem. However, we have not had time to look

further into this subject.

For our particular application where the nonlocal coupling in time comes

about from our semi-analytical model for the dynamics in the diﬀusion layer,

the convolution kernel does not depend explicitly on time but only on the

time delay, i.e., G = G(t − t

′

), and it is originally deﬁned in terms of an

inﬁnite Fourier series

G(t −t

′

) =

∞

¸

n=−∞

ˆ

G

n

e

inω(t−t

′

)

. (A.39)

Substituting this and the truncated series Eq. (A.33) for U(t) into the con-

volution integral we obtain

1

T

T

0

G(t

n

−t

′

)U(t

′

) dt

′

=

1

N

N

¸

m=1

G

N

(t

n

−t

m

)U(t

m

), (A.40)

where

G

N

(t −t

′

) =

N/2

¸

k=−N/2

′

ˆ

G

k

e

ikω(t−t

′

)

. (A.41)

This essentially tells us how to truncate the inﬁnite series Eq. (A.39) in a

consistent way. If the spectral time derivative is replaced by the leapfrog

rule, it is equivalent to replacing S

N

(t −t

′

) by a piecewise linear interpolant,

namely, the “hat” function P

N

(t −t

′

) = max¦1 −N[t −t

′

[, 0¦. In that case

it is more consistent to truncate the convolution kernel as

G

N

(t −t

′

) = N

T

0

G(t −t

′′

)P

N

(t

′′

−t

′

) dt

′′

. (A.42)

Implementation: FEMFREQ

We have implemented the relaxation scheme in a function femfreq. It com-

putes periodic solutions to simple ﬁrst order problems in time with a syntax

similar to Femlab’s standard time-stepping algorithm femtime. For prob-

lems involving convolution we abuse the syntax for second order problems by

interpreting the term E∂

2

t

Uinstead as E

T

0

((t−t

′

)U(t

′

) dt

′

, where ((t−t

′

)

is a some scalar function supplied by the user via the property ’Kernel’.

This approach is signiﬁcantly less ﬂexible than a general G(t, t

′

) that could

depend explicitly on both time and spatial position, but it serves our pur-

pose. We also support odd, even, and constant symmetries in the solution

which allows the computational cost to be reduced. We do not include the

code here but it is available at http://www.mic.dtu.dk/mifts.

130 FEMLAB code

Appendix B

Paper published in

Phys. Rev. E

Title

AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement, Faradaic

current injection, and nonlinear surface capacitance

Authors

Laurits Højgaard Olesen, Henrik Bruus, and Armand Ajdari

Reference

Phys. Rev. E 73, 056313 (2006) (16 pages)

132 Paper published in Phys. Rev. E

ac electrokinetic micropumps: The effect of geometrical conﬁnement, Faradaic current injection,

and nonlinear surface capacitance

Laurits Højgaard Olesen,

1,2

Henrik Bruus,

1

and Armand Ajdari

2

1

MIC, Department of Micro and Nanotechnology, Technical University of Denmark,

DTU bldg. 345 east, DK-2800 Kongens Lyngby, Denmark

2

Laboratoire de Physico-Chimie Théorique, UMR CNRS-ESPCI 7083, 10 rue Vauquelin, F-75231 Paris Cedex 05, France

͑Received 9 October 2005; published 31 May 2006͒

Recent experiments have demonstrated that ac electrokinetic micropumps permit integrable, local, and fast

pumping ͑velocities ϳmm/ s͒ with low driving voltage of a few volts only. However, they also displayed many

quantitative and qualitative discrepancies with existing theories. We therefore extend the latter theories to

account for three experimentally relevant effects: ͑i͒ vertical conﬁnement of the pumping channel, ͑ii͒ Faradaic

currents from electrochemical reactions at the electrodes, and ͑iii͒ nonlinear surface capacitance of the Debye

layer. We report here that these effects indeed affect the pump performance in a way that we can rationalize by

physical arguments.

DOI: 10.1103/PhysRevE.73.056313 PACS number͑s͒: 47.65.Ϫd, 47.32.Ϫy, 85.90.ϩh

I. INTRODUCTION

Lab-on-a-chip systems require micropumps and valves to

manipulate small volumes of a liquid sample ͓1͔. Often large

external pumps, such as syringe pumps, are used to deliver

the necessary pressures, but for portable systems a number of

on-chip micropumps have been developed over the last de-

cade. The latter fall into two major categories: One category

comprises mechanical actuation and deﬂectable membranes

to create pumps and valves. These have proven to be versa-

tile and simple to fabricate and operate, but there are some

difﬁculties with further downscaling. The second category

involves no moving parts but uses electric ﬁelds to induce

electrokinetic pumping. However, classical dc electro-

osmosis requires a relatively large voltage and ﬁeld strength,

which is undesirable. Moreover, when a dc current is drawn

from electrodes integrated in a microﬂuidic system there are

problems with bubble formation due to electrolysis.

ac electro-osmosis has recently been observed to induce

ﬂuid motion over pairs of microelectrodes ͓2–4͔. Based on

general symmetry arguments, Ajdari predicted that the same

mechanism would generate a net ﬂow over an asymmetric

array of electrodes ͓5͔, which was soon after demonstrated

experimentally by Brown et al. ͓6͔ and later by several other

groups obtaining pumping velocities ϳmm/ s with driving

voltage of a few volts ͓7–14͔.

Theoretically, Ramos et al. performed a more detailed lin-

ear response investigation of the pumping mechanism and

found a speciﬁc pumping direction, a pumping velocity scal-

ing as the square of the driving voltage, and a well-deﬁned

frequency for maximal pumping ͓15͔.

However, the above analysis is strictly valid only at low

driving voltage V

0

Ӷk

B

T/ eϳ25 mV, whereas experiments

are usually performed with at least a few volts. As a conse-

quence, the theory disagrees with experimental observations

in several ways: e.g., experimentally the pumping velocity

does not always compare well with the predicted V

0

2

scaling,

but looks more like a linear scaling ͓10,11͔ or tends to satu-

rate at large voltage ͓13͔. Moreover, Studer et al. found a

reversal of the pumping direction at driving frequencies well

above that for maximal pumping in the “forward” direction

͓11͔ and Ramos and co-workers observed reversal of the

pumping direction on a traveling-wave device for driving

voltages above 2 V at 1 kHz driving frequency ͓13,14͔. The

existing theory does not give many clues as to the mecha-

nism for this reversal, except that in the original paper Ajdari

predicted pumping in the “reverse” direction at low fre-

quency when Faradaic electrode reactions were included in

the model ͓5͔. Indeed, several groups have observed bubble

formation and electrode degradation at low frequency and

high voltage which indicates that electrode reactions are ac-

tually taking place ͓8–11,14͔. Lastochkin et al. observed

ﬂuid motion in the reverse direction of that usually expected

for ac electro-osmosis, which was attributed to strong Fara-

daic current injection ͓16͔. However, their experiments were

performed with driving frequencies in the MHz range which

is of the order of the Debye frequency for the electrolyte,

whereas ac electro-osmosis generally occurs around the in-

verse RC time of the device.

These discrepancies between experimental observations

and existing theory demonstrate the need for a more com-

plete theoretical understanding. As a ﬁrst step we here ad-

dress a few generalizations of the existing linear response

theory to the weakly nonlinear regime by taking into ac-

count, within the thin Debye layer approximation, the effect

of Faradaic currents both in a linearized scheme and using

the full nonlinear Butler-Volmer reaction kinetics. We also

include the nonlinear surface capacitance of the Debye layer

as described in Gouy-Chapman theory.

The paper is organized as follows: In Sec. II we ﬁx the

device geometry and in Sec. III we describe our model for

the electrokinetic system. In Sec. IV we extend the linear

analysis for low driving voltage to study in more detail the

effects of the device geometry and vertical conﬁnement of

the system, and we investigate the effect of Faradaic current

injection in a linearized scheme. This provides a ﬁrm starting

point when in Sec. V we study the nonlinear model both with

and without Faradaic current. In Sec. VI we summarize our

results and compare to experiments reported in the literature,

and ﬁnally, in Sec. VII we conclude the paper.

PHYSICAL REVIEW E 73, 056313 ͑2006͒

1539-3755/2006/73͑5͒/056313͑16͒ ©2006 The American Physical Society 056313-1

II. DEVICE GEOMETRY

We consider a two-dimensional ͑2D͒ geometry like that

shown in a side view in Fig. 1, similar to devices used ex-

perimentally ͓6–11,14͔. It consists of a substrate on which an

asymmetric array of interdigitated microelectrodes is depos-

ited. Each electrode pair in the array consists of a narrow

electrode of width W

1

and a wide electrode of width W

2

separated by a narrow and a wide gap G

1

and G

2

, respec-

tively, and the full period is L=W

1

+G

2

+W

2

+G

2

. Notice that

in order to break the left-right symmetry of the array it is

necessary that both W

1

W

2

and G

1

G

2

. On top of the

electrode array a microﬂuidic channel of height H is placed

and we assume that the device extends sufﬁciently far into

the third dimension that a 2D description in the xy plane is

appropriate.

The channel is ﬁlled with an electrolyte, and the narrow

and wide electrodes are biased with an ac voltage V

ext

= ±V

0

cos͑t͒, respectively; i.e., the applied voltage differ-

ence is 2V

0

cos͑t͒. This induces the formation of a Debye

layer on the electrodes that acts to partially screen the elec-

tric ﬁeld. The nonuniform partial screening gives rise to tan-

gential electric ﬁeld in the Debye layer and electro-osmotic

ﬂow. This ﬂow takes the form of ﬂuid rolls above the elec-

trode edges as indicated in Fig. 1.

III. ELECTROKINETIC MODEL

Our model for the electrohydrodynamics of the system is

similar to that of Ramos et al. ͓15͔ and is based on the

following classical approximations ͓17–20͔: ͑i͒ The bulk

electrolyte is assumed to be charge neutral with uniform salt

concentration, such that the ionic transport can be described

as Ohmic current. ͑ii͒ The Debye layer is assumed to be in

local equilibrium with the electrolyte immediately outside

the layer, such that the charge distribution and potential

variation in the Debye layer can be described by Gouy-

Chapman theory. ͑iii͒ Moreover, we assume that the thick-

ness of the Debye layer is much smaller than the size of the

electrodes, and we neglect surface diffusion and migration of

charge. ͑iv͒ The bulk ﬂuid motion is described by Stokes

ﬂow with a slip condition on the electrodes set by the elec-

troosmotic ﬂow induced in the Debye layer. ͑v͒ Finally, we

also assume that the bulk concentration of the reactants in

the Faradaic electrode reaction is constant; i.e., we neglect

the effect of mass transfer.

Upon these approximations, the only dynamical variables

we are left to consider are the potential distribution ͑x, y, t͒

in the bulk and the local surface charge density q͑x, t͒ accu-

mulated in the Debye layer on the electrodes. The instanta-

neous value of q determines the potential drop from the elec-

trode to the electrolyte immediately outside the Debye layer,

whereas the time evolution of q is determined by the balance

between the Ohmic current from the bulk and Faradaic cur-

rent from electrochemical reactions on the electrodes.

We emphasize that this simple electrokinetic model is ap-

propriate only in the weakly nonlinear regime of not-too-

high driving voltage ͑Շ125 mV͒. Beyond this, the assump-

tion of uniform bulk electrolyte concentration breaks down,

the effect of mass transfer on the Faradaic reaction kinetics

cannot be neglected, and the Debye layer may be driven

out of local equilibrium. This is discussed in more detail in

Sec. VI.

A. Bulk electrolyte

In the bulk electrolyte we assume charge neutrality such

that the charge continuity equation reduces to

١ · J = 0. ͑1͒

The electric current J is described simply as the Ohmic cur-

rent J=−١ where is the conductivity of the electrolyte.

Since we assume uniform salt concentration and conductivity

throughout the bulk, we then simply end up with a Laplace

problem for the potential ͑r, t͒.

B. Debye layer

The Debye layer is assumed to be in local equilibrium

with the electrolyte immediately outside the layer. This im-

plies that the driving frequency should be well below the

Debye frequency

D

=/ ⑀ of the electrolyte and that the

Faradaic current injection cannot be too strong. The total

charge accumulated in the Debye layer can then be directly

related to the potential drop ͑x, t͒ across it ͓18͔

q͑x, t͒ =

͵

Debye layer

͑x, y, t͒dy ͑2͒

=− sgn͑͒

ͱ

2⑀k

B

T

͚

n

c

n

*

͑e

−z

n

e/k

B

T

− 1͒, ͑3͒

where is the charge density, ⑀ the permittivity of the sol-

vent, k

B

Boltzmann’s constant, T the temperature, c

n

*

the bulk

concentration ͑number density͒ of the nth ionic species, and

z

n

its valence. In the Debye-Hückel limit Ӷk

B

T/ ze, Eq. ͑3͒

can be linearized to

q = −

⑀

D

, ͑4͒

where

D

=

ͱ

⑀k

B

T/ ͚

n

c

n

*

z

n

2

e

2

is the Debye length and ⑀/

D

is

the Debye layer capacitance ͑per unit area͒ at low voltage.

FIG. 1. Sketch of the device geometry. The interdigitated elec-

trode array is biased with an ac voltage that induces a buildup of a

Debye screening layer and electro-osmotic ﬂuid motion.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-2

For larger potentials we focus on the simple case of a

monovalent symmetric ͑1:1͒ electrolyte where Eq. ͑3͒ re-

duces to

q = −

⑀

D

2k

B

T

e

sinhͩ

e

2k

B

T

ͪ. ͑5͒

As in the classical Gouy-Chapman-Stern model we assume

that the compact ͑Stern͒ layer on the electrode simply gives

rise to an additional surface capacitance C

s

in series with the

Debye layer ͓18͔. Alternatively C

s

could model an oxide

layer grown intentionally on the electrodes to inhibit electro-

chemical reactions ͓4,12,20͔. In either case we shall for sim-

plicity assume that C

s

is independent of potential. The total

potential drop across the double layer ͑Debye and Stern͒ is

V

ext

− = −

1

C

s

q, ͑6͒

where V

ext

is the electrode potential and is the potential in

the electrolyte immediately outside the Debye layer. The

overall capacitance ͑per unit area͒ of the double layer is de-

ﬁned as

C͑͒ = −

q

V

ext

−

. ͑7͒

In the Debye-Hückel limit this reduces to

C

0

=

ͫ

D

⑀

+

1

C

s

ͬ

−1

=

1

1 + ␦

⑀

D

, ͑8͒

where we introduced the parameter ␦=⑀/

D

C

s

for the sur-

face capacitance ratio. At larger potential the Debye layer

capacitance becomes very large and C͑͒ is dominated by the

Stern layer only. Notice that C͑͒ always satisﬁes C

0

ഛC͑͒ ഛC

s

.

The charging of the Debye layer is due to both Ohmic

current from the bulk and Faradaic current from electro-

chemical reactions at the electrode:

ץ

t

q = − n · J + j

ext

, ͑9͒

where n is a unit normal pointing out of the electrode, −n· J

is the Ohmic current running into the Debye layer, and j

ext

is

the Faradaic current from the electrode reaction. We neglect

any surface diffusion and migration of charge in the Debye

layer.

C. Insulating walls

On the channel substrate between the electrodes and on

the channel lid we assume no normal current:

n · J = 0. ͑10͒

This is justiﬁed because the amount of charge required to

screen any ﬁeld normal to the channel wall is both very small

and builds up on the Debye time scale

D

=⑀/ .

D. Electrochemistry

For the Faradaic electrode reaction we consider a simple

one-step, one-electron redox process of the form

O + e R, ͑11͒

where z

O

=z

R

+1 to ensure charge conservation. Activated-

complex theory predicts an electric current density j

ext

from

this process given by ͓18͔

j

ext

= e͑k

R

c

R

s

e

͑1−␣͒e⌬

s

/k

B

T

− k

O

c

O

s

e

−␣e⌬

s

/k

B

T

͒, ͑12͒

where k

O

and k

R

**are the forward and backward standard rate
**

constants for the reaction, c

O

s

and c

R

s

are the concentrations of

the oxidized and reduced species directly at the electrode

͑Stern layer͒ surface, ␣ is the transfer coefﬁcient, and ⌬

s

=−q/ C

s

is the potential drop across the Stern layer. Assum-

ing local equilibrium, the concentrations directly at the sur-

face are related to those immediately outside the Debye layer

by a Boltzmann factor, c

n

s

=c

n

e

−z

n

e/k

B

T

. Then Eq. ͑12͒ can be

reexpressed as the current-overpotential equation

j

ext

= j

0

e

−͑z

O

−␣͒e/k

B

T

ͩ

c

R

c

R

*

e

͑1−␣͒e/k

B

T

−

c

O

c

O

*

e

−␣e/k

B

T

ͪ

,

͑13͒

where j

0

=ek

͑c

R

*

͒

␣

͑c

O

*

͒

1−␣

is the exchange current, k

is the

standard rate constant, c

O

and c

R

are the concentrations im-

mediately outside the Debye layer, and c

O

*

and c

R

*

are the

corresponding bulk values. =͑V

ext

−͒ −͑V

ext

−͒

eq

is the

difference between the actual potential drop across the

double layer and the thermal equilibrium for the given redox

process and bulk concentration; is also termed the overpo-

tential. Appreciating that c

n

s

in Eq. ͑12͒ differs from c

n

by a

Boltzmann factor is known as the Frumkin correction to the

standard Butler-Volmer equation.

Linearization for small and for c

O

and c

R

close to their

bulk values yields

j

ext

= j

0

e

−͑z

O

−␣͒e

eq

/k

B

T

ͩ

c

R

c

R

*

−

c

O

c

O

*

+

e

k

B

T

ͪ

, ͑14͒

where

eq

is the intrinsic equilibrium potential drop across

the Debye layer.

For simplicity we shall take

eq

=0 such that =V

ext

− in

Eqs. ͑13͒ and ͑14͒, set the transfer coefﬁcient ␣ equal to 1/ 2,

and assume that c

O

and c

R

are virtually at their bulk values;

i.e., we neglect the effect of mass transfer. Moreover, for

surface potentials outside the Debye-Hückel limit we shall

focus on a monovalent symmetric ͑1:1͒ electrolyte with z

O

=z

R

+1=1. Then Eqs. ͑13͒ and ͑14͒ reduce to

j

ext

= j

0

e

−e/2k

B

T

2 sinh

ͫ

e͑V

ext

− ͒

2k

B

T

ͬ

͑15͒

and

j

ext

= j

0

e͑V

ext

− ͒

k

B

T

=

V

ext

−

R

ct

, ͑16͒

respectively, deﬁning the ͑area speciﬁc͒ charge-transfer resis-

tance R

ct

=k

B

T/ j

0

e of units ͓⍀m

2

͔.

E. Fluid dynamics

Since we are considering a microsystem where the Rey-

nolds number is usually very low—i.e., where viscosity

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-3

dominates over inertia—the ﬂuid motion in the electrolyte is

described by the Stokes equation

m

ץ

t

u = − ١p + ١

2

u, ͑17͒

together with the incompressibility constraint

١ · u = 0. ͑18͒

Here

m

is the ﬂuid mass density, u the velocity, p the pres-

sure, and the dynamic viscosity. On the insulating walls

the no-slip boundary condition applies, whereas on the elec-

trodes we impose a tangential slip condition based on the

Helmholtz-Smoluchowski slip velocity:

u

s

= −

⑀

E

t

, ͑19͒

where E

t

=−ץ

x

is the tangential ﬁeld. This classical result

for the slip at a ﬂat surface with an externally applied tan-

gential ﬁeld also holds in general for thin Debye layers in

quasiequilibrium at a metal-electrolyte interface ͓21͔.

The ﬂuid ﬂow pattern is complex with rolls above the

electrode edges as sketched in Fig. 1. However, we are

mostly concerned with the net pumping ͗Q͘ =͐

0

H

͗u

x

͘dy. A

Fourier analysis shows that on average in time and space the

ﬂow is a simple Couette ﬂow driven by the average slip

velocity U on the bottom wall ͓15͔

U =

1

L

ͭ

͵

x

1

x

2

͗u

s

͘dx +

͵

x

3

x

4

͗u

s

͘dx

ͮ

, ͑20͒

in terms of which ͗Q͘ =HU/ 2—that is, assuming zero back-

pressure on the device. Here x

i

denote the positions of the

electrode edges ͑cf. Fig. 1͒ and the insulating walls do not

contribute due to the no-slip condition.

Notice that u

s

and hence U are determined entirely by the

solution to the electrical problem. Therefore we can study

pumping without resolving any details of the actual ﬂow

pattern above the electrodes—i.e., without solving Eq.

͑17͒—which is very convenient.

F. Characteristic dimensions

Before going into a detailed analysis it is convenient to

discuss the overall properties of the electrokinetic system.

Very roughly speaking the system is equivalent to the simple

RC circuit shown in Fig. 2. The characteristic length scale ᐉ

0

of the device is the narrow electrode gap G

1

, which deter-

mines the magnitude of the electric ﬁeld strength. The char-

acteristic time

0

is the RC time for charging the Debye layer

through the bulk electrolyte

0

=R

0

C

0

, where R

0

is the ͑area

speciﬁc͒ bulk electrolyte resistance. However, at high volt-

age the Stern layer dominates the capacitance of the double

layer and the relaxation time changes to

ϱ

=R

0

C

s

=

0

͑1

+␦

−1

͒. The characteristic time for ͑de͒charging the Debye

layer through the Faradaic reaction is

ct

=R

ct

C

0

. When the

electrode reaction is very facile and R

ct

ӶR

0

this can be sig-

niﬁcantly faster than the Ohmic charging, acting effectively

as a “short-circuit” on the Debye layer. Finally, the charac-

teristic ﬂuid velocity obtained with ϳV

0

/ ͑1+␦͒ and E

t

ϳV

0

/ ᐉ

0

is u

0

=⑀V

0

2

/ ᐉ

0

͑1+␦͒.

IV. LINEAR ANALYSIS

We ﬁrst study the problem in the Debye-Hückel limit

Ӷk

B

T/ ze. Here the electrokinetic problem can be linearized

which simpliﬁes the analysis signiﬁcantly. Indeed the entire

problem is characterized by only two dimensionless groups

measuring the ratios of the two relaxation times

0

and

ct

and the period of the driving voltage. Of course, the solution

also depends on the geometrical parameters W

1

, W

2

, G

1

, G

2

,

and H, of which four can be varied independently once a

characteristic length scale is ﬁxed. We do not vary all these

parameters simultaneously, though.

In the simplest case with negligible conﬁnement ͑HӷL͒

and no Faradaic electrode reaction we recover the earlier

results of Ramos et al. We also determine the optimal param-

eters for the geometry of the electrode array. Then with the

relative size of the electrodes ﬁxed we consider how vertical

conﬁnement of the ﬂuidic channel to HՇL affects the sys-

tem and we investigate the behavior when electrode reac-

tions occur. Finally, we discuss how part of the linear analy-

sis can be extended to cover systems with arbitrary intrinsic

zeta potential

eq

on the electrodes but low driving voltage

V

0

Ӷ͑1+␦͒k

B

T/ ze.

A. Dimensionless form for linear analysis

In preparation of our numerical analysis of the problem

we rescale the variables based on the characteristic dimen-

sions deﬁned in Table I, denoting dimensionless variables by

a tilde:

r = ᐉ

0

r ˜, t =

0

t

˜

, =

0

−1

˜ ,

= V

0

˜

, q = C

0

V

0

q ˜, u = u

0

u ˜ . ͑21͒

In terms of these, the linearized relations ͑4͒, ͑6͒, ͑9͒, and

͑16͒ for the charging of the Debye layer on the electrodes

reduce to

ץ

t

˜q ˜ = n · ١

˜

˜

+ K͑V

˜

ext

−

˜

͒, ͑22͒

V

˜

ext

−

˜

= − q ˜, ͑23͒

where K=R

0

/ R

ct

=

0

/

ct

is a measure of the facility of the

electrode reaction—i.e., a Faradaic conductance. The har-

R

0

C

0

R

ct

FIG. 2. Equivalent circuit diagram for the electrokinetic system:

R

0

is the bulk electrolyte Ohmic resistance, C

0

is the double-layer

capacitance, and R

ct

is the charge-transfer resistance for the elec-

trode reaction. The characteristic time for charging the Debye layer

through the electrolyte is

0

=R

0

C

0

, and the time scale for decharg-

ing the layer through Faradaic electrode reaction is

ct

=R

ct

C

0

.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-4

monic time dependence is most conveniently dealt with by

introducing complex variables

˜

͑r ˜, t

˜

͒ =

1

2

ˆ

͑r ˜͒e

i˜ t

˜

+ c.c. ͑24͒

and similarly for V

˜

ext

and q ˜. Then q ˜ in Eq. ͑22͒ can be elimi-

nated using Eq. ͑23͒ to obtain

n · ١

˜

ˆ

= ͑i˜ + K͒͑

ˆ

− V

ˆ

ext

͒ ͑25͒

for the charge balance in the Debye layer. On the insulating

walls,

n · ١

˜

ˆ

= 0, ͑26͒

whereas in the bulk Eq. ͑1͒ reduces to a Laplace problem

− ١

˜ 2

ˆ

= 0. ͑27͒

The time average electro-osmotic slip velocity can be ex-

pressed as

͗u ˜

s

͘ = ͗͑V

˜

ext

−

˜

͒ץ

x ˜

˜

͘ = −

1

2

ץ

x ˜

͗͑V

˜

ext

−

˜

͒

2

͘ = −

1

4

ץ

x ˜

͉V

ˆ

ext

−

ˆ

͉

2

,

͑28͒

where we used that ץ

x ˜

V

˜

ext

=0 across the electrodes. The spa-

tial average is then ͓15͔

U

˜

=

1

L

˜

ͭ

͵

x ˜

1

x ˜

2

͗u ˜

s

͘dx ˜ +

͵

x ˜

3

x ˜

4

͗u ˜

s

͘dx ˜

ͮ

= −

1

4L

˜

͉͑V

ˆ

ext

−

ˆ

͉

x ˜

2

2

− ͉V

ˆ

ext

−

ˆ

͉

x ˜

1

2

+ ͉V

ˆ

ext

−

ˆ

͉

x ˜

3

2

− ͉V

ˆ

ext

−

ˆ

͉

x ˜

4

2

͒. ͑29͒

B. Validation of the numerical scheme

We solve Eqs. ͑25͒–͑27͒ numerically with the ﬁnite-

element method, using the commercial software package

FEMLAB. We use second-order Lagrange elements to approxi-

mate the solution and a nonuniform ﬁnite-element mesh with

increased resolution at the electrode edges.

In the simple case without any Faradaic current, the De-

bye layer is charged by Ohmic current through the bulk only.

Figure 3 shows the time average slip velocity ͗u ˜

s

͘ across the

electrodes for ˜ =1 and K=0 in a geometry with W

˜

1

=1.5,

W

˜

2

=7, G

˜

2

=5, and H

˜

ӷL

˜

. The slip velocity is strongest close

to the edges of the electrodes and is always directed from the

edges towards the center of the electrodes. To understand this

behavior, notice that close to the electrode edges the electric

ﬁeld and hence the Ohmic current are stronger, which makes

the charging of the Debye layer faster and the screening of

V

˜

ext

more efﬁcient. Then Eq. ͑28͒ clearly shows that ͗u ˜

s

͘ is

always directed from regions where the screening is good

and ͉V

ˆ

ext

−

ˆ

͉

2

is large towards regions where the screening is

less efﬁcient and ͉V

ˆ

ext

−

ˆ

͉

2

is smaller—i.e., from the elec-

trode edges towards the center.

The slip velocity on the electrodes gives rise to bulk ﬂuid

motion with rolls above the electrode edges as sketched in

Fig. 1. However, the net pumping is determined only by the

average slip U

˜

which is shown in Fig. 4 as a function of

frequency ˜ . When ˜ Ӷ1 the screening is almost complete at

all times. Hence, there is no tangential ﬁeld and no electroos-

motic ﬂow. Conversely, when ˜ ӷ1 the driving is too fast for

any signiﬁcant screening to occur; i.e., there is no charge in

the Debye layer and no ﬂow. As a consequence, in Fig. 4 we

see that the pumping velocity U

˜

is maximized around ˜

ϳ1, whereas it falls off as ˜

2

at low frequency and slightly

faster than ˜

−2

at high frequency.

In the low-frequency limit ˜ Ӷ1 we can verify our nu-

merical results analytically by expanding the problem in

powers of ˜ ͓22͔. It then turns out that the leading-order time

average slip ͗u ˜

s

͘ scales as ˜

2

with a corresponding pumping

velocity U

˜

shown with a dashed line in Fig. 4. It is interest-

ing to note that the overall magnitude of u ˜

s

falls off only

linearly with ˜ at low frequency. This results from the uni-

form surface charge density to zeroth order in ˜ interacting

TABLE I. Characteristic dimensions of the electrokinetic system

and typical experimental values for a 10

−4

M KCl working electro-

lyte, assuming ␦=0.1.

Geometric length ᐉ

0

G

1

5 m

Debye length

D

ͱ

⑀k

B

T/ ͚

n

c

n

*

z

n

2

e

2

30 nm

Ohmic relaxation time

0

R

0

C

0

70 s

at large voltage

ϱ

R

0

C

s

770 s

Faradaic relax. time

ct

R

ct

C

0

— s

Debye relaxation time

D

⑀/ 0.5 s

Double-layer cpt. C

0

⑀/

D

͑1+␦͒ 20 mF m

−2

Capacitance ratio ␦ ⑀/

D

C

s

0.1 —

Bulk resistance R

0

ᐉ

0

/ 3.3 m⍀ m

2

Charge-transfer rst. R

ct

k

B

T/ j

0

e — ⍀ m

2

Thermal voltage — k

B

T/ e 25 mV

Fluid velocity u

0

⑀V

0

2

/ ᐉ

0

͑1+␦͒ — ms

−1

0 2 4 6 8 10 12 14

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

˜ x

˜u

s

**FIG. 3. Time average slip velocity ͗u ˜
**

s

͘ across the electrodes

obtained for ˜ =1 and K=0 ͑no Faradaic current͒ in a geometry

with W

˜

1

=1.5, W

˜

2

=7, G

˜

2

=5, and H

˜

ӷL

˜

.

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-5

with the ﬁrst-order electric ﬁeld, which gives rise to oscillat-

ing ﬂow but no time average. A similar expansion in the

high-frequency limit is more difﬁcult. Still, our numerical

results show that the magnitude of u ˜

s

falls off as ˜

−2

at high

frequency for ﬁxed x ˜, but also that it remains ﬁnite in a

narrow region of width ˜

−1

from the electrode edges. The

latter appears to be an artifact due to our simpliﬁed geometry

model with inﬁnitely thin electrodes: For electrodes of ﬁnite

thickness and radius of curvature R

˜

at the edges, the magni-

tude of u ˜

s

falls of as ˜

−2

everywhere on the electrodes for

˜ ӷR

˜−1

; see also ͓23͔.

One particular question that one might ask from Fig. 4 is

why the maximum net pumping velocity U

˜

max

ϳ0.003 is so

low: If the chosen velocity scale u

0

is appropriate, should we

not expect U

˜

max

ϳ1? Starting from Fig. 3 we note that al-

ready the time average slip ͗u ˜

s

͘ is less than O͑1͒ because the

double-layer potential drop ͉V

ˆ

ext

−

ˆ

͉ in Eq. ͑28͒ is only

around half the total applied voltage V

ˆ

ext

at the relaxation

frequency. The spatial average over the full period of the

array includes both positive and negative values of ͗u ˜

s

͘ and

passive sections of insulating channel wall, which reduces

the value of U

˜

further compared to unity. Interestingly, the

average slip across only the narrow electrode,

U

˜

1

=

1

L

˜

͵

x ˜

1

x ˜

2

͗u ˜

s

͘dx ˜, ͑30͒

changes sign for ˜ ϳ1. At low frequency, U

˜

1

and the corre-

sponding average U

˜

2

across the wide electrode are both posi-

tive with U

˜

1

ϳ0.1U

˜

2

ϰ˜

2

. Above the relaxation frequency

we have −U

˜

1

ϳU

˜

2

ϰ

−1

, although their sum U

˜

=U

˜

1

+U

˜

2

falls

off as ͑or slightly faster than͒ ˜

−2

.

In order to further validate our numerical scheme we

compare to the results obtained by Ramos et al. using a

boundary element method based on the Greens function for a

unit charge line source ͓15͔. This is shown in Fig. 5, where

the time average slip ͗u ˜

s

͘ from both electrodes is shown on

the same graph for ˜ =1.24 in a geometry with W

˜

1

=1, W

˜

2

=G

˜

2

=10/ 3, and H

˜

ӷL

˜

. The results from the two different

numerical schemes are hardly distinguishable, except close

to the electrode edges where our ﬁnite-element solution dis-

plays a slightly larger slip velocity ͓24͔.

We also checked our results for convergence with mesh

reﬁnement. The potential

ˆ

͑r͒ is everywhere smooth, but the

tangential ﬁeld is singular at the electrode edges due to the

abrupt change in boundary condition between Eqs. ͑25͒ and

͑26͒ ͓25͔. Of course, our ﬁnite-element solution cannot rep-

resent an inﬁnite slope, but by choosing a strongly nonuni-

form mesh with very ﬁne resolution near the edges we obtain

fairly good quality of the solution even close to the singular-

ity. Moreover, we are mostly concerned with the net pump-

ing U

˜

which is an integral quantity and therefore less sensi-

tive to the details at the electrode edges ͓26͔.

C. Optimal device geometry

An important question in relation to actual device design

is what electrode geometry would maximize the pumping

velocity. It is immediately clear that since the characteristic

velocity u

0

is inversely proportional to the characteristic

length scale ᐉ

0

, the pumping is increased by downscaling the

overall size of the electrodes. We focus again on the limit

HӷL, where the results are independent of H. Still, there

remains four different length scales describing the electrode

geometry—namely, W

1

, W

2

, G

1

, and G

2

—of which three can

be varied independently once a characteristic length scale is

ﬁxed.

Ramos et al. also investigated how the pumping velocity

depends on the electrode geometry ͓15͔. They rescaled the

problem using the width of the wide electrode W

2

as their

characteristic length scale, but found that the pumping veloc-

ity would increase without bounds as G

1

/ W

2

→0. Therefore

they ﬁxed the ratio G

1

/ W

2

at 0.1751, as in the experiment by

Brown et al. ͓6͔, and found that the pumping velocity was

then maximized when W

1

/ W

2

=0.24 and G

2

/ W

2

=0.8.

10

−2

10

−1

10

0

10

1

10

2

10

−6

10

−5

10

−4

10

−3

10

−2

˜ ω

˜

U

FIG. 4. Pumping velocity U

˜

as a function of frequency ˜ .

Dashed line: analytical result at low frequency where U

˜

ϰ˜

2

.

0 0.2 0.4 0.6 0.8 1

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

x/W

i

˜u

s

**FIG. 5. Time average slip velocity ͗u ˜
**

s

͘ for both electrodes

against the relative position on the electrode x/ W

i

. Symbols: results

of Ramos et al. ͓15͔ for the narrow ͑᭝͒ and wide ͑ᮀ͒ electrodes,

respectively. Solid lines: our numerical solution.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-6

However, the choice G

1

/ W

2

=0.1751 is arbitrary. It is

more appropriate to identify the narrow electrode gap G

1

as

the characteristic length scale and look for an optimal device

geometry as a function of W

˜

1

, W

˜

2

, and G

˜

2

in the parameter

ranges 0ϽW

˜

1

ഛW

˜

2

Ͻϱ and 1ഛG

˜

2

Ͻϱ. Notice that when

W

˜

2

and G

˜

2

attain their lower limits, the net pumping U

˜

van-

ishes due to left-right symmetry; i.e., the optimum cannot be

on the parameter range boundaries. Therefore we can use a

simple unconstrained optimization algorithm such as the

MATLAB function FMINSEARCH to determine the global opti-

mum:

U

˜

max

= max

͓˜ ,W

˜

1

,W

˜

2

,G

˜

2

͔

͕U

˜

͑˜ , K = 0, W

˜

1

, W

˜

2

, G

˜

2

͖͒. ͑31͒

In this way we ﬁnd that the pumping velocity is maximized

for W

˜

1

=1.51, W

˜

2

=6.55, and G

˜

2

=4.74, where U

˜

max

=0.003 59 at ˜

max

=0.80.

Given that we have tracked down the global optimum, it

is natural to ask how much better this geometry performs as

compared to that suggested by Ramos et al. and those used

experimentally. This is summarized in Table II: Almost all

experimental geometries come within 10% of the maximal

pumping velocity, and the theoretical result of Ramos et al. is

less than 1% off. Yet the frequency ˜

max

at which pumping is

maximized for the various geometries differs by almost a

factor of 2.

The overall conclusion to be drawn is that the pumping

velocity is fairly insensitive to the particular choice of elec-

trode geometry. Henceforth, we shall therefore continue to

use a geometry with rounded numbers W

˜

1

=1.5, W

˜

2

=7, and

G

˜

2

=5 which is still close to optimal; see also Fig. 6.

D. Conﬁned geometry

In several experimental studies the height of the pumping

channel H has been comparable to the period of the electrode

array L ͓8,10,11͔. It is not immediately clear how this would

affect the properties of the device compared to the simple

case of no conﬁnement HӷL.

Figure 7 shows the pumping velocity U

˜

as a function of

frequency ˜ and degree of conﬁnement L/ H. Our previous

results correspond to the base line L/ HӶ1, whereas the con-

ﬁnement becomes signiﬁcant for L/ Hտ2. In particular, the

pumping velocity increases by roughly a factor of 2 and the

optimal driving frequency ˜

max

becomes proportional to H

˜

for L/ Hӷ1. The latter is easily understood from our simple

circuit model of the system in Fig. 2: While the double-layer

capacitance C

0

is largely independent of conﬁnement, the

resistance of the bulk electrolyte changes from R

0

to R

0

/ H

˜

because the cross section of the conducting channel is de-

creased. The relaxation frequency therefore becomes

H

˜

/ R

0

C

0

—i.e., proportional to H

˜

.

Our ﬁndings here are important in relation to actual de-

vice design because they indicate that working with a

strongly conﬁned geometry does not, as one might have wor-

ried, mask the asymmetry and destroy the ability to pump.

Thus the channel height H and the period of the electrode

array L can be chosen independently. It is important to real-

ize that while the pumping velocity U does not depend much

on conﬁnement, the maximal ﬂow rate Q

max

=HU/ 2 that the

pump can deliver ͑per unit width of the channel͒ and the

maximal backpressure ⌬p

max

=6NLU/ H

2

it can sustain

͑where N is the total number of electrode pairs͒ certainly do.

TABLE II. List of geometries used experimentally and our linear model prediction for their dimensionless

maximum pumping velocity and optimal driving frequency. Of course, the actual pumping velocity depends

on the absolute size of the device through u

0

.

W

˜

1

W

˜

2

G

˜

2

U

˜

max

˜

max

Refs.

Brown et al. 0.93 5.7 3.5 0.00331 1.12 ͓6,7,11͔

Studer et al. 0.67 4.33 2.67 0.00293 1.47 ͓8͔

Mpholo et al. 1.0 5.0 3.0 0.00332 1.12 ͓9͔

Debesset et al. 0.8 5.2 3.2 0.00317 1.26 ͓10͔

Ramos et al. 1.37 5.71 4.57 0.00357 0.88 ͓15͔ ͑theory͒

Global optimum 1.51 6.55 4.74 0.00359 0.80 Present work

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

0

2

4

6

8

10

12

14

16

18

20

0

.

0

0

3

5

0

.0

0

3

0

.0

0

2

5

0

.0

0

2

˜

W

1

˜

W

2

S

ym

m

etry

FIG. 6. Contour plot of the maximal pumping velocity over

frequency max

˜

͕U

˜

͑˜ , . . . ͖͒ as a function of W

˜

1

and W

˜

2

for G

˜

2

=5.

When W

˜

1

=W

˜

2

the net pumping vanishes due to symmetry. There is

a unique optimum close to W

˜

1

=1.5 and W

˜

2

=7 with U

˜

max

=0.00358 and ˜

max

=0.8.

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-7

E. Faradaic current injection

When we include electrode reactions and Faradaic current

in our model, one of the most important new features is that

the complete screening of the electrodes at low frequency is

unbalanced. When the reaction is slow and the Faradaic con-

ductance K=R

0

/ R

ct

Ӷ˜ , the effect is small and only perturb

the result from purely Ohmic charging slightly. Conversely,

when Kӷ˜ the reaction is fast enough to establish a steady

state where the Ohmic and Faradaic currents are balancing.

For either Kӷ1 and/or ˜ ӷ1 the Debye layer is effectively

“short-circuited;” cf. Fig. 2. On the basis of these observa-

tions we would expect to see pumping even in the low-

frequency limit when the Faradaic reaction speed is moderate

and Kϳ1.

Figure 8 shows the pumping velocity U

˜

as a function of ˜

and K. For KӶ˜ we recover our previous results for pure

Ohmic charging. However, for Kӷ˜ we observe no pump-

ing at all, not even for Kϳ1. Figure 9 shows the result of a

more detailed investigation in the zero-frequency limit ˜

ϳ0, where we consider a device with asymmetric surface

properties on the electrodes, such that the Faradaic conduc-

tance takes the values K

1

and K

2

on the narrow and wide

electrodes, respectively. The ﬁgure shows that whenever K

1

K

2

the device does actually pump: When K

1

ϾK

2

it pumps

in the forward direction—i.e., in the same direction as ob-

served for purely Ohmic charging—but for K

2

ϾK

1

the

pumping direction is reversed.

This reversal of the pumping direction can be understood

qualitatively by inspecting the form of the time average slip

͗u ˜

s

͘ for the four different points marked ͑a͒ to ͑d͒ in Fig. 9,

the result of which is shown in Fig. 10. When K

1

=K

2

=1 ͑c͒

the slip velocity is clearly nonzero on both electrodes, but the

average across the narrow electrode exactly cancels that

across the wide electrode. When K

1

ӷK

2

=1 ͑d͒ the Debye

layer on the narrow electrode is essentially short-circuited by

the fast electrode reaction, leaving no charge and no elec-

troosmotic ﬂow there except from in a narrow region of

width ϳK

1

−1

from the electrode edges. Therefore the pump-

ing velocity U

˜

is dominated by the contribution from the

wide electrode. Similarly, when K

2

ӷK

1

=1 ͑a͒ the Debye

layer on the wide electrode is essentially short-circuited and

the pumping velocity is dominated by the contribution from

the narrow electrode—which tends to be negative.

In conclusion, the absence of pumping in the low-

frequency limit when K

1

=K

2

is a special case due to high

symmetry; whenever K

1

K

2

there is some pumping and this

falls off linearly with K

i

for K

i

Ӷ1. More generally we can

prove that for a device with spatially varying capacitance

ratio ␦͑x ˜͒ and Faradaic conductance K͑x ˜͒, there can be no

pumping in the low-frequency limit when the product ͓1

+␦͑x ˜͔͒K͑x ˜͒ is a constant on all electrodes. The details of the

proof are outlined in Appendix A.

0

.

0

0

1

0

.

0

0

2

0

.

0

0

3

0

.

0

0

4

0

.0

0

5

0

.0

0

6

0

.0

0

7

0

.

0

0

8

10

−2

10

−1

10

0

10

1

10

0

10

1

10

2

˜ ω

L

/

H

FIG. 7. Contour plot of pumping velocity U

˜

as a function of

frequency ˜ and degree of conﬁnement L/ H.

0

.

0

0

1

0

.

0

0

2

0

.

0

0

3

10

−2

10

−1

10

0

10

1

10

2

10

−3

10

−2

10

−1

10

0

10

1

10

2

˜ ω

K

FIG. 8. Contour plot of the pumping velocity U

˜

as a function of

frequency ˜ and Faradaic conductance K.

−

0

.

0

0

4

−

0

.

0

0

2

0

.

0

0.002

0.004

0.006

0.008

0.010

0.012

10

−3

10

−2

10

−1

10

0

10

1

10

2

10

3

10

−3

10

−2

10

−1

10

0

10

1

10

2

10

3

K

1

K

2

a b

c d

FIG. 9. Contour plot of U

˜

in the zero-frequency limit ˜ ϳ0 with

asymmetric current injection: K

1

and K

2

denote the Faradaic con-

ductance on the narrow and wide electrodes, respectively. The slip

velocity ͗u ˜

s

͘ at the points ͑a͒–͑d͒ is shown in Fig. 10.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-8

F. Linear response at nonzero intrinsic zeta potential

Many electrode-electrolyte systems spontaneously form a

Debye layer and possess an intrinsic zeta potential typically

of the order

eq

ϳ100 mV which is well beyond the Debye-

Hückel limit Ӷk

B

T/ zeϳ25 mV. Hence, our foregoing lin-

ear analysis cannot be applied immediately to such systems.

But then it is standard procedure in electrochemical imped-

ance spectroscopy to study the linear response of the electro-

kinetic system to a small applied voltage ͓18͔.

Consider a system where the electrolyte forms a Debye

layer on the electrodes with an intrinsic zeta potential

eq

. In

equilibrium there is a uniform surface charge density q

eq

given by Eq. ͑3͒ on the electrodes and a constant potential

eq

=−

eq

+q

eq

/ C

s

in the bulk. If the electrode array is biased

with a small-amplitude ac voltage V

ext

= ±V

0

cos͑t͒, we can

linearize around equilibrium and writing =

1

2

͑

eq

+

ˆ

Јe

it

͒

+c.c. and similarly for q and we recover the form of Eqs.

͑26͒ and ͑27͒ for the linear response

ˆ

Ј. Only now the char-

acteristic time used for the rescaling in Eq. ͑21͒ becomes

0

= R

0

C

d

, ͑32͒

where C

d

͑

eq

͒ is the differential capacitance of the double

layer:

C

d

= −

dq

d͑V

ext

− ͒

. ͑33͒

For a symmetric z: z electrolyte this can be written as

C

d

=

⑀

D

͓␦ + sech͑ze

eq

/2k

B

T͔͒

. ͑34͒

The relaxation time for the Faradaic electrode reaction be-

comes

ct

= R

ct

C

d

, ͑35͒

where R

ct

͑

eq

͒ is the differential charge-transfer resistance

given by

1

R

ct

=

dj

ext

d͑V

ext

− ͒

=

j

0

e

k

B

T

e

−͑z

O

−␣͒e

eq

/k

B

T

, ͑36͒

and the Faradaic conductance is redeﬁned as K=R

0

/ R

ct

͑

eq

͒

accordingly. The interaction of the linear response tangential

ﬁeld with the equilibrium Debye layer charge gives rise to a

linear response slip velocity

u

s

Ј =

⑀

eq

ץ

x

Ј. ͑37͒

However, because Ј is harmonic in time, this does not give

rise to any net pumping. The leading-order time average slip

is to second order in V

0

, where the interaction of the induced

charge to ﬁrst order with the tangential ﬁeld gives rise to a

slip velocity of

͗u

s

Љ͘ =

⑀

͗Јץ

x

Ј͘ = −

⑀

4

d

d͑V

ext

− ͒

ץ

x

͉V

ˆ

ext

−

ˆ

Ј͉

2

. ͑38͒

This is fully equivalent to Eq. ͑28͒, and the net pumping

produced can be obtained directly from the analysis pre-

sented earlier in the text. Only now the characteristic ﬂuid

velocity becomes

u

0

=

⑀V

0

2

ᐉ

0

d

d͑V

ext

− ͒

, ͑39͒

where, for a symmetric z: z electrolyte,

d

d͑V

ext

− ͒

=

1

1 + ␦ cosh͑ze

eq

/2k

B

T͒

. ͑40͒

Notice that u

0

͑

eq

͒ falls off exponentially with

eq

because Ј

becomes negligible.

The linear response dominates over the higher-order com-

ponents only at low applied voltage. Speciﬁcally, we must

require that both V

0

ӶC

d

/ ͓dC

d

/ d͑V

ext

−͔͒ and V

0

ӶR

ct

/ ͓dR

ct

/ d͑V

ext

−͔͒, and a sufﬁcient condition for this is

V

0

Ӷ͑1+␦͒k

B

T/ ze. However, because the time average slip

velocity vanishes to ﬁrst order in V

0

, one cannot neglect the

second-order tangential ﬁeld −ץ

x

Љ, which gives rise to an

additional component ͗u ¯

s

Љ͘ to the slip velocity given by

͗u ¯

s

Љ͘ =

⑀

eq

ץ

x

͗Љ͘. ͑41͒

In the absence of Faradaic currents, the time average poten-

tial ͗͘ vanishes to second and all higher orders within our

simple electrokinetic model, as is easily seen by taking the

time average of Eq. ͑9͒. Otherwise, ͗Љ͘ is determined as a

solution to the Laplace equation with the boundary condition

0 = n · ١͗Љ͘ −

1

R

ct

͗Љ͘ +

1

4

d

2

j

ext

d͑V

ext

− ͒

2

͉V

ˆ

ext

−

ˆ

Ј͉

2

͑42͒

at the electrodes, where

2 4 6 8 10 12

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

2 4 6 8 10 12

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

˜ x ˜ x

˜u

s

˜u

s

(c) (d)

(a) (b)

˜

U =0

˜

U >0

˜

U <0

˜

U =0

FIG. 10. Time average slip velocity ͗u ˜

s

͘ on the electrodes for the

parameter values marked in Fig. 9: ͑a͒ K

1

=1, K

2

=10

2

; ͑b͒ K

1

=K

2

=10

2

; ͑c͒ K

1

=K

2

=1; ͑d͒ K

1

=10

2

, K

2

=1.

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-9

d

2

j

ext

d͑V

ext

− ͒

2

= −

e

k

B

T

2

R

ct

ͫ

͑z

O

− ␣͒

d

d͑V

ext

− ͒

+ ␣ −

1

2

ͬ

.

͑43͒

This demonstrates that when Faradaic electrode reactions oc-

cur, the actual computation of the time average ﬂuid ﬂow and

net pumping for ﬁnite intrinsic zeta potential ͑on the elec-

trodes and/or the insulating channel walls͒ requires a sepa-

rate and more elaborate study, which we have to leave open

despite its experimental relevance.

V. NONLINEAR ANALYSIS

While theoretically it is convenient to work in the low-

voltage limit where the system can be characterized by rela-

tively few dimensionless parameters, most experimental

work has been done with driving voltages of a few volts in

order to obtain an appreciable ﬂuid motion. However, even

with the simple nonlinear model we consider here, the pa-

rameter space is large, and we therefore focus on the case of

a monovalent symmetric ͑1:1͒ electrolyte with no intrinsic

zeta potential formed on the electrodes—i.e.,

eq

=0. More-

over, for the Faradaic electrode reaction we take z

O

=z

R

+1

=1 and assume a symmetric transfer coefﬁcient ␣=1/ 2.

We ﬁrst investigate the system behavior without the pres-

ence of the electrode reaction so that the only nonlinearity in

our model arises from the nonlinear capacitance of the De-

bye layer. Then we turn on the electrode reaction and study

the full nonlinear model.

A. Dimensionless form for nonlinear analysis

Beyond the Debye-Hückel limit, the smallest characteris-

tic voltage of the electrokinetic system is the thermal voltage

k

B

T/ eϳ25 mV and not the driving voltage V

0

. We therefore

introduce a new rescaling for the electrical part of the prob-

lem:

=

k

B

T

e

˜

, V

0

=

k

B

T

e

V

˜

0

, q = C

0

k

B

T

e

q ˜. ͑44͒

Apart from this, the scaling remains as in Eq. ͑21͒; in par-

ticular, we retain u

0

=⑀V

0

2

/ ᐉ

0

͑1+␦͒ for the characteristic

ﬂuid velocity. Then Eqs. ͑5͒, ͑6͒, ͑9͒, and ͑15͒, governing the

charging of the Debye layer on the electrodes, reduce to

V

˜

ext

−

˜

=

˜

− q ˜␦/͑1 + ␦͒, ͑45͒

ץ

t

˜q ˜ = n · ١

˜

˜

+ j

˜

ext

, ͑46͒

where j

˜

ext

is the dimensionless Faradaic current given by

j

˜

ext

= 2Ke

−

˜

/2

sinh͓͑V

˜

ext

−

˜

͒/2͔, ͑47͒

with K=R

0

/ R

ct

as before, and where the zeta potential is

directly related to the Debye layer charge by

˜

͑q ˜͒ = − 2 sinh

−1

͓q ˜/2͑1 + ␦͔͒. ͑48͒

On the insulating walls we have again

n · ١

˜

˜

= 0 ͑49͒

and, in the bulk,

− ١

˜ 2

˜

= 0. ͑50͒

The problem, Eqs. ͑45͒–͑50͒, is solved numerically with the

ﬁnite-element method using FEMLAB. Due to the nonlinear-

ity, it is not possible to solve the problem with a single com-

plex variable for the base frequency component. Instead we

employ a relaxation method, where we represent the periodic

solution by a set of equispaced points over one period in

time, and calculate the time derivative in Eq. ͑46͒ using the

leapfrog ﬁnite-difference scheme; see Appendix B for de-

tails.

B. Nonlinear Debye layer capacitance

In the absence of Faradaic electrode reactions the only

nonlinearity in the model arises from the nonlinear surface

capacitance in the Debye layer. Figure 11 shows the pumping

velocity U

˜

as a function of the driving frequency ˜ and

voltage V

˜

0

for ␦=0.1 and K=0. At the base line for V

˜

0

Ӷ1

we recover the results from the linear analysis with U

˜

peak-

ing for ˜ ϳ1. However, at higher voltage when the double-

layer capacitance is dominated by the Stern layer, the fre-

quency at which U

˜

is maximized drops to ˜ ϳ˜

ϱ

, where

˜

ϱ

=

0

/

ϱ

=␦/ ͑1+␦͒ is the relaxation frequency at high volt-

age; cf. Table I.

Figure 11 also shows that the ͑dimensionless͒ pumping

velocity U

˜

falls off at high voltage. This is simply due to the

fact that the ͑physical͒ electro-osmotic slip velocity u

s

ϰץ

x

no longer scales as V

0

2

. When the driving frequency is

not too large, Eqs. ͑45͒ and ͑48͒ allow us to estimate the zeta

potential from

V

˜

0

ϳ

˜

+ 2 sinh͑

˜

/2͒␦. ͑51͒

At low voltage this reduces to

˜

ϳV

0

/ ͑1+␦͒ whereas at high

voltage

˜

ϳ2 log͑V

˜

0

/ ␦͒. The transition between the two re-

gimes occurs for V

˜

0

ϳ␦ when ␦ӷ1 and for V

˜

0

ϳ−2 log ␦

when ␦Ӷ1. Hence, for V

˜

0

ӷ␦−2 log ␦ we expect u

s

to scale

as V

0

log V

0

rather than V

0

2

. This is conﬁrmed by Fig. 12,

which shows U

˜

max

as a function of V

˜

0

; the data are scaled

with V

˜

0

2

to recover the voltage dependence of the physical

pumping velocity:

U =

⑀

ᐉ

0

͑1 + ␦͒

ͩ

k

B

T

e

ͪ

2

U

˜

V

˜

0

2

. ͑52͒

We must emphasize at this point that the upper voltage limit

V

˜

0

=10

3

in Figs. 11 and 12, corresponding to V

0

=25 V, is far

outside the range of validity of our simple electrokinetic

model: In Sec. VI we discuss a number of strongly nonlinear

phenomena that become signiﬁcant for V

0

տ125 mV in a

typical experiment. Beyond this voltage our results can

therefore only provide some characteristic features of the

system and should not be regarded as an accurate descrip-

tion.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-10

C. Nonlinear Faradaic current injection

While the nonlinearity induced by the Debye layer capaci-

tance manifests itself relatively slowly as the voltage is in-

creased, we expect a more dramatic effect from the Faradaic

current due to the exponential voltage dependence in Eq.

͑47͒. Figure 13 shows the pumping velocity U

˜

as a function

of driving frequency ˜ and voltage V

˜

0

for ␦=0.1 and K

=0.1. At the base line for V

˜

0

Ӷ1 we recover the results from

the linear analysis in Fig. 8 for K=0.1—i.e., no pumping in

the low-frequency limit and maximal pumping for ˜ ϳ1,

with a peak value of U

˜

less than 0.003. However, for V

˜

0

տ1 we do observe pumping at low frequency, whereas U

˜

falls off rapidly for all ˜ when V

˜

0

ӷ10. The ﬁgure also

shows that at low voltage the optimal driving frequency ˜

max

follows the same pattern as in Fig. 11, but for V

˜

0

տ10 it

bends sharply up towards higher frequency.

Qualitatively it is not so difﬁcult to see how the pumping

at low frequency arises: At low voltage the ͑magnitude of

the͒ charge density in the Debye layer on the narrow elec-

trode is a factor of W

2

/ W

1

larger than on the wide electrode

to ensure overall charge conservation. Correspondingly, also

the double-layer potential drop V

˜

ext

−

˜

is larger on the nar-

row electrode, such that the exponential increase in the Fara-

daic current ͓cf. Eq. ͑47͔͒ starts off earlier there. Effectively,

this makes the Faradaic conductance larger on the narrow

electrode than on the wide one, and according to Fig. 9 from

our linear analysis this should indeed give rise to pumping in

the forward direction at low frequency.

Figure 14 shows in more detail the voltage dependence of

the pumping velocity at low frequency ˜ Ӷ1. For V

˜

0

Շ1 the

net pumping grows rapidly with V

˜

0

. Apart from the intrinsic

V

˜

0

2

scaling, the time average asymmetry in effective Faradaic

conductance between the two electrodes grows roughly as V

˜

0

2

as well, yielding an overall V

˜

0

4

voltage dependence. At larger

voltage we ﬁnd that the pumping levels off to become almost

independent of V

˜

0

. This behavior can be partially understood

by considering the scaling of

˜

and the tangential ﬁeld in this

limit: When the frequency is low enough that a steady state

is established with the Ohmic and Faradaic currents balanc-

ing, we can write the applied voltage V

˜

0

as the sum of the

potential drop across the double layer V

˜

ext

−

˜

and the poten-

0

.

0

0

1

0

.

0

0

2

0

.

0

0

3

10

−2

10

−1

10

0

10

1

10

−1

10

0

10

1

10

2

10

3

˜ ω

˜

V

0

FIG. 11. Contour plot of the pumping velocity U

˜

as a function

of frequency ˜ and driving voltage V

˜

0

with capacitance ratio ␦

=0.1 and no Faradaic current, K=0. The white dashed line shows

the frequency ˜

max

at which U

˜

is maximized as a function of V

˜

0

. At

large driving voltage ˜

max

shifts from ϳ1 to ϳ˜

ϱ

, reﬂecting the

larger relaxation time when the double-layer capacitance is domi-

nated by the Stern layer.

10

−1

10

0

10

1

10

2

10

3

10

−4

10

−2

10

0

10

2

˜

V

0

˜

U

m

a

x

˜

V

2

0

˜

V

2

0

˜

V0

˜

V0

log

˜

V0

FIG. 12. Maximal pumping velocity U

˜

max

as a function of V

˜

0

, as

evaluated along the white dashed line in Fig. 11. The data are scaled

with V

˜

0

2

to recover the physical dependence on driving voltage. For

comparison, the dashed lines show linear and quadratic scaling with

V

˜

0

, while the dash-dotted line is a ﬁt to V

˜

0

log V

˜

0

.

0

.

0

0

1

0

.

0

0

2

0

.0

0

3

10

−2

10

−1

10

0

10

1

10

−1

10

0

10

1

10

2

10

3

˜ ω

˜

V

0

FIG. 13. Contour plot of the pumping velocity U

˜

as a function

of driving frequency ˜ and voltage V

˜

0

with capacitance ratio ␦

=0.1 and Faradaic conductance K=0.1. The dashed line shows the

optimum driving frequency ˜

max

as a function of V

˜

0

. Notice the

pumping in the low-frequency limit for V

˜

0

տ1.

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-11

tial drop across the bulk electrolyte when an Ohmic current

equal to j

˜

ext

is passed through it—i.e.,

V

˜

0

ϳ ͑V

˜

ext

−

˜

͒ + j

˜

ext

ϳ ͑V

˜

ext

−

˜

͒ + 2Ke

−

˜

/2

ϫsinh͓͑V

˜

ext

−

˜

͒/2͔. ͑53͒

If we neglect for simplicity the

˜

-dependent exponential

prefactor, we ﬁnd immediately that V

˜

ext

−

˜

ϳV

˜

0

/ ͑1+K͒ at

low voltage V

˜

0

ӶK−2 log K, changing to V

˜

ext

−

˜

ϳ2 log͑V

˜

0

/ K͒ at higher voltage. In effect, both the tangential

ﬁeld ץ

x ˜

˜

=−ץ

x ˜

͑V

˜

ext

−

˜

͒ and the zeta potential

˜

Ϸ͑V

˜

ext

−͒ / ͑1+␦͒ go as log V

˜

0

at large voltage, and we would ex-

pect the ͑physical͒ electro-osmotic slip velocity u

s

ϰץ

x

to

grow as slowly as ͑log V

0

͒

2

with driving voltage.

Finally, we note that since the electrode reaction is very

fast at large voltage, the “low-frequency” limit where the

Faradaic and Ohmic currents balance extends higher and

higher in frequency as the driving voltage is increased.

Therefore, the characteristic behavior from Fig. 14 with the

pumping velocity leveling off dominates even for ˜ ϳ1

when V

˜

0

տ10

2

. And when the optimal driving frequency in

Fig. 13 shifts towards higher frequency for V

˜

0

տ10, it is

essentially a trace of the point at which the Faradaic current

starts to dominate.

VI. DISCUSSION

Our numerical study of the simple electrokinetic model

has brought about a number of interesting theoretical predic-

tions that we brieﬂy summarize below and compare to ex-

perimental observations reported in the literature. In particu-

lar, we comment on the extent to which our results agree or

disagree with the experiments and discuss a number of ͑non-

linear͒ effects not included in our model that could account

for some of the more exotic experimental observations.

A. Conﬁnement

Our analysis has shown that vertical conﬁnement of the

pumping channel does not mask the asymmetry of the device

and destroy the ability to pump, as one might have worried,

but instead enhances the pumping velocity by roughly a fac-

tor of 2 and causes the optimal driving frequency to shift

down inversely proportional to L/ H. In all experimental

studies to date, the conﬁnement has been relatively weak

with L/ HՇ2, whereas our results indicate that conﬁnement

becomes signiﬁcant only for L/ HϾ2. It is important to keep

in mind, though, that while the pumping velocity U does not

depend much on conﬁnement, the maximal ﬂow rate Q

max

=HU/ 2 that the pump can deliver and the maximal back-

pressure ⌬p

max

ϰU/ H

2

it can sustain certainly do. The opti-

mal choice of channel height therefore depends strongly on

whether the pump is intended to work as a pressure or cur-

rent generator ͓27͔.

B. Nonlinear double-layer capacitance

In the absence of Faradaic electrode reactions, where the

sole nonlinearity in our model is due to the nonlinear capaci-

tance of the Debye layer, we ﬁnd that as the driving voltage

V

0

is increased well beyond the thermal voltage k

B

T/ e, the

relaxation frequency for the electrokinetic system, and hence

also the optimal frequency for pumping, drops from

0

=1/ R

0

C

0

to

ϱ

=1/ R

0

C

s

=

0

␦/ ͑1+␦͒, where ␦=⑀/

D

C

s

is

the surface capacitance ratio. At the same time, the scaling of

the pumping velocity with driving voltage changes from V

0

2

to V

0

log V

0

because the zeta potential entering the

Helmholtz-Smoluchowski slip velocity becomes only a small

fraction of the overall potential drop across the double layer.

Experimentally, Brown et al. reported that the optimal

driving frequency f

max

dropped from 3 kHz at

0.2 V

rms

to 1 kHz at 1.2 V

rms

͓6͔. For their particular elec-

trolyte and electrode geometry the relaxation frequency at

low voltage was expected to be f

0

=

0

/ 2Ӎ͑1+␦͒

ϫ2 kHz. The observed downshift in optimal frequency

therefore ﬁts well with our model predictions if we assume a

capacitance ratio of ␦Ӎ0.5. Likewise, Studer et al. found the

optimal driving frequency around 1 kHz for driving voltages

larger than 1 V

rms

with similar electrolyte and electrode ge-

ometry ͓11͔. The electrode material in those experimental

studies was uncoated gold and platinum, respectively, con-

sistent with a large Stern layer capacitance and hence small

␦.

Green et al. mapped out the entire ﬂuid velocity ﬁeld over

a single pair of gold electrodes with titanium coating ͓4͔. The

titanium spontaneously forms a thin oxide layer which was

intended to inhibit Faradaic electrode reactions in the experi-

ment. The observed velocity magnitude and frequency de-

pendence at 500 mV driving voltage was matched with lin-

ear theory predictions by assuming a capacitance ratio of ␦

=3. It is remarkable that this value was obtained for both a

0.15 mM and a 0.6 mM KCl solution: Naively, assuming a

constant value for the Stern layer capacitance C

s

one would

expect ␦=⑀/

D

C

s

to depend on the electrolyte concentration

through

D

͓28͔.

Cahill et al. observed ac electroosmotic ﬂuid motion due

to traveling-wave electric ﬁelds induced on a four-phase in-

10

−1

10

0

10

1

10

2

10

3

10

−8

10

−6

10

−4

10

−2

10

0

˜

V

0

˜

U

˜

V

2

0

˜

V

4

0

FIG. 14. Pumping velocity U

˜

as a function of V

˜

0

in the low-

frequency limit ˜ Ӷ1. The data are scaled with V

˜

0

2

to recover the

physical dependence on driving voltage.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-12

terdigitated electrode array and found an accurate V

0

2

scaling

for 100−500 mV driving voltage, although the pumping ve-

locity peak position and magnitude were smaller than the

linear theory predictions ͓12͔. In their experiments the elec-

trodes were coated with a 50-nm-thick Teﬂon-like insulating

layer, yielding a capacitance ratio of ␦=75–150 depending

on the electrolyte concentration. Hence, they remained in the

linear regime over the entire voltage range applied. Ramos et

al. performed a similar study using titanium-coated gold

electrodes ͓13͔. Their optimal driving frequency around

2 kHz was matched with the linear theory by assuming ␦

=1.5, while the maximum pumping velocity at low fre-

quency was an order of magnitude below the prediction from

linear theory.

Generally, the experimentally observed pumping veloci-

ties tend to be smaller than the linear theory prediction and

sometimes resemble more an afﬁne dependence than a qua-

dratic scaling with V

0

͓10,11͔. This is at least in qualitative

agreement with our nonlinear model results because it is dif-

ﬁcult to distinguish V

0

from V

0

log V

0

over the limited volt-

age range used experimentally.

C. Faradaic current injection

Our linear analysis at low driving voltage shows that

when Faradaic electrode reactions occur, the pumping is

strongly suppressed for R

ct

ӶR

0

due to “short-circuiting” of

the double layer, where R

ct

is the charge-transfer resistance

characterizing the electrode reaction and R

0

is the bulk elec-

trolyte Ohmic resistance. When R

ct

ϳR

0

the Faradaic current

injection merely imbalances the otherwise perfect screening

at low frequency and therefore induces electro-osmotic ﬂuid

motion. Moreover, we predict that the direction of the pump-

ing in the low-frequency limit depends on the relative mag-

nitude of R

ct

on the narrow and wide electrodes, respectively:

If R

ct

is smaller on the narrow electrodes than on the wide

ones, the pumping at low frequency is in the forward direc-

tion, whereas it is in the reverse direction otherwise. In the

symmetric special case when R

ct

is equal on both electrodes

there is no net pumping at low frequency.

When the driving voltage is increased, the nonlinearity in

the Faradaic current injection induces an effective asymme-

try in the charge-transfer resistance and we obtain pumping

in the forward direction at low frequency, even if the intrin-

sic surface properties are symmetric. Moreover, at even

larger voltage our nonlinear analysis shows that the pumping

velocity levels off to a constant value. Using the typical pa-

rameter values from Table I and R

0

/ R

ct

=0.1 as in Fig. 14, the

magnitude of the pumping velocity at this plateau is roughly

Uϳ100 m/ s.

The observation of electrode degradation and bubble for-

mation from electrolysis is experimental evidence that Fara-

daic reactions do occur above 1.5–4.0 V

rms

, depending on

the driving frequency and electrode material ͓8–11,14͔.

Moreover, Studer et al. observed that the pumping direction

was reversed when the driving frequency was increased

above 20 kHz and the voltage above 3 V

rms

͓11͔. Ramos et

al. also found reversal of the pumping direction on an asym-

metric electrode array, as well as a similar reversal on a

traveling-wave array for driving voltages above 2 V and fre-

quencies around 1 kHz ͓13,14͔. And Lastochkin et al. re-

ported that the direction of the time average slip velocity on

an array of T-shaped electrodes was reversed as compared to

the direction usually expected for ac electroosmosis ͓16͔.

We note that the reversal of the pumping direction that we

predict at low frequency is essentially due to an imbalance

between net pumping contributions of opposite signs from

the two electrodes. We never ﬁnd reversal of the time aver-

age slip direction within our simple electrokinetic model. It

is possible that the pumping reversal observed by Studer et

al. could be due to imbalance between the net pumping con-

tributions from the two electrodes: Although our nonlinear

analysis predicts pumping in the forward direction, there

could be other nonlinear effects leading to imbalance into the

reverse direction. However, Ramos et al. observed pumping

reversal on a traveling-wave array where all the electrodes

are identical—this rules out any imbalance between the con-

tributions from individual electrodes in their setup. And the

reversal of the time average slip direction observed by Las-

tochkin et al. is certainly a different phenomenon.

Lastochkin et al. argued that strong Faradaic current in-

jection in their system would reverse the polarity of the

charge in the Debye layer and hence the direction of the slip.

Based on the simple circuit model for the electrokinetic sys-

tem from Fig. 2, it is difﬁcult to see how the potential drop

across ͑and hence the charge in͒ the double layer could

change sign, regardless of how small R

ct

becomes. However,

since the driving frequency in their experiments was in the

MHz range, which is of the order of the Debye frequency for

the electrolyte, the Debye layer cannot be assumed to be in

local equilibrium and Eqs. ͑4͒ and ͑16͒ are no longer valid.

Finally, we recall that for simplicity we have been focus-

ing here on the case with no intrinsic zeta potential

eq

on the

electrodes and/or the insulating channel walls. Our analysis

in Sec. IV F indicates that when Faradaic electrode reactions

are present, the behavior for nonzero

eq

needs to be studied

more carefully, especially in the nonlinear regime.

D. Effect of mass transfer on reaction kinetics

One important effect that we have neglected in our model

is the inﬂuence of mass transfer on the electrode reaction

kinetics. When the oxidized and reduced species O and R are

consumed or produced at the electrode by the Faradaic reac-

tion, their local concentrations c

O

and c

R

become different

from the bulk concentrations c

O

*

and c

R

*

, respectively. Now,

the lower c

O

drops at the electrode, the more difﬁcult it be-

comes to run the reaction forward, and vice versa for c

R

.

This can be modeled as an additional impedance Z͑͒, called

the Warburg impedance, in series with the charge-transfer

resistance R

ct

in Fig. 2 ͓18͔. The transport of O and R from

the bulk to the surface is by diffusion, and the characteristic

distance that the nth species can diffuse over one cycle of the

driving voltage is ᐉ

n

͑͒ =

ͱ

D

n

/ , where D

n

is the diffusion

constant. If this distance is small compared to the character-

istic dimension of the electrodes ᐉ

0

, the diffusion process is

essentially one dimensional and the ͑area speciﬁc͒ Warburg

impedance takes the form

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-13

Z͑͒ =

k

B

T

ͱ

2e

2

ͫ

1

c

O

*

ͱ

D

O

+

1

c

R

*

ͱ

D

R

ͬ

1 − i

ͱ

. ͑54͒

Notice that Z͑͒ increases at low frequency because the re-

actants need to diffuse longer. However, at very low fre-

quency when ᐉ͑͒ ӷᐉ

0

—i.e., ӶD/ ᐉ

0

2

—the diffusion layer

extends much longer than the separation between the elec-

trodes and the Warburg impedance levels off to a constant

Z

0

ϳ

k

B

T

e

2

ͫ

ᐉ

0

c

O

*

D

O

+

ᐉ

0

c

R

*

D

R

ͬ

. ͑55͒

We note that Z

0

տR

0

=ᐉ

0

/ , where the electrolyte conductiv-

ity is given by =͚

n

D

i

c

n

*

z

n

2

e

2

/ k

B

T. Therefore, even if R

ct

ӶR

0

, it is not possible to short-circuit the double layer for

ӶD/ ᐉ

0

2

, because it is the larger of R

ct

and Z

0

that domi-

nates. More generally, the effect of mass transfer is to make

the short-circuiting of the double layer discussed in Secs.

IV E and V C less dramatic. Still, we do expect the predic-

tion of the pumping direction based on Fig. 9 to hold for ˜

Ӷ

0

and R

ct

տR

0

.

As the driving voltage is increased, the Faradaic current

can become so strong that the concentrations of the reactants

at the electrodes differ signiﬁcantly from their equilibrium

values. Ultimately, as one of them approaches zero, the re-

action stagnates. The current at which this occurs is termed

the limiting current ͓18͔, and it is given by the maximal rate

at which the reactants can be transported from the bulk to the

electrodes by diffusion. Hence, for the forward reaction the

limiting current is j

O

lim

ϳeD

O

c

O

*

/ ᐉ

O

͑͒. Assuming the fre-

quency is low enough that the Ohmic and Faradaic currents

are balancing—i.e., j

ext

ϳV

0

/ ᐉ

0

—we ﬁnd that the limiting

current is reached around V

0

ϳeD

O

c

O

*

ᐉ

0

/ ᐉ

O

͑͒. For the

typical parameter values from Table I and ϳ

0

, D

O

ϳ10

−9

m

2

/ s, and c

O

*

ϳ0.1 mM, this corresponds to V

0

ϳ100 mV. Beyond this voltage, the concentration c

O

tends

to zero at the electrode and the Faradaic current levels off at

the limiting value. If there is no supporting electrolyte, this

makes the conductivity in the diffusion layer differ signiﬁ-

cantly from the bulk value and the Debye layer is driven out

of local equilibrium and expands in width to ϳᐉ

0

1/3

D

2/3

͓30,31͔. However, if there is an excess of supporting electro-

lyte that does not participate in the electrode reaction, the

diffusion layer should remain charge neutral and the conduc-

tivity largely unaffected. In this case the system will be

dominated by capacitive charging and behave as if no Fara-

daic reaction is taking place. This underlines the potential

role of the speciﬁcs of the electrolyte and electrode material,

known to play a decisive role in a few electrically generated

phenomena ͓32,33͔.

E. Surface conduction in the Debye layer

At large voltage the Debye layer can accumulate enough

charge that the mean ion density, and hence the conductivity,

becomes signiﬁcantly larger than in the bulk electrolyte. In

this case one cannot neglect lateral surface currents in Eq.

͑9͒. This is quantiﬁed by the Dukhin number Du=

s

/ ᐉ

0

,

where

s

is the surface conductivity in the Debye layer, in-

corporating both migration and electro-osmotic convection

of charge, is the bulk conductivity, and ᐉ

0

is the character-

istic geometrical length scale. For a symmetric monovalent

͑1:1͒ electrolyte the Dukhin number can be expressed as ͓17͔

Du =

ͫ

1 + ͩ

k

B

T

e

ͪ

2

2⑀

D

ͬ

D

ᐉ

0

4 sinh

2

ͩ

e

4k

B

T

ͪ. ͑56͒

Taking ϳV

0

/ ͑1+␦͒ and using typical parameter values

from Table I, we ﬁnd that surface conductance becomes sig-

niﬁcant and Duտ1 for V

0

տ250 mV. Generally, we expect

surface currents to smear out the charge distribution across

the electrodes, reducing the tangential ﬁeld and hence the

electroosmotic slip and pumping velocity. We do not antici-

pate that this alone could be the mechanism responsible for

reversal of the time average slip velocity or the direction of

net pumping.

F. Local salt depletion at electrodes

Another issue relating to the massive accumulation of

ions in the Debye layer at large voltage is discussed by Ba-

zant et al. ͓19͔—namely, that of where those ions come

from. For a symmetric monovalent ͑1:1͒ electrolyte we de-

ﬁne the excess amount of neutral salt absorbed in the Debye

layer as

w͑x, t͒ =

1

2

͵

Debye layer

͑c

+

+ c

−

− 2c

*

͒dy ͑57͒

=c

*

D

4 sinh

2

ͩ

e

4k

B

T

ͪ. ͑58͒

This salt is periodically taken up and released again between

the Debye layer and a diffusion zone of width ᐉ͑͒ =

ͱ

D/ .

However, when the amount of salt absorbed in the Debye

layer approaches the total amount c

*

ᐉ͑͒ available in the

diffusion zone, we get local depletion of salt and our as-

sumption of uniform electrolyte concentration throughout the

bulk breaks down. This occurs for

ϳ

k

B

T

e

2 log͑ᐉ͑͒/

D

͒ =

k

B

T

e

log͑

D

/͒, ͑59͒

where we wrote

D

=D/

D

2

. Taking again ϳV

0

/ ͑1+␦͒ and

using typical parameter values from Table I, we ﬁnd that salt

depletion around the relaxation frequency

0

becomes an

issue for V

0

տ125 mV.

There are several interesting phenomena arising when the

approximation of uniform bulk electrolyte concentration and

conductivity breaks down. Among other things, this would

change the local Debye length as seen from the electrode

surface and give rise to space charge like ϳ⑀ץ

y

ץ

y

log͑͒

in the diffusion layer ͓32͔. While the amount of space charge

in the diffusion layer would typically be much smaller than

in the Debye layer, it interacts with the ﬂuid in a place more

remotely from the wall and is therefore more efﬁcient in

setting the ﬂuid in motion globally. We are currently study-

ing in more detail the role of this nonlinear mechanism for

inducing ﬂuid motion.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-14

VII. CONCLUSION

We have extended existing theory for ac electrokinetic

pumping to account for vertical conﬁnement of the pumping

channel, nonlinear surface capacitance of the double layer,

and current injection from Faradaic electrode reactions in

both a linear and a nonlinear scheme. For our particular

model system of an asymmetric electrode array micropump

we have obtained a number of results that we subsequently

have interpreted using simple physical arguments. As these

arguments are more general than the speciﬁc model, we ex-

pect our results to hold with some generality for other similar

electrokinetic systems. We therefore believe that our work

will be useful for design of electrokinetic micropumps.

Our results compare well with experiments in many as-

pects but there still remains unexplained phenomena—e.g.,

the reversal of the net pumping or the time average slip di-

rection. This points out the need for further studies of ͑other͒

nonlinear phenomena to fully understand the complex elec-

trohydrodynamic system.

APPENDIX A

We give here a proof that there can be no net pumping at

low frequency when the electrode surface properties are

symmetric and we stay within the Debye-Hückel limit. Con-

sider a device where the capacitance ratio ␦͑x ˜͒ and inverse

charge-transfer resistance K͑x ˜͒ vary across the electrodes.

The electro-osmotic slip velocity becomes

u ˜

s

=

1 + ␦

0

1 + ␦͑x ˜͒

͑V

˜

ext

−

˜

͒ץ

x ˜

˜

, ͑A1͒

where ␦

0

is the capacitance ratio used in the rescaling of the

problem, and Eq. ͑25͒ describing the charge balance in the

Debye layer generalizes to

n · ١

˜

ˆ

=

ͩ

i˜

1 + ␦

0

1 + ␦͑x ˜͒

+ K͑x ˜͒

ͪ

͑

ˆ

− V

ˆ

ext

͒. ͑A2͒

Combining these we can write the time average slip as

͗u ˜

s

͘ =

1 + ␦

0

1 + ␦͑x ˜͒

1

2

Re͓͑V

ˆ

ext

−

ˆ

͒ץ

x ˜

ˆ

*

͔

= −

1

2

Reͫ

n · ١

˜

ˆ

i˜ + ͑x ˜͒

ץ

x ˜

ˆ

*

ͬ, ͑A3͒

where ͑x ˜͒ =K͑x ˜͓͒1+␦͑x ˜͔͒ / ͑1+␦

0

͒. In the high-symmetry

case where ͑x ˜͒ is a constant independent of x ˜, the net pump-

ing velocity U

˜

can be manipulated as follows:

U

˜

=

1

L

˜

͵

electrodes

͗u ˜

s

͘dx ˜ =

1

2L

˜

͵

ץV

− Reͫ

n · ١

˜

ˆ

i˜ +

ץ

x ˜

ˆ

*

ͬds ˜,

͑A4͒

where ץV denotes the boundary of the computational domain

and we used the fact that n· ١

˜

ˆ

vanishes on insulating walls

and cancels over periodic boundaries. Then, using Gauss’

law and recalling that n is a unit normal pointing out of the

electrodes and into V, we ﬁnd

U

˜

=

1

2L

˜

Re

ͫ

1

i˜ +

͵

V

١

˜

· ͑١

˜

ˆ

ץ

x ˜

ˆ

*

͒dr ˜

ͬ

͑A5͒

=

1

2L

˜

Re

ͫ

1

i˜ +

͵

V

١

˜

ˆ

· ١

˜

͑ץ

x ˜

ˆ

*

͒dr ˜

ͬ

͑A6͒

=

1

2L

˜

͵

V

1

2

ץ

x ˜

͑١

˜

ˆ

· ١

˜

ˆ

*

͒ + ˜ Im͓١

˜

ˆ

· ١

˜

͑ץ

x ˜

ˆ

*

͔͒

2

+ ˜

2

dr ˜ ͑A7͒

=

1

2L

˜

˜

2

+ ˜

2

Im

ͫ

͵

V

١

˜

ˆ

· ١

˜

͑ץ

x ˜

ˆ

*

͒dr ˜

ͬ

. ͑A8͒

Here Eq. ͑A6͒ follows from ١

˜ 2

ˆ

=0, Eq. ͑A7͒ follows from

simple complex arithmetics, and ﬁnally, Eq. ͑A8͒ holds be-

cause the integral of the term ץ

x ˜

͑١

˜

ˆ

· ١

ˆ

*

͒ cancels out due to

periodicity.

The ﬁnal result in Eq. ͑A8͒ clearly shows that regardless

of , the net pumping U

˜

goes to zero in the low-frequency

limit ˜ →0. The apparent singular behavior for =0 is only

apparent because the screening is then complete and ١

˜

ˆ

zero

in the bulk. Hence, our analysis rules out any net pumping in

the low-frequency limit of our simple model when the elec-

trode surface properties are symmetric in the sense of mak-

ing the parameter independent of x ˜.

APPENDIX B

The nonlinear problem is converted to weak form by mul-

tiplying a test function ͑r ˜͒ for

˜

on Eq. ͑50͒, integrating

over the computational domain V, and using Gauss’ law with

Eqs. ͑46͒ and ͑49͒ to eliminate n· ١

˜

˜

. Hence, for

˜

to be a

weak solution we require that

͵

V

١

˜

· ١

˜

dr ˜ +

͵

electrodes

͓ץ

t

˜q ˜ − j

˜

ext

͑q ˜͔͒dx ˜ = 0 ͑B1͒

for all and at all times. The constraint, Eq. ͑45͒, on the

potential drop across the double layer is satisﬁed by multi-

plying a test function ͑x ˜͒ for q ˜ and requiring that

͵

electrodes

͓V

˜

ext

−

˜

−

˜

͑q ˜͒ + q ˜␦/͑1 + ␦͔͒dx ˜ = 0 ͑B2͒

for all and at all times. The weak problem is discretized

using the Galerkin method; i.e., we expand

˜

on a ﬁnite set

of basis functions ͕

n

͑r ˜͖͒

n=1

N

as

˜

͑r ˜, t

˜

͒ Ϸ

͚

n=1

N

˜

n

͑t

˜

͒

n

͑r ˜͒ ͑B3͒

and similarly for q ˜ on ͕

n

͑x ˜͖͒

n=1

N

q

, and we use those same

basis sets as test functions in Eqs. ͑B1͒ and ͑B2͒. Further, the

solution is periodic in time, so we discretize on a set of M

equispaced points t

˜

m

=m⌬t

˜

, m=1, 2, . . . , M, over one period,

and use the leapfrog ﬁnite-difference scheme ץ

t

˜q ˜

n

͑t

˜

m

͒

Ϸ͓q ˜

n

͑t

˜

m+1

͒ −q ˜

n

͑t

˜

m−1

͔͒ / 2⌬t

˜

to evaluate the time derivative in

Eq. ͑B1͒.

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-15

The full coupled problem for all M time steps is solved

numerically by damped Newton iterations, using the com-

mercial ﬁnite-element software FEMLAB to deﬁne the basis

functions

n

͑r ˜͒ and

n

͑x ˜͒ and to evaluate the Jacobian ma-

trix of the discretized problem. In the frequency-voltage

maps in Figs. 11 and 13 there are ﬁve data points per decade

for both frequency and voltage. The convergence of the

Newton iterations is accelerated by a continuation scheme

where the converged solution at one frequency is used as

initial guess for the next solution at higher frequency and the

same voltage; at the lowest frequency, we use the solution

from lower voltage as initial guess when stepping to higher

voltage.

With nonlinear Debye layer capacitance but no Faradaic

currents, we obtain fairly accurate results with a coarse time

resolution of M=16: When compared to a ﬁne resolution

result at M=64, the maximal relative error on U

˜

is less than

1% ͑3%͒ for V

˜

0

ഛ10

1

͑10

3

͒, whereas the maximal relative

error on q ˜ is about twice as large. Figures 11 and 12 show the

results for M=64. With Faradaic current injection, the charg-

ing and decharging of the Debye layer becomes very rapid at

high voltage, which makes the solution more demanding:

Comparing results for M=64 and M=128 we ﬁnd that the

maximal relative error on U

˜

is less than 0.1% ͑1.5%͒ for

V

˜

0

ഛ10

1

͑10

3

͒, but that the corresponding error on q ˜ is as

large as 0.5% ͑20%͒. However, one should keep in mind that

at high voltage our weakly nonlinear model also becomes

physically inadequate for reasons discussed in Secs.

VI D–VI F. Figures 13 and 14 show results for M=128.

͓1͔ H. A. Stone, A. D. Stroock, and A. Ajdari, Annu. Rev. Fluid

Mech. 36, 381 ͑2004͒.

͓2͔ N. G. Green, A. Ramos, A. Gonzalez, H. Morgan, and A.

Castellanos, Phys. Rev. E 61, 4011 ͑2000͒.

͓3͔ A. Gonzalez, A. Ramos, N. G. Green, A. Castellanos, and H.

Morgan, Phys. Rev. E 61, 4019 ͑2000͒.

͓4͔ N. G. Green, A. Ramos, A. Gonzalez, H. Morgan, and A.

Castellanos, Phys. Rev. E 66, 026305 ͑2002͒.

͓5͔ A. Ajdari, Phys. Rev. E 61, R45 ͑2000͒.

͓6͔ A. B. D. Brown, C. G. Smith, and A. R. Rennie, Phys. Rev. E

63, 016305 ͑2000͒.

͓7͔ S. Debesset, C. J. Hayden, C. Dalton, J. C. T. Eijkel, and A.

Manz ͑unpublished͒.

͓8͔ V. Studer, A. Pepin, Y. Chen, and A. Ajdari, Microelectron.

Eng. 61-62, 915 ͑2002͒.

͓9͔ M. Mpholo, C. G. Smith, and A. B. D. Brown, Sens. Actuators

B 92, 262 ͑2003͒.

͓10͔ S. Debesset, C. J. Hayden, C. Dalton, J. C. T. Eijkel, and A.

Manz, Lab Chip 4, 396 ͑2004͒.

͓11͔ V. Studer, A. Pepin, Y. Chen, and A. Ajdari, Analyst ͑Cam-

bridge, U.K.͒ 129, 944 ͑2004͒.

͓12͔ B. P. Cahill, L. J. Heyderman, J. Gobrecht, and A. Stemmer,

Phys. Rev. E 70, 036305 ͑2004͒.

͓13͔ A. Ramos, H. Morgan, N. G. Green, A. Gonzalez, and A.

Castellanos, J. Appl. Phys. 97, 084906 ͑2005͒.

͓14͔ P. Garcia, A. Ramos, N. G. Green, and H. Morgan ͑unpub-

lished͒.

͓15͔ A. Ramos, A. Gonzalez, A. Castellanos, N. G. Green, and H.

Morgan, Phys. Rev. E 67, 056302 ͑2003͒.

͓16͔ D. Lastochkin, R. Zhou, P. Wang, Y. Ben, and H.-C. Chang, J.

Appl. Phys. 96, 1730 ͑2004͒.

͓17͔ R. J. Hunter, Foundations of Colloid Science ͑Oxford Univer-

sity Press, New York, 1991͒.

͓18͔ A. Bard and L. Faulkner, Electrochemical Methods ͑Wiley,

New York, 1980͒.

͓19͔ M. Z. Bazant, K. Thornton, and A. Ajdari, Phys. Rev. E 70,

021506 ͑2004͒.

͓20͔ N. A. Mortensen, L. H. Olesen, L. Belmon, and H. Bruus,

Phys. Rev. E 71, 056306 ͑2005͒.

͓21͔ I. Rubinstein and B. Zaltzman, Math. Models Meth. Appl. Sci.

11, 263 ͑2001͒.

͓22͔ L. H. Olesen ͑unpublished͒, following an approach initially

suggested in Y. Ben ͑unpublished͒.

͓23͔ F. Nadal, F. Argoul, P. Krestener, B. Pouligny, C. Ybert, and A.

Ajdari, Eur. Phys. J. E 9, 387–399 ͑2002͒.

͓24͔ Ramos et al., used the width W

2

of the wide electrode as their

characteristic length scale rather than G

1

. Hence in ͓15͔ both

the dimensionless frequency and slip velocity are larger than

ours by a factor of W

2

/ G

1

.

͓25͔ A well-known instance of this singular behavior is that at high

frequency, where the screening is negligible, the potential

close to the sharp electrode edge goes like ϳr

1/2

cos͑/ 2͒,

and the electric ﬁeld diverges as ٌ͉͉ ϳr

−1/2

. Here r is the

distance from the edge and the polar angle. At ﬁnite fre-

quency the normal ﬁeld is linked to the potential drop across

the double layer through Eq. ͑25͒ and therefore cannot be sin-

gular. However, the tangential ﬁeld diverges logarithmically. In

particular, at low frequency it can be shown analytically that

the potential to ﬁrst order in goes like

ϳIm͓Li

2

͑e

−2͑i͑x−x

j

͒+y͒/L

͔͒ close to the electrode edge ͓22͔ and

the tangential ﬁeld diverges as ͉ץ

x

͉ ϳ−log͉x−x

j

͉. Here

Li

2

͑z͒ =͚

n=1

ϱ

z

n

/ n

2

is the dilogarithm function.

͓26͔ At high frequency u ˜

s

is nonzero only in a region of width ˜

−1

from the electrode edges. Therefore our numerical solution be-

comes inaccurate for ˜ տ10

3

and likewise for Kտ10

3

in the

case of Faradaic current injection.

͓27͔ A. Ajdari, C. R. Phys. 5, 539 ͑2004͒.

͓28͔ Experimentally, it is often found even for perfectly polarizable

solid electrodes that the double-layer impedance does not

match that of a perfect capacitor ͓29͔. Some authors therefore

ﬁnd it efﬁcient to use the empirical so-called constant-phase

element to model the double layer, but we prefer not to do so

due to the lack of structural justiﬁcation for it.

͓29͔ Z. Kerner and T. Pajkossy, Electrochim. Acta 46, 207211

͑2000͒.

͓30͔ W. H. Smyrl and J. Newman, Trans. Faraday Soc. 63, 207

͑1967͒.

͓31͔ K. T. Chu and M. Z. Bazant, SIAM J. Appl. Math. 65, 1485

͑2005͒.

͓32͔ P. J. Sides, Langmuir 19, 2745 ͑2003͒.

͓33͔ J. A. Fagan, P. J. Sides, and D. C. Prieve, Langmuir 21, 1784

͑2005͒.

OLESEN, BRUUS, AND AJDARI PHYSICAL REVIEW E 73, 056313 ͑2006͒

056313-16

Appendix C

Paper published in

Int. J. Numer. Meth. Eng.

Title

A high-level programming-language implementation of topology optimiza-

tion applied to steady-state Navier–Stokes ﬂow

Authors

Laurits Højgaard Olesen, Fridolin Okkels, and Henrik Bruus

Reference

Int. J. Num. Meth. Eng. 65, 975-1001 (2006) (27 pages)

150 Paper published in Int. J. Numer. Meth. Eng.

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2006; 65:975–1001

Published online 1 September 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/nme.1468

A high-level programming-language implementation of topology

optimization applied to steady-state Navier–Stokes ﬂow

Laurits Højgaard Olesen

∗, †

, Fridolin Okkels and Henrik Bruus

MIC—Department of Micro and Nanotechnology, Technical University of Denmark,

DK-2800 Kongens Lyngby, Denmark

SUMMARY

We present a versatile high-level programming-language implementation of non-linear topology opti-

mization. Our implementation is based on the commercial software package FEMLAB, and it allows

a wide range of optimization objectives to be dealt with easily. We exemplify our method by studies

of steady-state Navier–Stokes ﬂow problems, thus extending the work by Borrvall and Petersson on

topology optimization of ﬂuids in Stokes ﬂow (Int. J. Num. Meth. Fluids 2003; 41:77–107). We

analyse the physical aspects of the solutions and how they are affected by different parameters of the

optimization algorithm. A complete example of our implementation is included as FEMLAB code in

an appendix. Copyright ᭧ 2005 John Wiley & Sons, Ltd.

KEY WORDS: topology optimization; Navier–Stokes ﬂow; inertial effects; FEMLAB

1. INTRODUCTION

The material distribution method in topology optimization was originally developed for stiffness

design of mechanical structures [1] but has now been extended to a multitude of design

problems in structural mechanics as well as to optics and acoustics [2–5]. Recently Borrvall

and Petersson introduced the method for ﬂuids in Stokes ﬂow [6]. However, it is desirable to

extend the method to ﬂuids described in a full Navier–Stokes ﬂow; a direction pioneered by

the work of Sigmund and Gersborg-Hansen [7–9].

In the present work we present such an extension by introducing a versatile high-level

programming-language implementation of non-linear topology optimization, based on the com-

mercial software package FEMLAB. It has a wider range of applicability than the Navier–Stokes

∗

Correspondence to: Laurits Højgaard Olesen, MIC—Department of Micro and Nanotechnology, DTU Bldg. 345

East, Technical University of Denmark, DK-2800 Kongens Lyngby, Denmark.

†

E-mail: lho@mic.dtu.dk

Contract/grant sponsor: The Danish Technical Research Council; contract/grant number: 26-03-0037

Received 16 November 2004

Revised 24 May 2005

Copyright

᭧

2005 John Wiley & Sons, Ltd. Accepted 26 June 2005

976 L. H. OLESEN, F. OKKELS AND H. BRUUS

problems studied here, and moreover it allows a wide range of optimization objectives to be

dealt with easily.

Extending the topology optimization method to new physical domains generally involves

some rethinking of the design problem and some ‘trial and error’ to determine suitable design

objectives. It also requires the numerical analysis and implementation of the problem, e.g.

using the ﬁnite element method (FEM). This process is accelerated a lot by using a high-level

FEM library or package that allows different physical models to be joined and eases the tasks

of geometry setup, mesh generation, and postprocessing. The disadvantage is that high-level

packages tend to have rather complex data structure, not easily accessible to the user. This

can complicate the actual implementation of the problem because the sensitivity analysis is

traditionally formulated in a low-level manner.

In this work we have used the commercial ﬁnite-element package FEMLAB both for the

solution of the ﬂow problem and for the sensitivity analysis required by the optimization

algorithm. We show how this sensitivity analysis can be performed in a simple way that is

almost independent of the particular physical problem studied. This approach proves even more

useful for multi-ﬁeld extensions, where the ﬂow problem is coupled to, e.g. heat conduction,

convection–diffusion of solutes, and deformation of elastic channel walls in valves and ﬂow

rectiﬁers [10].

The paper is organized as follows: In Section 2 we introduce the topology optimization

method for ﬂuids in Navier–Stokes ﬂow, and discuss the objective of designing ﬂuidic devices

or channel networks for which the power dissipation is minimized. In Section 3 we express

the Navier–Stokes equations in a generic divergence form that allows them to be solved with

FEMLAB. This form encompasses a wide range of physical problems. We also work out the

sensitivity analysis for a class of integral-type optimization objectives in such a way that the

built-in symbolic differentiation tools of FEMLAB can be exploited. In Section 4 we present our

two numerical examples that illustrates different aspects and problems to consider: The ﬁrst

example deals with designing a structure that can guide the ﬂow in the reverse direction of an

applied pressure drop. The general outcome of the optimization is an S-shaped channel, but

the example illustrates how the detailed structure depends on the choice of the parameters of

the algorithm. The second example deals with a four terminal device where the ﬂuidic channel

design that minimizes the power dissipation shows a Reynolds number dependence. As the

Reynolds number is increased a transition occurs between two topologically different solutions,

and we discuss how the position of the transition depends on the choice of initial conditions.

Finally in the appendix we include a transcript of our FEMLAB code required for solving

the second numerical example. The code amounts to 111 lines—excluding the optimization

algorithm that can be obtained by contacting Svanberg [11–13].

2. TOPOLOGY OPTIMIZATION FOR NAVIER–STOKES FLOW IN STEADY STATE

Although our high-level programming-language implementation is generally applicable we have

chosen to start on the concrete level by treating the basic equations for our main example: the

full steady-state Navier–Stokes ﬂow problem for incompressible ﬂuids.

We consider a given computational domain with appropriate boundary conditions for the

ﬂow given on the domain boundary *. The goal of the optimization is to distribute a certain

amount of solid material inside such that the material layout deﬁnes a ﬂuidic device or

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 977

channel network that is optimal with respect to some objective, formulated as a function of

the variables, e.g. minimization of the power dissipated inside the domain.

The basic principle in the material distribution method for topology optimization is to replace

the original discrete design problem with a continuous one where the material density is allowed

to vary continuously between solid and void [2]. Thus in our ﬂow problem we assume the design

domain to be ﬁlled with some idealized porous material of spatially varying permeability. Solid

wall and open channels then correspond to the limits of very low and very high permeability,

respectively.

In the ﬁnal design there should preferably be no regions at intermediate permeability since

otherwise it cannot be interpreted as a solution to the original discrete problem. Alternatively it

may be possible to fabricate the device from polymeric materials such as PDMS that naturally

have a ﬁnite permeability to the ﬂuid [14].

2.1. Governing equations for ﬂow in idealized porous media

We assume that the ﬂuid ﬂowing in the idealized porous medium is subject to a friction force f

which is proportional to the ﬂuid velocity v, c.f. Darcy’s law. Thus f = − :v, where :(r) is

the inverse of the local permeability of the medium at position r. These properties of the

idealized porous medium may only be approximately valid for an actual medium. However,

the assumptions are not in conﬂict with any fundamental physical law, and since the converged

solutions contain only solid walls and open channels, the speciﬁc nature of the idealized porous

medium is of no consequence.

The ﬂow problem is described in terms of the ﬂuid velocity ﬁeld v(r) and pressure p(r).

The governing equations are the steady state Navier–Stokes equation and the incompressibility

constraint

,(v · ∇)v =∇ · − :v (1)

∇ · v =0 (2)

where , is the mass density of the ﬂuid. For an incompressible Newtonian ﬂuid the compo-

nents o

ij

of the Cauchy stress tensor are given by

o

ij

= − po

ij

+ n

*v

i

*x

j

+

*v

j

*x

i

(3)

where n is the dynamic viscosity. The formalism is valid in three dimensions, but for simplicity

we shall consider only two-dimensional problems, i.e. we assume translational invariance in

the third dimension and set r =(x

1

, x

2

) and v =(v

1

(r), v

2

(r)). The boundary conditions will

typically be either Dirichlet type specifying the velocity ﬁeld v on the boundary or Neumann

type specifying the external forces n · .

It is convenient to introduce a design variable ﬁeld ¸(r) controlling the local permeability

of the medium. We let ¸ vary between zero and unity, with ¸ =0 corresponding to solid

material and ¸ =1 to no material. Following Reference [6] we then relate the local inverse

permeability :(r) to the design ﬁeld ¸(r) by the convex interpolation

:(¸) ≡:

min

+ (:

max

− :

min

)

q[1 − ¸]

q + ¸

(4)

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

978 L. H. OLESEN, F. OKKELS AND H. BRUUS

where q is a real and positive parameter used to tune the shape of :(¸). Ideally, impermeable

solid walls would be obtained with :

max

=∞, but for numerical reasons we need to choose a

ﬁnite value for :

max

. For the minimal value we choose :

min

=0.

‡

For a given material distribution ¸(r) there are two dimensionless numbers characterizing

the ﬂow, namely the Reynolds number

Re =

,v

n

(5)

describing the ratio between inertia and viscous forces, and the Darcy number

Da =

n

:

max

2

(6)

describing the ratio between viscous and porous friction forces. Here is a characteristic length

scale of the system and v a characteristic velocity.

Almost impermeable solid material is obtained for very low Darcy numbers, in practice

Da 10

−5

. Further insight into the meaning of the Darcy number is gained by considering

Poiseuille ﬂow in a channel or slit of width between two inﬁnite parallel plates of porous

material. In this case the ﬂuid velocity inside the porous walls decays on a length scale

Da

,

where

Da

=

√

Da =

√

n/:

max

. See also Section 4.1.1 for details on how the ﬂow depends

on Da.

2.2. Power dissipation

In the pioneering work by Borrvall and Petersson [6] the main focus was on minimizing the

power dissipation in the ﬂuid. The total power dissipated inside the ﬂuidic system (per unit

length in the third dimension) is given by [15]

(v, p, ¸) =

¸

1

2

n

¸

i, j

*v

i

*x

j

+

*v

j

*x

i

2

+

¸

i

:(¸)v

2

i

dr (7a)

In steady-state this is equal to the sum of the work done on the system by the external forces

and the kinetic energy convected into it

(v, p, ¸) =

*

¸

i, j

¸

n

i

o

ij

v

j

− n

i

v

i

1

2

,v

2

j

¸

ds (7b)

Here n is a unit outward normal vector such that n· is the external force acting on the system

boundary and n · · v is the work done on the system by this force. Moreover, in the common

case where the geometry and boundary conditions are such that the no-slip condition v =0

‡

Borrvall and Petersson suggest a model for plane ﬂow between two parallel surfaces of varying separation h(r).

The power dissipation due to out-of-plane shears is modelled by an absorption term −:v, where v(r)

is the average velocity between the surfaces and :(r) =12n/h(r)

2

. In their model it is therefore natural to

operate with a non-zero :

min

=12n/h

2

max

in Equation (4).

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 979

applies on all external solid walls, while on the inlet and outlet boundaries v is parallel to n

and (n · ∇) v =0,

m

Equation (7b) reduces to

(v, p, ¸) =

*

−n · v

p +

1

2

,v

2

ds (7c)

Borrvall and Petersson showed that for Stokes ﬂow with Dirichlet boundary conditions

everywhere on the boundary *, the problem of minimizing the total power dissipation inside

the ﬂuidic device subject to a volume constraint on the material distribution is mathematically

well-posed. Moreover it was proven that in the case where :(¸) is a linear function, the optimal

material distribution is fully discrete-valued.

When :(¸) is not linear but convex then the solid/void interfaces in the optimal solution

are not discrete zero/unity transitions but slightly smeared out. Convexity implies that the

(negative value of the) slope of : at ¸ =0 is larger than at ¸ =1; therefore there will be a

neighbourhood around the discrete interface where it pays to move material from the solid side

to the void. Using the interpolation in Equation (4) we have :

(0) =(:

min

− :

max

)(1 + q)/q

and :

(1) =(:

min

−:

max

)q/(1 +q). For large values of q the interpolation is almost linear and

we expect almost discrete interfaces, whereas for small q we expect smeared out interfaces in

the optimized solution.

Consider the case when Equation (7c) applies. If the system is driven with a prescribed

ﬂow rate then minimizing the total power dissipation is clearly equivalent to minimizing the

pressure drop across the system. Conversely, if the system is driven at a prescribed pressure

drop, then the natural design objective will be to maximize the ﬂow rate which is equivalent

to maximizing the dissipated power, c.f. Equation (7c). In either case the objective can be

described as minimizing the hydraulic resistance of the system.

For problems with more complex design objectives, such as a minimax problem for the

ﬂow rate through several different outlets, there will typically be no analog in terms of total

dissipated power. In such cases there is no guarantee for the existence of a unique optimal

solution and one has to be extra careful when formulating the design problem.

3. GENERALIZED FORMULATION OF THE OPTIMIZATION PROBLEM

For a given material distribution we solve the Navier–Stokes ﬂow problem using the commercial

ﬁnite element software FEMLAB. It provides both a graphical front-end and a library of high-

level scripting tools based on the MATLAB programming language, and it allows the user to

solve a wide range of physical problems by simply typing in the strong form of the governing

equations as text expressions. The equations must then comply with a generic divergence form

that eases the conversion to weak form required for the ﬁnite element solution. However, that

is not a severe constraint since this is the natural way of expressing most partial differential

equations originating from conservation laws.

Since we have chosen ﬂuidics as our main example, we begin by expressing the incompress-

ible Navier–Stokes ﬂow problem in divergence form. Then we state the optimization problem

m

In particular this is the case when the inlets and outlets are chosen as straight channels sufﬁciently long that

prescribing a parabolic Poiseuille proﬁle can be justiﬁed, see Figures 1 and 6.

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

980 L. H. OLESEN, F. OKKELS AND H. BRUUS

with a general form of the design objective function and perform the discretization and sen-

sitivity analysis based on this generalized formulation. We stress that although for clarity our

examples are formulated in two dimensions only, the method is fully applicable for 3D systems.

3.1. The ﬂow problem in divergence form

We ﬁrst introduce the velocity–pressure vector u =[v

1

, v

2

, p] and deﬁne for i =1, 2, 3 the quan-

tities

i

and F

i

as

1

≡

¸

o

11

o

21

,

2

≡

¸

o

12

o

22

,

3

≡

¸

0

0

(8)

and

F

1

≡,(v · ∇)v

1

+ :(¸)v

1

, F

2

≡,(v · ∇)v

2

+ :(¸)v

2

, F

3

≡∇ · v (9)

Using this, Equations (1) and (2) can be written in divergence form as

∇ ·

i

=F

i

in , Governing equations (10a)

R

i

=0 on *, Dirichlet b.c. (10b)

−n ·

i

=G

i

+

3

¸

j=1

*R

j

*u

i

j

j

on *, Neumann b.c. (10c)

where

i

and F

i

are understood to be functions of the solution u, its gradient ∇u, and of the

design variable ¸. The quantity R

i

(u, ¸) in Equation (10b) describes Dirichlet type boundary

conditions. For example, ﬂuid no-slip boundary conditions are obtained by deﬁning R

1

≡v

1

and

R

2

≡v

2

on the external solid walls. The quantity G

i

(u, ¸) in Equation (10c) describe Neumann

type boundary conditions, and j

i

denote the Lagrange multiplier necessary to enforce the

constraint R

i

=0, e.g. the force with which the solid wall has to act upon the ﬂuid to enforce

the no-slip boundary condition. Of course, it is not possible to enforce both Dirichlet and

Neumann boundary conditions for the same variable simultaneously. Only when the variable u

i

is not ﬁxed by any of the Dirichlet constraints R

j

does the Neumann condition G

i

come into

play, as all *R

j

/*u

i

vanish and the Lagrange multipliers j

j

are decoupled from Equation (10c).

Inactive Dirichlet constraints can be obtained simply by specifying the zero-function R

i

≡0,

that also satisﬁes Equation (10b) trivially.

3.2. The objective function

In general the design objective for the optimization is stated as the minimization of a certain

objective function (u, ¸). We shall consider a generic integral-type objective function of the

form

(u, ¸) =

A(u, ¸) dr +

*

B(u, ¸) ds (11)

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 981

In particular, we can treat the design objective of minimizing the power dissipation inside the

ﬂuidic domain by taking, c.f. Equation (7a)

A≡

1

2

n

¸

i, j

*v

i

*x

j

+

*v

j

*x

i

2

+

¸

i

:(¸)v

2

i

in and B ≡0 on * (12)

Alternatively, the objective of maximizing the ﬂow out through a particular boundary seg-

ment *

o

is obtained by choosing

A≡0 in and B ≡

¸

−n · v on *

o

0 on *\*

o

(13)

and objectives related to N discrete points r

k

can be treated using Dirac delta functions as

A≡

N

¸

k=1

A

k

(u, ¸)o(r − r

k

) in and B ≡0 on * (14)

Finally we stress that not all optimization objectives lend themselves to be expressed in the form

of Equation (11)—an example of which is the problem of maximizing the lowest vibrational

eigenfrequency in structural mechanics.

3.3. Optimization problem

The optimal design problem can now be stated as a continuous constrained non-linear opti-

mization problem

min

¸

(u, ¸) (15a)

subject to:

**¸(r)dr − [|| 0, Volume constraint (15b)
**

: 0 ¸(r) 1, Design variable bounds (15c)

: Equations (10a)–(10c), Governing equations (15d)

With the volume constraint we require that at least a fraction 1 − [ of the total volume ||

should be ﬁlled with porous material.

The very reason for replacing the original discrete design problem with a continuous one by

assuming a porous and permeable material, is that it allows the use of efﬁcient mathematical

programming methods for smooth problems. We have chosen the popular method of moving

asymptotes (MMA) [11, 12], which is designed for problems with a large number of degrees-

of-freedom and thus well-suited for topology optimization [2]. It is a gradient-based algorithm

requiring information about the derivative with respect to ¸ of both the objective function

and the constraints. Notice that for any ¸ the governing equations allow us to solve for u;

therefore in effect they deﬁne u[¸] as an implicit function. The gradient of is then obtained

using the chain rule

d

d¸

[(u[¸], ¸)] =

*

*¸

+

*

*u

·

*u

*¸

dr (16)

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

982 L. H. OLESEN, F. OKKELS AND H. BRUUS

However, because u[¸] is implicit, it is impractical to evaluate the derivative *u/*¸ directly.

Instead, we use the adjoint method to eliminate it from Equation (16) by computing a set of

Lagrange multipliers for Equations (10a)–(10c) considered as constraints [16]. For details see

Section 3.4.

The optimization process is iterative and the kth iteration consists of three steps:

(i) Given a guess ¸

(k)

for the optimal material distribution we ﬁrst solve Equations

(10a)–(10b) for u

(k)

as a ﬁnite element problem using FEMLAB.

(ii) Next, the sensitivity analysis is performed where the gradient of the objective and con-

straints with respect to ¸ is evaluated. In order to eliminate *u/*¸ from Equation (16)

we solve the adjoint problem of Equations (10a)–(10c) for the Lagrange multip-

liers ˜ u

(k)

, also using FEMLAB.

(iii) Finally, we use MMA to obtain a new guess ¸

(k+1)

for the optimal design based on

the gradient information and the past iteration history.

Of the three steps, (i) is the most expensive computationalwise since it involves the solution

of a non-linear partial differential equation.

3.4. Discretization and sensitivity analysis

The starting point of the ﬁnite element analysis is to approximate the solution component u

i

on a set of ﬁnite element basis functions {m

i, n

(r)}

u

i

(r) =

¸

n

u

i, n

m

i, n

(r) (17)

where u

i, n

are the expansion coefﬁcients. Similiarly, the design variable ﬁeld ¸(r) is expressed

as

¸(r) =

¸

n

¸

n

m

4,n

(r) (18)

For our incompressible Navier–Stokes problem we use the standard Taylor–Hood element pair

with quadratic velocity approximation and linear pressure. For the design variable we have

chosen the linear Lagrange element.

¶

The problem Equations (10a)–(10c) is discretized by the Galerkin method and takes the

form

L

i

(U, ) −

3

¸

j=1

N

T

ji

j

=0 and M

i

(U, ) =0 (19)

where U

i

,

i

, and are column vectors holding the expansion coefﬁcients for the solution u

i, n

,

the Lagrange multipliers j

i, n

, and the design variable ﬁeld ¸

n

, respectively. The column vec-

tor L

i

contains the projection of Equation (10c) onto m

i, n

which upon partial integration

¶

Another common choice is the discontinuous and piecewise constant element for the design variable. Notice

that for second and higher order Lagrange elements the condition 0 ¸

n

1 does not imply 0 ¸(r) 1 for

all r because of overshoot at sharp zero-to-unity transitions in ¸. This in turn can result in negative :, c.f.

Equation (4), which is unphysical and also destroys the convergence of the algorithm.

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 983

is given by

L

i, n

=

(m

i, n

F

i

+∇m

i, n

·

i

) dr +

*

m

i, n

G

i

ds (20)

The column vector M

i

contains the pointwise enforcement of the Dirichlet constraint

Equation (10b)

M

i, n

=R

i

(u(r

i, n

)) (21)

Finally, the matrix N

ij

= − *M

i

/*U

j

describes the coupling to the Lagrange multipliers in

Equation (10c). The solution of the non-linear system in Equation (19) above corresponds to

step (i) in kth iteration. The sensitivity analysis in step (ii) requires us to compute

d

d

[(U(), )] =

*

*

+

3

¸

i=1

*

*U

i

*U

i

*

(22a)

which is done using the standard adjoint method [16]. By construction we have for any

that L

i

(U(), ) −

¸

3

j =1

N

T

ji

j

() =0 and M

i

(U(), ) =0. Therefore also the derivative of

those quantities with respect to is zero, and adding any multiple, say

˜

U

i

and

˜

i

, of them to

Equation (22a) does not change the result

d

d

[(U(), )] =

*

*

+

3

¸

i=1

*

*U

i

*U

i

*

+

3

¸

i=1

¸

˜

U

T

i

*

*

L

i

−

3

¸

j=1

N

T

ji

j

−

˜

T

i

*

*

(M

i

)

=

*

*

+

3

¸

i=1

˜

U

T

i

*L

i

*

−

˜

T

i

*M

i

*

+

3

¸

i=1

¸

*

*U

i

+

3

¸

j=1

˜

U

T

j

*L

j

*U

i

+

˜

T

j

N

ji

*U

i

*

−

3

¸

i=1

¸

3

¸

j=1

˜

U

T

j

N

T

ij

*

i

*

(22b)

Here we see that the derivatives *U

i

/* and *

i

/* of the implicit functions can be eliminated

by choosing

˜

U

i

and

˜

i

such that

3

¸

j=1

(K

T

ji

˜

U

j

− N

T

ji

˜

j

) =

*

*U

i

and

3

¸

j=1

N

ij

˜

U

j

=0 (23)

where we introduced K

ij

= − *L

i

/*U

j

. This problem is the adjoint of Equation (19) and

˜

U

and

˜

are the corresponding Lagrange multipliers.

In deriving Equation (22b) we implicitly assumed that N

ij

is independent of , i.e. that the

constraint R

i

(u, ) is a linear function. If this is not true then the gradient */* computed from

Equation (22b) is not exact, which may lead to poor performance of the optimization algorithm

if the constraints are strongly non-linear. In order to avoid such problems it is necessary to

include the non-linear parts of the constraint vector M into L and move the corresponding

Lagrange multipliers from into U. While this is beyond the scope of the divergence form

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

984 L. H. OLESEN, F. OKKELS AND H. BRUUS

discussed in Section 3.1, it is certainly possible to deal with such problems in FEMLAB.

Also the sensitivity analysis above remains valid since it relies only on the basic form of

Equation (19) for the discretized problem.

3.5. Implementation aspects

We end this section by discussing a few issues on the implementation of topology optimization

using FEMLAB.

Firstly there is the question of how to represent the design variable ¸(r). The governing

equations as expressed by

i

and F

i

in Equation (10a) depend not only on the solution u

but also on ¸, and the implementation should allow for this dependence in an efﬁcient way.

Here our simple and straightforward approach is to include ¸ as an extra dependent variable

on equal footing with the velocity ﬁeld and pressure, i.e. we append it to the velocity–pressure

vector, redeﬁning u as

u ≡[v

1

, v

2

, p, ¸] (24)

This was already anticipated when we denoted the basis set for ¸ by {m

4,n

(r)}. By making ¸

available as a ﬁeld variable we can take full advantage of all the symbolic differentiation, matrix,

and postprocessing tools for analysing and displaying the material distribution. Appending ¸ to

the list of dependent variables we are required to deﬁne a fourth governing equation. However,

since we are never actually going to solve this equation, but rather update ¸ based on the

MMA step, we simply deﬁne

4

≡

¸

0

0

, F

4

≡0, G

4

≡0, R

4

≡0. (25)

It is crucial then that the ﬁnite element solver allows different parts of the problem to be

solved in a decoupled manner, i.e. it must be possible to solve Equations (10a)–(10c) for u

i

for i =1, 2, 3 while keeping u

4

, i.e. ¸, ﬁxed.

In FEMLAB the non-linear problem Equation (19) is solved using damped Newton itera-

tions [17]. Therefore, the matrices K

ij

= − *L

i

/*U

j

and N

ij

= − *M

i

/*U

j

appearing in the

adjoint problem Equation (23) are computed automatically as part of the solution process and

can be obtained directly as MATLAB sparse matrices. They are given by

K

ij,nm

= −

m

i, n

¸

*F

i

*u

j

m

j, m

+

*F

i

*∇u

j

· ∇m

j, m

¸

+∇m

i, n

·

¸

*

i

*u

j

m

j, m

+

*

i

*∇u

j

· ∇m

j, m

¸

dr −

*

m

i, n

*G

i

*u

j

m

j, m

ds (26)

and

N

ij,nm

= −

*R

i

*u

j

r

i, n

m

j, m

(r

i, n

) (27)

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 985

Regarding the right-hand side vector */*U

i

in Equation (23), notice that for a general objective

as Equation (11), it has the form

*

*u

i, n

=

*A

*u

i

+

*A

*∇u

i

· ∇

m

i, n

dr +

*

*B

*u

i

m

i, n

ds (28)

It is not in the spirit of a high-level ﬁnite element package to program the assembly of this

vector by hand. Instead we employ the built-in assembly subroutine of FEMLAB. We construct a

copy of the original problem sharing the geometry, ﬁnite element mesh, and degree-of-freedom

numbering with the original. Only we replace the original ﬁelds

i

, F

i

, and G

i

with

˜

i

≡

*A

*∇u

i

,

˜

F

i

≡

*A

*u

i

and

˜

G

i

≡

*B

*u

i

(29)

Assembling the right-hand-side vector

˜

L

i

with this deﬁnition yields exactly Equation (28), c.f.

Equation (20). An extra convenience in FEMLAB is that we can rely on the built-in symbolic

differentiation tools to compute the derivatives *A/*u

i

, etc. In order to try out a new objective

for the optimization problem, the user essentially only needs to change the text expressions

deﬁning the quantities A and B.

After solving the adjoint problem Equation (23) for

˜

U

i

and

˜

i

to eliminate *U

i

/*

and *

i

/* for i =1, 2, 3 in Equation (22b) we can evaluate the sensitivity

d

d

[(U, )] =

*

*

+

3

¸

j=1

*L

j

*

T

˜

U

j

−

*M

j

*

T

˜

j

=

˜

L

4

−

3

¸

j=1

(K

T

j4

˜

U

j

− N

T

j4

˜

j

) (30)

where K

i4

= − *L

i

/*, N

i, 4

= − *M

i

/*, and

˜

L

4

=*/* in accordance with U

4

≡. Since

the fourth variable ¸ is treated on equal footing with the other three variables, all expressions

required to compute the matrices K

i, 4

and N

i, 4

come out of the standard linearization of the

problem. This is yet another advantage of including ¸ as an extra dependent variable.

When dealing with a problem with a volume constraint as in Equation (15b), it is necessary

to compute the derivative of the constraint with respect to

*

*¸

n

¸

1

||

¸(r) dr − [

¸

=

1

||

m

n, 4

(r) dr (31)

which can be obtained by assembling

ˆ

L

4

with

ˆ

4

≡0,

ˆ

F

4

≡1, and

ˆ

G

4

≡0. In the appendix we

have included a transcript of the code required to set up and solve the example from Section 4.2

below with FEMLAB. It amounts to 111 lines of code, of which the majority are spent on setting

up the actual Navier–Stokes ﬂow problem. Only a minor part goes to set up the adjoint problem

and perform the sensitivity analysis. Moreover, this part contains almost no reference to the

actual physical problem being solved, and therefore it should apply for any multi-ﬁeld problem

expressed in the divergence form Equations (10a)–(10c) with an objective function of the form

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

986 L. H. OLESEN, F. OKKELS AND H. BRUUS

of Equation (11). The code example employs, but does not include, a MATLAB implementation

of the MMA optimization algorithm [11–13].

3.5.1. Mesh dependence and regularization techniques. It is well known that many topology

optimization problems have trouble with mesh dependence, e.g. in stiffness design of mechanical

structures it often pays to replace a thick beam with two thinner beams for a given amount of

material. As the ﬁnite element mesh is reﬁned, smaller and smaller features can be resolved and

therefore appear in the optimized structure. In that sense the ﬂow problem that we consider

here is atypical because it is generally unfavourable to replace a wide channel with two

narrower channels; hence the proof for the existence of a unique optimal solution with respect

to minimization of the total power dissipation in Reference [6].

The problem with mesh dependence can be overcome by various regularization techniques

based on ﬁltering of either the design variable ¸(r) or the sensitivity d/d¸ [2]. The regular-

ization works by deﬁning a certain length scale r

0

below which any features in ¸(r) or d/d¸

are smeared out by the ﬁlter; in both cases this results in optimized structures with a minimal

feature size ∼ r

0

independent of the mesh reﬁnement. Unfortunately FEMLAB does not come

with such a ﬁlter, and hence its implementation is an issue that has to be dealt with before

our methodology here can be successfully applied to problems that display mesh dependence.

One strategy is to implement the convolution operation of the ﬁlter directly [2]. If the

computational domain is rectangular and discretized by square ﬁnite elements this is both

efﬁcient and fairly easy to program, if not one simply uses a standard ﬁlter from the MATLAB

Image Processing Toolbox. For an unstructured mesh of triangular elements the programming

is more involved and slow in MATLAB due to the need to loop over the design variable nodes

and searching the mesh for neighbouring nodes within the ﬁlter radius. Therefore an explicit

matrix representation of the ﬁlter would often be preferred [18].

Another possible strategy is to solve an artiﬁcial diffusion problem for the design variable ¸(r)

over some period in ‘time’ t =r

2

0

/k where k is the ‘diffusion’ constant. The diffusion equation

could be included into the ﬁelds of Equation (29) that are otherwise unused, and the ‘time’

evolution solved using the built-in timestepper in FEMLAB. This procedure is equivalent to

the action of a ﬁlter with Gaussian kernel of width r

0

, and it conserves the total amount of

material during the ﬁlter action. The same approach could be used to smooth out the sensitivity.

However, because d/d¸

n

is sensitive to the local element size one would need to rescale it

with

m

4,n

(r) dr before application of the ﬁlter—actually this is true for any ﬁlter acting

on d/d¸ whenever the mesh is irregular and

m

4,n

(r) dr not constant for all n.

The major disadvantage of this strategy is that it involves solving a time evolution problem

in each design iteration which could easily turn out to be the most time-consuming step.

Alternatively the timestepping algorithm could be implemented by hand, e.g. deciding on the

Crank–Nicholson algorithm with a ﬁxed stepsize ot t . The mass and stiffness matrices for

the diffusion problem can be obtained from FEMLAB, and the corresponding iteration matrix

need only be factorized once for the given stepsize and could thus be reused in all subsequent

design iterations, making this approach relatively cheap, although more cumbersome than using

the built-in timestepper.

3.5.2. Large-scale problems. For large-scale problems and three-dimensional modelling it is

often necessary to resort to iterative linear solvers because the memory requirements of a

direct matrix factorization becomes prohibitive. In that case the strategy we have outlined

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 987

here of obtaining the K and N matrices directly as sparse matrices in MATLAB and simply

transposing K before the solution of the adjoint problem may not be practical. Alternatively, if

the original physical problem is expressed in divergence form then the FEMLAB representation

of that problem contains the symbolic derivatives of

i

, F

i

, and G

i

appearing in Equation (26).

These ﬁelds can be transposed and set in the auxiliary copy of the original problem such that it

effectively deﬁnes

˜

K

ij

=K

T

ji

, while retaining the deﬁnitions in Equation (29) for the right-hand

side vector

˜

L

i

. Then the adjoint problem Equation (23) can be solved without direct handling

of the matrices in MATLAB, and using the same iterative solver algorithm as would be employed

for the original physical problem. Ultimately we still require an explicit representation of the

matrices K

i, 4

and N

i, 4

to evaluate the sensitivity d/d in Equation (30).

From our point of view the major advantage of using FEMLAB in its present stage of

development for topology optimization is not in solving large scale problems, though, but

rather in the ease of implementation and the ability to handle problems with coupling between

several physical processes.

4. NUMERICAL EXAMPLES

In this section we present our results for topology optimization of Navier–Stokes ﬂow for

two particular model systems that we have studied. These systems have been chosen because

they illustrate the dependence of the solution on the two dimensionless numbers Re and Da,

measuring the importance of the inertia of the ﬂuid and the permeability of the porous medium,

respectively, relative to viscosity. Moreover we discuss the dependence of the solution on the

initial condition for the material distribution.

For simplicity and clarity we have chosen to consider only two-dimensional model systems.

We note that the dimensionality of the problems has no fundamental consequence for the method

and the numerics, but only affects computer memory requirements and the demand for CPU

time. Our 2D examples can therefore be viewed as idealized test cases for our implementation

of topology optimization. Yet, the 2D models are not entirely of academic interest only as they

represent two limits of actual 3D systems. Due to planar process technology many contemporary

lab-on-a-chip systems have a ﬂat geometry with typical channel heights of about 10 m and

widths of 1 mm, i.e. an aspect ratio of 1:100 [14]. One limit is the case where the channel

width is constant and the channel substrate and lid are patterned with a proﬁle that is translation

invariant in the transverse (width) direction. In the limit of inﬁnitely wide channels the 2D-

ﬂow in the plane perpendicular to the width-direction is an exact solution, while it remains

an excellent approximation in a 1:100 aspect ratio channel. This is the model system we have

adopted for the numerical examples in the present work. The other important limit is when

the channel width is not constant, but the channel height is sufﬁciently slowly varying that

the vertical component of the ﬂuid velocity can be neglected. Then writing the Navier–Stokes

equation for the velocity averaged in the vertical (height) direction, the out-of-plane shear

imposed by the channel substrate and lid gives rise to an absorption term −:v. This approach

was studied by Borrvall and Petersson [6], see also the footnote in Section 2.1. Thus, if

one is interested in optimizing the height-averaged ﬂow ﬁeld in a ﬂat channel the 2D model

is sufﬁcient.

When solving the Navier–Stokes ﬂow problem we use the standard direct linear solver

in FEMLAB in the Newton iterations. Typically we have around 6000 elements in the mesh,

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

988 L. H. OLESEN, F. OKKELS AND H. BRUUS

corresponding to 30 000 degrees-of-freedom. The constrained optimization problem is solved

using a MATLAB implementation of the MMA algorithm kindly provided by Svanberg [11, 13],

except that we modiﬁed the code to use the globally convergent scheme described in

Reference [12]. The example script included in the appendix employs only the basic algo-

rithm mmasub, though. The design iterations are stopped when the maximal change in the

design ﬁeld is ¸

(k+1)

−¸

(k)

∞

0.01, at which point we typically have |

(k+1)

−

(k)

| < 10

−5

.

4.1. Example: a channel with reverse ﬂow

Our ﬁrst numerical example deals with the design of a structure that at a particular point inside

a long straight channel can guide the ﬂow in the opposite direction of the applied pressure drop.

The corresponding problem with a prescribed ﬂow rate was ﬁrst suggested and investigated

by Gersborg-Hansen [8]. We elaborate on it here to illustrate the importance of the choice of

permeability for the porous medium.

The computational domain is shown in Figure 1. It consists of a long straight channel of

height and length L=10; the actual design domain, inside which the porous material is

distributed, is limited to the central part of length 5. The boundary conditions prescribe a

pressure drop of p from the inlet (left) to the outlet (right), and no-slip for the ﬂuid on the

channel side walls.

The optimization problem is stated as a minimization of the horizontal ﬂuid velocity at the

point r

∗

at the centre of the channel, i.e. the design objective is

=v

1

(r

∗

) (32)

In terms of the general objective Equation (11) this is obtained with A≡v

1

(r)o(r − r

∗

) and

B≡0. There is no explicit need for a volume constraint because neither of the extreme solutions

of completely ﬁlled or empty can be optimal. When the design domain is completely ﬁlled

with porous material we expect a ﬂat ﬂow proﬁle with magnitude below p/(5:

max

). In the

other extreme case when the channel is completely devoid of porous material the solution is

simply a parabolic Poiseuille proﬁle with maximum

v

0

=

2

8n

p

L

(33)

However, a structure that reverses the ﬂow such that v

1

(r

∗

) becomes negative will be superior

to both these extreme cases in the sense of minimizing .

2.5 5 2.5

p

0

+∆p p

0

r

∗

Figure 1. Computational domain for the reverse ﬂow example. The design domain (grey) has length 5

and height , and the ﬂuid enters and leaves the design domain through leads of length 2.5. The

boundary conditions prescribe a pressure drop of p across the system, and the design objective is

to reverse the ﬂow direction at the point r

∗

at centre of the channel.

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 989

4.1.1. Reverse ﬂow in the Stokes limit, Re =0. We ﬁrst consider the Stokes ﬂow limit of

small p where the inertial term becomes negligible. The problem is then linear and the

solution is characterized by a single dimensionless parameter, namely the Darcy number Da,

Equation (6). We have solved the topology optimization problem for different values of Da.

The initial condition for the material distribution was ¸

(0)

=1, and the parameter q determining

the shape of :(¸) in Equation (4) was set to q =0.1. Anticipating that the structural details

close to r

∗

should be more important than those further away we chose a non-uniform ﬁnite

element mesh with increased resolution around r

∗

.

Figure 2 shows the optimal structures obtained for Da =10

−3

, 10

−4

, 10

−5

and 10

−6

. They

all consist of two barriers deﬁning an S-shaped channel that guides the ﬂuid in the reverse

direction of the applied pressure drop. At Da =10

−3

the two barriers are rather thick but leaky

with almost all the streamlines penetrating them; as the Darcy number is decreased the optimal

structures become thinner and less penetrable. This result can be interpreted as a trade-off

between having either thick barriers or wide channels. Thick barriers are necessary to force the

ﬂuid into the S-turn, while at the same time the open channel should be as wide as possible

in order to minimize the hydraulic resistance and maximize the ﬂuid ﬂow at the prescribed

pressure drop.

Notice that if we had chosen to prescribe the ﬂow rate through the device rather than the

pressure drop, then the optimal solution would have been somewhat different. When the ﬂow

rate is prescribed, it pays to make the gap between the barriers very small and the barriers

very thick in order to force the ﬁxed amount of ﬂuid ﬂow through the narrow contraction. The

optimal structure is therefore one with a very large hydraulic resistance. In Reference [8], this

problem was circumvented by adding a constraint on the maximal power dissipation allowed

at the given ﬂow rate.

Figure 2. Optimized structures (black) and streamlines at 5% intervals for Stokes ﬂow (Re =0) at

Darcy numbers decreasing from 10

−3

to 10

−6

. Only the central part of length 3 of the design

domain is shown. The structures consist of two barriers deﬁning an S-shaped channel that reverses

the ﬂow at the central point r

∗

. As the Darcy number is decreased, the optimized structures become

thinner and less permeable: (a) Da =10

−3

; (b) Da =10

−4

; (c) Da =10

−5

; and (d) Da =10

−6

.

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

990 L. H. OLESEN, F. OKKELS AND H. BRUUS

In order to validate the optimality of the structures computed by the topology optimization

we do as follows: For each of the optimized structures from Figure 2 we freeze the material

distribution and solve the ﬂow problem for a range of Darcy numbers. The resulting family of

curves for v

1

(r

∗

) vs Da is shown in Figure 3 where it is seen that each of the four structures

from Figure 2 do indeed perform better in minimizing v

1

(r

∗

) than the others at the value

of Da for which they are optimized.

For Da 10

−5

the optimal value of v

1

(r

∗

) tends to saturate because the thin barriers are

then almost completely impermeable and the open channel cannot get much wider. In this

limit the thickness of the optimized barrier structures approach the mesh resolution as seen in

Figure 2(d). When the optimal barrier thickness gets below the mesh size we have observed

the appearance of artiﬁcial local optima for the barrier structure. The problem is that the thin

barriers cannot continuously deform into another position without going through an intermediate

structure with barriers that are thicker by at least one mesh element. Depending on the initial

condition, the optimization algorithm can therefore end up with a sub-optimal structure. We

have tried to work around this problem by decreasing the value of q in order to smear out the

solid/void interfaces and thus reduce the cost of going through the intermediate structure. This

did not work out well; the reason may be that the smearing property of a convex :(¸) was

derived for the objective of minimizing the power dissipation subject to a volume constraint. In

the present example we are dealing with a different objective and have no volume constraint.

However, when the barrier structures are resolved with at least a few elements across them the

artiﬁcial local optima tend to be insigniﬁcant. Thus the problem can be avoided by choosing

a sufﬁciently ﬁne mesh, or by adaptively reﬁning the mesh at the solid/void interfaces.

Returning to Figure 3 we notice that as Da increases all the structures perform poorly in

minimizing v

1

(r

∗

), as they all approach v

0

. Extrapolating this trend one might suspect that

the S-turn topology will cease to be optimal somewhere above Da =10

−3

simply because the

Figure 3. Comparing the performance of the structures from Figure 2 optimized

at Da

opt

for different values of Da. The objective v

1

(r

∗

) is normalized with the

velocity in an empty channel, v

0

, c.f. Equation (33).

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 991

Figure 4. Optimized structure (black) and streamlines for Stokes ﬂow at Da =10

−2

; only the central

part of length 3. The design domain is completely ﬁlled with porous material, except immediately

above and below r

∗

where two empty regions emerge. These voids divert the ﬂow away from r

∗

,

resulting in a low velocity v

1

(r

∗

) =0.1v

0

.

porous material becomes too permeable to make reversal of the ﬂow direction possible. We have

tested this hypothesis by performing an optimization at Da =10

−2

, resulting in the structure

shown in Figure 4 where the value of the objective is v

1

(r

∗

) =0.1v

0

. It is seen to display a

different topology from those of Figure 2, with the design domain is almost completely ﬁlled

with porous material blocking the ﬂow through the channel. Only immediately above and below

the point r

∗

we see two empty regions emerging that act to guide the ﬂow away from r

∗

.

Actually, in all four cases from Figure 2, starting from an empty channel the design iterations

initially converge towards a symmetric structure blocking the ﬂow like that in Figure 4. However,

at a certain point in the iterations an asymmetry in the horizontal plane is excited and the

structure quickly changes to the two-barrier S-geometry. Whether the optimization converge to

an S- or an inverted S-turn depends on how the asymmetry is excited from numerical noise

or irregularity in the ﬁnite element mesh; in fact the structure in Figure 2(b) originally came

out as an inverted S but was mirrored by hand before plotting it to facilitate comparison with

the three other structures.

4.1.2. Reverse ﬂow at ﬁnite Reynolds number. We now consider ﬂow at ﬁnite Reynolds number,

characterized by the two dimensionless numbers Re and Da. The geometry and boundary

conditions remain unchanged, for convenience we introduce a non-dimensional pressure drop

˜ p =p ,

2

/n

2

, and ﬁnally we ﬁx the Darcy number at Da =10

−5

. We note from Figure 2

that this Darcy number allows some but not much ﬂuid to penetrate the walls. We have

nevertheless chosen this Darcy number for practical reasons, as the walls are ‘solid’ enough

and a lower value (more ‘solid’ wall) would increase the calculation time.

We have solved the topology optimization problem for different values of ˜ p, always using an

empty channel as initial condition. The results are shown in Figures 5(a)–(c) for ˜ p =0.2, 0.5,

and 1.0 ×10

5

, where only a few streamlines are seen to penetrate the barriers. For comparison

we also consider the ﬂow ﬁeld obtained when the structure optimized for Stokes ﬂow at

Da =10

−5

is frozen and exposed to the three different elevated pressure drops. This is shown

in Figures 5(d)–(f): As the pressure drop is increased, more and more streamlines penetrate the

barriers. Moreover, we ﬁnd a recirculation region emerging behind the second barrier which

reduces the pressure drop over the neck between the barriers.

Returning to Figures 5(a)–(c), we ﬁnd that the structures that have been optimized for the

corresponding pressure drops are generally thicker than that optimized for Stokes ﬂow, which

reduces the number of streamlines penetrating them. Also a beak-like tip grows on the second

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

992 L. H. OLESEN, F. OKKELS AND H. BRUUS

Figure 5. Optimized structures (black) and streamlines for Navier–Stokes ﬂow; only a part of

length 3.25 near the centre of the channel is shown. Panels (a) –(c) to the left show the optimized

structures for different values of the control parameter ˜ p =p,

2

/n

2

. For comparison the ﬂow ﬁeld

when the optimized structure from Figure 2(c) is frozen and exposed to the elevated pressure drops is

shown in panels (d) –(f) to the right. The Reynolds number is deﬁned as Re =,v

max

/n where v

max

is the maximal velocity measured at the inlet; note that for a particular value of ˜ p, the Reynolds

number is not ﬁxed but differs slightly between left and right column: (a) ˜ p =0.2 ×10

5

[Re =23];

(b) ˜ p =0.5 × 10

5

[Re =42]; (c) ˜ p =1.0 × 10

5

[Re =64]; (d) ˜ p =0.2 × 10

5

[Re =26];

(e) ˜ p =0.5 × 10

5

[Re =47]; and (f) ˜ p =1.0 × 10

5

[Re =71].

barrier that acts to bend the ﬂuid stream down. Finally, on the back of the second barrier a

wing- or spoiler-like structure appears that removes the recirculation.

In summary, our ﬁrst example has demonstrated that our implementation of topology opti-

mization works, but that the optimal design and performance may depend strongly on the choice

of the Darcy number. In particular, the zero Da limit solution contains zero thickness and yet

impermeable barriers deﬂecting the ﬂuid. In order to approximate this solution at ﬁnite Da and

on a ﬁnite resolution mesh it is important to choose the Darcy number small enough that even

thin barriers can be almost impermeable, but large enough to avoid difﬁculties with artiﬁcial

local optima in the discretized problem when the barrier thickness decreases below the mesh

resolution.

4.2. Example: a four-terminal device

Our second numerical example deals with minimization of the power dissipation in a four-

terminal device subject to a volume constraint. The problem is found to exhibit a discrete

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 993

Figure 6. Schematic illustration of the four-terminal device. Two inlet and two outlet leads

(white areas) of height and length 2 are attached to the design domain (grey) of height 5

and length L. The ﬂow is characterized by the Reynolds number Re =,v

max

/n, where v

max

is

the maximal velocity at the inlets.

change in optimal topology driven by the inertial term. The four-terminal device is related to

one considered by Borrvall and Petersson for Stokes ﬂow in Reference [6]; the present example

demonstrates that the optimization algorithm has difﬁculties in ﬁnding the optimal topology

when there are two strong candidates for the global optimum.

The computational domain, shown in Figure 6, consists of a rectangular design domain (grey)

to which two inlet and two outlet leads (white) are attached symmetrically. The boundary

conditions prescribe parabolic proﬁles for the ﬂow at the inlets, zero pressure and normal ﬂow

at the outlets, and no-slip on all other external boundaries. Choosing the height of the leads as

our characteristic length scale, we deﬁne the Reynolds number as Re =,v

max

/n, where v

max

is the maximal velocity at the inlets. The Darcy number is ﬁxed at Da =10

−4

to obtain

reasonably small leakage through the porous walls.

The optimization problem is stated as a minimization of the total power dissipation inside

the computational domain, given by Equation (7a), subject to the constraint that at most a

fraction [ =0.4 of the design domain should be without porous material, c.f. Equation (15b).

Figures 7(a) and (b) shows the two optimal structures obtained for Re =20 and 200, respec-

tively, in a geometry with L=3.5. At Re =20 the optimal structure turns out to be a pair of

U-turns connecting the inlets to the outlets on the same side of the design domain, while at

Re =200 the optimal structure is a pair of parallel channels. In order to minimize the power

dissipation at low Re, the channel segments should be as short and as wide as possible, which

favours the U-turns in Figure 7(a). However, as the Reynolds number is increased, the cost

of bending the ﬂuid stream grows. When inertia dominates, larger velocity gradients appear in

the long ‘outer lane’ of the U-turn. This increases the dissipation compared to low Re, where

more ﬂuid ﬂows in the shorter ‘inner lane’. At a certain point it will exceed the dissipation in

the parallel channels solution

0

=

96

9

4 +

L

nv

2

max

(34)

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

994 L. H. OLESEN, F. OKKELS AND H. BRUUS

Figure 7. Optimal structures (black) and streamlines at 10% intervals for the

four-terminal device at Reynolds number Re =20 and 200, respectively, in a

geometry with L=3.5: (a) Re =20; and (b) Re =200.

as estimated from Poiseuille ﬂow in two straight channels, each of length L+4 and height .

This number is independent of inertia due to translation symmetry, and we use

0

as a natural

unit of power dissipation (per unit length in the third dimension) in the following.

Clearly the Reynolds number at which the transition between the two classes of solutions

occurs will depend strongly on the ratio L/. For short lengths L2 the parallel channels

solution is expected to be optimal at all Re, whereas for long lengths L3 the U-turn

solution should be signiﬁcantly better than the parallel channels solution at low Re.

4.2.1. Dependence on the Reynolds number. In the following we investigate more closely the

transition between the U-turns and the parallel channels solution as a function of the Reynolds

number for the particular geometry L=3. The topology optimization problem is solved for

different Re in the range 0–200, using a homogeneous material distribution ¸

(0)

=0.4 as initial

condition. For the parameter q determining the shape of :(¸) in Equation (4) we use a two-step

solution procedure as suggested in Reference [1]. First the problem is solved with q =0.01

in order to obtain a solution with slightly smeared-out solid/void interfaces. Next this material

distribution is used as initial guess for an optimization with q =0.1 which generates fully

discrete solid/void interfaces at the resolution of our ﬁnite element mesh.

Figure 8(a) shows the result for the normalized power dissipation /

0

obtained as a

function of Re. At low Reynolds numbers the optimized solutions correctly come out as

U-turns with a power dissipation that is clearly less than

0

. However, at high Reynolds

numbers Re > 90 the method fails because the optimized solutions continue to come out as

U-turns even though this yields /

0

> 1. For Re 160 the solution jumps from the simple

U-turns to a hybrid structure, as shown in the inset. The full lines in Figure 8(a) show the

result when the material distributions optimized for Re =0, 50, and 180, respectively, are frozen

and the power dissipation evaluated at different Re. It is seen that the optimized solutions,

marked (◦), all fall on or below the full lines which conﬁrms that they are indeed superior to

the other solutions of the U-turn family. This also holds for Re > 90, except for the hybrid

structures at Re 160, that are actually inferior to the U-turns. Moreover, at Re =160 the

optimized solution falls slightly above that optimized at Re =180. This could be an indication

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 995

Figure 8. Power dissipation in structures optimized for different Reynolds numbers; normalized

with the Poiseuille ﬂow result

0

(dashed line): (a) markers (◦) show results when ¸

(0)

=0.4 is used

as initial condition, failing to ﬁnd the optimal solution for Re > 90. Full lines show the performance

of the structures optimized at Re =0, 50, and 180, when evaluated at different Reynolds numbers.

As expected, all points fall on or below the full lines, except the hybrid solutions for Re 160; and

(b) comparison between the two different initial conditions ¸

(0)

=0.4 (◦) and ¸

(0)

=1 (×), showing

the success of the empty channel initial condition in ﬁnding the optimal solution. The crosses (×) fall

slightly below /

0

=1 due to leakage through the porous walls (see the text).

that the hybrid structures are not local optima in design space after all, but rather a very narrow

saddle point that the optimization algorithm has a hard time getting away from.

The difﬁculty is that the two families of solutions, the U-turns and the parallel channels, are

both deep local minima for the power dissipation in design space. Using ¸

(0)

=0.4 as initial

condition, the initial permeability is everywhere very low, such that the porous friction almost

completely dominates the inertia and viscous friction in the ﬂuid. Therefore the iteration path

in design space is biased towards low Reynolds numbers and the U-turns solution.

In order to circumvent this problem we have tried using a completely empty design domain

with ¸

(0)

=1 as initial condition. This should remove the bias towards the U-turns and allow

the optimization algorithm to take inertia into account from iteration one. The result is shown

in Figure 8(b). For Re 80 the solutions are still U-turns, whereas for Re 90 they come out

as parallel channels. Notice that /

0

for the parallel channels solution is actually slightly less

than unity, namely 0.98. This is due to a small amount of ﬂuid seeping through the porous

walls deﬁning the device, which lowers the hydraulic resistance compared to the Poiseuille

ﬂow result derived for solid walls.

**Strictly speaking the initial condition ¸
**

(0)

=1 is not a feasible solution because it violates

the volume constraint that at least a fraction 1 −[ =0.6 of the design domain should be ﬁlled

**The ﬂow in a straight channel of height bounded by two porous walls of thickness can easily be found
**

analytically. At Da =10

−4

the hydraulic resistance of this system is 94% of that for a channel of height

bounded by solid walls, and it approaches this zero Da limit only as

√

Da. When L=3 we therefore expect

a power dissipation /

0

=(3×0.94+4)/7 =0.97 for the parallel channels solution, including the leads. This

is close to the observed 0.98.

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

996 L. H. OLESEN, F. OKKELS AND H. BRUUS

Figure 9. Comparison between structures optimized with the initially non-feasible material distribution

¸

(0)

=1 for different penalty parameters in the MMA optimization algorithm, revealing the difﬁculty in

choosing the condition for ﬁnding the global optimum. Full line: the successful result from Figure 8(b)

with moderate penalty; (+) lower penalty yielding wrong result for 40 < Re < 80; () higher

penalty yielding wrong results for 80 < Re < 190.

with porous material. However, the MMA optimization algorithm penalizes this and reaches

a feasible solution after a few iterations. This is controlled by choosing a penalty parameter.

If the penalty for violating the constraint is small, the material is added slowly and only where

it does not disturb the ﬂow much. If the penalty is large, the material is added quickly and

almost homogeneously until the constraint is satisﬁed. The successful result from Figure 8(b)

was obtained with a moderate penalty. In Figure 9 this is compared with results for smaller

and larger penalty parameters, respectively. The ﬁgure shows that with the small penalization,

the solution jumps to the parallel channels already at Re =50 which is not optimal. For the

large penalization, the solution does not jump until Re 190. Also we observe hybrid structures

similar to those in Figure 8(a) for Re 130. We have thus not full control over the convergence

towards the global optimum.

4.2.2. Discussion of problems with local optima. Further insight into the problem of local versus

global optima is gained by inspecting the ﬂow ﬁeld in the initial material distribution ¸

(0)

. This

is shown in Figure 10 for the Stokes ﬂow limit, Re =0. The streamlines are drawn as 10%

contours of the streamfunction, and Figure 10(a) shows that for ¸

(0)

=0.4 the streamline density

is largest between the two leads on the same side of the design domain. Based on the sensitivity

d/d¸ the optimization algorithm therefore decides to remove material from these strong-ﬂow

regions in order to reduce the porous friction. The iteration path in design space is therefore

biased towards the U-turn solution. This remains true even at ﬁnite Reynolds numbers as long

as the porous friction initially dominates inertia.

Figure 10(b) shows that when ¸

(0)

=1 the streamline density is largest between the leads

on the opposite side of the design domain. Because the volume constraint is violated the

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 997

Figure 10. Flow distributions at Re =0, L=3 and q =0.01: (a) initial design ﬁeld ¸

(0)

=0.4;

(b) initial design ﬁeld ¸

(0)

=0.4; and (c) the optimal design ﬁeld ¸

(∗)

obtained at Da =10

−2

.

optimization algorithm has to place material somewhere, which it does in the weak-ﬂow regions.

The solution is therefore biased towards the parallel channels. Indeed if the penalty is chosen

very small, the optimized solution comes out as parallel channels even for Stokes ﬂow at

L=3, which is far from optimal. When the penalty is larger and the material is added faster,

we move away from this adiabatic solution and closer to the situation for ¸

(0)

=0.4.

The additional complexity associated with making a proper choice of the penalty parameter

is somewhat inconvenient. We have therefore attempted to construct a more convex problem by

increasing the initial permeability. This can be done either by increasing the Darcy number, or

by decreasing the parameter q, c.f. Equation (4). Figure 10(c) shows the optimal solution ¸

(∗)

obtained for Da =10

−2

and q =0.01 at Re =0. At this level the problem is convex because

the solution is independent of the initial condition. Using this material distribution as initial

guess and gradually decreasing the permeability to Da =10

−4

and q =0.1 we correctly end

up in the U-turn solution. However, it is evident from Figure 10(c) that ¸

(∗)

has a fair amount

of parallel channels nature. Using the same procedure of gradually decreasing the permeability

at higher Reynolds numbers therefore result in a transition to the parallel channels solution

already for Re 30, which is not optimal. Moreover, when the Reynolds number is increased

and the inertia starts to play in, the system tends to loose convexity even at the initial high

permeability.

In summary, the topology optimization has difﬁculties in ﬁnding the global optimum for

the problem. There are two strong candidates for the optimal structure, and the solution found

is sensitive to the initial condition for the material distribution. Using an empty channel as

initial condition, the method is able to ﬁnd the correct solution for all Reynolds numbers.

However, this successful result depends on a particular choice of the penalty parameter in the

MMA algorithm. By using a high initial permeability of the porous medium, it is possible to

convexify the problem at low Reynolds numbers, but continuation of this solution to the desired

low permeability does not generally lead to the global minimum of the non-convex problem.

In the original paper Reference [6] it was argued that in Stokes ﬂow the true optimal design

should be rather insensitive to the choice of the Darcy number, although the dissipated power

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

998 L. H. OLESEN, F. OKKELS AND H. BRUUS

may deviate quite a lot from the zero Da limit. In our work we have observed that the

actual solution found by the topology optimization may depend a great deal on the choice of

the Darcy number, whereas the dissipated power should approach the zero Da limit roughly

as

√

Da.

5. CONCLUSION

Based on the work of Borrvall and Petersson [6] we have extended the topology optimization

of ﬂuid networks to cover the full incompressible Navier–Stokes equations in steady-state. Our

implementation of the method is based on the commercial ﬁnite element package FEMLAB,

which reduces the programming effort required to a minimum. Formulating the problem in

terms of a general integral-type objective function and expressing the governing equations

in divergence form makes the implementation very compact and transparent. Moreover, the

code for performing the sensitivity analysis should remain almost the same for any problem

expressed in this way, whereas that required for describing the physical problem of course

changes. Topology optimization of multi-ﬁeld problems can therefore be dealt with almost as

easy as a single realization of the underlying physical problem.

We would like to mention that our methodology is not as such restricted to the (large) class

of physical problems that can be expressed in divergence form. FEMLAB also allows problems

to be stated directly in weak form, e.g. for systems with dynamics at the boundaries. This

does in fact not invalidate the sensitivity analysis worked out in Section 3.4, since this analysis

only relies on the basic structure of the discretized non-linear problem and the availability of

the Jacobian matrix. It is therefore possible to apply our methodology to even larger classes

of physical problems than the ones comprised by the divergence form.

Our implementation of topology optimization has been tested on two ﬂuidics examples in 2D,

both illustrating the inﬂuence of different quantities and conditions on the efﬁciency of the

optimization method.

The ﬁrst example, a channel with reversed ﬂow, illustrates the inﬂuence of the Reynolds

number Re and the Darcy number Da on the solutions. We have shown that the choice of Da

has a strong impact on the solution when the structure contains barriers to deﬂect the ﬂuid

stream.

The second example, minimization of the power dissipation in a four-terminal device, reveals

the problems of determining the global minimum when two strong minima are competing. This

problem is highly non-convex, and we have shown that the solution depends on the initial

condition. For an initial homogeneous material distribution, the porous friction dominates and

the solution does not come out as the global optimum in all cases. Using an empty channel

as the initial state, inertia plays a role from the beginning, and better results can be obtained.

However, this initial condition in fact violates the volume constraint, and the part of the

optimization routine correcting this depends on a penalty factor. Unfortunately, the particular

value chosen for this factor strongly inﬂuences the results. Increasing the Darcy number makes

the problem more convex, but continuation from large to small Da, i.e. from high to low

permeability of the porous material, does not generally end up in the global optimum.

In conclusion, we have shown that our implementation of topology optimization is a useful

tool for designing ﬂuidic devices.

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 999

APPENDIX

% FEMLAB CODE FOR THE 4-TERMINAL DEVICE EXAMPLE OF SECTION 4.2

clear fem femadj

% DEFINE REYNOLDS NUMBER, DARCY NUMBER, LENGTH OF DESIGN DOMAIN, AND VOLUME FRACTION

Re = 50;

Da = 1e-4;

L0 = 3.0;

beta = 0.4;

% DEFINE GEOMETRY, MESH, AND SUBDOMAIN/BOUNDARY GROUPS [SEE FIGURE 6]

fem.geom = rect2(0,L0,0,5) + rect2(-2,0,1,2) + rect2(-2,0,3,4) + rect2(L0,L0+2,1,2) ...

+ rect2(L0,L0+2,3,4);

fem.mesh = meshinit(fem,’Hmaxsub’,[3 0.125]);

% subdomain groups 1:design domain 2:inlet/outlet leads

fem.equ.ind = {[3] [1 2 4 5]};

% boundary groups 1:walls 2:inlets 3:outlets 4:interior

fem.bnd.ind = {[2:3 5:8 10 12:14 16:18 20:22] [4 23] [1 24] [9 11 15 19]};

% DEFINE SPACE CO-ORDINATES, DEPENDENT VARIABLES, AND SHAPE FUNCTIONS

fem.sdim = {’x’ ’y’};

fem.dim = {’u’ ’v’ ’p’ ’gamma’};

fem.shape = [2 2 1 1];

% DEFINE CONSTANTS

fem.const.rho = 1;

fem.const.eta = 1;

fem.const.umax = Re;

fem.const.alphamin = 0;

fem.const.alphamax = 1/Da;

fem.const.q = 0.1;

Phi0 = 96*fem.const.eta*(L0+4)*fem.const.umaxˆ2/9;

% DEFINE EXPRESSIONS ON SUBDOMAIN AND BOUNDARY GROUPS

fem.equ.expr = {’A’ ’eta*(2*ux*ux+2*vy*vy+(uy+vx)*(uy+vx))+alpha*(u*u+v*v)’ ...

’alpha’ {’alphamin+(alphamax-alphamin)*q*(1-gamma)/(q+gamma)’ ’0’}};

fem.bnd.expr = {’B’ ’0’};

% DEFINE GOVERNING EQUATIONS AND INITIAL CONDITIONS [SEE EQUATIONS (8) AND (9)]

fem.form = ’general’;

fem.equ.shape = {[1:4] [1:3]}; % only define gamma on subdomain group 1

fem.equ.ga = {{{’-p+2*eta*ux’ ’eta*(uy+vx)’} {’eta*(uy+vx)’ ’-p+2*eta*vy’} {0 0} {0 0}}};

fem.equ.f = {{’rho*(u*ux+v*uy)+alpha*u’ ’rho*(u*vx+v*vy)+alpha*v’ ’ux+vy’ 1}};

fem.equ.init = {{0 0 0 beta}};

% DEFINE BOUNDARY CONDITIONS

fem.bnd.shape = {[1:3]}; % do not define gamma on any boundaries

fem.bnd.r = {{’u’ ’v’ 0 0} ... % walls: no-slip

{’u*nx+4*umax*s*(1-s)’ ’v’ 0 0} ... % inlets: parabolic profile

{0 ’v’ 0 0} ... % outlets: normal flow

{0 0 0 0}}; % interior: nothing

fem.bnd.g = {{0 0 0 0}}; % zero prescribed external forces everywhere

% PERFORM LINEARIZATION, DEGREE-OF-FREEDOM ASSIGNMENT, AND ASSEMBLE INITIAL CONDITION

fem = femdiff(fem);

fem.xmesh = meshextend(fem);

fem.sol = asseminit(fem);

% DEFINE STRUCTURE FOR COMPUTING RIGHT-HAND-SIDE IN ADJOINT PROBLEM [SEE EQUATION (29)]

femadj = fem;

femadj.equ.ga = {{{’diff(A,ux)’ ’diff(A,uy)’} {’diff(A,vx)’ ’diff(A,vy)’} ...

{’diff(A,px)’ ’diff(A,py)’} {’diff(A,gammax)’ ’diff(A,gammay)’}}};

femadj.equ.f = {{’diff(A,u)’ ’diff(A,v)’ ’diff(A,p)’ ’diff(A,gamma)’}};

femadj.bnd.g = {{’diff(B,u)’ ’diff(B,v)’ ’diff(B,p)’ ’diff(B,gamma)’}};

femadj.xmesh = meshextend(femadj);

% GET INDICES OF DESIGN VARIABLE IN THE GLOBAL SOLUTION VECTOR (fem.sol.u)

i4 = find(asseminit(fem,’Init’,{’gamma’ 1},’Out’,’U’));

% COMPUTE VOLUME BELOW DESIGN VARIABLE BASIS FUNCTIONS

L = assemble(fem,’Out’,{’L’});

Vgamma = L(i4);

Vdomain = sum(Vgamma);

% GET INDICES OF VELOCITY-PRESSURE VARIABLES

i123 = find(asseminit(fem,’Init’,{’u’ 1 ’v’ 1 ’p’ 1},’Out’,’U’));

% DEFINE VARIABLES AND PARAMETERS FOR MMA OPTIMIZATION ALGORITHM [SEE REFERENCES [11,12,13]]

a0 = 1;

a = 0;

c = 20;

d = 0;

xmin = 0;

xmax = 1;

xold = fem.sol.u(i4);

xolder = xold;

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

1000 L. H. OLESEN, F. OKKELS AND H. BRUUS

low = 0;

upp = 1;

% DESIGN LOOP FOR THE ACTUAL TOPOLOGY OPTIMIZATION

for iter = 1:100

% SOLVE NAVIER-STOKES FLOW PROBLEM TO UPDATE VELOCITY AND PRESSURE

fem.sol = femnlin(fem,’Solcomp’,{’u’ ’v’ ’p’},’U’,fem.sol.u);

% SOLVE ADJOINT PROBLEM FOR LAGRANGE MULTIPLIERS

[K N] = assemble(fem,’Out’,{’K’ ’N’},’U’,fem.sol.u);

[L M] = assemble(femadj,’Out’,{’L’ ’M’},’U’,fem.sol.u);

femadj.sol = femlin(’In’,{’K’ K(i123,i123)’ ’L’ L(i123) ’M’ zeros(size(M)) ’N’ N(:,i123)});

% SENSITIVITY ANALYSIS

gamma = fem.sol.u(i4);

Phi = postint(fem,’A’,’Edim’,2) + postint(fem,’B’,’Edim’,1);

dPhidgamma = L(i4) - K(i123,i4)’*femadj.sol.u;

% PERFORM MMA STEP TO UPDATE DESIGN FIELD

x = gamma;

f = Phi/Phi0; g = gamma’*Vgamma/Vdomain - beta;

dfdx = dPhidgamma/Phi0; dgdx = Vgamma’/Vdomain;

d2fdx2 = zeros(size(gamma)); d2gdx2 = zeros(size(gamma’));

[xnew,y,z,lambda,ksi,eta,mu,zeta,s,low,upp] = mmasub(1,length(gamma),iter, ...

x,xmin,xmax,xold,xolder,f,dfdx,d2fdx2,g,dgdx,d2gdx2,low,upp,a0,a,c,d);

xolder = xold; xold = x; gamma = xnew;

% TEST CONVERGENCE

if iter >= 100 | max(abs(gamma-xold)) < 0.01

break

end

% UPDATE DESIGN VARIABLE

u0 = fem.sol.u; u0(i4) = gamma;

fem.sol = femsol(u0);

% DISPLAY RESULTS FOR EACH ITERATION STEP

disp(sprintf(’Iter.:%3d Obj.: %8.4f Vol.: %6.3f Change: %6.3f’, ...

iter,f,xold’*Vgamma,max(abs(xnew-xold))))

postplot(fem,’arrowdata’,{’u’ ’v’},’tridata’,’gamma’,’trimap’,’gray’)

axis equal; shg; pause(0.1)

end

ACKNOWLEDGEMENTS

We are grateful to Ole Sigmund and Allan Gersborg-Hansen for illuminating discussions on the

topology optimization method, and to Krister Svanberg for providing us with the MATLAB code

for the MMA algorithm. This work is partly supported by The Danish Technical Research Council,

Grant No. 26-03-0037.

REFERENCES

1. Bendsøe MP, Kikuchi N. Generating optimal topologies in structural design using a homogenization method.

Computer Methods in Applied Mechanics and Engineering 1988; 71(2):197–224.

2. Bendsøe MP, Sigmund O. Topology Optimization-Theory, Methods and Applications. Springer: Berlin, 2003.

3. Eschenauer HA, Olhoff N. Topology optimization of continuum structures: a review. Applied Mechanics

Reviews 2001; 54(4):331–389.

4. Jensen JS. Phononic band gaps and vibrations in one- and two-dimensional mass-spring structures. Journal

of Sound and Vibration 2003; 266(5):1053–1078.

5. Jensen JS, Sigmund O. Systematic design of photonic crystal structures using topology optimization: low-loss

waveguide bends. Applied Physics Letters 2004; 84(12):2022–2024.

6. Borrvall T, Petersson J. Topology optimization of ﬂuids in Stokes ﬂow. International Journal for Numerical

Methods in Fluids 2003; 41:77–107.

7. Sigmund O, Gersborg-Hansen A, Haber RB. Topology optimization for multiphysics problems: a future

FEMLAB application? In Proceedings for the Nordic Matlab Conference 2003, Gregersen L (ed.). Comsol

A/S: Søborg, Denmark, 2003; 237–242.

8. Gersborg-Hansen A. Topology optimization of incompressible Newtonian ﬂows at moderate Reynolds numbers.

M.Sc. Thesis, Technical University of Denmark, Department of Mechanical Engineering, Lyngby, 2003.

9. Gersborg-Hansen A, Sigmund O, Haber RB. Topology optimization of channel ﬂow problems. Structural

and Multidisciplinary Optimization 2005. (Published online: DOI 10.1007/s00158-004-0508-7.)

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

TOPOLOGY OPTIMIZATION OF STEADY-STATE NAVIER–STOKES FLOW 1001

10. Okkels F, Olesen LH, Bruus H. Applications of topology optimization in the design of micro- and nanoﬂuidic

systems. In Proceedings of Nanotech, vol. 1. Anaheim: CA, U.S.A., 2005; 575–578.

11. Svanberg K. The method of moving asymptotes—a new method for structural optimization. International

Journal for Numerical Methods in Engineering 1987; 24:359–373.

12. Svanberg K. A class of globally convergent optimization methods based on conservative convex separable

approximations. SIAM Journal on Optimization 2002; 12(2):555–573.

13. A MATLAB implementation, mmasub, of the MMA optimization algorithm [11] can be obtained (free of

charge for academic purposes) from Krister Svanberg, KTH, Sweden. E-mail: krille@math.kth.se

14. Geschke O, Klank H, Telleman P (eds). Microsystem Engineering of Lab-on-a-chip Devices. Wiley-VCH

Verlag: Weinheim, 2004.

15. Landau LD, Lifshitz EM. Course of Theoretical Physics: Fluid Mechanics, vol. 6 (2nd edn). Butterworth

and Heinemann: Oxford, 2000.

16. Michaleris P, Tortorelli DA, Vidal CA. Tangent operators and design sensitivity formulations for transient

non-linear coupled problems with applications in elastoplasticity. International Journal for Numerical Methods

in Engineering 1994; 37(14):2471–2500.

17. FEMLAB Reference Manual. COMSOL AB: Stockholm, 2004.

18. A simple implementation of a cone ﬁlter for dealing with mesh dependence problems on an unstructured

triangular mesh in two dimensions with linear Lagrange elements for the design variable can be downloaded

from http://www.mic.dtu.dk/research/MIFTS

Copyright ᭧ 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 65:975–1001

178 Paper published in Int. J. Numer. Meth. Eng.

Bibliography

[1] A. Manz, N. Graber, and H. M. Widmer, ‘Miniaturized total chemical analysis

systems. A novel concept for chemical sensing’, Sens. Actuator B-Chem. 1, 244

(1990).

[2] D. R. Reyes, D. Iossiﬁdis, P. A. Auroux, and A. Manz, ‘Micro total analy-

sis systems. 1. Introduction, theory, and technology’, Anal. Chem. 74, 2623

(2002).

[3] H. A. Stone, A. D. Stroock, and A. Ajdari, ‘Engineering ﬂows in small devices:

Microﬂuidics toward a lab-on-a-chip’, Annu. Rev. Fluid Mech. 36, 381 (2004).

[4] T. M. Squires and S. R. Quake, ‘Microﬂuidics: Fluid physics at the nanoliter

scale’, Rev. Mod. Phys. 77, 977 (2005).

[5] M. A. Unger, H.-P. Chou, T. Thorsen, A. Scherer, and S. R. Quake, ‘Monolithic

microfabricated valves and pumps by multilayer soft lithography’, Science 288,

113 (2000).

[6] D. J. Laser and J. G. Santiago, ‘A review of micropumps’, J. Micromech. Mi-

croeng. 14, R35 (2004).

[7] S. H. Yao and J. G. Santiago, ‘Porous glass electroosmotic pumps: theory’, J.

Colloid Interface Sci. 268, 133 (2003).

[8] S. H. Yao, D. E. Hertzog, S. L. Zeng, J. C. Mikkelsen, and J. G. Santiago,

‘Porous glass electroosmotic pumps: design and experiments’, J. Colloid In-

terface Sci. 268, 143 (2003).

[9] S. L. Zeng, C.-H. Chen, J. C. Mikkelsen, and J. G. Santiago, ‘Fabrication and

characterization of electroosmotic micropumps’, Sens. Actuator B-Chem. 79,

107 (2001).

[10] D. S. Reichmuth, G. S. Chirica, and B. J. Kirby, ‘Increasing the performance

of high-pressure, high-eﬃciency electrokinetic micropumps using zwitterionic

solute additives’, Sens. Actuator B-Chem. 92, 37 (2003).

[11] A. Brask, J. P. Kutter, and H. Bruus, ‘Long-term stable electroosmotic pump

with ion exchange membranes’, Lab Chip 5, 730 (2005).

[12] A. Brask, ‘Electroosmotic micropumps’, PhD thesis, MIC – Dept. of Micro

and Nanotechnology, DTU (2005). http://www.mic.dtu.dk/mifts

[13] Y. Takamura, H. Onoda, H. Inokuchi, S. Adachi, A. Oki, and Y. Horiike,

‘Low-voltage electroosmosis pump for stand-alone microﬂuidics devices’, Elec-

trophoresis 24, 185 (2003).

180 BIBLIOGRAPHY

[14] A. Ajdari, ‘Pumping liquids using asymmetric electrode arrays’, Phys. Rev. E

61, R45 (2000).

[15] V. Studer, A. P´epin, Y. Chen, and A. Ajdari, ‘An integrated AC electrokinetic

pump in a microﬂuidic loop for fast and tunable ﬂow control’, Analyst 129,

944 (2004).

[16] A. Bard and L. Faulkner, Electrochemical Methods (Wiley, New York, 1980).

[17] R. J. Hunter, Foundations of Colloid Science (Oxford University Press, New

York, 1991).

[18] V. N. Shilov and S. S. Dukhin, ‘Theory of low-frequency dispersion of dielectric

permittivity in suspensions of spherical colloidal particles due to double-layer

polarization’, Colloid J. USSR 32, 245 (1970).

[19] N. I. Gamayunov, V. A. Murtsovkin, and A. S. Dukhin, ‘Pair interaction of

particles in electric-ﬁeld. 1. Features of hydrodynamic interaction of polarized

particles’, Colloid J. USSR 48, 197 (1986).

[20] N. I. Gamayunov and V. A. Murtsovkin, ‘Motion of disperse particles in a

uniform alternating electric ﬁeld’, Colloid J. USSR 49, 543 (1987).

[21] T. M. Squires and M. Z. Bazant, ‘Breaking symmetries in induced-charge

electro-osmosis’, J. Fluid Mech. 560, 65 (2006).

[22] A. Ramos, H. Morgan, N. G. Green, and A. Castellanos, ‘Ac electrokinetics:

a review of forces in microelectrode structures’, J. Phys. D Appl. Phys. 31,

2338 (1998).

[23] A. Gonz´alez, A. Ramos, N. G. Green, A. Castellanos, and H. Morgan, ‘Fluid

ﬂow induced by nonuniform ac electric ﬁelds in electrolytes on microelectrodes.

II. A linear double-layer analysis’, Phys. Rev. E 61, 4019 (2000).

[24] N. G. Green, A. Ramos, A. Gonz´alez, H. Morgan, and A. Castellanos, ‘Fluid

ﬂow induced by nonuniform ac electric ﬁelds in electrolytes on microelectrodes.

I. Experimental measurements’, Phys. Rev. E 61, 4011 (2000).

[25] A. B. D. Brown, C. G. Smith, and A. R. Rennie, ‘Pumping of water with ac

electric ﬁelds applied to asymmetric pairs of microelectrodes’, Phys. Rev. E

63, 016305 (2000).

[26] V. Studer, A. P´epin, C. Y, and A. Ajdari, ‘Fabrication of microﬂuidic devices

for AC electrokinetic ﬂuid pumping’, Microelectron. Eng. 61, 915 (2002).

[27] S. Debesset, C. J. Hayden, C. Dalton, J. C. T. Eijkel, and A. Manz, ‘A circular

AC electroosmotic micropump for chromatographic applications’, in ‘Proc.

MicroTAS 2002’, 655 – 657 (Kluwer, 2002).

[28] M. Mpholo, C. G. Smith, and A. B. D. Brown, ‘Low voltage plug ﬂow pumping

using anisotropic electrode arrays’, Sens. Actuator B-Chem. 92, 262 (2003).

[29] S. Debesset, C. J. Hayden, C. Dalton, J. C. T. Eijkel, and A. Manz, ‘An

AC electroosmotic micropump for circular chromatographic applications’, Lab

Chip 4, 396 (2004).

BIBLIOGRAPHY 181

[30] P. Garcia, A. Ramos, N. G. Green, and H. Morgan, ‘Experiments on AC

electrokinetic pumping of liquids using arrays of microelectrodes’, in ‘Proc.

IEEE Int. Conf. Dielectric Liquids’, 181 – 184 (IEEE, 2005).

[31] P. Garcia, A. Ramos, N. G. Green, and H. Morgan, ‘Experiments on AC

electrokinetic pumping of liquids using arrays of microelectrodes’, IEEE Trans.

Dielect. El. In. 13, 670 (2006).

[32] M. Gregersen, ‘AC asymmetric electrode micropumps’, MSc thesis, MIC –

Dept. of Micro and Nanotechnology, DTU (2005).

http://www.mic.dtu.dk/mifts

[33] A. Ramos, A. Gonz´alez, A. Castellanos, N. G. Green, and H. Morgan, ‘Pump-

ing of liquids with ac voltages applied to asymmetric pairs of microelectrodes’,

Phys. Rev. E 67, 056302 (2003).

[34] A. Ramos, H. Morgan, N. G. Green, A. Gonzalez, and A. Castellanos, ‘Pump-

ing of liquids with traveling-wave electroosmosis’, J. Appl. Phys. 97, 084906

(2005).

[35] D. Lastochkin, R. H. Zhou, P. Wang, Y. X. Ben, and H. C. Chang, ‘Electroki-

netic micropump and micromixer design based on ac faradaic polarization’,

J. Appl. Phys. 96, 1730 (2004).

[36] T. M. Squires and M. Z. Bazant, ‘Induced-charge electro-osmosis’,

J. Fluid Mech. 509, 217 (2004).

[37] M. Z. Bazant, K. Thornton, and A. Ajdari, ‘Diﬀuse-charge dynamics in elec-

trochemical systems’, Phys. Rev. E 70, 021506 (2004).

[38] K. T. Chu, ‘Asymptotic analysis of extreme electrochemical transport’, PhD

thesis, Dept. Mathematics, Massachusetts Institute of Technology (2005).

http://www.princeton.edu/∼ktchu

[39] K. T. Chu and M. Z. Bazant, ‘Nonlinear electrochemical relaxation around

conductors’, Phys. Rev. E 74, 011501 (2006).

[40] I. Rubinstein and L. Shtilman, ‘Voltage against current curves of cation-

exchange membranes’, J. Chem. Soc. Faraday Trans. 2 75, 231 (1979).

[41] S. S. Dukhin, ‘Electrokinetic phenomena of the 2nd kind and their applica-

tions’, Adv. Colloid Interface Sci. 35, 173 (1991).

[42] I. Rubinstein and B. Zaltzman, ‘Electro-osmotic slip of the second kind and

instability in concentration polarization at electrodialysis membranes’, Math.

Mod. Meth. Appl. S. 11, 263 (2001).

[43] I. Rubinstein, B. Zaltzman, and I. Lerman, ‘Electroconvective instabil-

ity in concentration polarization and nonequilibrium electro-osmotic slip’,

Phys. Rev. E 72, 011505 (2005).

[44] L. H. Olesen, H. Bruus, and A. Ajdari, ‘AC electrokinetic micropumps: The

eﬀect of geometrical conﬁnement, Faradaic current injection, and nonlinear

surface capacitance’, Phys. Rev. E 73, 056313 (2006).

[45] M. Z. Bazant and Y. Ben, ‘Theoretical prediction of fast 3D AC electro-osmotic

pumps’, Lab Chip 6, 1455 (2006).

182 BIBLIOGRAPHY

[46] L. H. Olesen, F. Okkels, and H. Bruus, ‘A high-level programming-language

implementation of topology optimization applied to steady-state Navier-Stokes

ﬂow’, Int. J. Num. Meth. Eng. 65, 975 (2006).

[47] COMSOL Multiphysics 3.2, the general purpose ﬁnite element software for-

merly known as FEMLAB, by COMSOL A/B, Sweeden.

[48] N. A. Mortensen, L. H. Olesen, L. Belmon, and H. Bruus, ‘Electro-

hydrodynamics of binary electrolytes driven by modulated surface potentials’,

Phys. Rev. E 71, 056306 (2005).

[49] N. A. Mortensen, L. H. Olesen, and H. Bruus, ‘Transport coeﬃcients for elec-

trolytes in arbitrarily shaped nano and micro-ﬂuidic channels’, New J. Phys.

8, 37 (2006).

[50] L. H. Olesen, F. Okkels, and H. Bruus, ‘Topology optimization of Navier–

Stokes ﬂow in microﬂuidics’, in ‘Proc. ECCOMAS 2004’, volume 2, 224 (2004).

[51] F. Okkels, L. H. Olesen, and H. Bruus, ‘Applications of topology optimization

in the design of micro- and nanoﬂuidic devices’, in ‘Proc. Nanotech 2005’,

volume 1, 575–578 (2005).

[52] L. H. Olesen, H. Bruus, and A. Ajdari, ‘AC electrokinetic micropumps: The

eﬀect of geometrical conﬁnement, Faradaic current injection, and nonlinear

surface capacitance’, in ‘Proc. MicroTAS 2005’, volume 1, 560–562 (2005).

[53] R. F. Probstein, PhysicoChemical Hydrodynamics, an introduction (John Wi-

ley and Sons, New-York, 1994).

[54] J. Newman, Electrochemical systems (Prentice-Hall, Inc., Eaglewood Cliﬀs,

New Jersey, 1991).

[55] I. Borukhov, D. Andelman, and H. Orland, ‘Steric eﬀects in electrolytes: A

modiﬁed Poisson–Boltzmann equation’, Phys. Rev. Lett. 79, 435 (1997).

[56] M. S. Kilic, M. Z. Bazant, and A. Ajdari, ‘Steric eﬀects in the dynamics of

electrolytes at large applied voltages: I. Double-layer charging’, Phys. Rev. E

75, 021502 (2007).

[57] M. S. Kilic, M. Z. Bazant, and A. Ajdari, ‘Steric eﬀects in the dynamics

of electrolytes at large applied voltages: II. Modiﬁed Poisson–Nernst–Planck

equations’, Phys. Rev. E 75, 021503 (2007).

[58] S. S. Dukhin, ‘Nonequilibrium electric surface phenomena’, Adv. Colloid In-

terface Sci. 44, 1 (1993).

[59] M. V. Dyke, Perturbation methods in ﬂuid mechanics (The Parabolic Press,

Stanford, 1975).

[60] W. H. Smyrl and J. Newman, ‘Double layer structure at limiting current’,

Trans. Faraday Soc. 63, 207 (1967).

[61] E. J. Hinch, J. D. Sherwood, W. C. Chew, and P. N. Sen, ‘Dielectric response

of a dilute suspension of spheres with thin double-layers in an asymmetric

electrolyte’, J. Chem. Soc. Faraday Trans. 2 80, 535 (1984).

[62] E. H. B. DeLacey and L. R. White, ‘The polarization impedance on an ideally

polarizable plane electrode’, J. Chem. Soc. Faraday Trans. 2 78, 457 (1982).

BIBLIOGRAPHY 183

[63] V. N. Shilov and S. S. Dukhin, ‘Theory of polarization of diﬀuse part of a thin

double layer at a spherical particle in an alternating electric ﬁeld’, Colloid J.

USSR 32, 90 (1970).

[64] S. S. Dukhin and V. N. Shilov, ‘Theory of static polarization of diﬀuse part

of thin electric double layer of spherical particles’, Colloid J. USSR 31, 564

(1969).

[65] B. V. Deryagin and S. S. Dukhin, ‘Theory of surface conductance’, Colloid J.

USSR 31, 277 (1969).

[66] K. T. Chu and M. Z. Bazant, ‘Electrochemical thin ﬁlms at and above the

classical limiting current’, SIAM J. Appl. Math. 65, 1485 (2005).

[67] A. Gonzalez, A. Ramos, and A. Castellanos, ‘Pumping of liquids using trav-

elling wave electroosmosis: a nonlinear analysis’, in ‘Proc. IEEE Int. Conf.

Dielectric Liquids’, 221 – 224 (IEEE, 2005).

[68] L. N. Trefethen, Spectral methods in matlab (SIAM, Philadelphia, 2000).

[69] J. P. Urbanski, J. A. Levitan, M. Z. Bazant, and T. Thorsen, ‘Fast AC electro-

osmotic pumps with non-planar electrodes’, Appl. Phys. Lett. 89, 143508

(2006).

[70] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical

recipes in C, the art of scientiﬁc computing (Cambridge University Press, New-

York, 1992).

[71] K. Svanberg, ‘The method of moving asymptotes - a new method for structural

optimization’, Int. J. Numer. Meth. Eng. 24, 359 (1987).

[72] K. Svanberg, ‘A Class of Globally Convergent Optimization Methods Based

on Conservative Convex Separable Approximations’, SIAM J. Optim. 12, 555

(2002).

[73] A Matlab implementation, mmasub, of the MMA optimization algorithm can

be obtained (free of charge for academic purposes) from Krister Svanberg,

KTH, Sweden. Email: krille@math.kth.se.

[74] M. P. Bendsøe and N. Kikuchi, ‘Generating optimal topologies in structural

design using a homogenization method’, Comp. Meth. Appl. Mech. Engng. 71,

197 (1988).

[75] M. P. Bendsøe and O. Sigmund, Topology Optimization: Theory, Methods and

Applications (Springer, 2003).

[76] H. A. Eschenauer and N. Olhoﬀ, ‘Topology optimization of continuum struc-

tures: a review’, Appl. Mech. Rev. 54, 331 (2001).

[77] J. S. Jensen, ‘Phononic band gaps and vibrations in one- and two-dimensional

mass-spring structures’, J. Sound Vib. 266, 1053 (2003).

[78] J. S. Jensen and O. Sigmund, ‘Systematic design of photonic crystal structures

using topology optimization: Low-loss waveguide bends’, Appl. Phys. Lett. 84,

2022 (2004).

[79] T. Borrvall and J. Petersson, ‘Topology optimization of ﬂuids in Stokes ﬂow’,

Int. J. Numer. Meth. Fluids 41, 77 (2003).

184 BIBLIOGRAPHY

[80] A. Gersborg-Hansen, ‘Topology optimization of incompressible Newtonian

ﬂows at moderate Reynolds numbers’, MSc thesis, Dept. of Mechanical Engi-

neering, DTU (2003).

[81] F. Okkels and H. Bruus, ‘Scaling behavior of optimally structured catalytic

microﬂuidic reactors’, Phys. Rev. E 75, 016301 (2007).

[82] P. Michaleris, D. A. Tortorelli, and C. A. Vidal, ‘Tangent operators and design

sensitivity formulations for transient nonlinear coupled problems with appli-

cations to elastoplasticity’, Int. J. Numer. Meth. Eng. 37, 2471 (1994).

[83] J. P. Urbanski, T. Thorsen, J. A. Levithan, and M. Z. Bazant, ‘Experiments on

nonlinear electrokinetic pumps in microﬂuidics’, Talk presented at American

Physical Society MAR06 meeting, Baltimore, MD.

[84] D. Burch, unpublished (2006).

ii

Abstract

The goal of the present thesis has been to investigate ways of controlling and manipulating ﬂuids and suspensions in microﬂuidic systems. In particular, we focus on a theoretical description of AC electroosmotic micropumps with asymmetric electrode arrays, that have recently been demonstrated to permit fast pumping (velocities ∼ mm/s) with low driving voltage of a few volt only. The dynamical description of electrokinetics and electrochemical transport at driving voltages of just a few volt is a theoretically challenging subject, and therefore simplifying assumptions such as the Debye–H¨ckel u approximation or linear response for weak applied ﬁeld have often been employed in the literature. We extend previous linear theory for AC electroosmotic ﬂow into the “weakly nonlinear” regime by accounting for nonlinear capacitance of the Debye screening layer, and also consider the eﬀect of Faradaic current injection from electrochemical electrode reactions. This allows us to explain why the frequency of maximum pumping is sometimes shifted down when the driving voltage is increased, but neither the linear nor weakly nonlinear models are able to account for the reversal of the pumping direction that has been observed experimentally. Therefore we also study the “strongly nonlinear” regime where classical circuit models with uniform bulk electrolyte concentration break down. We extend recent theoretical work in this regime, by accounting for dynamics in the diﬀusion layer developing when an AC voltage with driving frequency around the inverse RC time is applied, and by considering ﬂuid motion and convection of ions. Moreover, we attempt to include existing theory for double layers driven out of quasiequilibrium from problems of DC Faradaic conduction at “very large” (but experimentally relevant) voltage into our dynamical model. We solve the coupled electrohydrodynamical problem numerically for experimental micropump geometries and display contributions to the net pumping velocity from the diﬀerent ﬂow sources in the model: Our results indicate that both bulk electroconvection and electroosmotic ﬂow of the “second kind” may contribute a signiﬁcant fraction of the overall induced ﬂow already at driving voltages of a few volt. However, a rigorous account

iv

ABSTRACT

for nonequilibrium double layers in a dynamical setting would be necessary to justify several ad hoc assumptions in our model. Finally, we investigate diﬀerent ways of breaking the symmetry of an electrode array and determine in each case the optimal device geometry to maximize the pumping velocity at a given low driving voltage as described by the simple linear theory.

der tillader hurtig pumpning (hastigheder ∼ mm/s) ved lav spænding p˚ nogle f˚ volt. og ved at medregne væskestrømning og deraf følgende konvektion af ioner. Dermed kan vi forklare hvorfor den frekvens hvor pumpehastigheden maksimeres i nogle tilfælde falder n˚ spændingen øges. idet vi tager højde for ulineær kapacitans i Debye skærmningslaget. og derfor benyttes i a a litteraturen ofte forudsætninger s˚ asom Debye–H¨ckel approksimationen eller u lineært respons for svage p˚ atrykte felter. men hverken den lineære eller svagt ulineære teori kan ar forklare hvorfor pumperetningen vender. Desuden forsøger vi at inkludere eksisterende teori for elektriske dobbeltlag drevet ud af kvasiligevægt fra problemer med DC faradaisk strøm ved “meget høj” spænding i vores dynamiske model. og undersøger ogs˚ eﬀekten af faradaisk a strøm fra elektrokemiske elektrode reaktioner. Endelig undersøger vi forskellige m˚ ader at bryde symmetrien af et sæt af elektroder. a a Dynamisk beskrivelse af elektrokinetik og elektrokemisk transport ved spændinger p˚ bare f˚ volt er teoretisk udfordrende. Vi udvider den eksisterende lineære teori for AC elektroosmose til at gælde i det “svagt ulineære” regime. og bestemmer i hvert tilfælde den geometri der maksimerer .Resum´ e Form˚ med nærværende afhandling har været at undersøge m˚ alet ader at kontrollere og manipulere væsker og partikler i mikroﬂuide systemer. idet vi udvider tidligere teoretisk arbejde i dette regime ved at tage højde for dynamik i diﬀusionslaget der dannes n˚ en vekselspænding med frekvens omkring den inverse RC tid p˚ ar atrykkes. Især fokuserer vi p˚ en teoretisk beskrivelse af AC elektroosmotiske mikropumper a med asymmetriske sæt af elektroder. s˚ aledes som det ses eksperimentelt. Derfor studerer vi ogs˚ det “stærkt ulineære” regime hvor de klassiske a kredsløbsmodeller bryder sammen. og bestemmer bidrag til pumpehastigheden fra de forskellige strømningsled i modellen: Vores resultater indikerer at b˚ elektrokonvektion og elektroosmose af den “anden slags” ade kan bidrage med en betydelig del af den samlede væskestrøm allerede ved spændinger p˚ nogle f˚ volt. Vi løser det koblede elektrohydrodynamiske problem numerisk for eksperimentelt relevante mikropumpe geometrier. Imidlertid er en mere grundig behandling a a af dobbeltlag drevet dynamisk ud af kvasiligevægt nødvendig for at retfærdiggøre ﬂere ad hoc antagelser i vores model.

vi RESUME pumpehastigheden for en given lav spænding ved en beskrivelse med den simple lineære teori. .

and for sharing his long experience with AC electrokinetic experiments and recent observations. Vincent Studer digged up old impedance measurements for me to validate my theoretical model. Laurits Højgaard Olesen MIC – Department of Micro and Nanotechnology Technical University of Denmark 31 August 2006 . I would like to thank Henrik Bruus for good spirits and dedicated eﬀort as supervisor. The project was supervised by Prof. and for providing me invaluable insight on electrokinetic phenomena. The work has been carried out at MIC – Department of Micro and Nanotechnology in the Microﬂuidic Theory and Simulation Group (MIFTS) from September 2003 to August 2006. Finally. I thank my girlfriend Hanne for always being there and for reminding me that there is life beyond work. Armand Ajdari. Henrik Bruus.Preface The present thesis is submitted in partial fulﬁlment of the requirements for the PhD degree at the Technical University of Denmark (DTU). Also I would like to thank Nicolas Green for inviting me to see his lab in Southampton. Misha Gregersen never lost her temper when I once and again came up with just another interesting measurement to do. and while abroad by Prof. In the period January to May 2005 I worked at the Laboratoire de Physico-Chimie Th´orique at Ecole Sup´rieure de Physique e e et de Chimie Industrielles (ESPCI) in Paris. Much of the work I did in the ﬁnal part of my PhD studies depends on a theoretical model developed by Martin Bazant and Kevin Chu. I thank the people at MIC and in the MIFTS cave in particular for providing a very nice and inspiring working environment. I also thank Patrick Tabeling and his group at ESPCI for welcoming me and for exposing me to some of their interesting work and problems. My collaboration with Fridolin Okkels on topology optimization was an exciting sidetrack on the route I otherwise pursued. I would also like to thank Armand Ajdari for welcoming me at ESPCI. with whom I had many helpful discussions. and for allowing me great freedom to pursue my many wild ideas. I am grateful for many thoughtful comments on my work also after I left Paris.

viii PREFACE .

2. . . . . . . . . . .1 Dimensionless form .2. . . . . . . . . . . . . . . . . . . . . .2. 2. . . .1. . . . . . . . . . . . . .1. 1. . . .5 Surface conservation laws . . . . . . .1. . . . . . .2 Electroneutral bulk transport 2. .7 Weakly nonlinear model . . . . . .1 Classical theory . . xii xiii xv 1 1 3 5 7 9 9 10 11 15 18 19 22 24 26 29 30 33 37 39 42 49 . 57 . . . . . . . . . . .2 AC electrokinetics . . .2 Faradaic current injection . . . . . . . . . . . 2. . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . .Contents List of Figures List of Tables List of Symbols 1 Introduction 1. . . . . . . . 2. . . . . . . . . . . . . . 2. . . . . . . . . . . 2.2 Electrical double layer . .1. . 2. 2. . . . . . . . . .1. . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . .1 Capacitive charging .2. . . . . . . . .1 Linear analysis at low voltage . . . . . . . . . . . . . . . . . . .1 Microﬂuidics and the lab-on-a-chip . . . . . . . . . . . . . . . . . . . . 2. . . . . 2. . . . .4 Electroosmosis . . . . 53 3. . . . . . . . . . .8 Strongly nonlinear model . . 1. .6 Linear model . . . . . . . . . . .3 Summary . . . .3 Electrode reactions .2. . . . . . . . . . . . .4 Publications during the PhD studies 2 Electrokinetic theory 2. .9 Beyond quasiequilibrium . . . . . . . . . . . . . 2. . . . .1 Electrochemical transport . 3 Linear and weakly nonlinear analysis 51 3. . . . . . . 2. . . . .1. . . . . . . . . . . . . . . 1.3 Outline of the thesis . .3 Diﬀusion layer . . . . . . . . . . . . . . .4 Quasiequilibrium double layer 2. 54 3. . . . . . . . . . . . . . . .2 Mathematical model . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . .2. .

. . .2. . . . . . . . . . . . . . . . .2 2D model problems . . . . . . . . . 4. .3 Summary . .2. . . Numer. . . . . . . .1 Dynamical solutions . . . . . . . . .2. . . 5. . . Summary . . . . . . . J. Eng. . . .1 Planar asymmetric electrode pairs . . . . . . . . 3. .3 3D structured electrodes with regular shape . . . . . . . . . . . . .4 3D structured electrodes with arbitrary shape 5. . . . . . . . . . . . . . . . E C Paper published in Int. . . . . . . . . . . . . . . . . . . . . . . .2 Nonlinear Faradaic current injection . . . . 4. 5 Optimization 5. 3. . . . . . . . . . . . . . . 4.1.2 Planar electrodes with nonuniform coating . . . . . . A FEMLAB code 117 A. . the . . . 127 B Paper published in Phys. . . . . . . . . . . . . .2 Comparison to a full numerical solution of Poisson–Nernst–Planck equations . . .1. . . . . . . . 117 A. . . . .5 Thermodynamic eﬃciency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Meth. . . . .x 3. . CONTENTS potential . . .3 Periodic integro-diﬀerential equations in time . . . . . . .1. . . . . . . 4.2. . . . . . . . .2 Strongly nonlinear ﬁnite element model . . . . .1 Weakly nonlinear ﬁnite element model . . . . . . . . . . . 5. 4. . . . . . . . . . . . .1 Symmetric electrode array . . . .2 3. . .3 4 Strongly nonlinear analysis 4. . . . . . . . . .1 1D model problem . . . . .3 Linear response at nonzero intrinsic zeta Weakly nonlinear analysis . . 4. . .1 Nonlinear surface capacitance . 59 59 60 61 63 65 65 67 70 78 78 82 88 91 93 94 100 105 110 112 113 3. . . . . . 5. . . . . . . . . . . . . 5. . . . 6 Conclusion and outlook . . Rev. . .6 Summary . . . 122 A. 131 149 . . . . . .2 Asymmetric electrode array . . . . . . . . . . . . . . . . . . .

. . ﬂux. .17 4. . . . . Total angular momentum vs. . .15 4. . . . .3 3. Individual ion concentrations . .1 3. . . . .4 2. . . . . . . . Pumping velocity vs.4 4. . Streamlines on planar asymmetric electrode array . frequency and voltage . . . frequency and voltage . Pumping velocity vs. .5 3. . . Surface concentration and zeta potential vs. . . . . . . . . . . . . . . . Pumping velocity vs.3 2. . . . . . . . . . . . . . . . Slip velocity distribution on electrode . . . . . .1 2. . . . . .4 3. time . . . . . Time average electric ﬁeld vs. Bode plot of system impedance .7 4. . . . . .16 4. . .18 Schematic picture of electrical double layer . . . . Equivalent circuit diagram for electrochemical cell . . . . . . . . . . . . . . . and bulk electroconvection . . Pumping velocity vs.11 4. . .6 4. . . voltage . .8 4. . . . . . . . . . . . . and bulk electroconvection .12 4. . .3 4. . . . .5 4. frequency and charge transfer resistance Pumping velocity vs.List of Figures 2. . . . .14 4. . . . . . Salt concentration. frequency and voltage . . . . . . . . . electrode . . . Pumping velocity vs. . . . .1 4. . . . voltage . . . . . . . . . . . . . . . .6 3. voltage Concentration proﬁle in the diﬀusion layer . . Pumping velocity vs. . . Sketch of symmetric electrode array . . .2 3. . . . . Slip velocity distribution on electrode . . . Salt concentration. . . Electric ﬁeld on loglog scale . . . . . . . . . . . . . Individual ion concentrations on loglog scale . . .8 4. . . . . . . . . .13 4. Schematic picture of nested boundary layers . . . . . . . . . . . . . . . . . . . . . . . . . . .2 4. Sketch of 1D model geometry . .7 3. . . ﬂux. . . . . frequency and voltage . . . . . . . . . . . . . . . . . . . . . . . . . . Development of salt concentration proﬁle .5 3. . . . . . . .9 4. time . . . . . . . . . . . frequency and voltage . . . . .2 2. . . Dynamical solutions for potential vs. . Relative increase in anion concentration at positive Free energy landscape around transition state . 11 14 16 21 35 52 54 55 56 58 60 61 62 66 68 69 71 72 74 75 77 78 79 79 80 81 82 83 84 85 86 Sketch of planar pump with asymmetric pairs of electrodes . . . . . . . frequency . .10 4. . . Thickness of space charge layer vs. . .

. . . . . . . . electrode dimensions . . .10 5. Piecewise uniform coating .8 5. . . . . electrode dimensions . . .3 5. . . . . . . . . . . . . . . . . . .7 5. . Pumping velocity vs. . . . . . . . .1 5. . . . . . .1 Output from weakly nonlinear ﬁnite element model . . . . . . . . . . .xii LIST OF FIGURES 4. . . . .6 5. . . Sketch 3D geometry with stepped electrodes . . . . . . . . . . . . . . . Pumping velocity vs. . . . . . . Flowrate vs. . . overall channel height . . . . . Sketch of planar electrode array with asymmetric coating Optimal nonuniform coating . . . .4 5. . .2 5. 87 94 95 98 99 101 102 103 104 105 108 109 A. . . . . . . . . . . . . . . . 121 . . . voltage . . .9 5. . Optimized arbitrary 3D electrode shapes . 3D geometries with stepped and grooved electrodes . . .19 Pumping velocity vs. . . . . . .5 5. . . 5. .11 Optimal design of planar asymmetric electrode array . . . . . . Optimized arbitrary 3D electrode shapes . . . .

. .List of Tables 2. . .1 Typical experimental values and numbers for a 1 mM KCl solution in a device with 10 µm characteristic length . 24 .

xiv LIST OF TABLES .

Symbol c D D0 Da ∆D e E f fa F Fa j0 jext j J kB ℓ0 n O p q Q r R Rct Description (Molar) concentration Diﬀusivity Average diﬀusivity Ambipolar diﬀusion constant Diﬀusivity diﬀerence Elementary charge Electric ﬁeld Frequency Excess ambipolar ﬂux Flux density Ambipolar ﬂux Exchange current Faradaic current injection Surface excess current density Current density Boltzmann constant Geometrical length scale Surface outward normal Oxidized species Pressure Surface charge density Volume ﬂow rate Position vector Reduced species Charge transfer resistance Unit (mol) m−3 m2 s−1 m2 s−1 m2 s−1 m2 s−1 C V m−1 Hz [ ≡ s−1 ] (mol) s−1 m−2 (mol) s−1 m−2 (mol) s−1 m−2 A m−2 A m−2 A m−1 A m−2 J K−1 m – – Pa [ ≡ N m−2 ] C m−2 m3 s−1 m – Ω m2 1.381 × 10−23 . it should be obvious from the context whether a given quantity is dimensionless or not.602 × 10−19 1.List of Symbols To reduce the notational overload we denote dimensionless quantities by the same symbol as their dimensional counterparts.

xvi Symbol Rmt t T T u U V0 Vext w z γ Γi δ ǫ ǫ0 ε ζ ζ eq ς η ϑ λD µ ν ρ ρm σ σ τ0 τD φ Φ ψ ω Ω Description Warburg constant Time Period in time Temperature Fluid velocity Pumping velocity Driving voltage External potential on electrode Surface excess concentration Ion valence LIST OF SYMBOLS Unit Ω m2 s−1/2 s s K m s−1 m s−1 V V (mol) m−2 – (mol) m−3 (mol) s−1 m−1 – F m−1 F m−1 – V V V Pa s V m J m2 s−1 C m−3 kg m−3 S m−1 Pa s s V V V rad s−1 – Excess concentration Surface excess ﬂux density Surface capacitance ratio Permittivity Permittivity of vacuum 8.854×10−12 Dimensionless Debye length Zeta potential Intrinsic zeta potential Diﬀusion layer potential drop Dynamic viscosity Overpotential Debye length Electrochemical potential Kinematic viscosity Charge density Mass density Conductivity Stress tensor RC time Debye time Electrostatic potential Space-charge layer potential drop Excess potential Angular frequency Computational domain .

LIST OF SYMBOLS Accent a∗ a ¯ a ˜ a ˇ a ˘ a ˆ a∗ ˆ Description Equilibrium bulk reference value Bulk variable in asymptotic expansion Debye layer variable in asymptotic expansion Diﬀusion layer variable in asymptotic expansion Space-charge layer variable in asymptotic expansion Complex variable Complex conjugate xvii .

xviii LIST OF SYMBOLS .

inkjet printer heads. e. Also in more conventional analytical chemistry. reagent dispensers. because mass diﬀusion and heat transfer is very rapid on the microscale and the relative importance of surface to volume forces increases as the system is scaled down [3. and multilayer soft lithography provides a very ﬂexible platform for microﬂuidic plumbing with valves and peristaltic . that would automatize the complete analysis from sample to electrical read-out [1. Of course.Chapter 1 Introduction 1.g. the fabrication techniques originally developed by the microelectronics industry now allow accurate design of microﬂuidic circuits. miniaturization opens for new types of analysis that would not be possible in a macrosystem. mixers.1 Microﬂuidics and the lab-on-a-chip Microﬂuidics refers to hydrodynamics and ﬂow manipulation in systems and devices with characteristic dimensions less than a millimeter. 2]. advances in biology and biotechnology require manipulation of particles or living cells and chemical detection of very small quantities. Secondly.. separators. Depending on the application there can be very diﬀerent requirements for such a pump: In a laboratory test setup it is often suﬃcient to use external syringe pumps. 4]. for several reasons: First. but for more complex systems it is necessary to use active pumping to control the ﬂow. such systems have long been studied in biology and colloid science. Fabrication from polymer materials would lower the production cost and allow for cheap and portable devices to perform simple analyses. The ultimate goal is development of “lab-on-a-chip” systems with integrated pumps. Equally important. An important question when designing such microﬂuidic lab-on-a-chip systems is how to get the sample into the chip and manipulate it around. but recently there has been an enormous increase of attention on the subject. For a simple disposable device an elegant approach is the use of capillary forces. miniaturization has the advantage of low sample and reagent consumption. and detection units. and is being used to construct microelectromechanical systems (MEMS) and microﬂuidic devices.

but it allows local control of the ﬂow and could be used.or nanoporous medium. Another problem is that in order to maintain an electric ﬁeld down the pumping channel. the pumping velocity is inversely proportional to the size of the electrodes. but allows pumping velocities of the order of mm/s to be attained with just a few volt applied between the electrodes [15]. This requires driving the system with an AC voltage in the kHz range. into the channel which increases the hydraulic resistance dramatically but does not aﬀect the electroosmotic ﬂow. moving parts makes the fabrication delicate. a DC Faradaic current has to be drawn between electrodes submerged in the pumping liquid. to circulate sample between two chambers on a chip... and many valve designs are prone to failure if there are particles or bubbles in the liquid. On the other hand. This can be overcome by introducing a micro. .. a sintered glass frit. for disposable devices the production costs and ease of integration should be much more important. 12] or a gel salt bridge [13] at the expense of increased fabrication complexity. An integrated pump design that has been subject to intensive research over the years is the (piezo-actuated) membrane pump. 12]. Moreover. where a membrane is displaced to create pulsating ﬂow that is rectiﬁed by valves at the inlet and outlet to the pumping chamber [6]. in this way both a decent ﬂow rate and pressure can be generated at a thermodynamic eﬃciency of a few percent [7.g. i. which can be achieved by a single photolithography step followed by metal deposition and lift-oﬀ.2 Introduction pumps powered by external pressure transducers [5]. The fabrication is simple and only requires an array of interdigitated electrodes to be deposited on the channel substrate. 9. but their performance depends strongly on the pH and chemical composition of the pumping liquid. but only generate very small pressures so the thermodynamic (energy) eﬃciency is low [7]. 11. 10. Micropumps based on electroosmosis have no moving parts which makes them easy to fabricate. e. and especially if designed for implantation they should have very low power consumption. e. 8. However. An interesting alternative to standard DC electroosmotic pumps is the use of AC electroosmosis with an asymmetric array of microelectrodes to generate pumping [14]. This can be avoided by separating the electrodes from the pumping channel by means of ion exchange membranes [11.g.e. the AC electroosmotic micropump is very poor at generating pressure. the eﬃciency increases with miniaturization. Similar to the basic DC electroosmotic pump design. which in many cases generates bubbles due to electrolysis. The basic design where an electric ﬁeld is simply applied down a section of the microchannel can in principle pump the liquid in a plug ﬂow with velocities of a few mm/s in free ﬂow conditions. Devices for drug delivery need to be extremely fail-safe.

around the inverse RC time form the system [22. 24]. this was found for a dilute . 18].g. They were also able to match quantitatively the experimental pumping velocity and frequency of maximal pumping by assuming a compact Stern layer on the electrodes with roughly the same capacitance as the diﬀuse Debye layer.. In some sense this good agreement is surprising because their theory is based on the Debye–H¨ckel approximation and linearization which u is strictly valid only at low driving voltages V0 < kB T /e ≃ 25 mV.2 Vrms . [25] and later by several other groups [15.2 AC electrokinetics 3 1. and obtained the same direction of pumping as observed experimentally and similar streamline patterns [33]. An AC ﬁeld also gives rise to electrokinetic ﬂow around the particles but no net particle motion (if the ﬁeld is uniform and the particle is spherical). Ramos et al. where a small harmonic modulation is added to a (slowly varying) DC voltage and the current response of the electrochemical cell is measured. the electrophoretic velocity of the particles in an applied DC ﬁeld. changing to forward pumping above 1. 29. and one studies. is a standard and very powerful tool for determining reaction rate constants and also for probing the surface properties at the electrode [16]. 28. induce interaction between particles and inﬂuence their tendency to agglomerate [19.1. 23. Gregersen found reversal of the pumping direction at 2. 31. bubble formation and electrode degradation was observed because they used bare platinum electrodes that support electrochemical electrode reactions and Faradaic current. The ﬂow does. In colloid science the focus is most often on colloidal particles with a certain intrinsic surface charge and zeta potential. 32]. 20]. while experiments are typically performed with at least a few volt. 27.5 kHz and below 1. Studer et al. Squires and Bazant predict that a nonspherical egg-shaped metal colloid will align with the applied ﬁeld and “swim” due to induced-charge electroosmosis [21]. did a detailed theoretical study with realistic boundary conditions for the particular planar geometry used by Brown et al. While the general symmetry arguments of Ajdari predicts pumping in some direction for any locally asymmetric geometry.2 AC electrokinetics In electrochemistry impedance spectroscopy. Later. observed reversal of the pumping direction when driving the system at around 50 kHz and above 5 Vrms .. Applying the same voltage at 1 kHz. 30. well above the frequency for maximal pumping in the “forward” direction around 1 kHz [15]. or investigate their inﬂuence on the dielectric response when an AC ﬁeld is applied [17. e. which was soon after demonstrated experimentally by Brown et al. Interestingly. Ramos and co-workers ﬁrst observed electroosmotic ﬂow over pairs of microelectrodes when driven with an AC voltage in the kHz range. however. 26.2 Vrms [32]. Ajdari predicted that the same mechanism would give rise to pumping on an asymmetric electrode array [14]. Usually one stirs the solution to ensure uniform bulk reagent concentrations and simplify the experiment.

Chu extended this work to three dimensions by studying a spherical metal particle. where the capacitance of the Debye screening layer at the electrodes becomes nonlinear. 23. the same needs not be the case in AC because the reactions could be run reversibly [35]. because that problem is not analytically tractable they did not attempt to solve it. In particular they went beyond the linear theory to analyze the “weakly nonlinear” regime. We agree that for weakly nonlinear dynamics to hold. this theory is unable to account for the reversal of the pumping direction observed experimentally. Ramos and co-workers found reversal of the pumping direction on a travelling wave device at 1 kHz and above 2 V [34]. 36]. and by accounting also for electroosmotic ﬂuid motion and convection. with perturbations showing up only as higher order corrections. However. our guess is that many biochemists would be uncomfortable with uncontrolled electrode reactions taking place in same the channel where a delicate sample is to be pumped through. and investigated both the weakly nonlinear dynamical response to a suddenly applied DC ﬁeld and the steady-state solution to the strongly nonlinear problem arising at larger applied DC ﬁeld [38. the driving voltage V0 cannot greatly exceed the thermal voltage kB T /ze. and Ajdari recently studied diﬀuse-charge dynamics in a simple 1D model system analytically by means of matched asymptotic expansions [37]. We have taken this fruitful approach further by actually solving the strongly nonlinear dynamical model numerically in 2D. the total zeta potential indeed cannot greatly exceed the thermal voltage. They used titanium coated gold electrodes that do not support Faradaic current due to the formation of a thin natural titanium oxide layer. and also to be able to compare the theory more quantitatively with experiments as a function of both driving voltage and frequency. Thornton. However. but because the induced zeta potential . and on an asymmetric electrode array at 10 kHz and above 5 Vrms [31]. We note that while electrode reactions leading to bubble formation is a major challenge for DC electroosmosis. Further.4 Introduction KCl working electrolyte on bare platinum electrodes with the same device geometry as used by Studer. where the assumption of uniform bulk electrolyte concentration breaks down but bulk charge neutrality remains valid. 39]. however. Therefore it has been a major focus of our work to extend the existing theory to be able to describe these phenomena. which we denote “linear theory” even though the induced time-average electroosmotic ﬂow scales as the square of the driving voltage. They also derived a dimensionless criterion for the zeta potential at which the weakly nonlinear asymptotics break down. and argued that for weakly nonlinear dynamics to hold. Bazant. 33. but the bulk electrolyte concentration remains uniform and constant at leading order. they wrote the governing equations for the “strongly nonlinear” regime. Most previous work on AC or induced-charge electroosmosis has been based on the Debye–H¨ckel approximation with linearization of the electrical u problem [19.

1. the electrical double layer changes from quasiequilibrium to a nonequilibrium structure with an extended space charge layer. and to be able to handle it numerically we make some fairly crude approximations. nonlinear electrokinetic ﬂow is induced. In steady-state problems of Faradaic conduction it is well known that beyond the “diﬀusion-limited” current. In this we follow the work of Bazant. Further. This is a challenging problem. and because the driving voltages used experimentally are often large enough to make the strongly nonlinear model break down. which can account for overlimiting conductance [42. Our results indicate that electroosmosis of the second gives rise to net pumping in the same (“forward”) direction as the usual electroosmosis of the “ﬁrst kind”. but it also provides a ﬁrm starting point for going beyond the classical theory. Indeed the assumption of thin Debye layers underlies most classical analysis in electrochemistry. while bulk electroconvection contributes in the “reverse” direction. Thornton. even the strongly nonlinear asymptotics break down due to complete salt depletion in the diﬀusion layers close to the electrodes. as described ﬁrst by Rubinstein and Shtilman [40]. As the driving voltage is further increased. when tangential ﬁelds act upon this space charge layer. we try to incorporate the basic features of the nonequilibrium double layer and electroosmosis of the second kind into our dynamical model. but where they primarily focus on transient Chapter 2: Electrokinetic theory . before we proceed to construct a mathematical model based on matched asymptotic expansions. and Ajdari [37] and Chu [38] on diﬀuse charge dynamics. and studied it extensively in the context of nonlinear electrophoresis of conductive particles made from ion exchanger material [41].3 Outline of the thesis We ﬁrst give a brief overview of some of the classical electrokinetic theory. The latter is a perturbation technique where one identiﬁes the ratio of the Debye screening length to the characteristic geometrical length scale as a small parameter. More recently. Dukhin and coworkers introduced the term “electroosmosis of the second kind” to describe this phenomenon. 43].1. so the perturbation technique is often just a systematic way of obtaining already known results.3 Outline of the thesis 5 is related to the driving voltage in a nonlinear way we also ﬁnd that the latter can often be an order of magnitude larger than the thermal voltage at the point where the weakly nonlinear dynamics break down. Because even our strongly nonlinear model does not seem to be able to account for reversal of the pumping velocity. Rubinstein and Zaltzman theoretically predicted that second kind electroosmosis renders linearly instable the diﬀusion layer at a planar permselective (ion exchange) membrane and provides an eﬃcient mixing mechanism.

both to gain a better understanding of the dynamical solutions. Chapter 5: Optimization There are many ways of breaking the left-right symmetry of an electrode array to induce net pumping. We also include the nonlinear surface capacitance of the Debye layer as described by Gouy–Chapman–Stern theory. and from electroconvection due to concentration gradients in the electroneutral diﬀusion layers and bulk electrolyte. . In this chapter we consider both the “standard” pump with asymmetric pairs of planar microelectrodes. which has been published in the paper ‘A high-level programming-language implementation of topology optimization applied to steady-state Navier–Stokes ﬂow’ [46]. Faradaic current injection. and a 3D structured electrode design recently suggested by Bazant and Ben [45]. we study AC driving and steady-state periodic solutions.6 Introduction solutions for a suddenly applied DC voltage. Chapter 3: Linear and weakly nonlinear analysis We study the pumping from an array of asymmetric pairs of microelectrodes. Chapter 4: Strongly nonlinear analysis In this chapter we ﬁrst analyse the strongly nonlinear dynamics for a simple 1D geometry with parallel-plate blocking electrodes. In all cases the performance increases with miniaturization. but also to validate the asymptotic model by comparison to a full numerical solution of the Poisson–Nernst–Planck equations. but here we also take into account the eﬀects of mass transfer on the Faradaic electrode reaction. Chapter 6: Conclusion and outlook We present concluding remarks on our work on AC electrokinetic pumping and give some directions for future research. We have chosen not to present those results in this thesis but have included the paper in Appendix C. This work is related to our work done on topology optimization earlier during the PhD studies. Most of this work was published in our paper ‘AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement. so we investigate the optimal electrode design for a ﬁxed minimal linewidth or feature size in the geometry. see Appendix B. Next we study a 2D model of the planar AC electroosmotic pump and determine the contributions to the net pumping from both electroosmotic ﬂow of the ﬁrst and second kind. extending the existing linear theory for AC electroosmosis by accounting for Faradaic current injection both in a linearized scheme and using the full nonlinear Butler–Volmer reaction kinetics. and nonlinear surface capacitance’ [44]. a design with asymmetric coating of the individual electrodes.

A. N. Part of the reason may be the increased complexity involved with solving a time evolution problem with special variables deﬁned on the boundary. Belmon. Olesen. 056313 (2006). L. Phys. L. and to demonstrate this we have included a code example in Appendix A. Bruus. E 71. H. Olesen. 37 (2006). and A. 056306 (2005). AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement. A high-level programming-language implementation of topology optimization applied to steady-state Navier–Stokes ﬂow. L. not many attempts have been made in the weakly nonlinear regime.4 Publications during the PhD studies 7 Appendix A: FEMLAB code While the standard linear model for induced-charge electroosmosis has been studied numerically by several authors. Mortensen. Ajdari. L. and H. Transport coeﬃcients for electrolytes in arbitrarily shaped nano and micro-ﬂuidic channels. Num. Olesen. Electro-hydrodynamics of binary electrolytes driven by modulated surface potentials. H.4 Publications during the PhD studies Papers in peer reviewed journals 1. H. 1. Faradaic current injection. 4. 8. and nonlinear surface capacitance. However. Bruus. H. We also include the implementation of our strongly nonlinear model. Okkels. Olesen. . E 73. Phys. 65. Mortensen. J. F. Meth. Phys. H. Bruus. Bruus. and H. New J. A. using commercial ﬁnite element software like Femlab [47] this is actually straightforward to implement. N. 975 (2006). L. Rev. and discuss how to deal with the nonlocal coupling in time introduced by our semi-analytical solution for the dynamics in the diﬀusion layer.1. Eng. and H. Rev. 3. Int. 2.

and H. Bruus. N. H. ECCOMAS 2004. Bruus. p. . A. the Netherlands. vol. F. N. Olesen. Bruus. Nonlinear studies of AC electrokinetic micropumps. Transport Phenomena in Micro and Nanodevices. vol. June 2006.8 Conference proceedings Introduction 1. Olesen. proc. Applications of topology optimization in the design of micro. Olesen. a a 2. and H. Mass and charge transport in micro and nano-ﬂuidic channels. and nonlinear surface capacitance. Applications of topology optimization in microﬂuidics: ﬂuid ﬂows. Conf. Olesen. H. Olesen. and H. Olesen. 3. 7. Bruus. and A. USA. Utrecth. L. Faradaic currents. Euromech Fluid Mechanics Conf. H. 1. and A. Bruus. L. thermal transport. Finland.and nanoﬂuidic devices. 4. L. Boston. July 2005. March 2006. Anaheim. June 2006. Okkels. 2. F. 575–578. Sweden. Mortensen. H. F. L. October 2005. AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement. Mortensen. Nanotech 2005. Stockholm. vol. Ajdari MicroTAS 2005. Okkels. 2nd Int. and chemical reactions. USA. H. H. Italy. and L. p. 224.. 6. Bruus. H. p. H. F. Olesen. proc. H. and H. 1. USA. Universality in microﬂuidic phenomena inside microchannels with arbitrarily shaped cross-sections. proc. Ajdari APS MAR06 Meeting. Okkels. Jyv¨skyl¨. July 2004. May 2005. Baltimore. Barga. Topology optimization of Navier–Stokes ﬂow in microﬂuidics. Okkels. 6th EPS Liquid Matter Conf. 6. L. H. 5. 560–562. Bruus. L.

1) . even in a macroscopic electrochemical experiment. Therefore. A common feature underlying most of the classical analysis is the assumption that the system can be divided into an electroneutral bulk region and charged boundary layers that are very thin compared to the characteristic geometrical length scale in the system.Chapter 2 Electrokinetic theory In this chapter we start by giving a brief overview of the classical theory of electrochemical transport and electrode reaction processes. i ci zi e λD = (2. 2. and desalination by electrodialysis. electroplating. This simple picture can be formalized by perturbation theory.2 we go through a systematic derivation of a mathematical model based on matched asymptotic expansions.and macroscopic applications the continuum approximation is well justiﬁed. For a concentrated solution of 1 M the mean separation between ions is around 1 nm. The thickness of the screening layer is given by the Debye length ǫkB T ∗ 2 2. However. whereas for a dilute 1 mM solution the mean separation is around 10 nm. 2. but in true nanoﬂuidics it may not be. in most micro. and in natural processes such as corrosion and ion transport through ion channels in biological cells.1 Classical theory A fundamental approximation in the classical theory of electrochemical transport is the continuum approximation where the diﬀerent chemical species are represented by their concentration ﬁelds. the system spontaneously develops microscopic structure in the form of screening layers at electrodes or any other charged surfaces and interfaces in the electrolyte. and in Sec. This is a widely studied subject due to its importance both for industrial applications such as batteries and fuel cells.

c∗ is the bulk i concentration (number density) of the ith ionic species.3) for the ﬂux reduces to the Nernst–Planck equation zi e Fi = −Di ∇ci + ci ∇φ + ci u. e the elementary charge. Fi = −ci j where Lij are generalized mobilities that account for interactions between diﬀerent species. (2.5) kB T where we used the Einstein relation Di = kB T Lii to express the mobility in terms of the diﬀusivity Di . For a typical system λD is of the order of 1–10 nm. that is. (2. i (2. Nernst–Planck equation For an ideal dilute system the interaction between diﬀerent species can be neglected and the electrochemical potential can be decomposed into separate entropic and electrical terms as µi = kB T log ci + zi e φ (2. and zi the valence number.6) assuming constant permittivity ǫ of the solvent.1 Electrochemical transport Within a continuum model the transport of the individual chemical species is governed by conservation laws of the form ∂t ci = −∇ · Fi . . 54] (2.4) where is φ the mean-ﬁeld electrostatic potential.1. but upon proper averaging in time or in the transverse direction to the screening layer one should still be able to use continuum theory on the averaged ﬁelds. T the absolute temperature.2) where Fi is the ﬂux of the ith species and we disregard any bulk chemical reactions producing or consuming that species. and ci u represents transport by convection with the ﬂuid velocity u. 2. This means that the instantaneous “concentration proﬁle” over a particular point on the surface will be highly nonuniform and rapidly ﬂuctuating. (2.10 Electrokinetic theory where ǫ is the electrical permittivity of the solvent.3) Lij ∇µj + ci u. The electrostatic potential φ is determined by the charge distribution ρ through the Poisson equation −ǫ∇2 φ = ρ = e zi ci . kB is Boltzmanns constant. on average there are no more than a few ions across it. In general the ﬂux is a function of the electrochemical potential µj of all the species present in solution [53. Then Eq.

2. the mean-ﬁeld potential variation from the electrode to the charge neutral bulk.1: Schematic picture of the diﬀuse screening layer in the electrolyte next to a negatively biased metal electrode. This is shown schematically in Fig. The region between the OHP and the electrode is called the “compact layer”. 2.1 it is roughly two solvent molecules thick because the approach of the cations is obstructed both by their own solvation shell and by a monolayer of solvent molecules held at the surface by van der Waals forces. 2. and anions are repelled.2 Electrical double layer The common continuum model view of the screening layer is based on division into a compact and a diﬀuse part. and in Fig. The outer Helmholtz plane (OHP) deﬁnes the closest distance that the solvated cations can move towards the electrode. Cations in the electrolyte move towards the electrode to screen the negative charge.1. The ions in the screening layer outside the OHP are not static but undergo thermal motion.1 Classical theory IHP OHP Diﬀuse layer Bulk electrolyte 11 Solvated cation Metal Speciﬁcally adsorbed anion Solvent molecule Vext ζ φ(y) −λS 0 λD y Figure 2. and this region is therefore called the “diﬀuse layer”. Some ions in the electrolyte may interact more strongly with the electrode material and adsorb directly on the electrode.1 for the screening layer next to a negatively biased metal electrode. Also shown is φ(y). and this point of closest approach is called the outer Helmholtz plane (OHP). and the model is known as the Gouy–Chapman–Stern model [16]. and the locus of their electrical centers is denoted the inner Helmholtz plane (IHP). Cations are attracted to the electrode and anions are repelled so the interface appears charge neutral when seen from the bulk. In order to avoid confusion with the diﬀusion layer of width ℓi (ω) ∼ . The ﬁnite size of the ions can be taken into account by deﬁning a smallest distance that the solvated cations can move towards the electrode. Together the compact and diﬀuse layers are called the “double layer”. 2.

7) where c∗ is the reference concentration attained in the charge neutral bulk i electrolyte where φ = φ∗ ..10) . to and from which the ions have time to diﬀuse during one cycle of the driving voltage at angular frequency ω.14) = −sgn(ζ) 2ǫkB T =− 1 ǫ 2kB T zeζ . (2.12) − 1) (2. we prefer to denote the diﬀuse part of the double layer as the “Debye layer”. kB T (2. namely the Gouy–Chapman solution1 ψ(y) = zeζ 4kB T e−y/λD . (2. (2.13) (2.6) we get the Poisson– Boltzmann equation −∇2 ψ = e ǫ c∗ zi exp − i zi eψ .1. sinh λD ze 2kB T Gouy also showed that Eq.7). tanh−1 tanh ze 4kB T (2. and for a symmetric binary z:z electrolyte this reduces to the more familiar form 1 kB T zeψ ∇2 ψ = 2 . ci = c∗ exp − i zi e(φ − φ∗ ) .e.12 Electrokinetic theory Di /ω.11) where y is the distance perpendicular to the surface and ζ is the potential attained at the outer Helmholtz plane where y = 0. Substituting into Eq. (2. The total amount of charge accumulated in the Debye layer per unit area is q= 0 ∞ ρ dy = ǫ∂y ψ y=0 ∗ −zi eζ/kB T i ci (e (2. Gouy–Chapman–Stern model In thermal equilibrium the electrochemical potential is constant thoughout the system and the ions in the Debye layer are Boltzmann distributed.9) sinh kB T λD ze The Poisson–Boltzmann equation constitutes a single equation to solve for the equilibrium electrostatic potential. Fig. (2. We also tend to denote the compact part as the “Stern layer”. For one-dimensional systems it has an analytical solution. 2. kB T (2. i. and mention that this can also be used to model an oxide layer or insulating coating on the electrode. the ion distributions then follow directly from Eq. where ψ(y) = kB T 3 log 1 − ze 1 + cosh y/λD + 2 tanh−1 (1 + 2ezeζ/kB T )/3 .8) i where we introduced ψ = φ − φ∗ . cf.8) can be integrated for an asymmetric binary electrolyte with z+ = −2z− = 2z.

1 Classical theory 13 where Eq.7) eﬀectively limits the range of zeta potentials that makes sense . we stick to the simplest form of the Gouy–Chapman–Stern model and always assume a constant Stern layer capacitance and neglect speciﬁc adsorption. (2. (2.14) for a symmetric z:z electrolyte only. However. At larger potential the Debye layer capacitance becomes very large and the overall double layer capacitance is dominated by the Stern layer. i.e. the double layer capacitance reduces to λD −1 ǫ 1 .01.3. The overall potential drop across the double layer is then Vext − φ∗ = ζ − q . Neglecting speciﬁcally adsorbed ions. whereas for a titanium electrode with a natural TiO2 oxide of thickness λS ∼ 4 nm and ǫS = 110ǫ0 we ﬁnd δ ∼ 0.17) + = Cdl = CS ǫ λD (1 + δ) where we introduced the parameter δ = ǫλS /ǫS λD for the capacitance ratio between the Debye and Stern layers. The relevant parameter range for δ is large: If we compare the Debye layer capacitance for a dilute 1 mM aqueous electrolyte with λD = 10 nm and ǫ = 80ǫ0 to the molecular Stern layer deﬁned by the distance closest approach λS ∼ 1 ˚ of an ion to a bare metal electrode then we get a very A small capacitance ratio δ ∼ 0. and the speciﬁc adsorbtion and desorption of ions give rise to an additional component in Eq.15) and the diﬀerential capacitance per unit area becomes Cdl = − dq λD 1 = + ∗) d(Vext − φ CS ǫ cosh(zeζ/2kB T ) −1 .2. the region −λS ≤ y ≤ 0 behind the outer Helmholtz plane is free of charge and the potential varies linearly. Modiﬁed Poisson–Boltzmann equation The Gouy–Chapman–Stern model is widely used because it is simple and yet captures the generic features observed in typical electrochemical experiments.. (2. because we are not concerned with any particular material system but wish to study electrokinetics in more general. The exponential increase in ion concentration with potential implied in Eq.16) In the Debye–H¨ckel limit ζ ≪ kB T /e where the potential variations are u small compared to the thermal voltage. In real electrochemical systems the Stern layer on bare metal electrodes often does not behave exactly as an ideal capacitor but shows some voltage dependence in CS . CS (2.13) reduces to Eq. (2. (2. the Stern layer ideally behaves as a capacitor of capacitance CS = ǫS /λS . (2. and for even thin polymer coatings we easily obtain δ ≫ 1.16).

If we consider again a dilute 1 mM KCl solution then the bulk concentration is 6 × 1023 ions per m3 . cf. The ﬁgure shows that the ion concentration grows as (Vext − φ∗ )2 /δ2 at large voltage. the accumulated charge depends linearly on the applied voltage q ≃ −CS (Vext − φ∗ ). 2. One problem with this reasoning. the A “purpose” of the Stern layer is exactly to supress this unphysical behaviour.1 it occurs already at 500 mV while for δ = 10 the limit is reached around 10 V. where almost the entire potential drop across the double layer is placed on the Stern layer. which is far exceeded if we put 1 V across a 1 ˚ molecular Stern layer. However.32 ˚ for Cl− .2: Double logarithmic plot of the relative increase in anion concentration at the outer Helmholtz plane as a function of the overall potential drop Vext − φ∗ across the double layer for diﬀerent capacitance ratios δ in the Gouy–Chapman–Stern model. (2.0 δ = 10. so the steric limit of a close packed (hydrated) K A phase of ions corresponds to around 1028 ions per m3 . whereas the zeta potential grows only logarithmically with q.33 (2. .14 10 8 Electrokinetic theory −1 10 6 δ = 0. Eq. the concentration predicted at the outer Helmholtz plane exceeds 1032 ions per m3 or 100 ions per ˚3 which is ridiculously large. The ions have radii of a few ˚ angstrøm. or even a 5 nm titanium oxide A layer.2 where the relative increase in the anion concentration at the outer Helmholtz plane is plotted as a function Vext − φ∗ . The voltage at which this limit is reached depends on the Stern layer capacitance: Assuming a low capacitance ratio of δ = 0.13). exp ze ζ/kB T 10 4 10 2 10 0 10 0 10 1 10 2 10 3 ze (Vext − φ∗ )/kB T Figure 2. Therefore the increase in ion concentrations in the Debye layer also becomes less dramatic. is that most insulators suﬀer dielectric breakdown around 1–100 MV/m.1 δ = 1. with 1. At large voltage when the double layer capacitance is dominated by the Stern layer. but also that this large voltage regime is not reached until Vext − φ∗ exceeds ≈ 1 V. as shown in Fig. physically: Substituting ζ = 500 mV for a bulk dilute 1 mM KCl solution.32) ˚ for a A + ion and 2.

1.762 V. the neglect of correlations between diﬀerent ions inherent in the use of Eq. (2.g. This is very a interesting direction which may allow to explain. δc± (2. However. The ion concentrations c± cannot then exceed the steric limit 1/a3 . and Ajdari have recently extended this approach to the nonequilibrium case by deriving a modiﬁed Nernst–Planck equation based on Eq.4) for the electrochemical potential obviously becomes questionable long before the limit of steric exclusion sets in.21) (2. (2.20) (2.000 V. (2. the vanishing of the AC electroosmotic pumping at electrolyte concentrations above 10 mM [15].. essentially because the entropy associated with excluding the last solvent molecule from a condensed phase of ions is inﬁnite. Borukhov et al. In this regime the Gouy–Chapman–Stern model systematically underestimates the zeta potential.2.20) involves transfer of two electrons from the electrode to the Zn2+ ion. Such . ◦ E− = 0. but the violation of the steric limit is perhaps the most oﬀending feature in the Gouy–Chapman– Stern model. Bazant. Their modiﬁed Poisson–Boltzmann equation becomes ∇2 ψ = zeψ sinh kB T 1 kB T λ2 ze 1 − 2c∗ a3 + 2c∗ a3 cosh D zeψ kB T . They include the entropy of the solvent molecules in the expression for the free energy F . (2. and obtain for a symmetric binary z:z electrolyte with ions of characteristic size a the following expression for the electrochemical potential µ± = δF = kB T log(c± a3 ) + z± eφ − kB T log(1 − c+ a3 − c− a3 ). but may still give a good approximation to the overall capacitance provided the capacitance ratio δ is not too low and the nominal bulk concentration is not too high.18) where a3 is the volume occupied by a single ion.18) [56. e.19) Kilic.1 Classical theory 15 Further. ◦ E− = 0.22) ◦ where E− is the standard reduction potential. (2. have suggested a simple modiﬁcation of the Poisson–Boltzmann equation to account for the ﬁnite size of ions [55]. 2.3 Electrode reactions Electrode reactions are chemical redox reactions that involve transfer of one or more electrons from an electrode to the reactants. we also like stress that the ability of the double layer to absorb ions and act as a capacitor persist long after the steric limit is reached. 57].444 V. Familiar examples are Zn2+ + 2 e− (aq) 2 H+ + 2 e− (aq) O2(aq) + 2 H2 O(ℓ) + 4 e − ⇋ ⇋ ⇋ Zn(s) H2(g) 4 OH(aq) − ◦ E− = −0. The reaction in Eq.

(2. Likewise. the reaction in Eq. .16 State ‡ Free energy Electrokinetic theory ∆G‡ f State I ∆G‡ b State II Reaction coordinate Figure 2. single electron process of the form O + e− ⇋ R. (2. where zO = zR + 1 to ensure charge conservation.25) ◦ ◦ ◦ ◦ Here ∆G− = G− − G− = −eE− is the increase in standard free energy R O associated with the reaction. a tunneling process is very unlikely. To determine the reaction kinetics it is then suﬃcient to focus on this rate-limiting step whereas the other steps are considered as reaction equilibria. Often one step in such a multistep reaction is much slower than the rest.24) from which we obtain the Nernst equation for the equilibrium potential drop from the electrode to the electrolyte c∗ kB T ◦ (Vext − φ∗ )eq = E− + log O . and Eq. The activated complex (or transition state) ‡ is the conﬁguration of maximal free energy. O O (2.21) must also be a multistep reaction consisting of one step where H+ is reduced to form a single hydrogen atom adsorbed on the electrode.23) (2. In the present study we therefore assume a generic one-step. Nernst equation In thermal equilibrium the change in electrochemical Gibbs free energy associated with the reaction is zero ◦ ∆G = G− + kB T log c∗ + zR eφ∗ R R − + k T log c∗ + z eφ∗ − eV ◦ − GO B ext = 0. e c∗ R (2. followed by a second step where two hydrogen atoms join and desorb from the electrode as H2 .20) must therefore represent a multistep reaction where the Zn2+ is ﬁrst reduced to Zn+ and then Zn.3: Free energy landscape along reaction path from state I (reactant) to state II (product).

26) and similarly for the backward reaction νb νb ∝ exp ∆G‡ b − . Substituting into Eqs.26) and (2.28) Here the transfer coeﬃcient α is simply a measure of how close the transition state is to either the reactant or product state in terms of electrostatic energy.27) the net reaction rate becomes s s ◦ ◦ νnet = νf − νb = k− cs e−αe(Vext −φ )/kB T − k− cs e(1−α)e(Vext −φ )/kB T . (2.1 Classical theory Butler–Volmer kinetics 17 Out of equilibrium the forward and backward reaction rates no longer balance and a net reaction takes place..2.27) Further it is assumed that the electrochemical free energy at the transition state can be decomposed into a chemical term that is essentially independent of the applied potential. and an electrical term (G‡ )e that is a weighted average of the electrostatic energy of the reactant and product. cf.e. f b s and cs are the concentrations of the oxidized and reduced species at and cO R the surface.30) If we assume quasiequilibrium in the double layer we can relate the concentrations at the electrode surface to those immediately outside the Debye . i. The state ‡ of maximal free energy along the reaction path in conﬁguration space is denoted the transition state or activated complex.. i. (2. The Faradaic current injected from the electrode into the electrolyte by the reaction is jext = −eνnet . The forward reaction rate νf is proportional to the probability of getting from the reactant state into the transition state.29) f O b R ◦ ◦ where k− and k− are the forward and backward standard rate constants. 2. (2. According to transition state theory the rate limiting factor is associated with getting across a bump in the free energy landscape. (G‡ )e = (1 − α)(zO eφs − eVext ) + αzR eφs . and this probability drops exponentially with the diﬀerence in electrochemical free energy between the two states.e. (2. and φs is the potential in the electrolyte at the electrode surface (or outer Helmholtz plane). kB T (2. This result and its implications is known as the Butler–Volmer formulation of reaction kinetics [16]. νf ∝ exp − ∆G‡ f kB T . Fig.3. (2.

. Reversibility of the electrode reactions In our model calculations we always assume that the electrode reaction is reversible. Linearizing for small overpotentials we get jext = (z − α)eζ eq j0 e 1 exp − O ϑ. O R and the overpotential ϑ = (Vext − φ) − (Vext − φ)eq is the diﬀerence between the actual potential drop across the double layer and the thermal equilibrium for the given redox process and bulk concentration. . This phenomenon is called electroosmosis.31) ◦ ◦ where j0 = ek− (c∗ )α (c∗ )1−α is the exchange current. k− is the standard R O rate constant. by a change in pH. if the device is pumping from an inlet with “fresh” electrolyte to an outlet that is sent to “waste”.32) deﬁning the charge transfer resistance Rct . (2. If the device is pumping the electrolyte in a closed loop and for a long time then this should be a good approximation. However. In the bulk electrolyte the charge density ρ is small. While it may or may not be feasible or desirable to operate a pump in this way we do not account for it in our model. or by gradually dissolving the electrode material and ﬂushing it out to waste.29) and (2. cO and cR are the concentrations immediately outside the Debye layer. (2. but when a tangential ﬁeld acts on the charge in the Debye layers it can induce a ﬂow which in turn drags the bulk ﬂuid along. η the dynamic viscosity.1.4 Electroosmosis The ﬂuid motion in the electrolyte is governed by the Navier–Stokes equation ρm [∂t u + (u · ∇)u] = −∇p + η∇2 u + ρE. ρm is the mass density. ϑ= kB T kB T Rct (2. e. e − ∗e c∗ cO R (2. Then Eqs. This seems natural because our primary focus is on a micropump based on an array of electrode pairs that is driven with an AC voltage: We model this as an inﬁnite periodic system in space and look for steady-state periodic solutions in time after all transients have died out. p is the pressure.g.30) i can be re-expressed as the current-overpotential equation jext = j0 e−(zO −α)eζ/kB T cR (1−α)eϑ/kB T cO −αeϑ/kB T . c∗ and c∗ are the corresponding bulk values at equilibrium. 2. then the chemical composition in the electrolyte might change upon passage through the pump.18 Electrokinetic theory layer by a Boltzmann factor. by absorbtion of a small amount of electrolytic gasses. cs = ci e−zi eζ/kB T .33) where u is the ﬂuid velocity. and ρE is the electrical body force acting the ﬂuid.

On the other hand. which eﬀectively acts as a slip velocity for the bulk ﬂuid motion. rewriting the full Poisson–Nernst–Planck problem as a perturbation expansion in the small parameter λD ε= . 58].36) outside the Debye layer. (2.35) The ﬂuid velocity increases from no-slip at the surface to a constant value ux = − ǫζ Ex η (2.2 Mathematical model The solution of the full Poisson–Nernst–Planck equations is a challenging task.g. In deriving Eq. impermeable. and that only the screening layers at the boundary display structure on the Debye length scale. 2.37) ℓ0 .36) we associated the zeta potential with the noslip plane. if the charged surface has a thin adsorption layer of macromolecules or polymer gel then the electroosmotic ﬂow would be greatly suppressed due to the large hydrodynamic resistance of the polymer fragments [58]. 53]. This classical result is known as the Helmholtz–Smoluchowski formula [17.2 we deﬁned ζ as the potential at the outer Helmholtz plane. η (2.. (2. whereas in Sec. (2. but experiments for some particularly smooth. porous.2 Mathematical model 19 If we consider a uniformly charged ﬂat surface at y = 0 in contact with an electrolyte and with a tangential ﬁeld Ex acting on the charge in the screening layer then the tangenial component of the Navier–Stokes equation reduces to a balance between electrical and viscous forces 2 2 0 = η∂y ux − ǫ∂y ψ Ex . e. This simple picture can be formalized by perturbation theory. Yet our physical intuition is that the bulk electrolyte is simply charge neutral. even numerically the solution is made diﬃcult because the intrinsic length scale λD in the electrolyte is often several orders of magnitude smaller than the relevant geometrical length scale ℓ0 . Apriori there is no guarantee that the no-slip and the outer Helmholtz planes should coincide.34) Integrating twice and using that ux = 0 and ψ = ζ at the surface while ∂y ux = 0 and ∂y ψ = 0 in the bulk we obtain ux = − ǫ(ζ − ψ) Ex . and well-deﬁned surfaces indicate that they do [17.2. or inhomogeneous surfaces we cannot expect this.1. 2. for rough.

. The method of matched asymptotic expansions is a generalization of Prandtl’s viscous boundary layer theory. lim c± (x. However. but the higher order ¯± corrections c(i) can also contain important information. even if an inﬁnite number of terms are included [59]. where the inner approximation in the boundary layer is obtained by scaling the spatial variables to remove the singular perturbation before solving the problem. The primary interest is on the leading order term c(0) . (2. the driving voltage V0 for a truncation of the series to produce accurate results [37]. and it is a method for treating singular perturbations [59]. ¯ ¯± ¯± ¯± (2. ˜ ˜ ¯ (2. . The expansions need ¯± not always take the form Eq.38) y →∞ ˜ y→0 where y = y/ε is the rescaled spatial variable normal to the boundary in the ˜ inner region.. However. e. and logarithmic terms like ε log ε sometimes “pop up” in the process of deriving higher order terms [59]. and strictly there is no guarantee that the series will converge at all..20 Electrokinetic theory The Poisson equation turns out to be singularly perturbed in the sense that the small parameter ε multiplies the highest (second) order derivative. Essentially the method works by computing separate “outer” and “inner” approximations to the solution in the bulk and boundary regions. Because we focus on the periodic solution for an AC driving voltage after all transients have died out. and mathematically this singular perturbation can be seen as the origin of the disparate length scales appearing in the solution. We denote “weakly nonlinear” the regime where the leading order term in the standard asymptotic expansion remains a good approximation to the solution. there are actually three diﬀerent charac- . y. respectively. The standard procedure begins by seeking regular expansions in the form of power series c± = c(0) + ε c(1) + ε2 c(2) + . The two solutions are matched by requiring that the outer limit of the inner approximation c± ˜ should be asymptotically equal to the inner limit of the outer approximation ¯ c± .39) but can also contain fractional powers of ε.g.39) substituting into the governing equations. . experimentally there are several interesting and intriguing phenomena that show up at larger voltage. for any ﬁxed ε > 0 there could be ε dependent restrictions on. t). t) = lim c± (x. but by reordering the terms in the expansion to reﬂect the dominant balance we can develop another “strongly nonlinear” model valid in the joint limit ε → 0 and V0 → ∞. The approximations to the solution are “asymptotic” in the sense that they converge to the true solution in the limit ε → 0 with all other parameters in the problem held ﬁxed. y . so we would like to extend the validity of the model as far up in driving voltage as possible. and collecting like powers of ε. It turns √ out that for V0 ∼ kB T /ze ε the weakly nonlinear solution breaks down.e. i.

p c.2 Mathematical model The asymptotic model in a nutshell Nested boundary layers y c. ˇ ˇ ˇ γ . the Debye screening length λD . and the diﬀusion length ℓ(ω) = D0 /ω which is the typical distance that an ion can diﬀuse during one period of the driving signal. Likewise we can identify three different characteristic time scales. The potential drop ζ across the Debye ˇ ˜ layer determines the electroosmotic contribution to the slip velocity us . Here the diﬀusion length √ becomes ℓ(ω = 1/τ0 ) = λD ℓ0 . φ. check. the geometrical length scale ℓ0 . and tilde accents. p ˜ ˜ ˜ ˜ w± . u. φ. the “middle” diﬀusion layer of width ε. and the cell relaxation (RC) time τ0 = λD ℓ0 /D0 . u. ς . and the “outer” bulk region at the unit characteristic length scale [38]. p ¯ ¯ ¯ ¯ Variables 21 1 Bulk electrolyte Electrolyte √ ε ε Electrode Substrate √ y = y/ ε ˇ ˇ u. namely. the excess salt concentration γ = c − c is determined by a simple 1D diﬀusion problem. Applying the method of matched asymptotic expansions to study problems in electrochemistry and electrokinetics is not a new idea: Much work has been done on steady-state solutions for electrochemical cells operated at . AC electroosmosis occurs primarily for driving frequencies around the inverse RC time ω ∼ 1/τ0 . The Debye layer is in ˇ quasiequilibrium at the RC time scale with a structure that is completely determined by ∞ the excess amount of each ion w± = 0 (˜± − c) d˜ accumulated in the layer together with ˜ c ˇ y ˜ the salt concentration c seen from the Debye layer. namely.2. us ˇ ˇ ˇ y = y/ε ˜ c± . which means that in the asymptotic analysis we need to account for three nested regions: The “inner” screening layer of √ (dimensionless) width ε. teristic length scales in the system that we can identify. Variables associated with the diﬀerent regions are denoted by bar. The bulk is electroneutral with a salt concentration c that is constant in time but not necessarily in space. the charge relaxation (Debye) time 2 2 τD = λD /D0 . respectively. The diﬀusion layer is also elec¯ troneutral but has a dynamically varying salt concentration. us ˜ ˜ ˜ Diﬀusion layer Debye layer Electrode Figure 2. φ.4: Schematic picture of the nested boundary layers in the matched asymptotic expansion: The “outer” bulk region is connected via the “middle” diﬀusion layer to the “inner” Debye layer. The excess potential drop ˇ ˇ ¯ ˇ ς across the diﬀusion layer induces an eﬀective slip velocity us . ζ. the bulk diﬀusion time τℓ = ℓ0 /D0 .

We extend their work by taking into account dynamics in the diﬀusion layers when the system is driven into the strongly nonlinear regime by a large AC voltage around the inverse RC time. c0 φ ´ φ= . u0 (2. where the Debye– H¨ckel approximation ζ ≪ kB T /ze can be employed or the driving voltage u or applied ﬁeld is weak enough that the problem can be linearized. e.42) λD ℓ0 D0 (2. has been studied extensively. In the derivation we retain all terms that are important in the strongly nonlinear regime. τ0 c± = ´ c± . φ0 ´ u= u . The structure is sketched in Fig. ´= r r ..1 Dimensionless form For simplicity we assume a symmetric binary electrolyte with z+ = −z− = z. diﬀusion layer. Finally we discuss the conditions for validity or breakdown of both the weakly and strongly nonlinear models. they primarily focus on transient solutions for a suddenly applied DC voltage.2.41) is the average diﬀusivity. 2. Below we ﬁrst cast the problem into dimensionless form and identify the relevant dimensionless groups. and the low voltage regime. φ0 = kB T ze (2. and state the surface conservation laws governing the dynamics of the solution in the boundary layers. 61. Then we proceed determine the leading order solution in the bulk. ℓ0 t ´ t= . e.4. 58. 2. respectively. However. denoting dimensionless variables by an acute accent.g.22 Electrokinetic theory DC conditions. and strongly nonlinear dynamical models. weakly nonlinear.40) where ℓ0 is the characteristic geometrical length scale.g. τ0 = is the RC time for the system. [40. and propose an ad hoc modiﬁcation to the strongly nonlinear model when the double layer is driven out of quasiequilibrium at “very large” voltage. 60]. c0 is the nominal bulk salt concentration. 38. D0 = D+ + D− 2 (2. but we allow for asymmetry in the diﬀusivity D+ = D− . Dynamical solution have been looked at recently by Bazant and co-workers [37. and Debye layer. 39]. [23. Also we include electroosmotic ﬂuid motion into our model and account for bulk ion transport by convection.43) . we focus exclusively on the leading order solution and therefore drop the superscript (0) to simplify the notation.. 62]. The problem is cast into dimensionless form. which makes it easy in subsequent sections to extract the linear (low voltage).

50) (2.46) (2. By ´ ´ construction they add up to D+ + D− = 2.2. and 23 ǫφ2 0 (2. (2.52) D0 1 + Dr D0 1 + Dr where Dr = D+ /D− is the ratio of the cation to the anion diﬀusivity.53) 2 Dr + 1 ´ and is always in the range −1 < ∆D < 1.44) ηℓ0 is the characteristic electroosmotic velocity.2 Mathematical model is the thermal voltage. Here ε = λD /ℓ0 is the dimensionless thickness of the Debye layer. ´ 2 c ´ (2. (2. (2.51) Sc = ρm D0 D0 is the Schmidt number deﬁned as the ratio of momentum diﬀusion to ionic diﬀusion in the electrolyte. 2 1 ´´ 1 1 ´ p ´ 2´ u ´ u ∂t u + Pe(´ · ∇)´ = −∇´ + ∇ u − 2 ρ∇φ ´ ´´ Sc ε ε ´ ´ 0 = ∇ · u. (2.48) arise because we have chosen to focus on the ´ ¯ ´ dynamics at the RC time t and not the bulk diﬀusion time t = εt.45) (2.47) (2. Pe = u0 ℓ 0 ǫφ2 0 = D0 ηD0 (2.55) Da = D+ D− = (1 + Dr )2 . ´ ε t ´ ´ ´c ´´ F± = −D± ∇´± ± c± ∇φ + Pe uc± . The factors of 1/ε in front of the time derivatives in Eqs. and ν η = (2. whereas (half) their diﬀerence becomes 1 ´ Dr − 1 ´ ´ ∆D = (D+ − D− ) = . We also introduced dimensionless diﬀusion constants D+ 2Dr D− 2 ´ ´ D+ = = and D− = = . The characteristic diffusion constant for this quantity turns out to be the harmonic mean of the individual diﬀusivities 4Dr ´ ´ ´ . In dimensionless form the governing equations become u0 = 1 ´ ´ ∂´c± = −∇ · F± .45) and (2.48) (2.54) which we often denote as the “salt” concentration.49) is the P´clet number deﬁned as the ratio of the convective to the diﬀusive e ion ﬂux. It is convenient to operate also with the average ion concentration c = 1 (´+ + c− ). ´ ´´ 1 2 ´ ´ ´ c ´ −ε2 ∇ φ = ρ = (´+ − c− ).

in the numerical example above this corresponds to |u| ≪ 0.1 m/s.47) and equating like powers in ε we ﬁnd that the bulk . and p. Table 2. even at large driving voltage.2. The nonlinear convection term in the Navier–Stokes equation can ´ be neglected if |Re u| ≪ 1 which is usually the case. Eq.97.001 1 0.25 5 × 102 5 × 10−4 10−6 Electrokinetic theory Unit m2 s−1 m2 s−1 mV m−3 µm nm µs µm s−1 Scaling kB T /ze ǫkB T /2c0 z 2 e2 λD ℓ0 /D0 ǫφ2 /ηℓ0 0 λD /ℓ0 ǫλS /ǫS λD ǫφ2 /ηD0 0 ν/D0 ǫφ2 /ην 0 √ ∝ 1/ c0 √ ∝ ℓ0 / c0 ∝ 1/ℓ0 √ ∝ 1/ℓ0 c0 √ ∝ c0 ≡ Pe/Sc Table 2.1: Typical numbers with and without dimensions. but the common working electrolyte KCl is actually very close to symmetric with Dr ≈ 0.39) and similarly for φ.1 < Dr < 10. and also independent of ℓ0 because the characteristic velocity scales as 1/ℓ0 . (2. corresponding to a dilute 1 mM KCl solution at room temperature in a device with characteristic geometrical length scale ℓ0 = 10 µm.2 Electroneutral bulk transport The standard procedure of matched asymptotic expansions begins by seeking regular expansions (denoted here by bar accents) in the bulk “outer” region. The relative diﬀusivity is typically in the range 0. u. 2.1 shows typical values for the diﬀerent numbers with or without dimension in the problem. ¯ ¯ cf. it should be obvious from the context whether a given quantity is dimensionless or not. For the remainder of this thesis we drop the acute accent on dimensionless variables to reduce the notational overload. corresponding to a dilute 1 mM KCl solution at room temperature in a device with 10 µm characteristic length.24 Symbol ν D0 φ0 c0 ℓ0 λD τ0 u0 ε δ Pe Sc Re Expression Value 1× 2 × 10−9 25 6 × 1023 10 10 50 50 0. (2. Substituting ﬁrst into the Poisson ¯ equation Eq. Notice that the Reynolds number Re = Pe/Sc = u0 ℓ0 /ν is small. This is denoted the ambipolar diﬀusion constant and it is always in the range ´ 0 < Da ≤ 1.

58) and (2.2. the ambipolar diﬀusion constant Da [54]. We focus on the periodic solution reached after all transients in the system have died out. .45) we ﬁnd at leading order that ∂t c = 0. (2.62) where we then obtain2 ¯ ¯ ¯ ¯¯ Fa = 1 (D− F+ + D+ F− ) = −Da ∇¯ + Pe u c. c+ = c− = c.. c 2 ¯ ¯ ¯ −∇ · Fa = ∇ · Da ∇¯ − Pe u c = 0. i. and with a diﬀusivity that is simply the harmonic mean of the individual ion diﬀusivities.56) From the Nernst–Planck equation Eq.e.57) i.2 Mathematical model 25 charge density ρ vanishes to both zeroth and ﬁrst order in ε. ¯ i. In order for a periodic ﬁrst order correction to exist. that ∇ · F± = 0. (2..59) to eliminate ∇2 φ.63) At leading order in the Navier–Stokes equation for the ﬂuid motion we obtain (2.64) 2 It is an interesting feature of binary electrolytes that the evolution of the leading order ¯ bulk concentration proﬁle – at the bulk diﬀusion time scale t = εt – is determined by a simple convection-diﬀusion equation with no migration term. ¯ (2.60) ρ = −ε2 ∇2 φ = ε2 ¯ c ¯ ¯ using Eqs.e. ¯ ¯ ¯ (2. where ¯ c ¯ c J = −¯∇φ − ∆D∇¯.59) In the case of equal diﬀusivities where ∆D = 0 this reduces to an Ohmic current with spatially varying conductivity c. ¯ ¯ with the dynamics showing up only as higher order corrections.e.61) (2. (2. (2.58) (2. For a binary ¯ electrolyte this implies that the leading order ion concentrations are equal. c ¯ ∂t u = 0. (2. Writing the mean ion ﬂux as 1 ¯ ¯ ¯ ¯ ¯ F = 2 (F+ + F− ) = Fa + ∆D J. but more generally the second ¯ term related to the diﬀusion potential must also be taken into account. (2. Bulk charge neutrality together with the Nernst–Planck equation also imply that the ¯ ¯ ¯ leading order electric current density J = 1 (F+ − F− ) is divergence free 2 ¯ ∇ · J = 0. With this. it is necessary that the leading order time average ﬂuxes are divergence free. namely.. the leading order bulk concentration c = c(r) is stationary in time. the leading order charge density (that is to second order in ε) can be determined as ¯ ∇¯ · ∇φ + ∆D∇2 c c ¯ ¯ .

so that again the leading order ion concentrations are equal c+ = c− = c.2. 2. It may therefore be more appropriate to state the leading order problem as 1 ¯ ¯ ¯ ¯ ∂t u = −∇¯ + ∇2 u + ∇2 φ∇φ. cf.63) remains the same. the leading order ﬂuid velocity u = u(r) is stationary and determined by a time average force balance ¯ ¯ ¯ 0 = −∇¯ + ∇2 u + ∇2 φ∇φ .1.g. ˇ ˇ ˇ (2. Substituting into the Poisson equation (2. because the bulk transport of neutral salt is eﬀectively a diﬀusion process with diﬀusion constant Da and we consider driving frequencies around the inverse RC time. p εSc (2. p with the pressure p determined by the incompressibility constraint ¯ ¯ ∇ · u = 0.65) We denote as “bulk electroconvection” that component of the overall ﬂuid ¯ ¯ motion that is induced by the bulk electrical body force ¯ = ∇2 φ ∇φ f to distinguish it from.26 Electrokinetic theory ¯ ¯ i. (2.. (2. the ﬁrst order correction of O(εSc) may be not-so-small for the ε of interest to a microsystem. Because momentum diﬀusion is typically much faster than ionic diﬀusion corresponding to a large Schmidt number Sc ≫ 1. √ Further.68) To determine the dynamical solution we introduce the excess concentration γ = c − c in the diﬀusion layer relative to the bulk. Assuming the electrode is smooth on the diﬀusion length scale we model it locally as a ﬂat surface and deﬁne x as the tangential and y as the normal direction.47) we ﬁnd immediately that the charge density vanishes to zeroth order in ε.66) (2.. e. In any case the time average velocity entering Eq. However. and form weighted ˇ ˇ ¯ .3 Diﬀusion layer The Debye layer at the electrodes periodically absorbs and expels an excess amount of ions.67) for εSc ∼ O(1).e. Table 2. we introduce a scaled spatial variable y = y/ ε and seek regular ˇ asymptotic expansions (denoted by check accents) in the “middle” diﬀusion layer. electroosmotic ﬂow induced in the Debye layer. the disturbances do not have time to propagate into the bulk region. and the leading order dynamics only √ occur in a diﬀusion layer of width O( ε) around the electrode.

2.2 Mathematical model averages of the ionic ﬂuxes similar to Eq. (2.62) to get ˇ ¯ ∂t γ = −∇ · (Fa − Fa ) ˇ

27

1 2 2 ˇ = Da ∂y γ + εDa ∂x γ − ε Pe ux ∂x c + √ uy ∂y c − ux ∂x c − uy ∂y c ˇ ˇ ˇ ˇˇ ¯ ¯ ¯ ¯ ˇˇ ε 2 = Da ∂y γ + O(ε). (2.69) ˇˇ

At leading order the excess concentration in the diﬀusion layer is therefore determined by a simple 1D diﬀusion problem for y ∈ [0, ∞[. We argue ˇ below that the normal velocity component uy vanishes at leading order, which means that convection is indeed negligible provided |Pe ux ∂x c| ≪ ε−1 . Matching to the bulk solution requires that γ = 0 at y = ∞ whereas the ˇ ˇ boundary condition at y = 0 is determined by the excess ﬂux fed into the ˇ diﬀusion layer from the Debye layer 1 ˇ ¯ − √ Da ∂y γ = Fy,a − Fy,a ˇˇ ε

y =0 ˇ

ˇ ≡ fa .

(2.70)

**The solution can be expressed as a Fourier series γ (ˇ, t) = ˇ y ε ωDa √ ˆ f ˇ √a,n einωt− inω/Da y + c.c. in n=1
**

∞

(2.71)

**ˆ ˇ where fa,n is the nth Fourier component of fa (t) ˆ fa,n =
**

0 T

ˇ fa (t) e−inωt dt.

(2.72)

**Equivalently, the solution can be written as a convolution integral γ= ˇ ε 1 ωDa T
**

T 0

ˇ G(ˇ, t − t′ )fa (t′ ) dt′ y

(2.73)

**where G the periodic 1D diﬀusion kernel (Green’s function) G(ˇ, t − t′ ) = y √ ′ 1 ˇ √ einω(t−t )− inω/Da y + c.c. in n=1
**

∞

(2.74)

ˇ In the weakly nonlinear regime where the excess ﬂux fa is O(1) the excess √ concentration γ only exist as an O( ε) small perturbation to the bulk conˇ √ centration; the strongly nonlinear regime at driving voltages V0 ∼ 1/ ε is then essentially deﬁned as the point where the perturbation reaches O(1). Since the diﬀusion layer is charge neutral, the current must be constant across it at leading order, i.e., 1 ˇ ¯ Jy = − √ c∂y φ + ∆D∂y c = Jy = − c∂y φ + ∆D∂y c . ˇ ˇˇ ¯ ¯ ¯ ˇˇ ε (2.75)

28

Electrokinetic theory

ˇ ˇ ¯ Introducing the excess potential ψ = φ − φ and rearranging a bit we obtain √ ¯ 1 − 1 − ∆D∂y log c/¯ , ∂y ψ = − ε Jy ˇ c (2.76) ˇ ˇ c ˇ c ¯ ¯ ¯ √ ¯ where we used ∂y φ = ε∂y φ and similarly for c. Integrating once we get ˇ ˇy ψ(ˇ) = √ ¯ ε Jy

∞ y ˇ

1 1 − dˇ′ − ∆D log c(ˇ)/¯ . y ˇy c ′) c(ˇ ˇy c ¯

(2.77)

The interpretation is straightforward: The last term is the diﬀusion potential due to the diﬀerence in concentration across the layer, whereas the ﬁrst term is just the excess potential drop required to drive the bulk current through the diﬀusion layer when the local resistivity 1/ˇ diﬀers from c the bulk value 1/¯. In the weakly nonlinear regime where c is only a small c ˇ perturbation to the bulk concentration c, the magnitude of the diﬀusion ¯ √ potential −∆D log(ˇ/¯) is only O( ε), whereas the excess Ohmic potenc c tial drop is O(ε). In the strongly nonlinear regime both terms reach O(1) √ ¯ because log(ˇ/¯) ∼ O(1) and Jy ∼ 1/ ε. 3 c c Concerning the ﬂuid motion induced in the diﬀusion layer, the incompressibility constraint immediately yields uy = 0. In the normal direcion ˇ the Navier–Stokes equation reduces to a force balance between the excess electrical body force and the pressure at leading order 0 = −∂y p + ˇˇ from which 1 1 2ˇ ˇ ˇ ˇ ∂y ψ ∂y ψ = ∂y − p + ∂y ψ ˇ ˇ ˇ ˇ ε 2ε p= ˇ

2

,

(2.79)

1 ˇ 2 ∂y ψ . (2.80) ˇ 2ε In the tangential direction we obtain a balance between the viscous and electrical forces and the tangential pressure gradient

2 2ˇ ˇ 0 = −ε∂x p + ∂y ux + ∂y ψ∂x φ. ˇ ˇˇ ˇ

(2.81)

where ς = ψ(0) is the total excess potential drop across the diﬀusion layer. ˇ ˇ

ˇ Neglecting terms proportional to ∂x ψ because these are much smaller than ¯ ∂x φ we arrive at a Helmholtz–Smoluchowski type tangential velocity component ˇ ¯ ux = (ˇ − ψ)∂x φ, ˇ ς (2.82)

3 To make a rough estimate of the integral factor we can approximate c by a piecewise ˇ linear proﬁle c ≈ co + (¯ − co ) ω/Da y for y ≤ Da /ω and c ≈ c for y ≥ Da /ω to get ˇ ˇ c ˇ ˇ ˇ ˇ ¯ ˇ ∞ 0

1 1 dˇ ≈ − y − c ˇ c ¯

1 c c Da log(ˇo /¯) ∼ O(1). + ω c − co ¯ ˇ c ¯

(2.78)

2.2 Mathematical model

29

2.2.4

Quasiequilibrium double layer

The singular perturbation in the Poisson equation gives rise to a boundary layer of width ε where the charge density is nonzero to zeroth order in ε. Introducing a scaled spatial variable y = y/ε to remove the singular ˜ perturbation, we can seek regular asymptotic expansions (denoted by tilde accents) in the Debye “inner” layer. Substituting into the Nernst–Planck equations Eqs. (2.45) and (2.46) and using ∂y = ε∂y we ﬁnd at leading order ˜ 0 = ∂y ∂y c± ± c± ∂y φ , ˜ ˜˜ ˜ ˜˜ (2.83)

that is, the Debye layer is in quasiequilibrium and the ions are Boltzmann distributed ˜ c± = ce∓ψ , ˜ ˇ (2.84) ˜ ˜ ˇ where ψ = φ − φ is the excess potential in the Debye layer relative to the bulk, and c is the limiting value of the concentration in the diﬀusion layer ˇ seen from the Debye layer at y = 0. Likewise for the charge density and ˇ average ion concentration in the Debye layer we have ˜ ρ = −ˇ sinh(ψ) ˜ c and ˜ c = c cosh(ψ). ˜ ˇ (2.85)

The quasiequilibrium arises because the Debye layer reaches local equilib˜ rium on the Debye time scale t = t/ε which is much faster than the RC time of the system. Strictly, Eq. (2.83) does not imply that the leading order ﬂuxes 1 ˜ Fy,± = − D± ∂y c± ± c± ∂y φ ˜ ˜˜ (2.86) ˜˜ ε vanish but only that they are constant; it is by matching with bulk ﬂuxes ≪ O(1/ε) that we arrive at the Boltzmann proﬁles. Substituting into the Poisson equation, the leading order problem is the dimensionless Poisson– Boltzmann equation 2˜ ˜ ∂y ψ = c sinh(ψ), ˇ (2.87) ˜ which has the well-known Gouy–Chapman solution ˜ ˜ ψ = 4 tanh−1 tanh ζ/4 e−

√ cy ˇ˜

.

(2.88)

˜ The integration constant ζ is simply the leading order zeta potential, and √ ˇ 1/ c is the local eﬀective Debye length. As in the diﬀusion layer, the incompressibility constraint immediately yields uy = 0 at leading order, and the balance between the electrical body ˜ force and the pressure in the normal direction that gives rise to an (excess) osmotic pressure of 1 ˜ 2 p = 2 ∂y ψ . ˜ (2.89) ˜ 2ε

30

Electrokinetic theory

**Substituting into the Navier–Stokes equation in the tangential direction we obtain at leading order
**

2 0 = −ε2 ∂x p + ∂y ux − ρ ∂x φ. ˜ ˜ ˜ ˜˜

(2.90)

2˜ ˜ ˇ ˜ Writing ∂x φ = ∂x φ+ ∂x ψ and using again ρ = −∂y ψ we arrive at the general ˜ ˜ expression for electroosmotic ﬂow of the ﬁrst kind 4

˜ ˜ ˇ ux = (ζ − ψ)∂x φ + 4 log ˜

˜ cosh(ψ/4) ∂x log c. ˇ ˜ cosh(ζ/4)

(2.91)

The ﬁrst term is the familiar Helmholtz–Smoluchowski velocity due to the action of the tangential ﬁeld from the diﬀusion layer, and the second term is the diﬀusioosmotic ﬂow due to tangential gradients in the osmotic pressure in the Debye layer.

2.2.5

Surface conservation laws

The redistribution of ions across the Debye layer in the normal direction is instantaneous on the (RC) time scale that we consider here, but the total amount of ions accumulated locally can change only by ﬂux in or out of the layer or by tangenial ﬂuxes within the layer. Following Bazant and Chu [37, 39] we quantify this by deﬁning εw± (x, t) as the excess amount of ˜ each ionic species accumulated locally in the Debye layer, such that w± = ˜ 1 ε c± − c± dy = ˜ ˇ

∞ 0

Debye layer

(˜± − c± ) d˜. c ˇ y

(2.92)

**Likewise we deﬁne the surface charge and average excess concentration by q= ˜
**

0 ∞

(˜ − ρ) d˜ ρ ˇ y

and

w= ˜

0

∞

(˜ − c) d˜, c ˇ y

(2.93)

respectively, where by construction w± = w ± q . At leading order the ˜ ˜ ˜ integrals can be evaluated using Eqs. (2.84) and (2.88) to get √ ˜ ˇ w± = 2 c e∓ζ/2 − 1 , ˜ from which also √ ˜ q = −2 c sinh(ζ/2) ˜ ˇ

4

(2.94)

and

√ ˜ w = 4 c sinh2 (ζ/4). ˜ ˇ

(2.95)

The notion “electroosmosis of the ﬁrst kind” was introduced by Dukhin and co-workers to distinguish it from their “electroosmosis of the second kind” in double layers driven out of quasiequilibrium [41, 42].

.. cf.± d˜ − Fy. we introduced ˜ the surface excess ﬂux Γx.± − Fy.101) At leading order we can evaluate Γx.102) = −D± w± ∂x µ± + Pe ˜ ˇ (˜± ux − c± ux ) d˜.91) we get ∞ 0 (˜± ux − c ux ) d˜ = − 2 ± w± + c ˜ ˇˇ y ˜ − 2 w± + ˜ √ √ ˜ ˜ c W‡ ζ + ζ ˇ ˜ ˜ c W† ζ ± ζ ¯ ˇ ∂x φ ∂x log c.104) 2 32 2048 .100) y →∞ ˜ y →0 ˇ ˜ and wrote Fy.ext .g. This is a powerful result because it is independent of the particular model for the double layer [39].± by rewriting the ﬂux in terms of the electrochemical potential µ± = log c± ±φ to get F± = −D± c±∇µ± +Pe u c± . Next we partially integrated the normal ﬂux. (2.± − Fy. Eq. (2. Then ˜ Γ±.x = −D± ∞ 0 (˜± ∂x µ± − c± ∂x µ± ) d˜ + Pe c ˜ ˇ ˇ y ∞ 0 ∞ 0 (˜± ux − c± ux ) d˜ c ˜ ˇ ˇ y (2. However.± − Fy. y (2. and using Eq.98) (2. Finally.± .± given by ˜ Γx.97) (2.45) to eliminate the time derivative. it remains valid when the dilute approximation breaks down at large voltage.± ∼ lim Fy. i.x − F±.96) d˜ y (2.3).e.ext for the ﬂux injected into the Debye layer at the ˜ electrode surface.± = 0 ∞ ˜ ˇ F±. ˇ (2.2.± |y =0 = F±. (2. in order to handle the convection term we do need a model. c ˜ ˇ ˇ y where we used that µ± is constant across the Debye layer because it is in quasiequilibrium at leading order. used the ﬂux matching condition ˜ ˇ lim Fy. e.± − Fx.2 Mathematical model The time evolution of w± is determined by ˜ ∂t w± = ˜ ∞ 0 31 ∂t c± − c± d˜ ˜ ˇ y ∞ (2.99) =− 0 ˇ ˜ ˜ ˇ ε∂x Fx.x d˜. Here we used the continuity equation Eq.± y ∞ y =0 ˜ ˜ ˇ = −ε∂x Γx.± + ∂y Fy.± + F±.103) Here W † and W ‡ are shorthands for the expressions 1 9 7 4 W † (ζ) = −8 log[cosh(ζ/4)] − Li2 [tanh2 (ζ/4)] ≈ − ζ 2 + ζ (2.± − Fx.± ˜ ∞ = −ε∂x 0 ˜ ˇ ˜ ˇ Fx. and may even be generalized to account for interactions between the diﬀerent species by replacing the diﬀusivities D± by generalized mobilities Lij . by electrochemical reaction. (2.

and Shilov in their studies of surface conductance and the polarization of thin but highly charged double layers around spherical particles in weak applied ﬁelds [18. (2. q ˜ ˜ q ˇ (2.x + Γ−. but the results in Refs. (2. Deryagin. in the English translation it appears that there is a sign error on their term for ﬂux by electroosmotic convection proportional to ∂x log c. 39]. these subtle changes will be very hard to detect experimentally due to the factor of ε multiplying the divergence of the surface ﬂuxes in Eq. Dukhin and co-workers also accounted for electroosmotic convection and our result is fully equivalent to their work except for two features: First. which in our notation ˇ ˜ − 8 log[cosh(ζ/4)] ∂x log c. (2. 64. 65].x = −(D± + 2Pe)w± ∂x µ± .x ) as 2 ˇ x = − (1 + 2Pe)w + ∆D˜ ∂x φ − (1 + 2Pe)˜ + ∆Dw ∂x log c. 63. The presence of the dilogarithm terms qualitatively changes the form of the convective ﬂux at low zeta potential: The slope of the coeﬃcient to ˇ ˜ ∂x φ becomes nonzero at ζ = 0. Recently Chu and Bazant revisited the problem using matched asymptotic expansions which allowed them to lift the requirements for weak applied ﬁelds and small deviations from equilibrium in the bulk [38. 2 144 (2. Secondly.32 and Electrokinetic theory 1 1 3 W ‡ (ζ) = Li2 [tanh(ζ/4)] − Li2 [− tanh(ζ/4)] ≈ ζ − ζ . 65] do not.107) (2.103) satisﬁes this property.108) . What we have contributed is to extend their work to account also for surface convection and evaluate the integral in Eq. For a symmetric electrolyte ˜ ˜ reads 2 ± w± + ζ ˇ ˇ and φ while retaining c unchanged must result in a ˜ ˇ a change of sign of φ symmetric interchange of the cation and anion ﬂuxes. In this regime the surface excess conductance is dominated ˜ by the factors w± that grow exponentially with ζ and the expression for the ˜ surface excess ﬂux simpliﬁes to ˜ ˜ ˇ Γ±. Only at high zeta potential is the excess amount of ions accumulated in the Debye layer large enough to make the surface ﬂux term signiﬁcant.x ) and ˜ ˜ ˜ ˜ neutral salt ﬂux Γ = 1 (Γ+. Our Eq. ˜ ˜ q q ˜ ˇ and ˇ ˜ Γx = − (1 + 2Pe)˜ + ∆D w ∂x φ − (1 + 2Pe)w + ∆D˜ ∂x log c.x − Γ−. or for electrodes or polarizable objects subjected to large external voltages or applied ﬁeld. However. n=1 Surface conservation laws of this form were used already by Dukhin. the dilogarithm terms are absent in their work.105) and Li2 (z) = ∞ z n /n2 is the dilogarithm.103).99). [18. and coeﬃcient to ∂x log c becomes a nonˇ ˜ monotonic function of ζ.106) 1 ˜ ˜ Using this we can write the surface excess current x = 2 (Γ+. (2. which can occur either for colloidal particles with a large intrinsic zeta potential determined by their surface chemistry.

D+ + D− = 2.113) The surface conservation law for charge accumulation in the Debye layer becomes ˆ iω q = ext + n · ∇φ.c. 2 and likewise for the ionic concentrations ci (r. and strongly nonlinear regimes.95) as √ √ ˜ w = 4 c sinh2 (ζ/4) = q 2 + 4ˇ − 4ˇ.109) where ς . and −˜ δ are the potential drops over the diﬀusion. (2. q and ¯ ˜ w in the Debye layer. w+ − w− = 2˜. ˜ ˇ With this we are ready to formulate in the following sections the leading order dynamical models corresponding to the linear.111) Because we focus on driving frequencies around the inverse RC time.2. ˆ i (2. Since we drive the system with a harmonic signal. ¯ To close the coupled electrical problem for φ and c in the bulk. The second condition is a requirement of consistency in Eq. we need another two ˜ ˇ boundary conditions. The ﬁrst is the potential drop across the boundary layers obtained from the Gouy–Chapman–Stern model ¯ ˇ ˜ ˜ Vext − φ = ς + ζ − q δ. and δ is the surface capacitance ratio.112) (2. respectively.6 Linear model ˜ In the Debye–H¨ckel limit ζ ≪ 1 where the potential is small compared to u the thermal voltage the electrokinetic problem can be linearized. so the bulk ˆ concentration is uniform and Eq.2 Mathematical model 33 where we used that w+ + w− = 2w. (2. at leading order all variables can be expressed as an equilibrum value with a small deviation that varies harmonically in time ˆ ˆ φ(r. and c in the diﬀusion layer. Debye. (2. This acts ¯ as a Dirichlet condition for the bulk potential φ. and ˜ ˜ ˜ ˜ ˜ q D+ − D− = 2∆D. and ˇ ˜ q Stern layers. ˆ ˆ (2. the perturbations do not have time to propagate into the bulk.2. ζ. ˜ ˇ ˜ c c (2. t) = Re φ(r)eiωt = 1 φ(r)eiωt + c.110) √ ˜ ˇ ˜ where we used ζ = −2 sinh−1 q /2 c to express w directly in terms of the ˜ dynamical variables q and c. This simpliﬁes the analysis signiﬁcantly and this regime has therefore been studied extensively in the literature. t) = c∗ + Re ci (r)eiωt .58) reduces to a Laplace problem for φ ˆ −∇2 φ = 0. 2.114) . weakly nonlinear. (2.

115) In the absense of Faradaic electrode reactions the boundary condition for the potential therefore becomes [23] iω ˆ ˆ ˆ Vext − φ . and similarly approximating fR by −ˆext we then arrive at ext = ˆ 1 1 ˆ ˆ (Vext − φ) − ∗ Rct cR ε 1 ext − ∗ ˆ iωDR cO ε ext . (2. 1+δ (2.2. Approximating the ﬂux fO into the diﬀusion layer ˆ by the external current ext . and cO and cR are the concentrations of the oxidized and reduced species ˆ ˆ immediately outside the Debye layer. (2. Since the bulk concentration is uniform there is no body force at leading order and the Stokes ﬂow problem for the bulk ﬂuid motion is driven entirely by the electroosmotic slip.119) ˆ and similarly for cR . ˆ iωDO (2. which is equivalent to neglecting accumulation ˆ ˆ of O in the Debye layer.120) .3 to get cO |y =0 = ˆ ˇ ε ˆ f iωDO O (2. 2.118) where Rct = σkB T /j0 eℓ0 is the dimensionless charge transfer resistance. Faradaic current injection The Faradaic current is determined from Eq.34 Electrokinetic theory ˆ where ext is the Faradaic current injection at the electrode and n · ∇φ is the ˆ Ohmic current running into the Debye layer from the bulk. −n · ∇φ = 1+δ (2.116) The slip velocity condition for the time average bulk ﬂuid motion reduces to [33] 1 1 ˜ s¯ ˆ sˆ ˆ sˆ ˆ ˜ us = ζ∇ φ = Re ζ ∗ ∇ φ = Re (Vext − φ)∗ ∇ φ 2 2(1 + δ) 1 ˆ2 ˆ =− ∇ Vext − φ .32) as ˆ ext = 1 c ˆ cO ˆ ˆ ˆ (Vext − φ) + R − ∗ . The capacitance of the double layer at low voltage is Cdl = 1/(1 + δ) so the accumulated charge is related to the potential drop over the double layer by q=− ˆ ˆ ˆ Vext − φ .117) s 4(1 + δ) ˜ ¯ where we used that ζ = (Vext − φ)/(1+ δ) at low voltage and that ∇ Vext = 0 s because the metal electrode is an equipotential surface. Rct c∗ cO R (2. Assuming we can neglect migration of O and R in the diﬀusion layer we employ the result from Sec.

5: Equivalent circuit diagram for the linear model: The double layer capacitance Cdl is connected in parallel with the charge transfer resistance Rct for the Faradaic ˆ electrode reaction and the Warburg impedance ZW (ω) describing the eﬀects of mass transfer.5 with the double layer capacitance Cdl = 1/(1 + δ) connected in parallel ˆ to the charge transfer resistance Rct that is in series with ZW . In such experiments it is common practice to add an excess amount of supporting electrolyte that is The Warburg constant is often denoted σ in electrochemistry but we prefer the nonstandard symbol Rmt to avoid confusion with the electrolyte conductivity. 2.2 Mathematical model Rb ZW Cdl Rct Vext 35 Figure 2. while all are connected in series to the bulk Ohmic resistance Rb = 1. This way of representing an electrochemical cell is a standard tool in electrochemical impedance spectroscopy which allows experimental parameters such as Rct to be determined with high accuracy [16]. All are connected in series to the bulk Ohmic resistance Rb .124) The equivalent circuit diagram for the entire system is shown in Fig.123) + ∗ Rmt = ∗ cO DO cR DR Finally.2.121) √ (2. (2. (2. The expression in the parenthesis is called the Warburg constant and we denote it by Rmt 5 √ √ ε ε . substituting into the surface charge conservation law we obtain a closed relation between the complex bulk current running into the double layer and the potential drop across it ˆ −n · ∇φ = iω 1 + ˆ 1 + δ Rct + ZW (ω) ˆ ˆ Vext − φ . This can be rearranged to ˆext = j where ˆ ZW (ω) = c∗ O ˆ ˆ Vext − φ .122) is known as the Warburg impedance [16]. ˆ Rct + ZW (ω) √ ε ε + ∗ DO cR DR 1 √ iω (2. 5 .

e.g. The term proportional to ∆D in the nominator in Eq. min{c∗ . with the only diﬀerence being the deﬁnition of Rmt . 2 (2.126) Using this we ﬁnd a relation similar to Eq. c∗ } ≪ ε the mass transfer resistance is O R so large that the Faradaic electrode reaction is negligible alltogether. for ω ≫ ε. i. However.. because AC O R electroosmosis is strongly supressed at high electrolyte concentration it is relevant to consider the case when. √O R at very low concentration min{c∗ .125) ˆ where Da is the ambipolar diﬀusion constant and the ﬂux fa is given by 1 ˆ ˆ ˆ ˆ ˆ fa = 2 (D− F+ + D+ F− ) = 1 D− ext − iω∆D q . O R (2.127) + ˆ 1+δ Rct + ZW (ω) ˆ where now ZW is given by ˆ ZW (ω) = √ c∗ R √ ε ε + √ DR 2 Dr 1 √ iω (2.124) between the bulk current and potential drop across the double layer √ ∆D 1 − iωε (1+δ)√D iω a ˆ= ˆ ˆ −n · ∇φ (Vext − φ). zO = zR + 1 = 1 and c∗ = 1.. we must keep in mind that our model with thin diﬀusion layers is valid only when ω is well above the inverse bulk diﬀusion time.129) However. c∗ } ≪ 1. (2. iωDa (2.128) and Dr = D+ /D− .36 Electrokinetic theory electrochemically inert but serves to minimize the bulk Ohmic resistance.e. c∗ }2 . On the other hand. and it is √ negligible at leading order due to the factor of ε. (2. This also justiﬁes the neglect of migration of O and R in the diﬀusion layer and of accumulation of those species in the Debye layer. O Performing exactly the same analysis as above but taking into account migration of O in the diﬀusion layer we get cO = ˆ ε ˆ fa . i. . The eﬀect of mass transfer is therefore signiﬁcant in our model only if one of the electrochemically active species is present at a concentration much lower than the supporting electrolyte.e. In our formulation this corresponds to the limit c∗ ≪ 1 and c∗ ≪ 1.127) is related to the accumulation of O in the Debye layer. the species O also plays the role of cation in the supporting electrolyte. Therefore the case when O plays the role of cation in the supporing electrolyte (or R plays the role of anion) is formally equivalent to the that of excess supporting electrolyte.. (2. The Warburg impedance reaches O(1) only at driving frequencies below 2 ω ∼ Rmt ∼ ε/ min{c∗ .

131) 2 ¯ 0 = −∇¯ + ∇ u. Our weakly nonlinear model is therefore only valid at relatively low voltage V0 ≪ δ/ε.130) (2. p (2.133) which is matched with the following set of dynamical boundary conditions in the Debye layer ¯ ∂t q = n · ∇φ + jext . ¯ 0 = ∇ · u. Femlab’s standard time-stepping algorithm.2. 2. 2¯ (2. the capacitance ratio δ.138) For blocking electrodes where the Faradaic current jext vanishes.135) (2. When these conditions are satisﬁed the solution again simpliﬁes signiﬁcantly.132) (2.134) (2. The electrical part of the problem is nonlinear. (2. The leading order bulk problem is linear and given by c = 1.g. ˜ ¯ ˜ ˜ Vext − φ = ζ − q δ. namely. . ¯ 0 = −∇ φ. the weakly nonlinear model is described in terms of four dimensionless parameters. ¯ • The diﬀusion layer around the electrodes exists only as a small perturbation to the bulk solution. namely the driving voltage V0 and frequency ω.2 Mathematical model 37 2. but local in time so it is straightforward to integrate using. (2.7 Weakly nonlinear model Our notion of the “weakly nonlinear” regime is characterized by two features in the solution.2. q (2.137) As we discuss below in Sec..2. ˜ s¯ ¯ ˜ u = us = ζ∇ φ .136) where the zeta potential is related to the charge accumulated in the Debye layer by ˜ ζ = −2 sinh−1 (˜/2). The ﬂow problem is linear and can be computed from the time average slip velocity once the electrical problem has been solved. e.9 both the surface ﬂuxes in the Debye layer and the perturbations in the local salt concentration in the diﬀusion layer reach O(1) at suﬃciently large driving voltage V0 ∼ δ/ε. • The bulk electrolyte concentration is uniform and given by c = 1. ˜ and the equilibrium zeta potential ζ eq on the electrodes.

3 we discuss how it can be handled by replacing the standard time evolution with a relaxation method to determine the steady-state periodic solution in time while employing the commercial ﬁnite element software Femlab to solve the spatial problem [47].142) At large positive overpotential the Faradaic current jext saturates at the R R √ limiting value jlim (ω) = DR c∗ /ℓR (ω) = 1/Rmt ω. and after some rearranging we arrive at 1 T T 0 ˇ G(t − t′ )jext (t′ ) dt′ = ˜ ˜ √ 2 sinh ϑ/2 − Rct eζ/2 jext (t) . and similarly for O at large negatively overpotential.3 to get cO = c∗ + ˇ O ε 1 ωDO T T 0 ˇ ˇ G(t − t′ )fO (t′ ) dt′ . (2.140) assuming the bulk concentration cO is at the equilibrium value c∗ . 2. where ℓR (ω) = εDR /ω R is the diﬀusion length for species R. The numerical solution of the problem is more complicated than for the simple case with blocking electrodes because the solution is nonlocal in time.38 Faradaic current injection Electrokinetic theory When electrode reactions do occur the leading order Faradaic current jext is determined from the current-overpotential equation Eq. the linear analysis in Sec. 2.2. (2. DR (2. If this is not satisﬁed O the Faradaic reaction would invalidate our assumption of small perturbations to the bulk concentration in the diﬀusion layer and make the problem “strongly nonlinear” already at relatively low voltage V0 ∼ O(log ε) due to the exponential voltage dependence in Eq. e − ∗e ∗ e Rct cR cO (2. and ¯ O ˇ ˇ similarly for cR . Using fO = jext and fR = −jext and substituting into ˇ Eq. and we assume a symmetric transfer coeﬃcient α = 2 and take zO = zR + 1 = 1. In Appendix A. ω ˜ R ˜ O Rmt eϑ/2 + Rmt e−ϑ/2 (2.141) O R Here the constants Rmt and Rmt are the contributions from O and R to the Warburg constant. To further simplify the solution we also assume that there is excess supporting electrolyte such that c∗ ≪ 1.6 O Rmt = 1 c∗ O ε DO and R Rmt = 1 c∗ R ε .139) we obtain a single integral equation for jext (t).139). (2. cO and cR are the conˇ centrations of the oxidized and reduced species immediately outside the 1 Debye layer. cf.31) jext = ˇ ˜ ˇ 1 −ζ/2 cR ϑ/2 cO −ϑ/2 ˜ ˜ . ¯ ¯ ˇ ϑ = (Vext − φ) − (Vext − φ )eq is the overpotential. For c∗ ≪ 1 we can neO glect migration of O in the diﬀusion layer and employ the 1D solution from Sec.139) where Rct = σkB T /j0 eℓ0 is the dimensionless charge transfer resistance. . (2.2.

t).147) (2. together with the normal ˜ ˜ ¯n and salt ﬂux Fn that act as Lagrange multipliers in the model ˇ current J ˇ ˜ ∂t w = −Fn − ε∇ · Γ. ˜ s ¯ ∂t q = −Jn − ε∇ · ˜. (2. c (2. ¯ ¯ 0 = −∇¯ + ∇ u + ∇ φ∇φ . . c ¯ c ¯ ¯¯ Fa = −Da ∇¯ + Pe u c. namely.2 Mathematical model 39 2.151) (2. c (2.8 Strongly nonlinear model Our notion of the “strongly nonlinear” regime is characterized by the following features in the solution ˜ • The surface excess ﬂux Γ in the Debye layer induces O(1) bulk electrolyte concentration gradients. • We disregard Faradaic electrode reactions because the problem is complicated enough without.149) (2. an additional eﬀective slip from space charge in the diﬀusion layer. that are governed ¯ ¯ by ¯ 0 = ∇ · J. concentra¯ tion c(r). ¯ ¯ • In the diﬀusion layer the perturbations to the local electrolyte concentration caused by the periodic massive uptake and release of ions in the Debye layer reach O(1). although the leading order bulk concentration is constant in time c = c(r).144) 2 2¯ ¯ 0 = ∇ · u. and ﬂuid velocity and pressure u(r) and p(r). ˜ s w= ˜ q 2 + 4ˇ − ˜ c √ 4ˇ. • The ﬂuid motion has components from both electroosmotic slip from the Debye layer. the potential φ(r. • The Debye layer remains in quasiequilibrium but the solution depends on the local electrolyte concentration in the diﬀusion layer immediately outside the Debye layer. the accumulated charge q and excess neutral salt w. p (2.2.148) In the Debye layer we also have four dynamical variables. ¯ 0 = ∇ · Fa . and bulk electroconvection.145) (2.150) (2. ¯ In the bulk we have four variables.146) ¯ ¯ where the electric current J and ambipolar ﬂux Fa are given by ¯ J = −¯∇φ − ∆D∇¯. namely.2.143) (2.152) ¯ ˜ ˜ Vext − φ = ζ − q δ.

160) With this.154) (2. and ˇ s s ¯ because ς is small.159) ˇ ¯ ˇ where the ambipolar ﬂux entering fa is expressed in terms of Jn and Fn as ˇ ˇ ¯ ˇ ¯ fa = Fn − ∆D Jn − Fn − ∆D Jn .153) and we neglected the potential drop ς across the diﬀusion layer in the conˇ dition Eq.40 where Electrokinetic theory √ ˜ ˇ ζ = −2 sinh−1 q/2 c . t) = c + ˇy ¯ ε 1 ωDa T T 0 ˇ G(ˇ. the concentration c. that is determined by ˇ c(ˇ.162) ˇ Now. The . t − t′ )fa (t′ ) dt′ . the excess potential drop across the diﬀusion layer can be evaluated ∞ √ 1 1 ¯ dˇ − ∆D log(ˇ/¯).157) and the Debye layer also gives rise to an electroosmotic slip velocity of (2. namely. (2. ˜ s¯ ˜ us = ζ∇ φ . In the expressions for the surface exˇ q ˜ cess current ˜ and neutral salt ﬂux Γ we drop the diﬀusion term proportional ˇ to ∇ log c because this is negligible compared to ∇ φ at large voltage. hence moreover approximate the latter by ∇ φ ˇ s ˜ = − (1 + 2Pe)w + ∆D˜ ∇ φ. ˇ s¯ (2.158) where again we dropped the term proportional to ∇ log c and approximated ˇ s ¯ the tangential ﬁeld by −∇ φ.155) ¯ ˇ The Lagrange multipliers Jn and Fn determine the boundary conditions for the bulk current and ambipolar ﬂux running into the Debye layer ¯ ¯ n · J = Jn . ¯ ˇ ¯ n · Fa = Fn − ∆D Jn . (2. q ˜ s¯ (2. s In the diﬀusion layer we have just one dynamical variable. ˜ q s¯ ˜ Γ = − (1 + 2Pe)˜ + ∆D w ∇ φ. (2. y c c (2. we already argued that ς is small so one might ask if us is not negˇ ˜ ligible compared to the electroosmotic slip us from the Debye layer.156) (2.151) for the overall potential drop over the boundary layers because ς is small compared to −˜ δ.161) − ς = εJn ˇ c ˇ c ¯ 0 in terms of which the eﬀective slip velocity induced in the diﬀusion layer becomes ˇ us = ς ∇ φ . ˜ (2. y (2.

because the bulk concentration proﬁle is controlled only by ﬂux (Neumann) boundary conditions we need an additional constraint to ﬁx the total number of ions in the system. t+ ¯ T /2) = c(r.. e. t). ζ eq = 0. Once this problem is solved. which is also a relatively weak dependence. and the intrinsic ˜ zeta potential vanishes.164) ˜ where q eq = −2 sinh(ζ eq /2) is the equilibrium intrinsic surface charge. only the concentration c at y = 0 enters the coupled dynamical problem for ˇ ˇ ¯ ¯ ¯ ¯ ˜ ˜ ¯ ˇ ˇ φ. where the total amount of salt is ﬁxed by the initial condition.2. e.153). the excess concentration from the diﬀusion layer does not enter since it has no time average by construction. but it is necessary in our case because we compute the steady-state periodic solution directly using a relaxation method. ˇ ˇ we can compute the full proﬁle for c as a function of y and evaluate ς along ˇ the electrode surface. whereas the rest of the solution components have even symmetry. c dr + ε ¯ Ω w ds = ˜ ∂Ω Ω c∗ dr = |Ω|. namely.. c(r.g. w. however. φ(r. c. when the Gouy–Chapman solution in the Debye layer breaks down due to steric exclusion there may be viscoelectric eﬀects suppressing the electroosmotic ﬂow. c . In the ˜ special case where the ionic diﬀusivities are equal.151) and (2. Alternatively. Likewise we need to impose a constraint on the total amount of charge accumulated in the Debye layers q ds = ˜ ∂Ω ∂Ω q eq ds. and determine the slip velocity us and the corresponding bulk ﬂow component as a postprocessing step. (2. ˜ Eqs. u. t). if the system is not closed but open and connected to a large reservoir of electrolyte at unit concentration.g. Fn . t + T /2) = −φ(r. (2. ∆D = 0.2 Mathematical model 41 ˜ ˜ point is that ζ scales only as ζ ∼ 2 log q ∼ 2 log(V0 /δ) at large voltage. which is convenient. it is interesting to take into account other sources of electrokinetic ﬂow that could start to dominate in that situation. p.148). cf. assume that ˇ the component of the bulk ﬂow induced by us can be neglected in the convection of neutral salt in Eq. evaluating the integral over y by numerical quadraˇ ˇ ture. the problem has additional symmetry in ¯ ¯ time: The bulk potential φ and q and Jn in the Debye layer have odd ˜ ¯ ¯ symmetry across one half period in time. Upon this approximation. ˜ (2. Finally. Such a constraint would be redundant in an ordinary time evolution problem.163) because the nominal bulk concentration is exactly c∗ = 1. the total amount of ions is of course not ﬁxed. Moreover. but the salt concentration level is determined by a Dirichlet condition c = 1 ¯ at the reservoir boundary. (2. In order to simplify the numerical solution we do. ¯ . Jn . q . While a description of that is beyond the scope of this thesis.

which makes the assumption of local charge neutrality break down and with that also the leading order strongly nonlinear solution. for electrokinetic problems involving microelectrodes that are biased by an external voltage. At even larger voltage the double layer is driven out of quasiequilibrium and expands to form an extended space charge layer: We give a brief summary of the existing theory for the structure of this layer in the case of DC Faradaic conduction. ˜ (2. although certainly an experimentally relevant one.165) Bazant et al. and discuss how the key features in this solution could be incorporated into our strongly nonlinear model to extend its validity even beyond quasieuilibrium. As the driving voltage is increased the electrolyte concentration in the diﬀusion layer approaches zero. the induced zeta potential depends on the driving voltage in a nonlinear way and it is preferable to express the conditions for validity directly in terms of V0 rather than ˜ ζ. The common way of expressing the condition for validity of the weakly nonlinear description of an electrokinetic system is in terms of the Dukhin ¯ number Du = σs /σℓ0 ∼ ε˜ / J . V0 ∼ δ/ε.166) δ . Around this voltage both the bulk concentration gradients and the perturbations to the bulk concentration in the diﬀusion layer reach O(1). namely. (2. and also for initially uncharged polarizable particles subjected to an external DC or AC electric ﬁeld.9 Beyond quasiequilibrium In this section we ﬁrst develop an estimate for the driving voltage at which the weakly nonlinear model breaks down. (2.154) ˜ Du = ε(1 + 2Pe)w = ε(1 + 2Pe)4 sinh2 ζ/4 . cf. Eq.2. In colloid science it is natural to express the Dukhin number in terms of the zeta potential because for a colloidal particle the electrokinetic properties are often determined by the intrinsic zeta potential due to bound surface charges when the particle reacts chemically with the suspending electrolyte. deﬁned as the ratio of the double layer surface excess conductance σs to the bulk conductivity σ per geometrical length ℓ0 [58]: Surface conductance and distortion of the double layer can be neglected only when Du ≪ 1.42 Electrokinetic theory 2. showed that the same dimensionless group governs local salt depletion in the diﬀusion layer in the dynamical problem of a suddenly ˜ applied voltage. However. and they used the low voltage estimate ζ ≃ V0 /(1 + δ) to argue that for weakly nonlinear dynamics to hold the applied voltage cannot greatly exceed the thermal voltage [37]. To this end we observe that at large voltage the surface capacitance is dominated by the Stern layer so the accumulated charge can be estimated as V0 q∼ ˜ . so it would seem that the strongly nonlinear model applies only to a relatively narrow voltage range. The voltage where this occurs is also around V0 ∼ δ/ε. For a symmetric binary electrolyte it can be expressed as.

Likewise.169) reaches O(1). if an amount of ions εw (per unit area) is taken up by the Debye layer from a diﬀusion ˜ layer of width ε/ω.167) ˜ with ζ understood to be a function of q. (2.. (2. for V0 ∼ δ/ε assuming the system is driven around the inverse RC time with ω ∼ δ. and that the intrinsic surface charge ˜ q eq = −2 sinh ζ eq /2 can be disregarded. not before the driving voltage is well above the thermal voltage. The strongly nonlinear model both in the bulk and the diﬀusion layer is based fundamentally on local electroneutrality.01 and unit surface capacitance ratio δ = 1 [38. (2. because the bulk salt ﬂux F is only O(1). for an uncharged metal particle subjected to a DC electric ﬁeld where O(1) perturbations to ¯ the concentration were reached at a dimensionless ﬁeld strength of E ≃ 5 for dimensionless Debye layer thickness ε = 0. the surface excess salt ﬂux ˜ becomes signﬁcant when the ratio Γ ˜ εΓ εV 2 ∼ ε(1 + 2Pe)˜∇ φ ∼ 0 q s¯ ¯ δ F (2. This argument is consistent with our numerical results for pairs and arrays of microelectrodes. This shows that the surface excess e current should become signiﬁcant only at very large voltage V0 ∼ δ/ε.e. which occurs already for V0 ∼ δ/ε. the perturbation to the concentration in the diﬀusion layer can be estimated as √ εω V0 εw ˜ . The ˜ excess accumulated salt is bounded by w ≤ |˜| by construction and at large ˜ q voltage where the absorption is dominated by uptake of coions rather than expulsion of counterions we have w ≃ |˜|. Eq. (2. i. The leading order charge . or. How¯ ever.168) Du ∼ ε(1 + 2Pe)w ∼ ˜ δ assuming the P´clet number is O(1).166) is an upper bound on ˜ q where a more accutate estimate would be ˜ q∼ ˜ V0 ˜ δ + sech(ζ/2) .2 Mathematical model 43 provided that the driving frequency is at or below the inverse RC time such the screening is not negligible. 39]. If ε is small it is clear that the weakly nonlinear dynamics do not break down before the V0 signiﬁcantly exceeds unity. and also with those of Chu et al. If steric eﬀects are taken into ˜ account the Debye layer capacitance becomes a non-monotonic function of the accumulated charge. The Dukhin number can therefore ˜ q be estimated as εV0 .2.170) γ∼ ˇ ∼ δ ε/ω √ reaching O(1) for V0 ∼ δ/ εω . but q V0 /δ remains valid also in that case.

174) Nonequilibrium double layer structure When the electroneutral solution breaks down..171) Obviously. for ρ ∼ c. namely [42] • A quasiequilibrium “Debye” layer of width O(ε) at the electrode surface.172) In an electrochemical cell the “diﬀusion-limited” DC Faradaic current is O(1) which leads to the conclusion that the electroneutral bulk solution breaks down for c ∼ ε2/3 [60. .60) ¯ ρ = −ε2 ∇2 φ = −ε2 ¯ ∇¯ · J c ¯ . Rubinstein and Shtilman later showed that a region completely depleted of counterions can develop at an electrode or ion exchange membrane when driven above the diﬀusion-limited current [40]. cf. Eq.173) In eﬀect.44 Electrokinetic theory density in the bulk is nominally to second order in ε and for a symmetric electrolyte it can be expressed simply as. (2. |∂y cJy | ∼ ˇ¯ δ (2. 54]. which occurs when ¯ ¯ c ∼ ε2 ¯ ∇¯ · J c ¯ c2 ¯ or c ∼ ε2/3 ∇¯ · J ¯ c ¯ 1/3 . the structure of the double layer changes from quasiequilibrium to a nonequilibrium structure and expands in width from O(ε) to O(ε2/3 ). In this regime one can identify three sublayers within the nonequilibrium double layer. in our case of strong AC forcing ¯ applied to blocking electrodes the Ohmic currents involved and the average ion ﬂux fed into the diﬀusion layer at the electrode surface are much larger: ¯ Taking J ∼ V0 and using −∂y c = Fy ∼ ω w we obtain ˇ ˇ ˜ ωV02 . the leading order strongly nonlinear solution based on electroneutrality breaks down already for c ∼ (ωε)1/3 . (2. i. (2. ˇ taking V0 ∼ δ/ε. this was ﬁrst described by Smyrl and Newman [60]. c2 ¯ (2. that assumption breaks down when the charge density becomes comparable to the average concentration. • An extended “space charge” layer of width yo > O(ε2/3 ) that is completely depleted of counterions. as c approaches zero the charge density diverges which is incon¯ sistent with the underlying assumption of electroneutrality.e. However.

4 show that such a layer does indeed form. In the space charge layer migration dominates and the solution for the electric ﬁeld reduces to (denoting variables by a breve accent) ˘ ∂y φ = ± 1 ε ¯ ∓2Jn (yo − y)/D± . D± (yo − y) (2.176) and the (small) charge density from the coions by diﬀerentiation 2˘ ρ = −ε2 ∂y φ = ± ˘ ε 2 ¯ 2|Jn | . 66]. the Poisson–Nernst– Planck equations can be reduced to a single master equation for the electric ﬁeld whose solution can be expressed in terms of Painlev´ trancendents [42.175) where the positive sign applies to a space charge layer completely depleted of anions (˘+ = 0 and c− = 0) and the negative sign to depletion of cations. D± o (2.178) where a is determined by matching to the ﬁeld in the space charge layer a= ¯ |Jn |/2D± . (2. although it was argued by Bazant et al.179) and b is related to the total amount of charge q accumulated in both the ˜ Debye and space charge layers by q = ±2a coth(b).2 Mathematical model 45 • A “Smyrl–Newman” transition layer of width O(ε2/3 ) around y = yo connecting the space charge layer to the quasi-electroneutral diﬀusion layer.180) . In the existing theory for DC Faradaic conduction. The potential drop across the space charge layer is found by integrating Eq. e 60. ˜ (2.2. y (2. c ˘ and the thichkess yo of the layer arises simply as an integration constant.175) to get 2 ˘ ˘ ˘ Φ = φ(0) − φ(yo ) = ∓ 3ε ¯ 2|Jn | 2/3 y . To our knowledge this fundamental structure of the nonequilibrium double layer has only been studied for electrochemical systems in steady state with a DC Faradaic current. that a similar solution with a transient space charge layer should form also in systems driven dynamically at very large voltage [37]. Our numerical solution to the full Poisson–nernst–Planck equations in Chap. (2. (2.177) The ions in the inner O(ε) quasiequilibrium Debye layer are Boltzmann distributed and the solution for the potential takes the form [66] ˜ φ = ± log sinh2 (a˜ + b) + const.

During decharging the interpretation of the integration constant yo as the thickness of the space charge layer does not apply and matching to the solution in the diﬀusion layer seems impossible. ε2/3 (2. or vice versa at the positively biased electrode. (2. Moreover.159) even if that results in unphysical negative concentrations close to the electrode surface.g. Ad hoc modiﬁcation to strongly nonlinear model In order to incorporate the key features from the nonequilibrium double layer into our strongly nonlinear model we attempt the following modiﬁcations: • The breakdown of the quasiequilibrium solution is identiﬁed with the point where the concentration c in the diﬀusion layer drops to zero at ˇ the electrode surface. at the end of the charging ¯ period when Jn approaches zero before it changes sign.181) Finally. but not during decharging. This is indicated in Eq. e. 4 it appears that the space charge layer collapses rapidly when the current is reversed. • Based on the solution for c we determine yo (t) as the width of the ˇ region where c is negative during charging of the Debye layer. the thickness of the 2/3 /|J |1/3 ) Smyrl–Newman transition layer diverges. However.175) where the expression under the ¯ square root is positive only when Jn has the opposite sign of the coions. + 2 4a (2. ¯ i. but from our numerical solutions in Chap. Beyond this point we continue to solve a simple diﬀusion problem for the concentration proﬁle in the diﬀusion layer.182) When we try to apply this solution structure in a dynamical rather that steady-state setting there are several things we need to worry about. However. ˇ . It is not clear to ¯ O(ε n us how the dynamics at this point should be properly described.e. but at least the fundamental form of the solution is an invariant function of the rescaled spatial variable [42] y = (y − yo ) ` ¯ |Jn /D± |1/3 . developing during charging of the negatively ˘ ˘ biased electrode.. in the Smyrl–Newman transition layer |y − yo | ≤ O(ε2/3 ) all terms in the master equation play in.e.. the most important complication seems to be that the solution structure is only valid during the charging of the Debye layer. in the space charge layer the migration term must dominate over ∂t c± ˘ in the Nernst–Planck equation. (2.46 Electrokinetic theory The excess potential drop across the O(ε) Debye layer can be determined as 1 |˜| q ˜ ζ = ∓2 log . and the detailed structure of the Smyrl– Newman transition layer might change due to the dynamic nature of yo (t). we maintain Eq. i. when Jn < 0 for a space charge layer completely depleted of anions with c+ = 0 and c− = 0..

2.2 Mathematical model

47

we always set yo = 0 during decharging to model a rapid collapse of the space charge layer when the current is reversed. • The nonequilibrium zeta potential is modelled as ˜ ζ = −2 sinh−1 q ˜ , ¯ 2 c + 8|Jn /D± |yo ˇ (2.183)

where c is a strictly positive model for the concentration at the surface ˇ c = max{ˇ(0, t), ε1/3 }. ˇ c (2.184)

Here ε1/3 is a small number introduced to avoid division by zero at the point where c(0, t) and yo (t) vanish simultaneously, and the exponent ˇ ˜ is chosen in accordance with Eq. (2.174). This approximation for ζ is 1/3 and asympidentical to the quasiequilibrium solution for c(0, t) ≥ ε ˇ ¯n /D± |yo ≫ ε1/3 . totically equal to Eq. (2.181) for |J ˇ • When computing the potential drop ς across the diﬀusion layer in Eq. (2.161) we only consider the interval where c ≥ c. ˇ ˇ ˘ • We include the potential drop Φ across the space charge layer as given by Eq. (2.176) into the boundary condition for the bulk potential ¯ ˘ ˜ ˜ Vext − φ = Φ + ζ − q δ, but continue to exclude ς . ˇ • The relation between the surface excess accumulated charge and neutral salt remains ˜ w= q 2 + 4 max{ˇ, 0} − ˜ c 4 max{ˇ, 0}, c (2.186)

(2.185)

i.e., we simply impose w = |˜| in the nonequilibrium double layer ˜ q completely depleted of counterions. • The surface conservation laws can still be applied to the nonequilibrium double layer provided the thickness is small enough, yo ≪ 1. The (excess) surface ﬂux is dominated by the O(ε) Debye layer because the concentration in the space charge layer is small in comparison. Since the Debye layer is in quasiequilibrium the electrochemical potential of the coions is constant, which allows the diﬀusion and migration part of the surface ﬂux to be evaluated exactly as in Sec. 2.2.5 to get −w± ∇ µ± . However, the electrochemical potential is not constant ˜ s˘ across the space charge layer, so µ± = µ± . What we do assume is that ˘ ¯ ˘ µ± is dominated by the electrical term so that we can write µ± ≈ ±φ. ˘ ˘ The convection part of the surface ﬂux is more diﬃcult to evaluate, ˘ ˜ but for simplicity we continue to use Γ± = ∓w± (1 + 2Pe)∇ φ, which ˜ s is consistent with our assumption that the Helmholtz–Smoluchowski term dominates the electroosmotic ﬂow of the ﬁrst kind.

48

Electrokinetic theory • For the electrosmotic slip induced in the nonequilibrium double layer we have both the usual component of the “ﬁrst kind” from the O(ε) quasiequilibrium Debye layer ˜ s˘ ˜ us = ζ∇ φ, (2.187)

but also an additional component of the “second kind” from the space charge layer, for which we use the result of Rubinstein et al. [43] ˘ us = −

3 yo ¯ ˘ s¯ ∇ |Jn | + Φ∇ φ. s 9ε3 D±

(2.188)

Here the ﬁrst term is expected to dominate at very large voltage, but the second Helmholtz–Smoluchowski type term is signiﬁcant at more realistic ﬁnite voltages. Clearly, there are several crude approximations in this approach. Especially our model for the transport in the diﬀusion layer is oﬀending because of the region of negative concentration, which corresponds to ions absorbed by the double layer that were really not there, but should have been taken up from the region y ≥ yo . This eﬀectively means that we systematically underes˘ ˘ timate yo , and therefore also Φ and us . A more correct approach would be to solve a diﬀusion problem for γ in the dynamically changing domain ˇ √ ˇ yo (t) ≤ y < ∞ with the ﬂux boundary condition −Da ∂y γ = εfa applied ˇ ˇ ˇˇ √ at yo = εyo . However, this problem is signiﬁcantly more challenging to ˇ implement numerically than the simple form Eq. (2.159) so we have chosen not to pursue it. On the contrary, to avoid the complexity involved with determining yo from the spatial proﬁle of c dynamically and as a function ˇ of the position along the electrode surface, we make a simple estimate for it based on a quadratic extrapolation of c from the value at the surface, i.e., ˇ

2 c(ˇ, t) ≈ c(0, t) + ∂y c y + 1 ∂y c y 2 , ˇy ˇ ˇˇ ˇ 2 ˇˇ ˇ

(2.189)

**to get an algorithm of the form if c ≥ 0 or sign q Jn ≥ 0 ˇ ˜¯ yo ≈ 0, ˇ
**

2 elseif ∂y c < 0 ˇˇ

**yo ≈ −ˇ/∂y c, ˇ c ˇˇ else y0 ≈ − ˇ end ∂y c ˇˇ 2c + ∂y ˇ ˇ ∂y c ˇˇ 2c ∂y ˇ ˇ
**

2

(2.190)

−2

c ˇ 2ˇ . ∂y c ˇ

2.3 Summary

49

√ ˇ 2ˇ ˇ Because both c(0, t), ∂y c = − εfa /Da , and ∂y c = ∂t c/Da are readily availˇ ˇˇ ˇ able as solution variables at the surface, this estimate for yo is straightˇ forward and cheap to implement numerically. In Chap. 4 we compare our simple model to a full solution of the Poisson–Nernst–Planck equations in a simple 1D geometry, and ﬁnd that they agree remarkably well: Even at V0 = 200 for ε = 10−3 where the space charge layer expands a signiﬁcant √ depth yo ≈ 0.025 into the diﬀusion layer of characteristic length ε ≈ 0.03, the rough estimate from Eq. (2.190) comes within 10% of the true thickness of the space charge layer as observed from the full numerical solution, except just around the collapse of the space charge layer. Another issue is that the bulk concentration variations due to surface excess salt ﬂux in the Debye layer can become so large that c approaches ¯ zero. This occurs ﬁrst close to the electrode edges or other sharp features where the electric ﬁeld is particularly strong. We take this as an indication that space charge layers should develop that are steady, i.e., not transient, in space but with a charge density that changes with the polarity of the driving voltage. However, it is not clear to us how the dynamics in such a situation should be described. For the driving voltages that we consider in Chap. 4 the region where c approaches zero is small, so we simply choose ¯ neglect this problem except for another ad hoc modiﬁcation to our model: • We replace c with the strictly positive c = max{¯, ε1/3 } in Eqs. (2.59) ¯ ¯ c and (2.60) to avoid unphysically large predictions of bulk electroconvection and low conductivity.

2.3

Summary

In this chapter we ﬁrst gave a brief overview of some of the classical theory for electrochemical transport, electrode reactions, and the standard Gouy– Chapman–Stern model for the electrical double layer. Next we constructed a mathematical model based on matched asymptotic expansions in the three nested regions, namely, the “inner” Debye layer, “middle” diﬀusion layer, and “outer” bulk region. In this we build on recent work by Bazant et al. and Chu on dynamical solutions in electrochemical systems, where our contribution has been to write an explicit semi-analytical solution (in terms of a Fourier series) for the salt concentration in the diﬀusion layer developing when the system is driven with an AC voltage around the inverse RC time, and also to account for electroosmotic ﬂuid motion and convection of salt in the bulk region. We formulated separate model equations corresponding to the linear, weakly nonlinear, and strongly nonlinear regimes, which form the basis for our numerical results in Chap. 3 and 4. Finally, we derived an estimate for the driving voltage where the weakly and strongly nonlinear models break down, and also discussed how existing theory for nonequilibrium double layers in electrochemical cells driven above the diﬀusion-limited

50

Electrokinetic theory

current could be incorporated into our dynamical model. However, this attempt to extend the validity of our strongly nonlinear model beyond the breakdown of local electroneutrality in the diﬀusion layers has character of an ad hoc modiﬁcation and lacks the mathematical rigor otherwise pursued in this chapter.

[44]. 31]. The only free adjustable parameter is the surface capacitance ratio between the Debye and Stern layers. 30.1. 26. 29. 27. It was on a device like this that Studer et al. the theory is completely unable to account for reversal of the pumping velocity at higher frequency. . This discrepancy between experimental observations and predictions from existing theory has motivated us to extend the latter by including nonlinear capacitance in the Debye layer. This theory predicts a single maximum in the pumping velocity at a driving frequency around the inverse RC time. 32. This design was ﬁrst used by Brown et al. Another motivation for including electrode reactions into the theory was that the original results of Ajdari indicated reverse pumping in this case. 28. which allows the inverse RC time and to some extent also the magnitude of the pumping velocity to be ﬁtted to experiments. where the “forward” direction is deﬁned by the experimental observations of Brown et al.Chapter 3 Linear and weakly nonlinear analysis In this chapter we study the standard planar AC electroosmotic pump design with an asymmetric array of interdigitated thin microelectrodes on a planar substrate. but in the present chapter we take into account also the effect of mass transfer limitations on the Faradaic electrode reaction. Several groups have observed bubble formation and electrode degradation at driving frequencies below 1 kHz and voltages of a few volt [26. although in the low. However. and the predictions from the linear theory of Ramos et al. not high frequency limit [14]. [25] and later by several other groups [15. 15. 32]. which is experimental evidence that electrode reactions do actually occur. see Appendix B. with rapid and monotonical decrease in pumping at both lower and higher frequencies. and Faradaic electrode reactions in both a linearized scheme and using the full nonlinear Butler–Volmer reaction kinetics. 29. [33]. Most of our results have been published in the paper Ref. 31. observed pumping in the “reverse” direction [15]. as sketched in Fig. 3.

In order to determine the net pumping we observe that because the ¯ system is periodic in the x direction.1) . 3. As discussed in Chap.5. Brown et al. In order to reduce the number of geometrical parameters we assume an idealized geometry with inﬁnitely thin and ﬂat electrodes. originally used W1 = 4. Idealized geometry model As sketched in Fig. The ﬂuid ﬂow pattern is complex with vortices above the electrode edges. but in this chapter we study a design with (dimensionless) electrode widths and gaps W1 = 1. W2 = 7. 5 this choice is close to optimal in the sense of maximizing the pumping velocity for a given (low) driving voltage. but the qualitative features in the solution are not very sensitive to the particular choice of electrode geometry.5 µm. Each pair has a narrow electrode of width W1 and a wide electrode of width W2 separated by a narrow and wide gap of width G1 and G2 . but the spatial average in the x direction is a simple Couette ﬂow proﬁle.1 the device consist of an array of interdigitated electrodes. W2 = 25.7 µm.2 µm. and a channel height H ≫ L large enough to not inﬂuence the solution. where each pair has a narrow electrode of width W1 and a wide electrode of width W2 separated by a narrow and a wide gap of width G1 and G2 . and we assume that the transverse width of the channel is large enough that a 2D description in the xy plane is appropriate.6 µm. (3.5. Notice that in order to break the left-right symmetry it is necessary that both W1 = W2 and G1 = G2 . y) = ∞ n=0 ˆ un (y) ei2πnx/L + c. and the full period is L = W1 + G1 + W2 + G2 . respectively. On top of the array a microﬂuidic channel of height H is placed.c. G2 = 5.1: Sketch of the device geometry for the planar pump design with asymmetric pairs of interdigitated electrodes.52 L Linear and weakly nonlinear analysis Channel lid Average ﬂow is a simple Couette ﬂow H Electrolyte Fluid rolls over electrode edges U Electrodes Substrate y x W1 G1 x1 x2 x3 W2 x4 G2 x1 +L 2V0 cos(ωt) Figure 3. and L = 14. and L = 50 µm. G1 = 1. G1 = 4. the ﬂuid velocity u can be expressed as a Fourier series [33] ¯ u(x. G2 = 15.

The ˆ ˆ solution for u0 is simply a Couette proﬁle. (2. y) for an array of ﬂuid velocity line sources Ψ(x.4) where the Green’s function is given by H(x − x′ .1 Linear analysis at low voltage In the Debye–H¨ckel limit ζ ≪ kB T /ze the electrical problem can be linu earized. but also in the weakly nonlinear regime. cf. [33] Ramos et al. u0 = y U/H ex . whereas Q grows linearly with H.1 Linear analysis at low voltage ˆ where un is the nth Fourier component.5) We adopt this for calculating the streamfunction and visualizing the ﬂow pattern around the electrodes also at ﬁnite H ≫ L. ˜ electrodes (3.3) and the net ﬂow rate becomes Q = HU/2. 1 + δ Rct + Rmt / iω (3. In Ref. this Fourier analysis applies not only to the low voltage limit. or more generally to any slip velocity driven ﬂow above a planar surface. Moreover.124) ˆ −n · ∇φ = iω 1 ˆ ˆ √ + (Vext − φ). i. Eq.3. y) = y sinh 2πy/L . carried out a complete Fourier analysis in the ˆ limit H → ∞ and determined all components un (y) in terms of the slip velocity distribution us (x). ¯ (3.. y) us (x′ ) dx′ . 31]. To quantify the pump performance it is therefore convenient to report the “pumping velocity” U instead of the net ﬂow rate Q. Importantly. The net ﬂow rate is given by H 53 Q= 0 ux dy. only the ˆ zeroth order Fourier component u0 can contribute to the net ﬂow rate. When the channel height is suﬃciently large H ≫ L to not aﬀect the electric ﬁeld line distribution around the electrodes. the slip velocity and hence U becomes independent of H. and the dimensionless boundary condition for the Ohmic current running into the Debye layer becomes. y) = 1 L L 0 H(x − x′ . where U is the spatial average slip velocity U= 1 L us dx. 3.6) . cosh 2πy/L − cos 2π(x − x′ )/L (3. They also expressed the streamfunction Ψ for ˜ the ﬂow in terms of the Green’s function H(x − x′ .2) and due to incompressibility Q must be independent of x.e. the linearity of the Stokes ﬂow problem allows us to include (back) pressure driven ﬂow simply by superposition of a parabolic Poiseuille ﬂow proﬁle [25. ˜ (3.

whereas the time average electroosmotic slip velocity reduces to 1 ˜ s¯ ˆ2 ˆ ˜ us = ζ∇ φ = − ∇ Vext − φ . Only around the inverse RC time ω ∼ 1 when there is partial screening do we have both charge and . Also the inﬂuence of the capacitance ratio δ is almost trivial as it simply shifts the inverse RC time from ω ∼ 1 to ω ∼ 1 + δ. s 1+δ (3. with the understanding that the eﬀect of nonzero δ and other V0 1 is trivial.2: Pumping velocity U as a function of frequency ω. In our paper Ref. Conversely. 3.7) ˆ ˆ Here Vext − φ is the potential drop across the double layer. The dependence on the driving voltage in the model can be scaled out by rescaling the potential with V0 instead of the thermal voltage kB T /ze. so there is no tangential electric ﬁeld to drive the electroosmotic ﬂow. and the ﬂuid velocity by V02 . At low frequency ω ≪ 1 the Debye layers are fully charged and the screening is complete at all times. for a given electrode geometry there are only ﬁve dimensionless parameters describing the problem. the driving frequency ω and voltage V0 . and reduces the electroosmotic slip by a factor of (1 + δ)−1 . the surface capacitance ratio δ. Dashed line: Analytical result at low frequency where U ∝ ω 2 . The pumping is maximized around the inverse RC time ω ∼ 1 when the screening of the electrodes is partial.1. but here we simply set V0 = 1 and δ = 0 throughout the linear analysis. at high frequency ω ≫ 1 the screening is negligible. and Warburg constant Rmt .54 −2 Linear and weakly nonlinear analysis 10 10 −3 10 −4 U 10 −5 10 −6 10 −2 10 −1 10 0 10 1 10 2 ω Figure 3. charge transfer resistance Rct . so there is no charge in the Debye layers and again no ﬂow.1 Capacitive charging In the absense of Faradaic electrode reactions the only dimensionless parameter left in the problem is the driving frequency ω. namely. [44] we included this into the rescaling of the problem.

8 is shown in Fig. [33]..4). 3.0035.2 shows the pumping velocity U as a function of ω. the pumping U matches ω −2 for ω ≫ R−1 . and I is the total current running onto the electrodes through the power supply. 3. following Ramos et al. Fig. using more realistic boundary conditions [33]. tangential ﬁeld in the Debye layers.3: There is a uniform ﬂow from the left to the right far above the electrodes and more tortuous streamlines close to the electrodes.8 where the pumping velocity is maximized. including two recirculating vortices at the right edge of both electrodes. and the pumping falls oﬀ rapidly as ω 2 for ω ≪ 1 or as ω −2 for ω ≫ 1. (3. .1 Linear analysis at low voltage 55 Figure 3. The streamline pattern for the ﬂow at ω = 0. There is a unique maximum around ω ≃ 0. whereas the particular direction and magnitude was obtained by Ramos et al.e. Electrical impedance ˆ ˆ ˆ The impedance of the system can be determined as Z = ∆V /I. Notice that the stagnation point where the slip velocity changes sign is close to the center of the narrow electrode but is shifted more to the right on the wide electrode.3. 3. The streamlines are drawn as equidistant contours of the streamfunction Ψ.3: Streamlines for the time average ﬂuid ﬂow at ω = 0. Eq.1 This scaling of the pumping velocity at low and high frequency was predicted already in the original paper by Ajdari [14] for a simple generic model of an electrode array as a harmonic potential modulation at the bottom of the channel.8 where U = 0. where ∆V = 2 is the magnitude of the voltage diﬀerence applied between the ˆ electrodes. i. For a more realistic geometry with electrodes of ﬁnite thickness and radius of curvature R at the edges. cf. but this is an artefact due to our idealized geometry model with inﬁnitely thin electrodes. (3.2 it appears that U falls oﬀ faster than ω −2 at ω ≫ 1.8) 1 In Fig. I(t) = jext − ∂t (˜ + q electrode #1 ¯ ε n · ∇φ ) dx.

4: Bode plot of the dimensionless impedance Z as a function of frequency ω at zero capacitance ratio δ = 0: Magnitude (solid line) and phase angle (dashed line).2 ˆ For ω ∼ 1 we observe a slight frequency dispersion in Z as compared to a ˆ constant circuit element model Z0 = R∞ + 1/iωC0 . At low frequency the impedance is dominated by the capacitance C0 of the Debye layers on the two electrodes that are connected in series.10) whereas at high frequency the bulk Ohmic resistance R∞ = 0.. (3.24 (3.e. i. Here jext is the Faradaic current passed from the electrode into the electrolyte. = + = C0 W1 W2 1. −˜ is the charge on the electrode required to screen the accumulated q ¯ charge in the Debye layer.11) where φ is a solution to the Laplace equation with a Dirichlet boundary condition φ = Vext on the electrodes. 1 ˆ For comparison also the impedance Z0 = R∞ + iωC0 for a simple RC circuit with linear circuit elements is shown (dotted lines). 3. Likewise. 1 1 1 1 .8. which is due to a change 2 At high frequency the resistance R∞ can be computed with high accuracy from the power dissipation P = ∆V 2 /R∞ in the electrolyte 1 1 = R∞ ∆V 2 Ω ∇φ · ∇φ dr = 1.5 0 10 −2 −π/2 10 −1 10 0 10 1 10 2 ω ˆ Figure 3. −1 we can neglect the last term and obtain For ω ≪ ε ˆ I= 1 iω √ + 1 + δ Rct + Rmt / iω Vext − φ electrode #1 ˆ ˆ dx.56 2 Linear and weakly nonlinear analysis 0 1. The phase angle passes through −π/4 when the frequency is at the inverse RC time at ω ≈ 0.5 ˆ log10 |Z| 1 −π/4 0.88 dominates.9) Fig. and −ε n·∇φ is the dielectric polarization charge.125. the capacitance due to dielectric polarization at high frequency . ˆ θ(Z) (3.4 shows a Bode plot of the dimensionless impedance of the system in the absense of Faradaic electrode reactions and for zero capacitance ratio δ.

In general. 0. These diﬀerent regimes are illustrated in Fig. but holds for other choices of the electrode widths and gaps as well.46. that is. whereas for Rct ≤ 1 we see a minor downshift in the optimal frequency for pumping and a factor of 3 reduction in the magnitude of the pumping velocity. 3.2 Faradaic current injection When we include electrode reactions and Faradaic current into our model. (i) when capacitive charging dominates for ω ∼ 1 and either Rct 1 or Rmt 1. we expect AC electroosmotic ﬂow whenever there is partial screening of the electrodes. in particular because there is actually electroosmotic ﬂuid motion in this parameter range. and 0.1.2 for purely capacitive charging.125ε. one of the most important new features is that the complete screening at frequencies below the inverse RC time is unbalanced. Which is perhaps ˆ not surprising. (3.1. 3. where we see the peak around 2 ω ∼ Rct shifting down and even outside the ﬁgure window in Fig. given that both the impedance Z and the pumping velocity U are integral quantities over the electrodes. provided the electrode reaction is not limited by mass transfer. It turns out that the vanishing of the pumping is particular to the case when the charge transfer resistance takes the same value on both electrodes. 3.6) there are three regimes in parameter space where the surface impedance is O(1). (ii) when 2 the electrode reaction kinetics dominate for Rct ∼ 1 and Rmt ≤ ω ≤ 1. however. This is puzzling. In Fig. respectively. and 2 (iii) when mass transfer dominates for ω ∼ Rmt and Rct 1. namely. whereas if Rct is larger on the narrow electrode the pumping is in the “reverse” direction. so at the baseline Rct ≫ 1 we recover the previous results from Fig. Although we have not gained an intuitive understanding of why the 1 can be determined from the electric ﬁeld energy U = 2 C∞ ∆V 2 as C∞ = 2ε ∆V 2 Ω 1 ∇φ · ∇φ dr = 1. which coincides almost exactly with the frequency where the pumping velocity is maximized.5.8. The inverse ˆ RC time can be read oﬀ as the point when the phase angle of Z passes through −π/4.1 Linear analysis at low voltage 57 in the electric ﬁeld line pattern between low and high frequency.5(a) we have Rmt = 1. 3.5(b) to (d) show results for Rct = 0.5(d). is the peak in the pumping velocity for Rct ∼ 1 2 and Rmt ω 1. Inspecting the boundary condition Eq. 3. 3. If Rct is larger on the wide electrode than on the narrow electrode we do get pumping in the usual “forward” direction. 3. 2 (3.12) .1. where the pumping velocity U is shown as a function of ω and Rct for diﬀerent values of Rmt . but only no net pumping.01. see our paper in Appendix B for more details on this. namely at ω ≈ 0. Figs. What we do not see. This “coincidense” is not particular to our speciﬁc electrode geometry from Fig.3.

002 0. we do have a qualitative explanation for the direction of the pumping when Rct takes diﬀerent values on the two electrodes: The net pumping contribution from the narrow electrode tends to be in the reverse direction.003 100 101 102 ω ω Figure 3. Faradaic charging does not give rise to any pumping when Rct is uniform.1.58 10-2 Linear and weakly nonlinear analysis (a) 10-1 100 (b) 0.0 0.001 101 (c) (d) Rct 0. (c) Rmt = 0. or by geometrically lowering the reverse slipping sections from the channel surface as suggested by Bazant and Ben [45] and investigated in detail in Sec. whereas the wide electrode contributes in the forward direction.5: Contour plots of the pumping velocity U as a function of frequency ω and charge transfer resistance Rct for diﬀerent values of the Warburg constant Rmt : (a) Rmt = 1.003 0.. In the simple case with purely capacitive charging. 5. (d) Rmt = 0. one could inhibit the ﬂow from the reverse slipping electrode by modifying the intrinsic zeta potential on the electrodes as in Sec.g.003 0.001 102 103 10-2 10-1 100 0. by depositing a porous layer permable to ions but with a large hydraulic resistance. This idea can be useful in other designs as well where. e. which inhibits the electroosmotic ﬂow and leaves the other electrode dominating and determining the pumping direction. At lower frequency the Debye layer on the electrode where Rct is smaller tends to be “short-circuited” by the electrode reaction.46.0005 Rct 0. 3.002 0.3.002 0. the wide electrode dominates and the net pumping is in the forward direction.01.002 0. .001 10-2 10-1 100 101 102 10-3 10-2 10-1 0.001 101 102 103 10-3 005 0. (b) Rmt = 0.1.003 0.3 below.

This would give rise to an additional component to the ﬂow when the corresponding time average electric ﬁeld interacts with the constant intrinsic zeta potential. and yet there have been few theoretical investigations that go beyond the Debye–H¨ckel limit (or linear response) in the u context of AC electrokinetics. ˜eq (3. except that the RC time for the system is determined by the diﬀerential capacitance. (3.1. . However.3. which is well beyond the Debye–H¨ckel limit ζ ≪ kB T /ze ≃ 25 mV.2 Weakly nonlinear analysis 59 3.15) However.2. we also need to take into account the possibility that a time average current to second order in V0 could run between the electrodes. but we can still attempt a linear response analysis if the driving voltage is low enough V0 ≪ (1 + δ)kB T /ze.2. Of course. Our u previous linear analysis does not immediately apply to such systems.2 Weakly nonlinear analysis The weakly nonlinear model is a ﬁrst step towards better understanding of the dynamics at larger voltage.1. In the absense of Faradaic electrode reactions we obtain essentially the same results as in Sec.3 Linear response at nonzero intrinsic zeta potential Many electrode-electrolyte systems spontaneously form a Debye layer and possess an intrinsic zeta potential typically of the order ζ eq ∼ 100 mV.1. as we discuss at the end of Sec. When Faradaic electrode reactions do occur we need to replace Rct by the diﬀerential charge transfer resistance ˜ Rct (ζ eq ) = djext ¯ d(Vext − φ) −1 = Rct (0)e(zO −α)ζ . 3. and to demonstrate this we have included a code example in Appendix A.1.14) ˜ ˜ which can be come very small if ζ eq is large.13) Also the characteristic ﬂuid velocity is proportional. t) deﬁned only on the ˜ electrode boundary. t) with a special variable q(x. not to (1 + δ)−1 but to ˜ dζ 1 ¯ = 1 + δ cosh(ζ eq /2) ˜ d(Vext − φ) (3. using commercial ﬁnite element software like Femlab [47] this is straightforward to implement. at very large ζ eq where the Dukhin number reaches O(1) we would also have to include tangential surface ﬂuxes into the model. and the dimensionless frequency where the pumping occurs is therefore shifted down from ω ∼ 1 + δ to ω∼ − d˜ q ¯ d(Vext − φ) −1 ˜ = sech(ζ eq /2) + δ. 3. One reason for this may be that the numerical solution is slightly more complicated: Instead of a single linear problem for ˆ the complex potential φ(r) one has to solve a time evolution problem for ¯ the bulk potential φ(r. 3.

1 and ζ eq = 0. namely. Fig.6 that the pumping velocity falls oﬀ at large voltage. When the intrinsic zeta potential is nonzero. However.003 10-1 10-2 10-1 100 101 ω Figure 3.6 where V0 extends to 103 is far outside the range of validity of weakly nonlinear model for a typical microsystem.2.001 0.2 with a peak in the pumping velocity for ω ∼ 1 + δ. 3. the pumping velocity follows the linear response analysis from Sec. and the inverse RC time where the pumping is maximized drops to ω ∼ δ. the driving voltage V0 and frequency ω. 3. At the baseline V0 ≪ 1 we recover the result from Fig. cf. the results are largely unaﬀected by ζ We must emphasize that the upper voltage limit in Fig. but this ˜ s¯ ˜ is simply because the electroosmotic slip us = ζ∇ φ no longer scales as V02 ˜ ˜ at large voltage: There ζ ∼ 2 log(V0 /δ) rather than ζ ∼ V0 /(1 + δ).1. Nevertheless.6 shows the pumping velocity U as a function of V0 and ω for δ = ˜ 0. The dashed white line marks the frequency where the pumping velocity is maximized as a function of V0 . the discussion in Sec. 3. with the data rescaled to U/V02 to remove the inherent quadratic voltage dependence and ease comparison to the linear analysis. and well above that point ˜eq .9. 3. 2. at higher voltage the double layer capacitance is dominated by the Stern layer. 3.2. the conclusion that according to the weakly nonlinear model the only eﬀect of having a nonlinear surface . and the slip velocity grows only as V0 logV0 .6: Contour plot of the rescaled dimensionless pumping velocity U/V02 as a function of driving frequency ω and voltage V0 for a capacitance ratio of δ = 0. 3.60 103 Linear and weakly nonlinear analysis 102 V0 101 100 0.1 and no Faradaic current. and the intrinsic zeta potential ζ eq on the electrodes. It also appears from Fig.1 Nonlinear surface capacitance In the absense of Faradaic electrode reactions there are four dimensionless parameters in the model. ˜ the capacitance ratio δ.002 0.3 up to the point where the induced ˜ ˜ perturbations in ζ becomes comparable to ζ eq .

Although Rct is relatively large. Fig.1 and Faradaic O R current with charge transfer resistance Rct = 10 and Rmt = Rmt = 0.2 Nonlinear Faradaic current injection With electrode reactions included into our model. The dashed white line marks the frequency where the pumping velocity is maximized as a function of V0 .002 100 0.2 we know that if Rmt 1. capacitance is to downshift the inverse characteristic RC time and to change the scaling of U with driving voltage from V02 to V0 logV0 is important as a starting point for interpreting the results from the strongly nonlinear model in Chap.7: Contour plot of the rescaled dimensionless pumping velocity U/V02 as a function of driving frequency ω and voltage V0 . the exponential increase in Faradaic current with overpotential .1. so in order to actually see a nonvanishing eﬀect we consider O R Rmt = Rmt = 0. Further we take Rct = 10 and ζ eq = 0 on the electrodes. 3. 3. except that the scaling diﬀers by a factor of 1 + δ.5(c). and the dashed black line roughly marks the voltage where the limiting current is reached. 3.141). where again we rescaled to U/V02 to remove the inherent quadratic voltage dependence.3. and the intrinsic zeta potential ζ eq on the electrodes and channel substrate. From the analysis in Sec.0 03 0. the number of dimensionless parameters increases by the charge transfer resistance Rct .7 shows the pumping velocity U as a function of the driving voltage V0 and frequency ω. 3.001 10-1 10-3 10-2 10-1 100 101 ω Figure 3. mass O R ˜ transfer coeﬃcients Rmt and Rmt .1.2 Weakly nonlinear analysis 103 61 102 V0 101 0. for capacitance ratio δ = 0.1. the eﬀect of the Faradaic electrode reaction is always relatively small. (2. For V0 ∼ 5 there is a peak in the pumping velocity at low frequency which is due to Faradaic current injection. Some data points at large voltage V0 ≥ 102 are missing due to numerical diﬃculties with obtaining a self-consistent solution for jext (t) in Eq. 4.2. At the baseline for V0 ≪ 1 we recover the results from Fig.0 005 0.

Moreover. the nonlinearity induces an asymmetry between Rct on the two electrodes that leads to pumping in the forward direction.1. Speciﬁcally.. (3.7 roughly marks the voltage where the deviations in cO and ˇ cR relative to the bulk reach O(1).2 and in more detail in our paper [44] in Appendix B. beyond this voltage the diﬀusion-limited ˇ current is reached. 3.e. The peak fades out for ω ≪ 10−2 due to the mass-transfer limitations on the Faradaic current. always makes it signiﬁcant for V0 ≫ 1. the discussion in Sec. The corresponding DC tangential electric ﬁeld ¯ −∂x φ will interact with any intrinsic zeta potential on the electrodes or channel substrate and induce electroosmotic ﬂow. Eq. i.7. The nonlinearity in the reaction kinetics also implies that a DC Faradaic current can be running between the electrodes even when a harmonic AC voltage diﬀerence is applied..62 103 Linear and weakly nonlinear analysis 102 V0 101 0. this ﬂow does not give rise to any net pumping on a planar device because there is no global potential gradient along the electrode array.0004 0. i. the dashed black line in Fig. because the potential drop across the double layer is generally larger on the narrow electrode than on the wide.17). the Faradaic current saturates.8: Contour plot of the rescaled average tangential ﬁeld Es /V02 along the insulating substrate.. In the special case when ˜ ζ eq takes the same value on both electrodes and substrate. the eﬀective value of Rct is smaller. cf. L 0 ˜ ¯ ˜ ζ eq ∂x φ dx = ζ eq 0 L ¯ ˜ ∂x φ dx = ζ eq ¯ φ L x=0 = 0.e. cf. 3.0006 100 0. and for V0 ≫ 1 the system behaviour is again dominated by the capacitive charging.0002 0. 3. i.e.001 10-1 10-3 10-2 10-1 100 101 ω Figure 3. (3.16) . a feature which was missing in our earlier results in that paper. for the same parameters as in Fig. but above the diﬀusion-limited current Es saturates and the eﬀect becomes negligible compared to the “ordinary” induced-charge electroosmosis. Depending on the ˜ intrinsic zeta potential ζ eq this can induce pumping in either the “forward” or “reverse” direction.

and Fig.17) which gives rise to an additional component to the pumping velocity of ˜eq ˜eq ˜eq U ′ = ζe − ζs Es . if the (diﬀerence in) dimensionless intrinsic zeta potential is larger that 2.8 due to the rescaling with V02 .0012. Therefore. but we do not expect it to have much impact on the results in Fig.8 shows the average tangential ﬁeld along the channel substrate Es = 1 L − ∂x φ substrate ¯ dx. 3. As before the data are rescaled to Es /V02 to remove the inherent voltage dependence.e.7. the pumping induced by the time average ﬁeld becomes comparable in magnitude to the peak pumping velocity from the “ordinary” induced-charge electroosmotic ﬂow. 3. When the electrode reaction is fast enough to eﬀectively “short-circuit” the Debye layer. a DC Faradaic current implies a time average ﬂux of O and R between the two electrodes. Strictly. (3. the electroosmotic ﬂow is suppressed because it depends fundamentally on partial screening of the electrodes. Since the induced-charge electroosmotic ﬂow continues to grow as V0 logV0 . the electrode reaction can run no faster than the electrochemically active species O and R are transported to and from the electrode by diﬀusion. The ﬁgure shows that at low voltage and for ω ∼ 0. i. 3. this means that any net pumping from Es is negligible at large voltage. Taking such into account would reduce the value of the DC current and hence Es . However.3. where ζe is the constant intrinsic zeta potential on the ˜eq electrodes and ζs is the corresponding value on the channel substrate. Beyond the voltage where the diﬀusion-limited current is reached. ¯ ¯ O R which means that we neglected any bulk mass transfer limitations. 3.3 Summary 63 ˜ However. while the direction of this ﬂow component depends on ˜ ˜eq ˜eq the sign of ∆ζ eq = ζe − ζs .. where the nonlinear capacitance of the Debye layer causes the characteristic inverse RC . and depending on whether this mass-transfer limitation is weak or strong.1 there is a maximum in the average tangential ﬁeld along the substrate with Es = 0.3 Summary In this chapter we have studied the planar AC electroosmotic pump design used experimentally by several groups. In our model we assumed uniform bulk concentrations cO = c∗ and cR = c∗ of the electrochemically active species. Es saturates and therefore appears to fall oﬀ rapidly in Fig. extending the existing linear theory to account for Faradaic electrode reactions. We also extended the theory into the weakly nonlinear regime. the pumping velocity is either suppressed and shifted 2 down in frequency to ω ∼ Rmt or virtually unaﬀected by the electrode reaction. when the charge transfer resistance is small Rct ≪ 1. if ζ eq is nonuniform there can be pumping.

Neither the linear or weakly nonlinear models are able to account for the reversal observed experimentally at large voltage and frequencies above the inverse RC time. At larger voltage there is pumping at low frequency due to an asymmetry in the eﬀective charge transfer resistance between the narrow and wide electrode. and the electroosmotic slip to a scale as V0 logV0 with the driving voltage rather than V02 . entering only through λD in the surface capacitance ratio δ = ǫλS /ǫS λD . while otherwise the Faradaic current would induce bulk concentration gradients and strong concentration polarization which we consider “strongly nonlinear” phenomena. and their results are fully equivalent to ours [67]. Therefore these models are also unable to account for the experimental fact that AC electroosmosis is suppressed at concentrations above 10 mM [15]. and the direction of the induced pumping velocity depends on the sign of the (diﬀerence in) intrinsic zeta potential. have reported a nonlinear theoretical study of a travelling wave device. . In many aspects the weakly nonlinear model does not diﬀer qualitatively from the linear theory. Gonz´lez et al. Ultimately. the Faradaic current saturates and our weakly nonlinear model becomes dominated by capacitive charging. The nonlinearity in the Butler– Volmer electrode reaction kinetics gives rise to more dramatic results: A DC Faradaic current scaling as V02 at low voltage should interact with any intrinsic zeta potential on the electrodes or channel substrate. and could therefore be classiﬁed as a “circuit model” with nonlinear circuit elements. Another feature is that the bulk concentration almost completely scales out of the problem.64 Linear and weakly nonlinear analysis time to shift down from ω ∼ 1 + δ to ω ∼ δ. when the diﬀusion-limited current is reached and the concentration of O and R (dynamically) approaches zero at the electrodes. We should stress that this last conclusion is valid only because we assume an excess of inert supporting electrolyte.

2. The purpose of studying this system is both that it gives good understanding of the dynamical properties of the model. 4. and study the periodic response after all transients have died out. and electroconvection induced when a current is passed through concentration gradients in the electroneutral diﬀusion layer and bulk. but also that it allows us to test the validity by comparing with a dynamical solution of the full Poisson–Nernst–Planck equations. we investigate the net pumping induced in the asymmetric device. in Ref. namely.2 we study the electrokinetic ﬂow induced on both a symmetric and asymmetric electrode array.. and the (dimensionless) width of the electrode gap is L = 2. electroosmotic ﬂow of the ﬁrst and second kind. Finally. 4. In Sec. [37].1 1D model problem In this section we study the dynamics of the strongly nonlinear model in a one-dimensional geometry as shown in Fig. we neglect Faradaic electrode reactions.1. i. except that we apply an AC voltage rather than a step voltage. We ﬁrst apply the model to a simple 1D geometry in Sec.Chapter 4 Strongly nonlinear analysis In this chapter we present numerical results obtained using the strongly nonlinear model derived in Chap. The system is identical to that studied by Bazant et al. Of course. which is necessary when we later interpret the results in higher dimensions. the symmetric device does not pump. 4. but it helps us gain understanding of the system behaviour. Such a distinction between diﬀerent ﬂow components is artiﬁcial in the sense that experimentally only the overall ﬂow is observed. but it allows us to monitor the strength of the diﬀerent ﬂow components. 4. .1 with an electrolyte conﬁned between parallel-plate blocking electrodes.e.

1 so the only dynamical ¯ variables in the electroneutral bulk are the potential φ(y. (4.6) (4. some kind of electroconvective instability occurs. The electrolyte is conﬁned between parallelplate blocking electrodes separated by a gap of width L = 2.1: Sketch of 1D model problem. so the leading order bulk concentration is simply constant c(y. and a harmonic potential of ±V0 sin(ωt) is applied to the left and right electrode.163) to impose mass conservation ¯ ¯ across the entire system we obtain co = 1 − ǫ w .2) φ = φo (t) − co ¯ ¯ where φo is the potential at the centre of the cell. The bulk ﬁelds are matched to the solution in the boundary layers at y = ±1 by requiring ¯ ∂t q = ±J. but we do not consider that here.5) (4.3) (4. and using Eq.4) (4. there can be no time average ﬂux from the ¯ boundary layers into the bulk. By symmetry there can be no ﬂuid motion. t). ˜ (4. respectively.7) ¯ J ¯ ˜ ˜ Vext − φo ± = ζ − q δ. t) = co . In one dimension the solution to the leading order bulk problem simpliﬁes a lot. Moreover. so the overall voltage drop across the system is 2V0 sin(ωt). co ¯ √ ˜ ζ = −2 sinh−1 q /2 c . (2. ˜ ˇ √ w = q 2 + 4ˇ− 4ˇ. ¯ ˜ (4.1) ¯ The bulk electric current J (t) is constant in space and the potential is given by ¯ J(t)y ¯ ¯ . of course. ˜ ˜ c c in the Debye layers. ˜ ˇ ∂t w = ±F . 1 . t) and the average concentration c(y. where c is the concentration in the diﬀusion layer seen ˇ Unless.66 −1 0 Strongly nonlinear analysis 1 y Electrolyte 2V0 sin(ωt) Figure 4.

We do not include the excess potential drop ς across ˜ ˜ ˇ the diﬀusion layer into Eq. ˇ and ζ ˜ ˇ w(t + T /2) = w(t).8) The bulk potential must have odd symmetry upon reﬂection in space and translation by one half period in time ¯ ¯ φ(−y.4. δ = 1. and ∆D = 0. 4. except when ζ changes sign and the Debye layer capacitance shortly drops to 1.1 1D model problem from the electrode surface c = co ∓ ˇ ¯ ε 1 Da ω T T 0 67 ˇ ¯ G(t − t′ ) F (t′ ) − ∆D J (t′ ) dt′ . and using N = 128 grid points in time. The nonlinearity is not very pronounced and the system behaves almost as an ideal linear ˜ ˆ impedance Z = 1 − i. t).g.11) see Appendix A. and the zeta potential ζ q across the Debye layer. N/2 ¯ J(t) = ′ N ˆ Jm eimωt = n=1 m=−N/2 ¯ = J(tn )SN (t − tn ). (4. F .1. (4. At this driving voltage ˜ the Debye layer capacitance is large so ζ is small and the potential drop across the double layer is primarily on the Stern layer.10) where SN is the periodic sinc function 1 SN (t − t ) = N ′ N/2 ′ m=−N/2 eimω(t−t ) . ′ (4. and c have even symmetry. whereas w. e.9) ¯ ¯ which implies that also J and φo must have odd symmetry in time. at a relatively large voltage V0 = 50. namely. As usual for an RC circuit driven at a frequency around the inverse . (4. [68] for details. the Stern layer potential drop −˜ δ. The solution for the diﬀerent components in the overall potential drop across ¯ c (half) the system. e. is shown in Fig. (4. algebraic.5) because this vanishes at low voltage and typically remains small compared to Ohmic potential drop across the bulk at larger voltage.3 or Ref... ω = 1. and integral equations are solved numerically by substituting a truncated Fourier series and using a collocation scheme. the central potential φo vanishes and both q ˜ ˜ have odd symmetry.1 Dynamical solutions We solve the problem for ε = 0. The coupled diﬀerential.g.2(a). In the ¯ special case where ∆D = 0.001. the Ohmic potential drop J/¯o across the bulk ˜ electrolyte. t + T /2) = −φ(y. 4. which results in a small decrease in the current.

4. ς 0. Therefore the frequency ω = 0. but because the intrinsic capacitance ratio δ is small. and V0 = 25. the potential drop across the Debye and Stern layers are comparable.1. 4.2: Dynamical solution to the 1D model problem as a function of time for two sets of parameter values: (a) V0 = 50.68 (a) 40 Strongly nonlinear analysis (b) 20 20 10 Potential 0 Potential 0 −20 0.2(b) shows the solution for a diﬀerent set of parameter values where now ω = 0.5 0 20 40 60 c ˇ 1 c ˇ 1 2 3 4 5 6 1 0 0 0 0 10 20 30 40 50 60 Time t Time t Figure 4.2(a) also shows the surface concentration c that oscillates at ˇ the double frequency 2ω. Therefore we also expect the pertur˜ bations to the concentration in the diﬀusion layer to grow linearly with V0 .1.1 is actually also a little below the inverse RC time. the driving voltage V0 sin(ωt) (dotted ).2(a). the phase angle of the current is around π/4 ahead of the driving voltage. and δ = 0. whereas the charge accumulated in the Debye layer is π/2 ahead the current. The inset (solid ). Stern layer potential −˜ δ (dashed ). namely.5 0 −10 0. and zeta potential ζ√ q ¯ c shows the potential drop ς across the diﬀusion layer which is O( ε) smaller than J/¯o . 2. The top panel shows the diﬀerent components to the overall potential drop across (half) ¯ c the system. while at higher voltage our estimate ˜ ˇ w ∼ V0 /δ from Sec.5 0 1 2 3 4 5 6 −20 2 −0. The concentration variations in the diﬀusion layer are directly linked to the √ amount of salt accumulated in the Debye layer. (b) V0 = 25. ˇ RC time. The current has a signiﬁcant bump at the point ˜ where ζ changes sign and the double layer capacitance drops from ∼ 1/δ to 1/(1 + δ). as shown in the inset.4. bulk Ohmic potential J /¯o ˜ (dash-dot). ω = 0.5 0 −40 2 −0. Again.2. and δ = 1. 4. The maximal amount of charge q and salt w accumulated in ˜ ˜ the Debye layer is about a factor of 3 larger than in Fig. the diﬀusion ˇ layer potential drop is small. ˇ For comparison Fig. Fig.9 applies. ω = 1. At low voltage ˜ w = 4 c sinh2 ζ/4 scales as V02 . but the variations in the surface concentration c are similar because the diﬀusion ˇ √ length ℓ(ω) ∝ 1/ ω is also increased by roughly a factor of 3. at this voltage the Debye layer capacitance is large.1. ˇ The bottom panel shows the concentration c in the diﬀusion layer as seen from the surface.1. δ = 0. Although the variations in c are O(1).

4 0. and ˇ ˜ maximal values of the zeta potential maxt ζ (upper solid line) and diﬀusion layer potential drop maxt ς (lower solid line). as a function of the driving voltage V0 for ω = 1 and δ = 1.3. as suggested also in Sec. ˜ ςmax . When c approaches zero both ζ ˇ and ς diverge logarithmically. (4.6) and similarly when computing ς . and therefore scales as V03 at low ˜ voltage changing to V02 at higher voltage.6 0. ε1/3 } ˇ ˇ ˇ c in Eq. 4.6) and similarly when computing ς . cf. ˇ and the result is shown with dotted lines in Fig.6 1. ζmax ˇ .3.9. for V0 ≫ 1.4.2 0 0 V0 Figure 4. 4.9 the leading order strongly nonlinear solution based on local electroneutrality in the diﬀusion layer breaks down already when c ∼ ε1/3 .8 0. The excess potential drop across the diﬀusion layer depends on both the current passing though the layer and the perturbations to the concentration. 2. ε1/3 } in Eq. However. the ad ˇ c ˇ hoc modiﬁcations suggested in Sec.2 1 0. 2. this divergent behaviour of ζ discussed in Sec.8 1. Fig. and 10 there is a clear linear dependence. The divergence is suppressed by replacing c with c = max{ˇ. while at higher voltage the scaling is ˜ only logarithmic with ζ ∼ 2 log(V0 /δ).3 shows the minimal concentration mint c in the diﬀuˇ sion layer as a function of the driving voltage V0 at ω = 1 and δ = 1.2. 4.4 cmin ˇ 1. Around V0 ≈ 90 indeed for V0 the minimal concentration approaches zero.3: Minimal salt concentration mint c in the diﬀusion layer (dashed line). 2.9.2. which marks the breakdown of local electroneutrality in the diﬀusion layer close to the electrodes. The dotted lines show the result when this is suppressed by ˇ ˇ replacing c with c = max{ˇ. (4.1 1D model problem 20 18 16 14 12 10 8 6 4 2 20 40 60 80 0 100 69 2 1. ˇ Around V0 ≈ 90 the concentration c approaches zero which results in unphysically large ˇ ˜ values of both ζ and ς .2. which in turns prevents c from going negative ˇ ˇ ˜ and ς is unphysical: As in Fig. The ˜ ﬁgure also shows the maximal values of ζ and ς . At low voltage the zeta ˇ ˜ potential ζ ∼ V0 /(1 + δ) grows linearly.

4 as a number of snapshots over one half 1 period in time. or even ﬁner at larger voltage.47) reduce to 2 −ε2 ∂y φ = ρ.e.. t − co + c ˇ ˜ ¯ . t − c(0. t − c(0. In 1D this is straightforward to achieve with a graded mesh.45) to (2.2(a). 4.5% in the diﬀusion layer and around around 2.12) √ where y = (1 ± y)/ǫ and y = (1 ± y)/ ǫ are the rescaled spatial variables in ˜ ˇ the inner and middle regions.e. δ = 1.001 it is necessary to use a very ﬁne ﬁnite element mesh of hmin 10−4 close to the electrodes in order to resolve the highly compressed Debye layer structure.15) ∂t ρ = ε∂y (∂y ρ + c∂y φ). The ﬁrst frame at t = 16 T is just about when the charge 2 4 density in the Debye layer changes sign..14) (4.70 Strongly nonlinear analysis 4. i. t) = co + c ¯ ˇ 1+y 1+y √ . 4.2 Comparison to a full numerical solution of the Poisson–Nernst–Planck equations Based on the matched asymptotic expansions we can construct a uniformly valid approximation in space by adding the solutions from the inner. holding 5 them at t = 16 T . The agreement between the full numerical solution and the asymptotic expansion is very good. cf.17) (4. so for t = 16 − 16 T we see the Debye layer charging up and accumulating a lot of ions.2(a).5% in the Debye . When solving this problem for ε = 0.18) 0 = −∂y ρ − c∂y φ. t) . For a symmetric electrolyte in one dimension Eqs. t) is shown in Fig. ∂t c = ε∂y (∂y c + ρ∂y φ).16) (4.001. t) ˇ ε ε 1−y 1−y + c √ . and V0 = 50. middle.1. before it decharges and expels them back into the diﬀusion 6 8 layer for t = 16 − 16 T . In order to validate the asymptotic model we can compare this uniformly valid approximation to a full numerical solution of the Poisson–Nernst–Planck equations in time and space. but in 2D or 3D it can become a serious problem. with boundary conditions at y = ±1 Vext − φ = ∓εδ∂y φ. Fig. ε = 0. (4. (2. with a maximal relative error less than 0. c(y. 0 = −∂y c − ρ∂y φ. ω = 1. (4. and outer regions. The results for the average concentration c(y. We compare the two models for the same parameter values as in Fig.13) (4. t − co + c ˜ ¯ . ˇ ε ε (4. 4. i. and subtracting the overlaps.

5 1 0. Also the leading order approximation c in the diﬀusion layer alone ˇ is shown (solid line). [37].04 0.001.5 1 0. δ = 1.05 and δ = 0. and the agreement between the two is very good. the error for ω ∼ 0.5 t= 5 T 16 t= 2 T 16 2 1. ω = 1.5 1 0.5 2 1.02 0.5 1 0.4.1.4: Snapshots of the average concentration c in the diﬀusion layer over one half period in time for ε = 0. and V0 = 50.04 t= 1 T 16 2 1.. and.7 extends across the entire system. Notice that the vertical axis does not go to zero.5 2 1. e.5 t= 6 T 16 t= 3 T 16 2 1.5 1 0.5 t= 8 T 16 0. which are the parameter values used throughout Ref.1 at the inverse RC time is large because the diﬀusion layer of thickness ℓ(ω) ∼ ε/ω ≃ 0.02 0. for ε = 0. If we increase ε the quality of the leading order asymptotic approximation of course decreases.5 2 1. The ﬁgure compares a full numerical solution of the Poisson–Nernst–Planck equations (circles) to the uniformly valid approximation from the asymptotic model (dotted line).5 t= 7 T 16 t= 4 T 16 2 1.5 1 0.06 1+y 1+y Figure 4. which is only O( ε) smaller that ˇ the bulk Ohmic potential drop.g. even at very small ε the leading order strongly nonlinear solution breaks down when the driving voltage is .06 0 0.1 1D model problem 71 2 1. √ layer. The error is of the order O( ε) because we neglected the excess po√ tential drop across the diﬀusion layer ς . However.5 1 0.5 0 0.5 1 0.

ω = 1.05 0.125 1+y 1+y Figure 4. at a driving voltage V0 = 200 which is large enough to drive the double layer out of quasiequilibrium.125 0 0 0. The ﬁgure shows the full numerical solution for the concentration of the cations c+ (solid 6 3 line) and anions c− (dashed line).075 0.075 0. Between t = 16 T and 16 T a layer completely depleted of cations appear between the electrode and the electroneutral diﬀusion layer.001. and δ = 1.1 0.5: Individual ion concentrations for ε = 0.05 0. .1 0.72 Strongly nonlinear analysis 4 t= 3 1 T 16 4 t= 3 5 T 16 2 2 1 1 0 4 0 4 t= 3 2 T 16 t= 3 6 T 16 2 2 1 1 0 4 0 4 t= 3 3 T 16 t= 3 7 T 16 2 2 1 1 0 0 4 t= 3 4 T 16 4 t= 3 8 T 16 2 2 1 1 0 0 0.025 0.025 0.

2. and identify the unphysical region of negative “salt” concentration with the width yo of the space charge layer.4. we continue to solve a simple 1D diﬀusion problem for c even beyond the ˇ breakdown of local electroneutrality.1 is also low but ﬁnite. 4. This is the extended space charge layer discussed in Sec.2. Finally. while the anion concentration c− ∼ 0. For ε = 0. which was ﬁrst described by Rubinstein and Shtilman for a permselective membrane driven above the DC “diﬀusion-limited” current [40]. Fig. (4. At 5 4 t = 16 T and 16 we can clearly distinguish a region completely depleted of cations. ω = 1. (2. and for 8 7 t = 16 T and 16 T the diﬀusion layer is quickly reﬁlled by an equal amount of anions expelled from the Debye layer and cations transported by migration from the opposite electrode. and vice versa during the ˘ second half period when a space charge layer completely depleted of anions develops.001 and ω = δ = 1 this occurs around V0 ≈ 90. Eq. When the driving voltage is further increased beyond the point where the concentration at the surface approaches zero. δ = 1. 4. and a layer almost completely depleted of ions develops between the Debye layer and the electroneutral diﬀusion layer. respectively. again as a number of snapshots over one half period for ε = 0. ρ where the charge density ρ is given by Eq. from a full solution to the Poisson–Nernst–Planck equations.9. namely. The solid and dashed lines show the cation and anion concentrations. which is immediately before the current reverses. Fig. 2.6 q shows that the asymptotic model captures the overall dynamical behaviour . and V0 = 200. cf. Here we see the same basic structure of the nonequilibrium double layer developing and dynamically 6 propagating into the diﬀusion layer.5. The ﬁgure also shows the results from our strongly nonlinear asymptotic model when the ad hoc modiﬁcations suggested in Sec.19) ˜ sinh2 (a˜ + b) y ¯ cf.3. in the O(ε) quasiequilibrium “Debye” layer the leading order anion concentration is 4a2 2˜ c− = |2˜| = 2|∂y φ| = ρ . (2. followed by a rapid uptake of ions for t = 16 T to 16 T . we see the space charge layer retreating. the prefactor 3 is particular to the simple 1D geometry that we study here. 4. and for other values of ε and δ the breakdown occurs approximately at V0 ≈ 3 δ/ε. 4. while in the space charge layer c+ = 0 and c− = |2˘|. 1 2 Between t = 16 T and 16 T we see the charge density in the Debye layer 3 5 change sign. but the scaling with δ/ε should hold generally. This is shown in Fig. At t = 16 T . Fig.001. 4.1 1D model problem 73 increased enough that the solution in the diﬀusion layer approaches zero. 2. In the electroneutral diﬀusion layer for y ≥ yo (t) we have c+ = c− = c ˇ at leading order. the Debye layer does not stop charging but continues to take up ions from the diﬀusion layer.6 shows the same data as Fig. where a = |Jn /2D− and b = coth−1 (|˜|/2a).177).9 are employed.178).5 but on double logarithmic axis which better allow the diﬀerent boundary layers to be distinguished.

6: Individual ion concentrations: Same data as in Fig.5 but plotted on double logarithmic axis. Comparison of the full numerical solution for the concentration of the cations c+ (stars) and anions c− (circles) to the asymptotic model for the cations (solid line) and anions (dashed line). 4.74 Strongly nonlinear analysis 10 2 t= 1 T 16 10 2 t= 5 T 16 10 1 10 1 10 0 10 0 10 −1 10 −1 10 −2 10 −2 10 2 t= 2 T 16 10 2 t= 6 T 16 10 1 10 1 10 0 10 0 10 −1 10 −1 10 −2 10 −2 10 2 t= 3 T 16 10 2 t= 7 T 16 10 1 10 1 10 0 10 0 10 −1 10 −1 10 −2 10 −2 10 2 t= 4 T 16 10 2 t= 8 T 16 10 1 10 1 10 0 10 0 10 −1 10 −1 10 −2 10 −4 −2 10 10 −3 10 −2 10 −1 10 0 10 −4 10 −3 10 −2 10 −1 10 0 1+y 1+y Figure 4. .

.1 1D model problem 75 10 4 10 4 10 3 10 3 10 2 t= 1 T 16 10 2 t= 5 T 16 10 1 10 1 10 0 10 0 10 4 10 4 10 3 10 3 10 2 t= 2 T 16 10 2 t= 6 T 16 10 1 10 1 10 0 10 0 10 4 10 4 10 3 10 3 10 2 t= 3 T 16 10 2 t= 7 T 16 10 1 10 1 10 0 10 0 10 4 10 4 10 3 10 3 10 2 t= 4 T 16 10 2 t= 8 T 16 10 1 10 1 10 0 10 −4 0 10 10 −3 10 −2 10 −1 10 0 10 −4 10 −3 10 −2 10 −1 10 0 1+y 1+y Figure 4. 4.7: Magnitude of the electric ﬁeld E = −∂y φ for the same parameters as in Fig. Comparison of the full numerical solution for the ﬁeld (stars or circles depending in the sign of E) to the asymptotic model (solid or dashed lines).5. The agreement is good 6 7 except for t = 16 T and 16 T just before and after the space charge layer collapses.4.

part of the reason for which is ¯ simply that we have not included the O(ε2/3 /|Jn |1/3 ) Smyrl–Newman transition layer. And the quality of the asymptotic model within the space charge layer is generally not as good as in the other regions.03. Fig. and therefore the divergent tail of ρ close to yo is visible in ˘ the plots. having said this.7 compares the magnitude of the electric ﬁeld E = −∂y φ from the asymptotic model to the full numerical solution. there is a visible diﬀerence between the solutions inside the O(ε) Debye layer. both in the electroneutral bulk and diffusion layer. Fig. 4. to the width of the region of negative concentration from the asymptotic model. there is √ a signiﬁcant discrepancy in the diﬀusion layer around 1 + y ∼ ε ≃ 0.8 compares the actual thickness of the space charge layer from the full numerical solution.9 that the nonequilibrium solution breaks down close to the point where the current reverses. the agreement is good. but once the space charge layer has formed the asymptotic model is 6 7 spot on the full numerical solution in the Debye layer. except in an O(ε2/3 ) region around ˇ ¯ c yo where the ﬁeld from the diﬀusion layer −∂y φ = Jn /ˇ diverges. except close to the point where the space charge layer collapses. To further support this. A solution for c with the ambipolar ﬂux boundary condition ˇ ˇ applied on a dynamically changing domain yo ≤ y < ∞ would certainly imˇ ˇ prove the quality of the model. However. 2 At t = 16 T . and to the approximate expression for yo based on Eq. 4. The concentration exceeds 102 at 1 + y = 10−4 and increases to almost 104 for 1 + y ∼ 10−5 . We already argued in Sec.2. deﬁned (arbitrarily) as the width of the region where the cation or anion concentration drops below ε2/3 .76 Strongly nonlinear analysis of the full numerical solution well. because migration then no longer dominates in the space charge layer. Again. (2. we still ﬁnd that our simple model captures the qualitative features of the full numerical solution well. whereas the approximate expression tends to underestimate. For t = 16 T and 16 T . immediately before and after the collapse of the space charge layer. Both are within roughly 10 % of the full numerical solution. Finally. and in the quasiequilibrium or nonequilibrium double layer. allowing for negative values of c. Returning to Fig. and because the thickness of the Smyrl–Newman transition layer diverges. 2. just after the charge density in the Debye layer has changed sign. The ﬁgure shows that the width of negative concentration tends to overestimate yo . For a dilute electrolyte with nominal bulk concentration of 1 mM this corresponds to a concentration of about 1028 ions per m3 or 10 ions per nm3 at the surface which is approx- . and except 6 7 for t = 16 T and 16 T immediately before and after the current reverses and the space charge layer collapses.190). This accuracy is good enough that we dare employ the simple ad hoc modiﬁcations to the strongly nonlinear model and use it below to predict the direction and magnitude of electroosmotic ﬂow of the second kind. Also.6 we notice that the Debye layer is highly compressed at this large voltage. 4. much of the discrepancy is due to our crude model for salt diﬀusion.

1. In particular. This means that the Gouy–Chapman–Stern model or more generally the use of dilute solution theory to describe the structure of the O(ε) quasiequilibrium Debye layer at the electrode becomes ˜ questionable. 2.1 1D model problem 0. are still relevant.03 77 Thickness yo 0. cf. but as we argued also in Sec.1. . including the development of an extended space charge layer.02 0. our model for the zeta potential ζ breaks down.8: Thickness yo of the space charge layer as a function of time: Comparison of the space charge layer thickness from the full numerical solution (solid line) to the width of the region of negative concentration in the asymptotic model (dashed line) and to the approximation for yo based on Eq. imately at the steric limit. The surface conductance could also be aﬀected by strong interactions between the ions.190) (dash-dot).2 this does not necessarily imply that our entire model becomes invalid: Even beyond the steric limit the Debye layer continues to accumulate charge. the discussion in Sec. 2. While this is an interesting topic. and for more concentrated solutions this limit is reached already at lower voltages. (2. so we continue to investigate the dynamical behaviour in our model as derived from dilute solution theory. it is also beyond the scope of the present thesis. even though the Debye length is nominally λD = 10 nm for a 1 mM electrolyte. At the same time.2. the highly compressed Debye layer is eﬀectively no more than about an ˚ angstrøm wide.4.01 0 0 1 2 3 4 5 6 Time t Figure 4. so the eﬀects of salt depletion in the diffusion layer.

whereas at the electrode center the concentration is higher.e.1 at the edges. 4. driving pairs of ﬂuid vortices above the electrodes.78 Strongly nonlinear analysis 4. ω = 1.9 also shows the ﬂuid ﬂow pattern for V0 = 50 and no convective transport.11(a) shows the bulk concentration proﬁle as a contour plot: At the electrode edges the concentration is lower than the equilibrium bulk value c = 1. i.1 at the edges. Fig. δ = 1. The interdigitated electrodes all have width W = 2 and ﬁnite thickness and radius of curvature R = 0.001. driven by electroosmotic slip from the electrode edges towards the center. but the electroosmotic slip dominates because the zeta ˜ potential ζ is much larger than the excess Ohmic potential drop ς across the ˇ diﬀusion layer.. Pe = 0. 4..9. and to avoid introducing unphysically strong salt depletion due to the ﬁeld singularity at the edges we do not model the electrodes as being inﬁnitely thin but allow a ﬁnite thickness and radius of curvature R = 0. The two components have opposite sign.2. 4. i. This ¯ nonuniform concentration proﬁle is maintained by a continuous surface ﬂux in the Debye layer running from the electrode edges to the center. and transports it towards the center where it is expelled and diﬀuses back. so we observe the “usual” pattern with slip from the electrode edges towards the center. 4.10. The slip pattern is shown in detail in Fig. The dominant component in the ﬂow is electroosmosis. we take ε = 0.1 Symmetric electrode array We consider a symmetric array of interdigitated electrodes as sketched in Fig. 4. and the streamlines show the usual ﬂow pattern with a vortex above the electrode edge. on average in time the Debye layer sucks in salt at the electrode edges. and are separated by gaps of G = 1. Further.e. Fig. Dashed lines mark the computational domain or unit cell. and assume a symmetric electrolyte with ∆D = 0. this .2 2D model problems Figure 4. where the solid line shows electroosmotic slip and the dashed line is electroconvection induced in the diﬀusion layer. The electrodes have (dimensionless) width W = 2 and are separated by a gap G = 1.9: Sketch of the geometry of the symmetric electrode array.

Although for a typical aqueous electrolyte the nominal (a) (b) (c) 1 75 1. (c) Bulk electroconvection component of ﬂuid ﬂow. arising when a current is passed through a region of f nonuniform concentration.0 25 0. Fig.05 0.11(b). (b) Purely diﬀusive bulk salt ﬂux ¯ ¯ F = −∇¯. Now.8 1 Arc length s Figure 4. This ﬂow takes the form of a ﬂuid vortex with the same direction as that from electroosmotic slip. The electroosmotic slip (solid line) and electroconvection from the diﬀusion layer (dashed) have opposite directions but electroosmosis dominates.6 0. The ﬂow takes the form c of a ﬂuid vortex with the same direction as that from electroosmotic slip.4. 4. it turns out that neglecting bulk salt transport by convection is a poor approximation. but the magnitude is much smaller than the electroosmotic ﬂow at V0 = 50. .11(c) shows the bulk electroconvection q and ζ ˜ component of the overall ﬂuid motion due to the volume electrical body ¯ ¯ force ¯ = ∇2 φ∇φ.92 5 Figure 4. 4. The concentration at the electrode edges is lower than the equilibrium concentration c = 1 and higher at the electrode center. diﬀusive ﬂux is shown in Fig.10: Slip velocity us as a function of the distance s from the electrode center.9 0. (a) Contour plot of the bulk concentration proﬁle. The feature that the surface excess ˜ ﬂux Γ = −(1 + 2Pe)˜∇ φ is in the same direction as the electroosmotic q s¯ ˜ s¯ ˜ slip us = ζ∇ φ holds more generally for a symmetric electrolyte because ˜ have opposite sign.2 0.2 2D model problems 79 0 −10 −20 us −30 −40 −50 0 0.11: Solution at V0 = 50 and Pe = 0.95 1.4 0.

4. 4.8 and minr c ≈ 0. a transient space charge layer forms and gives rise to electroosmotic ﬂow of the second kind. 4. 4. 4. (c) The bulk electroconvection induced with the distorted concentration proﬁle features a second counterrotating vortex in the electrode gap.25. Interestingly. i.9 2 0.10. 4.05 Figure 4. re¯ ¯ spectively.12(b) is dominated by convection.12: Solution for V0 = 50 and Pe = 0. the bulk electroconvective ﬂow in Fig. the overall ﬂow pattern is still dominated by electroosmosis so this second small vortex does not show up in the overall picture. cf.80 (a) (b) Strongly nonlinear analysis (c) 1 0. In order to quantify the relative strength of the diﬀerent components in .11(a).25 is relatively small.. opposite of q.e.13 shows the slip velocity distribution along the electrode for V0 = 100 and Pe = 0. (a) The salt concentration proﬁle is distorted and (b) the bulk salt ﬂux is dominated by convection.25.25: The salt ﬂux in Fig. Fig.12(c) induced with the distorted concentration proﬁle has diﬀerent topology than for Pe = 0: There is again a large ﬂuid vortex rotating in the same direction as the electroosmotic ﬂow. so in general the Helmholtz–Smoluchowski ˜ ¯ ˘ s φ of the second kind electroosmotic slip should have the same component Φ∇ direction as that of the ﬁrst kind. which distorts the concentration proﬁle in Fig. The minimal concentration in the bulk region close to the electrode edge is around minr c ≈ 0. but a second smaller vortex appears in the gap between the electrodes. that drives the ﬂuid in the same direction as electroosmosis ˘ of the ﬁrst kind: The potential drop Φ across the space charge layer has the ˜ same sign as ζ. the true ratio of convective to diﬀusive ion transport Pe ∗ = ¯¯ |Pe uc| ∼ Pe|¯ | u |∇¯| c (4.12 we show the solution at Pe = 0. Fig.7 for Pe = 0 and Pe = 0. Peclet number Pe = 0.95 0. Upon a further increase in the driving voltage.20) is large because the dimensionless ﬂuid velocity is much larger than unity at V0 = 50.1.9 5 1. but the concentration in the diﬀusion layer is very close to dynamically reaching zero at V0 = 50. Table 2. cf. However.25.97 5 1.12(a) as compared to Fig.0 25 0. Therefore in Fig. 4. and in particular we notice a signiﬁcant contribution from electroosmotic ﬂow of the second kind close to the electrode edge.

However. while electroconvection from the diﬀusion layer (dashed) is opposite. which means that the accumulated charge becomes more uniformly distributed along the electrode. This explains why electroosmosis of the ﬁrst kind saturates already for .2 2D model problems 81 50 0 us −50 −100 −150 0 0. while electroconvection induced in the diﬀusion layer contributes positively. and both electroosmotic slip of the ﬁrst and second kind and electroconvection from the diﬀusion layer are driven by the tangential ﬁeld. 4.4.14(a) for Pe = 0 and in Fig. This slows down the charging close to the electrode edges.9. cf.6 0. Because of symmetry the linear momentum of the ﬂuid inside Ω vanishes. The mechanism behind this is as follows: The surface excess salt ﬂux from the electrode edges towards the center tends to produce a region of low salt concentration and therefore low conductivity in the gap between the electrodes.21) where Ω is the computational domain. 4.e. the overall ﬂuid ﬂow we compute the total angular momentum Ltot = Ω ¯ r × u dr. except bulk electroconvection. 4. they give rise to ﬂuid motion in the clockwise direction. bulk convection of salt tends to counteract the development of a low conductivity region between the electrodes by stirring the solution and bringing down unit concentration electrolyte into the gap.8 1 Arc length s Figure 4.2 0. That in turns suppresses the tangential electric ﬁeld because that is fundamentally dependent on nonuniform charging of the Debye layer. (4.25.4 0. which makes Ltot independent of the choice of origin for r.. Fig.13: Slip velocity distribution at V0 = 100. i. The contribution to Ltot from the diﬀerent ﬂow components is shown in Fig. In general it appears that bulk convection of salt for Pe = 0 enhances all components in the ﬂow. In both cases electroosmosis of the ﬁrst and second kind and bulk electroconvection all contribute negatively to Ltot .14(b) for Pe = 0. The electroosmotic ﬂow of the ﬁrst kind (solid line) and second kind (dash-dot) are in the same direction.

c 4. 4.12(c) grows so large that the overall contribution from the two vortices almost integrates to zero.5 and W2 = 7. The solid line shows the overall Ltot . and because bulk electroconvection is not induced at the surface.2 Asymmetric electrode array We now break the left-rigth symmetry of the electrode array and consider a planar device as that sketched in Fig.25. 3 to compute the pumping . 3.15 for V0 = 50.2. and Pe = 0 and 0. with pairs of electrodes of width W1 = 1. respectively. 4. Because the device is no longer completely ﬂat.1 and our ad hoc modiﬁcation. The arrows in Fig.25.14(b) as in Fig. For comparison the dotted line shows the corresponding result from the weakly nonlinear model.14(a) but not until above V0 175 in Fig.2 at the edges. separated by gaps of G1 = 1 and G2 = 5. 4. becomes active. but this is partly because the secondary vortex from Fig. The development of the low conductivity zone between the electrodes is shown in Fig.25. and consider ﬁrst a symmetric electrolyte with ∆D = 0. 4. Further.14(a). as well as the point around V0 ≈ 175 where the bulk concentration drops below ε1/3 = 0. we take ε = 0. 4. replacing c by c = ¯ ¯ max{¯. 100. and 200. 150.001 and δ = 1.1. respectively. 4. and from electroconvection from the diﬀusion layer (squares) and bulk (triangles). To avoid problems associated with the ﬁeld singularity at sharp edges we allow for electrodes of ﬁnite thickness and curvature R = 0.14: Total angular momentum in ﬂuid vortex as a function of driving voltage V0 for: (a) Pe = 0 and (b) Pe = 0. The two arrows mark the voltage where local electroneutrality ﬁrst breaks down in the diﬀusion layer and bulk region. ε1/3 } in the bulk conductivity and ﬂuid body force.14 marks the point around V0 ≈ 50 where local electroneutrality ﬁrst breaks down in the diﬀusion layer close to the electrode edge. and the symbols show contributions from electroosmosis of the ﬁrst kind (circles) and the second kind (stars). V0 100 in Fig. The bulk electroconvection only contributes very little to Ltot at P e = 0. respectively.82 Strongly nonlinear analysis Angular momentum 25 0 −25 −50 −75 −100 −125 0 (a) 25 0 −25 −50 −75 −100 −125 0 (b) 25 50 75 100 125 150 175 200 25 50 75 100 125 150 175 200 V0 V0 Figure 4.14(b). we cannot apply the simple Fourier analysis from Chap. 4. and why both electroosmosis of the second kind and electroconvection from the diﬀusion layer is about twice as strong in Fig.

which makes the charging of the Debye layer more uniform and reduces the tangential ﬁeld. 8 0.8 0.1 1 0.9 1 1. The low conductivity region between the electrodes locally reduces the current into the Debye layer at the electrode edges. 1.25 1 1 0. 1 1 0. 150.9 0.9 Figure 4. 9 1 0.8 1.8 0.9 0. 1.2 0.8 0.1 1. 1.8 0. 1.9 0. 8 1.9 0. 1 1 1 0. 100.1 0. 1 1 1.8 0.2 2D model problems 83 Pe = 0 V0 = 50 1 1 P e = 0. 9 9 1.4.2 0.9 0. 9 1. Bulk convection of salt when Pe = 0 counteracts this by ﬂushing “fresh” electrolyte down into the gap region.15: Salt concentration proﬁle for Pe = 0 (left) and 0. 9 1. 9 1 0. 1 1 0.9 0.9 1 1 1 1 1 1 0.2 0.8 0.25 (right) at V0 = 50.8 0. 1 . 1 0.9 0.2 0.1 1 V0 = 200 1.8 0.8 1 1.1 0. and 200.1 1 1 9 V0 = 150 1 1 1.9 9 0.9 V0 = 100 1 1 1 1.

and above this voltage the electroosmotic ﬂow of the second kind contributes to pumping in the reverse direction. Around V0 ≈ 50 the concentration in the diﬀusion layer dynamically approaches zero close to the right edge of the narrow electrode. (3. while bulk electroconvection pumps in the “reverse” direction with much smaller strength. When the driving voltage is further increased. The contribution from electroconvection falls oﬀ as V04 at low voltage. velocity from the spatial average slip. However. towards the center or the edges of the electrodes.2.2).25.16: Pumping velocity as a function of driving voltage V0 at ω = 1 for: (a) Pe = 0 and (b) Pe = 0.. that determines the net pumping. compute the net volume ﬂow rate Q by integrating over a cross section of the channel. so the model does not predict any reversal of the pumping direction for V0 200. and deﬁne the eﬀective pumping velocity as U = 2Q/H. Note that we might actually have expected the opposite. both with and without convective transport of salt. respectively.1 that bulk electroconvection induces vortices in the same direction as electroosmosis. for a planar asymmetric electrode array it is not only the slip direction. and the symbols show contributions from electroosmosis of the ﬁrst kind (circles) and the second kind (stars). 4.84 30 25 30 Strongly nonlinear analysis (a) 25 20 15 10 5 0 −5 0 (b) Pumping velocity 20 15 10 5 0 −5 0 25 50 75 100 125 150 175 200 25 50 75 100 125 150 175 200 V0 V0 Figure 4. and from electroconvection from the diﬀusion layer (squares) and bulk (triangles). Instead we determine the full bulk ﬂow proﬁle. cf. while electroconvection induced in the diﬀusion layer slips in the opposite direction. where H is the height of the channel. since we learned in Sec. Eq. where the ﬂow induced in the diﬀusion layer contributes in the “forward” direction. 4. For comparison the dotted line shows the corresponding result from the weakly nonlinear model. Fig. The solid line shows the net pumping U .16 shows the pumping velocity as a function of driving voltage V0 for ω = 1 at the inverse RC time. The two arrows mark the voltage where local electroneutrality ﬁrst breaks down in the diﬀusion layer and bulk region.e. the narrow electrode experiences breakdown of local electroneutrality and formation of transient space charge layers over the entire surface of the elec- . i. In both cases the net pumping is dominated by electroosmotic ﬂow of the ﬁrst kind. but also the balance between diﬀerent sections with opposite slip direction at diﬀerent strength.

(c) electroosmotic ﬂow of the second kind. (a) Net pumping velocity U and contributions from: (b) Electroosmotic ﬂow of the ﬁrst kind. The dashed lines mark the breakdown of local electroneutrality in the diﬀusion layer and bulk region. The dashed black and white lines mark the voltage where local electroneutrality breaks down in the diﬀusion layer and bulk region.17(a) the total pumping velocity U is maximized around ω ∼ 1 at the inverse RC time.16.2 2D model problems 200 175 150 125 85 12 12 10 (a) 6 (b) 8 10 6 8 V0 100 75 4 4 2 50 25 0 200 175 150 125 2 V0 100 75 50 25 0 10-1 100 101 10-1 100 101 0 ω ω Figure 4.4. and again bulk convection of salt counteracts the formation of such a region for Pe = 0.16(b) where Pe = 0. 4. but at large voltage there is a more complex frequency dependence due to the contribution from bulk electroconvection. 4. the arrows in Fig. respectively. while the wide electrode experiences this only close to the left edge. In Fig.17(b) shows the contribution 0 2 −2 (c) (d) −4 3 −0 .17 shows the pumping velocity on a contour plot as a function of the driving frequency ω and voltage V0 for Pe = 0. 4.25. Fig. 4. 4. As for the symmetric device this occurs because the surface excess ﬂux gives rise to a region of low conductivity in the gap between the electrodes which reduces the tangential ﬁeld. The electroosmotic ﬂow of the ﬁrst kind saturates for V0 125 in Fig. Fig. trode. but not in Fig. and (d) bulk electroconvection.16(a) with Pe = 0. 4. cf.5 2 1 0 .17: Contour plot of pumping velocity as a function of driving frequency ω and voltage V0 for Pe = 0. Therefore for V0 110 the overall contribution from electroosmosis of the second kind to the net pumping changes to be in the forward direction. respectively.

A similar feature with a secondary peak in the pumping velocity has been observed experimentally by Urbanski et al. The contribution from electroosmosis of the second in Fig. 4. but changes to positive at larger voltage. Finally. from electroosmosis of the ﬁrst kind. which are qualitatively very similar to those at Pe = 0. although for a diﬀerent electrode geometry [69].17(c) starts out negative.5 4 4 2 2 (c) (d) 3 2 1 −0.18 shows the results at Pe = 0. dashed black and white lines mark the voltage where local electroneutrality breaks down in the diﬀusion layer and bulk region. 4.25.86 200 175 150 125 Strongly nonlinear analysis 20 (a) 8 (b) 8 20 16 16 12 12 6 6 V0 100 75 50 25 0 200 175 150 −0 . The electroosmotic ﬂow of the second kind displays a twin peak at large voltage. respectively. 4. Again. (a) Net pumping velocity U and contributions from: (b) Electroosmotic ﬂow of the ﬁrst kind.e.. and the same is true for bulk electroconvection in the reverse direction. bulk electroconvection in Fig.18: Contour plot of pumping velocity U as a function of driving frequency ω and voltage V0 for Pe = 0. and (d) bulk electroconvection. Fig. but quantitatively the electroosmotic ﬂow of the ﬁrst kind is about twice as strong because convection of salt prevents formation of a low conductivity region in the electrode gap. (c) electroosmotic ﬂow of the second kind..17(d) contributes in the reverse direction around ω ∼ 1 but in the forward direction at larger frequency. i. in the reverse direction. We do not show the electroconvection induced in the diﬀusion layer because its contribution to the net pumping is small. featuring a single peak around ω ∼ 1+δ at low voltage that shifts down to ω ∼ δ at larger voltage.25. 0 125 .5 −2 −1 0 V0 100 75 50 25 0 10-1 100 101 10-1 100 101 0 ω ω Figure 4.

so we are not fully conﬁdent with our numerical results here.159) and (2. However.19: Pumping velocity as a function of driving voltage V0 at ω = 1 for an asymmetric electrolyte with ∆D = 0. around 5% or less. such that ∆D = 0.18. (2. Moreover. and they are very similar to those for a symmetric electrolyte from Fig.33: (a) Pe = 0 and (b) Pe = 0.2 2D model problems 10 10 87 Pumping velocity 8 (a) 8 (b) 6 6 4 4 2 2 0 0 −2 0 20 40 60 80 100 −2 0 20 40 60 80 100 V0 V0 Figure 4..16 is N = 32 points in time over one period. the convergence of the solution for large voltage and frequencies ω 3 was relatively slow. Limitations of the numerical scheme The temporal resolution in Fig.17. We compute the periodic solution . Until this point we have focused on the symmetric electrolyte case with ∆D = 0 in the strongly nonlinear model which is already farily complex. The results are shown in Fig.19 as a function of driving voltage at ω = 1. while in Figs. The solid line shows the net pumping U . and the concentration at the electrode approaches zero already around V0 ≈ 40.33 and Da = 0. but as a ﬁnal twist let us consider an asymmetric electrolyte with cations twice as mobile as the anions. Therefore electroosmotic ﬂow of the second kind also shows up earlier. The diﬀerence between the results obtained with N = 16 and N = 32 is small. but with magnitude similar to the symmetric electrolyte solution. and 4.160). which in turns is ˇ c ˇ linear in V0 due to the electric current component in the excess ambipolar ﬂux.4. and from electroconvection from the diﬀusion layer (squares) and bulk (triangles).e. i.88.25. Eqs. 4. For comparison the dotted line shows the corresponding result from the weakly nonlinear model. but also not completely negligible. D+ /D− = 2. The contribution from electroconvection induced in the diﬀusion layer scales as V02 at low voltage because the diﬀusion potential −∆D log(1 + γ /¯) is linear in the excess concentration γ . cf. 4. 4.19 it is only N = 16. Arrows mark the breakdown of local electroneutrality in the diﬀusion layer and bulk region.16. salt depletion in the diﬀusion layer is more pronounced due to the lower eﬀective diﬀusivity Da . 4. and the symbols show contributions from electroosmosis of the ﬁrst kind (circles) and the second kind (stars). 4.

allowing for a much better resolution in time and space than we have been able to achieve here. and electroconvection induced in the diﬀusion layer contributes only little to the net pumping. whereas the electroconvection induced in the electroneutral diﬀusion layer slips in the opposite direction of electroosmosis. 31. e. we have not had the time to pursue this. On the asymmetric array our model predicts that electroosmotic ﬂow of the ﬁrst and second kind contributes to pumping in the same “forward” direction at large voltage.. whereas bulk electroconvection contributes in the “reverse” direction. both for a simple 1D geometry and for 2D models of a symmetric and an asymmetric electrode array. and found excellent agreement at driving voltages below the point where local electroneutrality breaks down in the diﬀusion layer. this eﬀectively limits the time resolution that we can obtain. except close to the point where the current reverses and the space charge layer collapses. because the computer memory required to solve the problem grows as O(N 2 ). for the spatial resolution in the ﬁnite element mesh that we use. We compared the strongly nonlinear model to a full numerical solution of the Poisson–Nernst–Planck equations in one dimension. However.g. 34]. We also found that electroosmotic ﬂow is limited at large voltage by the formation of a region of low conductivity in the gap between the electrodes. 4. we run out of computer memory already at N = 64.3 Summary In this chapter we have studied numerical solutions to our strongly nonlinear dynamical model. Obviously. it would be a great advantage to develop a timestepping algorithm to compute the solution because the memory requirements for a single timestep would be O(1) only. but also admit that there is room for general improvement of the model. a dilute 1 mM KCl solution with λD = .001 and δ = 1 used throughout this chapter corresponds to. However. so we are not able to account for the reversal of the net pumping direction that has been observed experimentally [15. The particular parameter values of ε = 0. the electroosmotic ﬂow dominates over the entire voltage range that we have considered. For the symmetric electrode array we found that both electroosmotic ﬂow of the ﬁrst and second kind and bulk electroconvection gives rise to ﬂuid motion in the same direction.3.88 Strongly nonlinear analysis using a relaxation method as described in Appendix A. However. but bulk convection of salt tends to counteract this. At even larger voltage we found that the simple modiﬁcations to the model to account for formation of transient space charge layers captures the qualitative features of the full numerical solution well. which allows us to account for the nonlocal time dependence of the surface concentration on the ﬂux injected from the Debye layer.

For electrolytes of even lower concentration or devices of smaller feature size. electroosmosis of the second kind and electroconvection from the quasielectroneutral diﬀusion layers and bulk as studied in this chapter. In that case the other sources of electrokinetic ﬂow. with an electrode material such as titanium that forms a natural oxide layer and therefore has a moderate Stern layer capacitance.. Table 2. However.4. one might also envisage viscoelectric eﬀects from strong interactions in a dense phase of ions that would tend to inhibit ﬂuid motion.e. the amount of ions accumulated in the Debye layer is larger. i.3 Summary 89 10 nm in a device with 10 µm characteristic feature size. For electrolytes of higher concentration the use of dilute solution theory to describe the Debye screening layer also breaks down at lower driving voltage. i. which again makes the strongly nonlinear phenomena show up at lower voltage. for electrodes made from gold or platinum that can have very large Stern layer capacitance. which would in turn enhance the electroosmotic ﬂow of the ﬁrst kind. the relevant value of ε would be larger. cf.1. Likewise. The eﬀect of steric exclusion due to ﬁnite size of the ions is ﬁrst to decrease the capacitance of the Debye layer. and the strongly nonlinear phenomena would become signiﬁcant at lower driving voltage. to increase the zeta potential for a given amount of accumulated charge [56]..e. would become more signiﬁcant to the overall ﬂuid motion. .

90 Strongly nonlinear analysis .

which is a function of all the parameters x that can be adjusted in the optimization process. There are many trade-oﬀ’s to be made and it is often not easy to measure what is the “optimal” design.1) ≤ xj ≤ Depending on the nature of the problem there are several diﬀerent solution algorithms to choose from: For constrained linear problems the method of choice is the so-called simplex algorithm. provided the gradient of the objective function with respect to x can be computed [70]. price. .. G2 ]T . . including device performance. . for a simple planar array of asymmetric electrode pairs as studied in detail Chap. Hence a large class of optimization problems can be formulated as minimize f (x) subject to gi (x) ≤ 0 xmin j for i = 1. G1 . (5. In mathematics an optimization problem is formulated in a much more clean and abstract way: The goal of the optimization is to minimize or maximize some quantity f (x). In the present chapter we focus on optimizing the electrode geometry to maximize the pumping velocity U as a function of driving frequency and of the diﬀerent parameters describing the electrode shape. W2 . e. E. .g. xmax j for j = 1. 3 we have x = [ω. . . a geometrical length scale cannot be negative.Chapter 5 Optimization Device optimization is an important discipline in engineering. There may be constraints on the values that the parameters can take. . For low-dimensional nonlinear constrained optimization a popular method is sequential quadratic programming where in each iteration the objective and constraints are approximated by quadratic functions . whereas for unconstrained nonlinear problems one can apply either a conjugate gradient or a quasi-Newton method. m. ease of manufacturing. . robustness. etc.g. It is also a complex task with many objectives to be pursued simultaneously. W1 . the objective function. or more complex nonlinear constraints that depends on several of the parameters together which are typically written as gi (x) ≤ 0.. n.

We always stick to the low voltage regime V0 ≪ kB T /e.92 Optimization of x. photonic crystals [78].and induced-charge electroosmosis but it would also be extremely diﬃcult to fabricate. in particular topology optimization. either exactly or by ﬁnite diﬀerencing. including the Matlab Optimization Toolbox. 75. 72. The algorithm requires the gradient and Hessian (second derivative) of the objective and constraints to be evaluated. the electroosmotic slip is an inherent surface phenomenon which is incompatible with a fuzzy transition between solid electrode and ﬂuid electrolyte. which allows changes in topology to occur smoothly rather than discretely. the electrode topology is not really up for discussion: A device with free-hanging metal islands above the base electrodes would certainly be an interesting sort of combination of AC. to determine the optimal solution. . both with low.and highdimensional design spaces. The primary reason is that optimization is an expensive task numerically and therefore we prefer to work with the linear model which is in itself inexpensive to solve. The optimization algorithm is then allowed to distribute a ﬁxed amount of material in a continuous way. and using it to study a small ﬂuidic network whose optimal topology changes as a function of the Reynolds number [46]. Okkels continues our work by optimizing ﬂow in chemical enzyme reactors and cell metabolism [51. Topology optimization is an approach to designing mechanical structures where one does not a priori specify or even know the topology of the optimal geometrical structure [74. and ﬂuidic structures [79. In such cases it is preferable to use a simple gradient based algorithm. implementations can be found in many software libraries. extending the original results of Borrvall and Petterson for Stokes ﬂow [79] to account for inertial forces. 81]. However. there are two major obstacles that makes it impractical for optimizing the electrode geometry for AC electroosmotic pumping: First. 76]. for large-dimensional problems typical of structural optimization. We have used the so-called method of moving asymptotes (MMA) where the approximation for the objective and constraints is chosen to be both convex and separable [71. It was originally developed in mechanical engineering to optimize beams for maximal stiﬀness at a given amount of material. F. Second. like steepest descent. in the present chapter we use more classical shape optimization of the electrode structures. Often this improves the convergence towards the globally optimal topology. While topology optimization is a powerfull tool. 80]. Therefore. More recently the method has been extended to optimize acoustical systems [77]. We have been working with the method in the ﬂuidic domain. but for ﬂuidic systems we observed problems when there are two strong compeeting local optima [46]. 73]. The basic idea is to replace the discrete transition between solid structure and void with a smooth one by introducing an (artiﬁcal) porous medium whose permeability to the ﬂuid depends on the local material density. it becomes very expensive to compute the Hessian. see our paper in Appendix 5.

1 Planar asymmetric electrode pairs 93 Our experience from Chap.5 µm and G2 = 15. all the user has to do is write a function that takes ( log ω.1 Planar asymmetric electrode pairs Most of the AC electroosmotic pump designs so far tested experimentally [25. whereas both W2 and G2 must be strictly larger than W1 and G1 . W2 . 30] have been planar devices with asymmetric electrode pairs as studied in Chap. G1 . (5.2) where ω is the driving frequency. sets up and solves the electrokinetic problem. 3 is that the weakly nonlinear regime is not very diﬀerent from the low voltage regime. using Femlab. 1 ≤ W1 ≤ W2 . and 20 µm. 26. G2 ). because the net pumping vanishes for W1 = W2 or G1 = G2 due to symmetry. 5. In the strongly nonlinear regime we still need a better understanding of the system behaviour before we can start thinking of systematic optimization. 15. 2 In order to use the fminunc function from the Matlab Optimization Toolbox. (5. 29. W2 . Often the relative size of the electrodes has been chosen as in the original paper by Brown et al. tested devices with three diﬀerent overall periods L = 50 µm.2 µm and W2 = 25. 28. In practice the minimal size of the microelectrodes and gaps is limited by the lithography used to fabricate the device. this allowed them to conﬁrm that the pumping velocity increases with downscaling of the system. for a given minimal feature size one can still improve on the pumping velocity U by choosing the relative electrode sizes optimally. due to the truncation error and noise introduced by remeshing the geometry in each function call.1.5. 3 and sketched in Fig.1 Because the pumping velocity increases with downscaling we can be sure that in the optimal design at least one of W1 and G1 will attain their lower limit. W1 .3) 1 Optimizing for log ω eﬀectively removes any problems associated with zero or negative frequencies.1) writing maximize subject to U ( log ω. G2 ) 1 ≤ G1 ≤ G2 . separated by gaps of G1 = 4. so it seems reasonable that a device geometry optimized at low voltage will also perform well at larger voltage. e.6 µm. [25]. and returns (minus) the pumping velocity U . Mpholo et al. but used the same relative electrode sizes in each case.7 µm. 3. This problem can be expressed as a constrained optimization of the form of Eq.. with narrow and wide electrodes of width W1 = 4. W1 . W1 . and the exact location of the optimum can therefore be determined using a simple unconstrained optimization algorithm such as the Matlab function fminunc to solve the problem2 maximize U ( log ω. G1 = 1. 30 µm.g. G2 ) as arguments. respectively. It turns out that in the optimal design G1 = 1 and W1 > G1 . (5. W2 . respectively. indeed we would argue this is the natural way of reporting any frequency. However. it is important choose appropriate not-so-small values for the termination tolerance and for the step size used when fminunc approximates the . However.

1: Contour plot of the maximum pumping velocity max ω U as a function of W1 and W2 for G1 = 1 and G2 = 5. and G2 = 5. and G2 = 4. namely W1 = 1.5 2 2.5 and W2 = 7 with U = 0.0035 contour is fairly broad.5 3 3.51. 3 to work with a design with rounded numbers.74.00358 at ω = 0. This is demonstrated in Fig.8.5 4 4. A robust but also less eﬃcient alternative is direct search algorithms like the Matlab fminsearch Nelder–Mead simplex method.00358 at ω = 0. It is for this reason that we have chosen in Chap. W2 = 7.80. Another simple approach is to use pairs of planar electrodes that are symgradient and Hessian by ﬁnite diﬀerencing.8.1 where the maximum pumping velocity max ω U is shown as a function of W1 and W2 for G1 = 1 and G2 = 5. There is a unique optimum close to W1 = 1. where U = 0. The maximum is close to W1 = 1.2 Planar electrodes with nonuniform coating There are many ways that one can break the left-right symmetry of an electrode array.002 0 etry Symm 1. but it is close to the global optimum according to our present linear analysis. When W1 = W2 the net pumping vanishes due to symmetry. The pumping velocity is actually not very sensitive to the particular choice of planar electrode geometry. . This does not correspond to any device tested experimentally (yet). 5.0 03 0. 5.0 025 .5 and W2 = 7 where U = 0.5 5 W1 Figure 5.55.0 03 5 0. From a fabrication point of view the asymmetric pairs of planar electrodes are particularly simple because the device can be processed with just a single lithography step to deﬁne metal electrodes on the substrate. G1 = 1.5.94 20 18 16 14 12 Optimization W2 10 8 6 4 2 0 0 0. provided we start the algorithm somewhere in the interior of the feasible region. In this way we ﬁnd that the pumping velocity is maximized for W1 = 1. but the ﬁgure also shows that the range of designs that come within the 0.00359 at ω = 0.5 1 0. W2 = 6.

the capacitance ratio δ(x) is a function on the position with 0 ≤ δ(x) ≤ ∞. 5. see Fig. For our present purpose it turns out to be more convenient to parametrize the problem in terms of Λ = (1+δ)−1 . 2 4 (5.1. We consider the general problem of an array of blocking electrode where the Stern layer capacitance varies along the electrodes. 33]. This was actually one of the suggestions for breaking the symmetry in the original paper by Ajdari [14]. metric in size.2.5.5) (5.e. The time average slip velocity can be expressed as 1 1 ˆ ˆ ˆ2 ˆ ˆ us = Λ(x) Re (Vext − φ)∂x φ∗ = −Λ(x) ∂x Vext − φ . streamlines correspond to the optimal solution at driving frequency ω = 5. The coating thickness is not drawn to scale.2 Planar electrodes with nonuniform coating 95 W G Figure 5. because this stays in the unit interval 0 ≤ Λ(x) ≤ 1. i.6) . but with an oxide layer or coating that makes the Stern layer capacitance left-right asymmetric on the individual electrodes.2: Planar electrodes (black) with a spatially nonuniform thin oxide layer or coating (orange). which is simply the ratio of the overall capacitance of the double layer to the Debye layer capacitance [23. In the low voltage regime the dimensionless problem then takes the form ˆ −∇2 φ = 0 in the bulk with ˆ ˆ ˆ −n · ∇φ = iωΛ(x)(φ − Vext ) (5.4) ˆ on the electrodes and n · ∇φ = 0 on the channel walls..

we obtain the weak form of the problem ˆ ∇ϕn · ∇φ dr + iω ϕn Λ(x)(φ electrodes ˆ − Vext ) dx = 0 ˆ for all n. (5.10) ∇ϕn · ∇ϕm dr + iω ˆ Fn = iω ˆ ϕn Λ(x)ϕm electrodes dx.14) electrodes electrodes Since the number of parameters Λk describing the distribution of the coating material is large.4) by a test function ϕn and integrating over the computational domain Ω. where ˆ Knm = Ω (5.96 Discretization and sensitivity analysis Optimization ˆ We approximate both the complex potential φ and the design variable Λ by expanding on a ﬁnite basis set ˆ φ(r) = m ˆ ϕm (r)φm and Λ(x) = k χk (x)Λk . (5.13) We want to maximize the pumping velocity U . (5.15) .11) and ϕn Λ(x)Vext electrodes (5. ω). ω) = 0.12) ˆ Notice that K is symmetric but not Hermitian. (5.9) m Ω This can be written more compactly in matrix form ˆˆ ˆ KΦ = F. dω (5. We also deﬁne the residual ˆ = F − KΦ in terms of which the electrical problem becomes ˆ ˆˆ L ˆ ˆ L(Φ. we use the gradient based algorithm MMA [71. Λ. (5. 72] to determine the optimum. (5.7) Multiplying Eq.5) to eliminate n · ∇φ. Hence we need to compute the derivatives d ˆ U Φ(Λ. Λ .8) Ω ˆ where we used Gauss’ law with Eq. (5. dx (5. Λ dΛ and d ˆ U Φ(Λ. Inserting the ˆ ˆ expansion for φ we obtain a linear problem for the coeﬃcients φm ∇ϕn · ∇ϕm dr + iω ϕn Λϕm electrodes ˆ dx φm = iω ϕn ΛVext electrodes ˆ dx. and for an array of planar electrodes this is simply deﬁned as as the spatial average slip velocity U= 1 L us dx = − 1 4L ˆ ˆ ˆ∗ ˆ Λ (Vext − φ∗ )∂x (Vext − φ) dx + c. ω).c.

(5. gradient of L with respect to Λ and ω at a point in (Φ. = + Re WT ∂Λ ∂Λ ∂Λ (5.20) = ˆ ∂U ∂L ˆ ∂L . ˆ ˆ . deﬁned by ˆ ˆ the constraint Eq.16) ˆ ˆ where both the potential Φ and its complex conjugate Φ∗ are considered ˆ is a Lagrange multiplier for the constraint as independent variables. ω) = U (Φ. ˆ ∂W ˆ and the Lagrange multiplier W from ∂U 1ˆ ˆ ∂L = − KT W = 0.21) (5. for a given Λ and ω we ﬁrst compute the potential Φ from ∂L ˆ ˆˆ = F − KΦ = 0. ˆ ˆ.5. L(Φ.2 Planar electrodes with nonuniform coating 97 where Φ(Λ.15) can be computed as the ˆ ˆ. and W ˆ ˆ L(Φ. Λ dω Numerical implementation We build the physical model and solve the problem Eq.13). ω) is considered an implicit function of Λ and ω. W∗ Λ. 2 2 (5.17) electrodes electrodes . The derivatives in Eq. because Φ(Λ. 1 ˆ ˆ 1 ˆ ˆ ˆ ˆ. ω). we cannot simply use the chain rule. (5. Further we need to compute the partial ˆ derivatives of U with respect to Φ and Λ 1 4L n ∂U 1 =− ∂Λk 4L ∂U = ˆ ∂φ ˆ ˆ ˆ∗ ˆ∗ Λ (Vext − φ∗ )∂x ϕn + ϕn ∂x (Vext − φ∗ ) dx ˆ2 ˆ χk ∂x Vext − φ dx.18) (5. Φ∗ W. ω) is implicit and ∂ Φ/∂Λ ˆ and ∂ Φ/∂ω are unknown. However. (5. = Re WT ∂ω ∂ω (5. ω).22) = ˆ ∂L ˆ ∂L . ˆ ˆ . Λ. Λ dΛ and d ˆ U Φ(Λ. That is. Φ∗ W. Instead we use the adjoint method [82] and introduce the Lagrangian L. W∗ Λ.17) using Femlab. ˆ ˆ ∂Φ ∂Φ 2 Then the desired derivatives are given by d ˆ U Φ(Λ. Φ∗ Λ) + WT L + W∗T L∗ .19) (5. ω) ˆ and W and their complex ˆ space where L is stationary with respect to both Φ ˆ conjugates. (5. ω) = 0. ˆ ˆ ˆ ˆ Therefore the both system matrix K = −∂ L/∂ Φ and the matrix ∂ L/∂Λ are computed automatically as part of the solution process and can be obtained directly as Matlab sparse matrices.

and we use second order ˆ elements for the potential φ.2 (d) dU dΛ 0.0 Sensitivity 0.2 0.0 0.1 0. cf.011 to the solution for uncoated electrodes (dashed) at ω = 4.1 0.2 1.0 where U = 0: ˆ ˆ (a) Design variable Λ(x).5 2 0.05 0. 0) .23) (5.1 where U = 0.5 2 (c) Slip velocity us 0. ω) is a linear function of ω.8 0. Λ.5 2 x x Figure 5.24) =W electrodes ˆT ˆ ˆ ˆ ˆ where the last equality follows because L(Φ. (5. 1) − L(Φ. (c) Time-average ˆ ˆ and Im[φ] slip velocity us (x). Λ. i. This can also be done with Femlab although on the way we need to make an auxiliary copy of the original model. Fig. The design variable Λ is approximated with ﬁrst order Lagrange elements.0 −0.4 0.0 0.6 0. 5.e. it is piecewise linear. see our paper in Appendix C for details. Finally we need to evaluate ˆ ˆ ∂L = i WT ∂ω ˆ ˆ ˆ ΛW (Vext − φ) dx ˆ ˆ L(Φ.2.5 1 1.. The optimization algorithm is initialized with a uniform distribution Λ(x) = 1 corresponding to uncoated electrodes.0 −0.3: Comparison of the solution for optimal coating (solid) at ω = 5.5 2 0. The position of the electrode on the x-axis is marked in gray.2 0. (b) Potential φ(x. and with a frequency of ω = 4 which is about twice the relaxation frequency . Optimal solution We consider a symmetric device with electrodes of a ﬁxed width W = 1 and gap G = 1.5 1 1. (d) Sensitivity dU/dΛ(x).0 0 Optimization (b) 1.5 1 1.98 (a) Coating Λ 1.2 0 0.5 1 1. Λ. uncoated Re[φ] (dashed) and Im[φ] (dash-dot).4 0. y = 0) on electrodes.1 0 0.8 ˆ Potential φ 0. optimal Re[φ] (solid) ˆ (dotted) vs.2 0 −0.6 0.05 −0.1 −0.

the latter is not an option due to the bound Λ ≤ 1.0 0.011 at ω = 5.3: Close to the right edge of the electrode there is an uncoated section of width ∼ 0.2 corresponds to the optimal capacitance ratio δ = Λ−1 − 1.75 where Λ = 0. in optimization jargon this is known as the complementary slackness condition. 5. (5.4 0.1 −0.6 0.5 2 0. The initial solution is shown with dashed lines in ˆ Fig.5 2 x x Figure 5. ˆ ˆ (a) Potential on electrodes. the thickness of the coating in Fig. (b) Slip velocity us .1.5. The maximum pumping velocity is U = 0.4.0 −0. but these slowly fade away as the design iterations proceed.8 (b) Slip velocity us 0. (a) 1. Finally. The sensitivity dU/dΛ of the pumping velocity U to local changes in Λ shows that U can be increased by decreasing Λ on the left half of the electrode or by increasing Λ on the right half.0105 at ω = 5. However. Re[φ] (solid) and Im[φ] (dashed).1 ˆ Potential φ 0.2 0 0.4: Approximating the optimal solution from Fig. and the solution is shown with solid lines in Fig.1.25. the slip velocity decreases locally with decreasing Λ. 5. and the streamlines correspond to the optimal solution at ω = 5. but for the remaining part it turns out to be optimal to add coating. 5.5 1 1.25 and an uncoated section of width W2 = 0. The optimization algorithm adjusts Λ(x) according the sensitivity and quickly reaches a solution that is close to optimal. The maximal pumping velocity is U = 0.2 0.2 Planar electrodes with nonuniform coating 99 for the uncoated device. The potential varies almost linearly and the time-average slip velocity us is positive throughout the coated section. dU/dΛ = 0 only where Λ is limited by the upper bound.2 0. At ﬁrst this result may seem counterintuitive: From Eq. After 71 iterations the relative increase in pumping velocity is less than 10−5 for two successive iterations and we terminate the algorithm. except for some small wiggles at the resolution of the ﬁnite element mesh.3 by a design with electrodes with a uniformly coated section of width W1 = 0. From an engineering point of view it may be possible to fabricate devices .5 1 1.6) we know that at ﬁxed tangential ﬁeld and potential drop across the double layer.2 0 0.25 where Λ = 1. a decrease in Λ on the left half of the electrode also induces changes in the potential in such way that the overall result is an increase in the net pumping U .3: The potential φ has left-right symmetry so the induced electroosmotic slip velocity does not give rise to any net pumping. Notice that in the optimal solution. 5. however.0 −0.

. One may object to the above solutions because they have features below the (unit) characteristic length scale.25) subject to 1 ≤ W1 .0105 at ω = 5. 5. notice that the slip velocity is discontinuous at the point where Λ changes abruptly from Λ1 to Λ2 .0.5 [45]. and Λ2 = 1. W1 . where U = 0. Λ1 = 0. Therefore the coating thickness will have to be designed for a particular concentration. asymmetric coating allows the pumping velocity to be increased by a almost a factor of 3. 5. They also tested the 3D design experimentally and obtained already a factor of 7 larger pumping velocity than with planar devices of the same minimal feature size [83. Therefore. Λ2 = 1) 1 ≤ W2 . Compared to the device with asymmetric pairs of electrodes. but better suited for fabrication at the lower limit on resolution set by standard lithography. but with standard lithography techniques it is much easier to obtain piecewise constant coating thickness. The maximum pumping velocity for this simple design is U = 0. while oﬀ that concentration the performance is suboptimal. inspired by the optimal solution we consider a device where the electrodes are divided into two sections of width W1 = 0.3. This is about 15% lower than the optimum from Fig.9.0096 at ω = 1. 0 ≤ Λ1 ≤ 1. A disadvantage of the use an asymmetric coating to break the left-right symmetry is that the design variable Λ depends on the electrolyte concentration through the Debye layer capacitance. G.4. one may obtain a signiﬁcant gain in pumping eﬃciency by stepping into the third dimension: Bazant and Ben recently suggested the use electrodes like those shown in Fig. If the same lithography technique is used to deﬁne both the electrodes and the coating on them it is more appropriate to formulate the problem as maximize U ( log ω. 5. 5. 69]. at the expense of an additional lithography step to deﬁne the coating. G = 1.4. Here the optimal design is W1 = 1.31.25 where the design variable is Λ1 = 0.75 and W2 = 0. W2 = 1. W2 .4.3 3D structured electrodes with regular shape While most work so far have focused on planar designs with thin ﬂat electrodes on a substrate. which is merely 5% lower than the optimum from Fig. for which they predict pumping velocities an order of magnitude larger than the planar design of Brown et al. Λ1 .25 and Λ2 = 1. (5. respectively. 1 ≤ G.100 Optimization with a nonuniformly graded coating thickness. The solution is shown in Fig.3. 5.

the electrode array now acts as a ﬂuid “conveyor belt” where all sections help pumping the ﬂuid in the forward direction [45]. a base section of width W2 . Eﬀectively. or if it is appropriate for quantifying the pump performance at all. Streamlines correspond to the solution with W1 = W2 = H1 = G = 1 at driving frequency ω = 0.05.1. The result is shown in Fig. comes out of a simple Fourier analysis as the natural way of quantifying the pump performance. deﬁned as the spatial average slip velocity.8. The basic design principle that they propose is to raise those sections of the electrodes where the slip is in the “forward” direction and to lower the sections with “reverse” slip. the net pumping is maximized around ω = 0.5.8. To clarify this we compute the net ﬂow rate Q as a function of the channel height Htot .5 with electrodes that have both a raised section of width W1 and height H1 . where it is conﬁned by the vertical solid electrode surface rather than struggling up against the forward ﬂow.3 3D structured electrodes with regular shape 101 H1 W1 W2 G Figure 5. which clearly demonstrates the potential of 3D designs.5: Stepped 3D electrode geometry: Each electrode in the array has a raised section of width W1 and height H1 and a base section of width W2 and the gap between adjacent electrodes is G. By analogy with the planar pumps we calculate the eﬀective pumping velocity as U ≈ 2 Q/Htot ≈ 0. W1 = W2 = H1 = G = 1. 5. This is an order of magnitude larger than the optimal result from Sec.6 where we clearly see a linear dependence for . This enhances the forward ﬂow because the reverse vortex is living in a trench. We consider ﬁrst a geometry as shown in Fig. On the contrary: At the level of the forward slipping surface the recirculation from the vortex enhances the forward pumping instead of inhibiting it. 5. 5. For 3D structured electrodes it is not immediately obvious how the pumping velocity should actually be deﬁned. In the simple case when all dimensions are set to unity. and with a gap G between adjacent electrodes. where Htot ≫ 1 is the overall channel height measured from the substrate to the channel top wall. close to the electrode extra streamlines are plotted to show the details of the ﬂow pattern. For planar designs the pumping velocity U .

Htot →∞ Htot lim (5.e.6: Maximal ﬂowrate maxω Q vs. and ∂Q/∂Htot requires ﬁnite diﬀerencing. and therefore the pumping velocity that we report here can be up to a few percent lower than the sharp limit.7. For Htot = 1 the top wall touches the raised electrode section and the ﬂow is blocked. i. should it be measured from the substrate or from the top of the electrodes? For Htot = H1 = 1 the channel top wall touches the raised electrode section and the ﬂow is blocked. We use R = 0.15 0. we simply choose a “large” Htot = 40 and use U ≈ 2 Q/Htot throughout this section. . In order to compute the net ﬂow rate Q we need to solve the Stokes ﬂow problem with a slip condition on the electrodes. For a given electrode design we can therefore uniquely deﬁne the pumping velocity as U= 2 Q(Htot ) . which is perhaps not so surprising. Smaller R gives stronger ﬁeld at the protruding corners which actually tends to increase the pumping velocity U . To avoid numerical diﬃculties associated with inﬁnite slip velocity from the electric ﬁeld singularity at sharp corners on the electrode we use a geometry model with rounded corners of radius R. Htot ≫ 1.26) Extrapolating Q(Htot ) to zero we get a nonzero intercept at around H ∗ = 0.102 0. In order to estimate U from data at ﬁnite Htot we could either use U ≈ 2 Q/(Htot − H ∗ ) or U ≈ 2 ∂Q/∂Htot .1 0. However. total channel height Htot measured from the substrate to the channel top wall in a geometry with W1 = W2 = H1 = G = 1..05 0 0 2 4 6 8 10 Htot Figure 5. because H ∗ depends on H1 . For Htot ≫ 1 there is a linear dependence: Data points (circles) and linear ﬁt (dashed line).02 which is not very sharp.25 Optimization 0. but the limit of sharp corners is ﬁnite.2 Flowrate Q 0. given the ambiguity in the deﬁnition of the channel height.

05 0.5.5 1 1.01 0 0 W1 0. respectively. with the other parameters ﬁxed at unity. Optimal solution Using standard planar lithography to process the device.03 0.05 0. G} as the characteristic length scale. and H1 .06 0.07 0.03 0.04 0.04 0. (5. 5. Dashed and solid lines show the results using 1 and min{1.5 1 1.3 3D structured electrodes with regular shape 0. W2 . Therefore we formulate the optimal design problem as maximize subject to U ( log ω.04 0.03 0.02 0.07 0.6.07 0.5 3 103 Pumping velocity U 0.7 and H1 ≈ 0. W1 . W1 .8 showing the maximum pumping velocity as a function of both W2 and H1 : It is close to .06 0.08 0.5 2 G H1 Figure 5.08 0.27) 1 ≤ G. H1 . W2 .5 2 2. G. 5. W2 .5 2 0 0 0. The ﬁgure shows that the optimum for both W1 and G is at the lower limit W1 = G = 1 whereas for W2 and H1 it is around W2 ≈ 1.01 0. it is much easier to control the thickness of the raised section on the electrodes than the transverse dimensions.05 0.5 1 1. The global optimum can be determined from a contour plot as in Fig.02 0. G) 1 ≤ W2 .02 0.04 0. 1 ≤ W1 . when the other parameters are ﬁxed at unity.08 0. In Fig.7: Maximum pumping velocity maxω U as a function of each one of the individual geometrical parameters W1 .05 0. W2 .7 we investigate the dependence of the maximum pumping velocity maxω U on each the individual geometrical parameters W1 .06 0. G.07 0.01 0.5 2 0 0 0.08 0.02 0. respectively.06 0.01 0 0 1 1. and H1 .5 W2 Pumping velocity U 0.03 0.

0545 at ω = 0.6 0. the simple design with H1 = 0. where U = 0.2 1 0. W2 = 1.25 and H1 = 0.3. the ﬁgure also shows that the optimum is broad.8: Contour plot of the maximum pumping velocity maxω U as a function of W2 and H1 for W1 = G = 1.65. Yet another possibility is to add thin metal electrodes on top of a 3D structured substrate as shown in Fig.089 at ω = 2.05 0.7. Whether or not these more complex designs are actually superior to the simple design in Fig. H1 = 0.25 2. In particular. The optimal solution is. However.0 35 Optimization 1.055 0. A virtue of the simple 3D structured rectangular design in Fig.078 at ω = 0.5 W2 Figure 5.25 and H1 = 0.7 where U = 0. W2 = 1. and H2 = 1. say. 5.0566 at ω = 0. Burch reported that such a design would increase the pumping velocity over the solid metal electrode design [84]. the basic design principle in [45] speaks of both raising sections of “forward” slip and recessing sections of “reverse” slip on the electrode. where U = 0.04 5 4 H1 1. where U = 0. However.5 depends on how diﬃcult it is to pattern both the metal electrodes and the substrate: If it turns out that the minimal linewidth that one can (routinely) produce in the more complex designs is.8 1. Therefore we can generalize the simple design by allowing the base section of the electrode to be lowered by an amount H2 relative to the substrate level as shown in Fig.0. W2 = 1.5 1. The optimal solution to this problem turns out to be W1 = 1.0 0.9(b).8 0.4 1 1.6 1.9 which is within 2% of the global optimum.5 is that it can be fabricated with just two lithography steps to deﬁne the base and raised sections of the metal electrodes on top of a ﬂat substrate. then the simple design is actually superior due to the .104 2 0. G = 1.25 1.0566 at ω = 0.7 and W1 = W2 = G = 1 has U = 0.2.8. D. twice as large as for the simple design. and H2 = 0.75 2 2. 5. 5. roughly.4 0. 5. There is a unique optimum around W2 = 1. and our starting point W1 = W2 = H1 = G = 1 is within 12%. W1 = 1. H1 = 0.9(a). G = 1.8.1.4.9.

When we transform the weak form of the governing equations from the physical frame to the reference frame we make use of the transformation rules f (r) dr = Ω Ω∗ f r(R) detJ dR (5. H2 = 1. Y ). Because r the number of parameters to describe ¯(s) is large (inﬁnite) we will need to r compute the derivative dU/d¯ of the pumping velocity U with respect to r local changes in the electrode shape exactly rather than using ﬁnite diﬀerence approximation. H1 = 0. G = 1.4.4 3D structured electrodes with arbitrary shape (a) (b) 105 H1 H2 H1 H2 W1 W2 G W1 W2 G Figure 5. 5. One way to obtain this is by an arbitrary Lagrangian Eulerian (ALE) type formulation. fundamental scaling of the pumping velocity with one over the characteristic length scale. W2 = 1. H2 = 0.5. W2 = 1. y) to a ﬁxed reference frame (X.2.9. or if some other shape with sloped side walls could improve the performance even further. but keep a ﬁxed ﬁnite element mesh in the reference domain Ω∗ where we solve the problem numerically.089 at ω = 2.28) .4 3D structured electrodes with arbitrary shape A natural question to ask is whether the rectangular electrode structures we investigated in the previous section are (close to) optimal in general. The idea is to continuously deform the physical domain Ω to match the electrode shape.3.65.1. where U = 0. G = 1. H1 = 0. where U = 0. (b) Design with thin metal electrodes deposited on a 3D structured substrate. introducing a mapping from the physical frame (x. In order to answer this question we now consider electrodes with a general shape described by some continuous function ¯(s).078 at ω = 0.9: Stepped and grooved 3D electrode geometries: (a) Design with the “reverse” slipping section of the electrode put into a groove. The optimal solution is W1 = 1. The optimal solution is W1 = 1.

106 and Optimization ∂Xj ∂ ∂ = . but also sets a limit on the overall arc length.. since r Θm = N θm+1 − θm = N a ∆θm ≈ Ltot κm . (5. a (5. ﬂuid velocity u(r) and pressure p(r). tan θm = ∆¯m /∆¯m . in order to be r a valid shape ¯(s) cannot intersect itself. Then the nodal coordinates can be expressed as n−1 n−1 xn = x0 + a ¯ ¯ m=1 cos θm and yn = y0 + a ¯ ¯ m=1 sin θm . Discretization r Our bold statement that we consider a general continuous curve ¯(s) for the electrode shape is subject to a number of constraints. (5. comˆ plex potential φ(r). ∂xi ∂xi ∂Xj (5. This makes the . We discretize the problem by introducing a piecewise linear approximar r tion to ¯(s) with N line segments deﬁned by N + 1 nodes ¯n where the endpoints ¯0 and ¯N are ﬁxed. including the mesh deformation r(R). In the design process the optimization algorithm will therefore have to make a trade-oﬀ between curvature and arc length.29) where detJ is the determinant of the Jacobian matrix Jij = ∂xi /∂Xj for the mapping from the reference domain Ω∗ to the physical domain Ω. Moreover we want it to be fairly “nice” with not too many or too sharp wiggles.g.31). e.2. In the examples below we r control the shape of ¯ by limiting Θm to a ﬁnite interval Θmin ≤ Θm ≤ Θmax . Alternatively we r can parametrize in terms of the N − 1 relative angles Θm = N (θm+1 − θm ) between adjacent line segments. The shape of ¯ is uniquely determined by the y x r N − 1 ﬁrst angles whereas the ﬁnal angle θN and the segment length a are chosen to match the ﬁnal node with the ﬁxed endpoint ¯N . Eq. i.31) writing Ltot = N a for the overall arc length of ¯.e. 5. cf.. The gradient dU/d¯ is r computed using the adjoint method as described in Sec. r This eﬀectively limits the curvature of ¯ and determines how sharp features r can be on the electrode.30) where θm is the angle between the mth segment and the horizontal axis. Θ is also a measure of the local curvature κ = ∂s θ of ¯. but more cumbersome because there are more dependent variables in the present problem. and ∂Xj /∂xi is an element of the inverse Jacobian invJ. In order to gain more control over the shape r r of ¯ and avoid clustering of the nodes we force all line segments to have the r same length a. where s is the arc length in the physical frame. It is straightforward in principle. and a number of Lagrange multipliers that we account for explicitly to gain full control over which constraint forces act where in the governing equations.

. Θ) − Θmax ≤ Θn ≤ Θmax x0 ≤ xm ≤ xN ¯ ¯ ¯ for n = 1.32) The optimal solution to this problem for diﬀerent values of Θmax with N = 100 is shown in Fig.66 which is just enough to allow for two half circles of radius R ≈ 0. . N − 1. The maximal pumping velocity is U = 0. curvature κ = 1/R ≈ 2. and formulate the optimization problem as maximize subject to U ( log ω. The overall features in the solutions are the same as in Fig. Θ) − Θmax ≤ Θn ≤ Θmax − π/2 ≤ θm ≤ π/2 for m = 1. In the ﬁrst four panels (a)-(d) the optimal solution is everywhere limited either by |Θn | ≤ Θmax or by |θm | ≤ π/2. . where the “purpose” of the extra “pillar” dividing the groove into two seems to be to weaken the counter-rotating vortex. N − 1. (5.10: The width of the raised section decreases and a double . N. . but since we like the electrodes to have neither too sharp features nor too large a surface area we are still content with our approach. which is in accordance with the trend in Fig.e. .3. i. . Optimal solutions As a ﬁrst design problem. . we consider a device with a ﬁxed width W = 2 of the electrode and a ﬁxed gap of G = 1. but maintain that all nodes should conform with the overall electrode width W .078. In our second design we relax the constraint on backward sloping structures. .10. (5. 5. even though the present electrode shape completely lacks sharp corners to enhance the electric ﬁeld strength. we see the lowered section develop into a double groove. In panels (b)-(f) we gradually increase Θmax : We see ﬁrst the width of the raised section decrease. 5. .5. N − 1. 5. . 5. Secondly. In panel (a) we set Θmax = 6.11. . . . but in panels (e) and (f) this is no longer the case: In particular in panel (f) neither of the two “pillars” are capped with half circles but have a width and shape that start to be “optimal” in itself rather than limited by the particular constraints that we impose. inspired by the results for rectangular electrodes in the previous section. . again with N = 100 line segments to describe the electrode shape. 5. for n = 1. Thus we formulate the problem as maximize subject to U ( log ω.. we require −π/2 ≤ θ ≤ π/2. i. . To avoid diﬃculties with fabrication we do not allow undercut or backward sloping structures. and a total arc length around π.4 3D structured electrodes with arbitrary shape 107 interpretation of the results more complex.33) The result is shown in Fig..e.5.7(a) for W1 < 1. . for m = 1. which is almost as good as the corresponding optimal rectangular grooved structure from Sec.

(e) Θmax = 40.66.10: Electrode shapes (black) computed as optimal solutions to the problem Eq.108 (a) U = 0. The electrode width is ﬁxed at W = 2 and the gap at G = 1 and the electrode shape is described by N = 100 line segments (nodes in red).3. (5. (f) Θmax = 80. (d) Θmax = 20.111 (e) U = 0. (c) Θmax = 13.097 (d) U = 0.32) for diﬀerent values of the maximum curvature Θmax : (a) Θmax = 6.123 Figure 5.121 (f) U = 0.090 (c) U = 0. (b) Θmax = 10.078 (b) Optimization U = 0. .

11: Electrode shapes (black) computed as optimal solutions to the problem Eq.3.5.092 109 (c) U = 0.127 Figure 5. (e) Θmax = 40.101 (d) U = 0. (d) Θmax = 20.33) for diﬀerent values of the maximum curvature Θmax : (a) Θmax = 6. (f) Θmax = 80. . The shapes in (e) and (f) are not fully converged solutions but were terminated after 200 design iterations. (c) Θmax = 13.078 (b) U = 0.130 (f) U = 0.116 (e) U = 0.4 3D structured electrodes with arbitrary shape (a) U = 0. (b) Θmax = 10. (5.66.

5.11 we ﬁnd that. 9]. It is diﬃcult to compare the performance of our optimized electrode shapes from Figs. which is true for our linear and weakly nonlinear models but not in the strongly nonlinear regime.3 because strictly they are not optimal solutions to the same problem: In Sec. 5.35) where U is the pumping velocity and H is the height of the pumping channel.3 we optimized the geometry subject to a constraint on the minimal transverse linewidth allowed. 5.34) The electroosmotic contribution to the net ﬂow rate per unit width of the channel is Qeo = U H/2 (5. looking at the results in Fig. Provided convection can be neglected in the electrical part of the problem.110 Optimization groove develops when Θmax is increased.3 are indeed (close to) optimal because we constructed their low-dimensional design spaces exactly based on this basic principle together with our knowledge of what electrode structures can most easily be manufactured in practice.10 and 5. and the two last solutions in panels (e) and (f) look pathological in the sense that they contain very narrow parts in the electrode shape. It therefore seems likely that the simple rectr angular geometries from Sec. 5. (5. the thermodynamic eﬃciency [7.10 and 5. At the same time the extent of the backward sloping becomes more and more pronounced. Actually.36) . 5. The pressure driven ﬂow is given by Qp = −∆p/Rhyd (5. panels (e) and (f) do not correspond to fully converged optimal solutions because in both cases the design process was terminated after a total of 200 iterations.. even though it does not know of any such principle but only has information on the local gradient dU/d¯. However. largely. or the maximum amount of “useful work” that the pump can deliver at a given power consumption.e. i. Equally important might be the maximum pumping velocity for a given electrical power consumption used to drive the electrode array.11 with the rectangular geometries from Sec. whereas in the present section we ﬁx the total width of the electrode at W = 2 and optimize subject to a constraint on the minimal feature size as represented by the maximal curvature. 5. the optimization algorithm seems to follow the basic design principle of raising the forward slipping and recessing the reverse slipping sections on the electrodes.5 Thermodynamic eﬃciency Throughout this chapter we have focused on the maximum pumping velocity that can be obtained for a given driving voltage. the net ﬂow rate Q can be decomposed into separate and independent electroosmotic and pressure-driven components Q = Qeo + Qp .

41) ˆ The dimensionless function LU 2 /Re[1/Z] should be optimized geometrically.40) =β ˆ Pin Re[1/Z] where L = W1 + G1 + W2 + G2 is the array period and β is a dimensionless parameter given by β= 3τD u0 ǫV02 3 ǫ = . H = 50 µm. whereas the maximal backpressure ∆pmax = Qeo /Rhyd is deﬁned as the point where the pump stalls and Q vanishes.5 Thermodynamic eﬃciency 111 where ∆p is the backpressure that the pump is pumping up against.1. (5. but otherwise the scaling of β shows that it pays to pump an electrolyte of low conductivity σ because this lowers the power consumption at a given driving voltage. To optimize for thermodynamic eﬃciency one should therefore maximize the ratio Pout LU 2 .37) where Ltot is the total length of the channel.38) ˆ where ∆V = 2 is the (dimensionless) voltage applied between the electrodes when Vext = ±Re[eiωt ]. cf. 2 8 σH ηℓ0 (1 + δ) 8H(1 + δ)2 (5. and Z is the impedance for a single pair. It pays to squeeze down the channel lid and lower H because this increases the internal hydraulic resistance and enables to pump up against a more decent backpressure. Sec. I is the total current passed though the device. but does not (in theory) aﬀect the pumping velocity.5. and using U ≃ 0. which should hold also at larger voltage although the scaling is diﬀerent. The Q-∆p diagram for the pump is linear. according to the low-voltage linear theory. δ = 0.1. eo 4 The electrical power consumption is Pin = 1 1 ˆ ˆ Re[I ∗ ∆V ] = 2N Re ˆ 2 Z (5.0035 . and Rhyd is the (dimensionless) hydraulic resistance per unit width Rhyd = 12Ltot H3 (5. and V0 = 1 V. N is ˆ the number of electrode pairs. 3. with maximal ﬂow rate Qmax = Qeo obtained at zero backpressure. The work done when pumping up against an external backpressure is Pout = Q∆p which is maximized at ∆p = ∆pmax /2 where 1 Pout = Rhyd Q2 . And ﬁnally. it pays to increase the driving voltage V0 and decrease the capacitance ratio δ.39) (5. Substituting typical numbers for a 1 mM KCl solution in a device with ℓ0 = 5 µm. It pays scale down the device and decrease ℓ0 as much as possible.

42) What this tells us is that an AC electroosmotic pump may be employed to move liquid rapidly through a microchannel. The thermodynamic eﬃciency is maximized at a ≈ 2. 12]. we obtain a thermodynamic eﬃciency of Pout ≈ 10−8 . but not against any signiﬁcant backpressure. with eﬀective pore radius a from a few µm down to 100 nm [8. several authors have introduced a porous material or sintered glass frit into the pumping channel. Pin (5. 9.6 Summary In this chapter we have investigated diﬀerent strategies for breaking the leftright asymmetry of an electrode array to create an AC electroosmotic pump. and experimentally eﬃciencies of around 1 % have been obtained pumping deionized water [9] an up to 5. The thermodynamic eﬃciency is low. We believe it is more fruitful to think of AC electroosmosis as a means for local ﬂow control in a microﬂuidic system. Our results from Ref.5λD . [44] show that strong geometrical conﬁnement of the pumping channel down to H ≪ L does not.6 % for a buﬀer solution with the strongly dipolar additive tris(hydroxymethyl)aminomethane hydrocloride [10]. but the 3D structured design suggested by Bazant and Ben [45] could increase the pumping velocity by an order of magnitude at the expense of an additional lithography step. 11. . From a fabrication point of view the easiest device to produce is that based on asymmetric pairs of electrodes. 10. and to focus on ease integration rather than (lack of) energy eﬃciency. as one might have worried. so one should focus on the use of AC electroosmotic pumps for local generation of ﬂow but not for pumping up against large external backpressure or hydraulic resistance. This should actually be no surprise to us because the same limitation applies to ordinary inline DC electroosmotic pumps. 5. To overcome this. However. introducing porous material in the pumping channel does not enhance the eﬃciency of the AC electroosmotic pump. Hence geometrical conﬁnement could be a route for increasing the thermodynamic eﬃciency. mask the asymmetry of the electrodes and destroy the ability to pump. and determined in each case the design that maximizes the pumping velocity for a given (low) driving voltage. but there is a long way to go from 10−8 .112 Optimization and L ≃ 15 as corresponding to the optimized planar device.

if there is no inert supporting electrolyte the AC Faradaic current should give rise to strong concentration polarization above the diﬀusion-limited current with transient space charge layers and electroosmosis of the second kind. We conclude that fast electrode reactions act to “short-circuit” the Debye layer and thereby suppress electroosmotic ﬂow. provided that the electrode reaction runs reversibly such that electrolysis and bubble formation is avoided.Chapter 6 Conclusion and outlook Conclusion In the present thesis we have studied AC electrokinetics with focus on micropumps based on asymmetric electrode arrays. and the dynamical behaviour at larger voltage is dominated by capacitive charging. unless the reaction itself is limited by mass-transfer of the reactants. We have extended the existing linear theory for AC electroosmosis into the weakly nonlinear regime by accounting for nonlinear capacitance of the Debye layer and Faradaic current injection from electrochemical reactions at the electrodes. Depending on the application it may be feasible to operate a micropump with Faradaic current injection. that is a strongly nonlinear phenomenon that we have have not investigated. If the electrochemically active species are only present in small concentrations with an excess of inert supporting electrolyte. However. On the other hand. Building on recent theoretical work for diﬀuse-charge dynamics in electrochemical systems. The primary eﬀect of the nonlinear capacitance is to change the scaling of the pumping velocity with driving voltage. the Faradaic current from the reaction levels oﬀ when the diﬀusion-limited current is reached. we have constructed a strongly nonlinear model based on bound- . and to shift down the inverse RC time to lower frequency at large driving voltage. Neither the linear nor weakly nonlinear models are able to account for the reversal of the pumping velocity observed experimentally. and therefore we have also studied the strongly nonlinear regime where the classical assumption of uniform bulk electrolyte concentration breaks down.

However. Outlook There remains many open problems and possible directions for future research on strongly nonlinear electrokinetics.. for the voltage range that we have considered. we have contributed by coupling an explicit dynamical solution in the diﬀusion layer to the evolution of the quasiequilibrium Debye layer and to the bulk concentration proﬁle that is nonuniform but stationary in time. What makes this subject so challenging.114 Conclusion and outlook ary layer analysis and matched asymptotic expansions. In particular. but for simple geometries it can be useful for validating the strongly nonlinear theory. Finally. Because the diﬀerent eﬀects interact in a nontrivial way we can envisage . and the experimental results so diﬃcult to interpret. large surface excess ﬂux in the Debye layer leading to nonuniform bulk concentration with convection dominated bulk salt ﬂux. We also account for ﬂuid motion and convection of salt in the bulk. either. but not against a signiﬁcant backpressure. It is important to realize that AC electrokinetic micropumps can provide fast pumping of the order of ∼ mm/s. but the technology is still attractive due to ease of integration in a microsystem or lab-on-a-chip. i. salt depletion in the diﬀusion layer followed by breakdown of local electroneutrality and development of an extended space charge layer. The solutions are expected to be close to optimal also at larger voltage as long as electroosmosis of the ﬁrst kind dominates the electrokinetic ﬂow. A full solution of the Poisson–Nernst–Planck equations provides only little insight into the solution structure. so we are not able to account for the experimentally observed reversal of the pumping direction with our strongly nonlinear model. and even breakdown of dilute solution theory in the highly charged Debye layer. Thermodynamic eﬃciency will therefore always be low.e. We have obtained numerical solutions for this model in both a simple 1D geometry and in 2D for a symmetric and an asymmetric electrode array. is that there is a whole zoo of strongly nonlinear eﬀects showing up at roughly the same time. electroosmosis dominates. because even our strongly nonlinear model based on local electroneutrality breaks down at “very large” but still experimentally relevant voltage. Moreover. we have attempted to include existing theory for nonequilibrium double layers and electroosmosis of the second kind into our model. while bulk electroconvection contributes to pumping in the reverse direction. and because it oﬀers local control over the ﬂuid motion in the device. we have looked at various ways of breaking the symmetry of an electrode array and determined in each case the design that maximizes the pumping velocity at a given low driving voltage. We predict that electroosmotic ﬂow of the ﬁrst and second kind contributes to pumping in the forward direction.

Finally. In particular. and in particular the dynamics of the collapse of the layer when the current reverses is a challenging problem. However. ∆D. the dependence on the model parameters ε. before this can fully be exploited to optimize the pump performance or to engineer robust “forward” operation at large voltage. deserves more attention and systematic variation. it is necessary to gain a more detailed understanding of the dynamical behaviour. δ. e. and possibly ˜eq . where local electroneutrality breaks down in the bulk region close to the poles of the sphere due to large surface ﬂuxes towards the equator. In relation to the simple ad hoc modiﬁcations that we suggested. by studying simple geometries such as metal spheres or cylinders. Alternatively. The space charge layers developing in this case are immediately describable by the existing theory for DC Faradaic conduction.g. the quality of the model would certainly improve by solving a 1D semi-inﬁnite diﬀusion problem on the interval yo (t) ≤ y < ∞ for the ˇ ˇ concentration in the diﬀusion layer.115 complex dynamical behaviour in geometries such as the AC electroosmotic pump with 3D structured electrodes.. and it would be a great advantage to device a timestepping algorithm to allow for better resolution in time when computing the periodic solution. the existing theory for nonequilibrium double layers could deserve a rigorous account for the solution structure when the width of the space charge layer changes in time with the applied current. The theoretical description for DC rather than AC driving is simpliﬁed because the diﬀusion layer and bulk region merge and the number of nested boundary layers drop by one. To this end it would also be interesting to actually determine the experimental ˜ values of δ and ζ eq by ﬁtting the linear response theory to impedance spectra recorded as a function of frequency and with a DC voltage applied relative to a distant reference electrode. lower ζ values of the capacitance ratio δ are relevant to many experiments. . It would therefore be interesting to extend the work of Chu on a metal sphere in an externally applied DC ﬁeld to even larger ﬁelds.

116 Conclusion and outlook .

2) (A.2. with ¯ 0 = n · ∇φ.7 in the absense of Faradaic electrode reaction. ˜ ¯ ˜ ˜ Vext − φ = ζ − q δ.8) (A.4) The solution to the electrical problem determines the electroosmotic slip velocity that is boundary condition for the bulk time average Stokes ﬂow ¯ problem for the ﬂuid velocity u and pressure p ¯ ¯ 0 = ∇ · σ.3) (A.1 Weakly nonlinear ﬁnite element model In this section we give an example of our implementation of the weakly nonlinear model from Sec.1) on the insulating channel walls and a set of dynamical boundary conditions ¯ ∂t q = n · ∇φ. ¯ p (A. ¯ 0 = ∇ · u. 2.5) (A. The leading order electrical problem reduces to a Laplace problem in the bulk ¯ 0 = −∇2 φ. ¯ where σ is the stress tensor deﬁned by ¯ ¯ σij = −¯ δij + ∂i uj + ∂j ui . (A. q (A.Appendix A FEMLAB code A.7) .6) (A. on the electrodes. where the zeta potential is given by ˜ ζ = −2 sinh−1 (˜/2).

12) In Femlab we use so-called strong or ideal constraints to impose the no-slip condition on the channel walls. (A. whereas ¯ we want it to be a constraint on u only. 0=− = Ω Ω ¯ ¯ u · (∇ · σ) + ξ ∇ · u dr ∂Ω ¯ ¯ ∇u : σ − ξ ∇ · u dr + ¯ u · (n · σ) ds.118 The slip condition on the electrodes is ˜ s¯ ¯ u = ζ∇ φ . The electrical problem is converted to weak form by multiplying Eq.10) ¯ ∇ϕ · ∇φ dr + ϕ ∂t q ds. FEMLAB code (A. ˜ electrodes which is supposed to be satisﬁed for all test functions ϕ. i. and the incompressibility constraint by a test function ξ for the pressure. and using Eq.. (A. 0=− = Ω ¯ ϕ ∇2 φ dr = Ω Ω ¯ ∇ϕ · ∇φ dr + ∂Ω ¯ ϕ n · ∇φ ds (A.e.13) Unlike the usual notation with Gauss’ law our boundary normal n points out of the electrodes and channel walls and into the electrolyte and Ω. and use this to ¯ f ¯ eliminate n · σ in the boundary integral to get 0= Ω 1 ¯ ¯ ∇u : σ − ξ ∇ · u dr + electrodes ˜ s¯ ¯ u · ¯ + f · u − ζ∇ φ f ds.3) to ¯ eliminate n · ∇φ. The reason is that with an ideal constraint Femlab ¯ ¯ would consider Eq. but for the slip condition on the electrodes we need weak constraints.6) by a test function u for the velocity. i.e. and also want the reaction forces or Lagrange multipliers from that constraint to act in the ﬂow problem only and not in the electrical problem. (A.11) electrodes which is to be satisﬁed for all ̺..9) equally much as a constraint on φ as on u. .9) ¯ and on the channel walls we impose the no-slip condition u = 0.1 The Gouy– Chapman–Stern boundary condition for the potential drop across the double layer is imposed by multiplying a test function ̺ for the accumulated charge and integrating over the electrodes 0= ¯ ˜ ˜ ̺ Vext − φ − ζ + q δ ds. and integrating over Ω.1) by a test function ϕ for the potential and integrating over the volume Ω of the computational domain. employing Gauss’ law. (A. (A. The ﬂow problem is converted to weak form by multiplying the force balance (A. Therefore we introduce the reaction force ¯ = n · σ from the electrode surface as a separate variable. (A.

2. G2 = 5..weak = {{’phix_test*phix+phiy_test*phiy’ ’-p_test*(ux+vy)’ .shape = {[1:7] [1:7] [1:4] [1:4] [1:4]}.fillet(rect2(G2/2+W1+G1.R) .equ.. fem.05 0.const. fem.G2/2+W1. ’fx_test*(u-us)+fy_test*(v-vs)’ ’u_test*fx+v_test*fy’} . fem.. fem.mic.w = 1. % BOUNDARY CONDITIONS % group [1 2] narrow and wide electrode % group [3] insulating walls % group [4] master periodic boundary % group [5] slave periodic boundary fem.-1. W2 = 7.geom = rect2(0.weak = {{’phi_test*qt’ ’q_test*(Vext-phi-zeta+q*delta)’ ..equ.dtu.const. W1 = 1.mesh = meshinit(fem.form = ’weak’.shape = [2 2 2 1 2 2 2].A..L.’Radii’.05. .2]).5. % GOVERNING EQUATIONS fem..R). fem. fem.V0 = 10.dk/mifts % AC ELECTOOSMOTIC PUMPING BY WEAKLY NONLINEAR MODEL clear fem % GEOMETRY G1 = 1. fem.R).1 Weakly nonlinear ﬁnite element model 119 ¯ for all test functions u.dim = {’phi’ ’u’ ’v’ ’p’ ’q’ ’fx’ ’fy’}. f .bnd. L = W1+W2+G1+G2. ξ. In this form the problem can be typed directly into Femlab [47] as shown in the code example below: Code also available on http://www.ind = {[4 9 10] [6 11 12] [2 3 5 7] [1] [8]}. fem.fillet(rect2(G2/2.’Hmaxedg’. ’ux_test*(-p+2*ux)+vy_test*(-p+2*vy)+(uy_test+vx_test)*(uy+vx)’}}. 0.’Hmax’.. and subject to the no-slip condition u = 0 on the channel walls.R). fem. . H = 25.0.bnd. % driving frequency % driving voltage % surface capacitance ratio % VARIABLES fem.’Radii’..[4 6.shape = {[1:4]}.delta = 1.. R = 0. ’zeta’ ’-2*asinh(q/2)’ ’us’ ’zeta*phiTx’ ’vs’ ’zeta*phiTy’}.expr = {’Vext’ {’V0*sin(w*t)’ ’-V0*sin(w*t)’ {} {} {}} .bnd. % PARAMETERS fem..G2/2+W1+G1+W2.const.H) . % % % % % % % gap between electrodes = reference dimension width of narrow electrode width of wide electrode gap between adjacent electrode pairs height of channel period of electrode array electrode corner radius fem. % MESH fem.sdim = {’x’ ’y’}.bnd.-1.

xmesh = meshextend(fem).m % from www.sol = asseminit(fem).elem and clean up fem.C.. map. fem.’Tlist’. .shape = {[4] [6 7]}.map = {map}.ind = {{’1’}}. src. % implement no-slip and periodic boundary conditions as ideal constraints fem. fem.com/support/knowledgebase/942.elem = ’elcplextr’.bnd.map = {’0’ ’0’ ’0’ ’0’}.geomdim = {{{} dst {}}}. % POINT CONDITIONS % clamp pressure at a single point to fix the pressure level % clamp constraint force [fx fy] where the electrodes and substrate meet % and [fx fy] are overruled by the ideal constraint Lagrange multipliers fem. % declare {1} as map *from* destination *to* intermediate mesh dst.sg = ’1’.g = {’1’}. dst..sol = femtime(fem.[0 4*pi]. map. cpl. map.’Solcomp’.src = {{{} src {}}}..sv = {’11’.{’u’ ’v’ ’p’ ’fx’ ’fy’}.. cpl. For an explanation see: elcplextr_example.type = ’linear’. % This is black magic.ind = {[1] [3 6 7 10]}.elem = {cpl}.constr = {{} {} {’u’ ’v’} {} .expr = {{’phi’} {’u’} {’v’} {’p’}}.sol. fem.constr = {’p’ {’fx’ ’fy’}}. % COUPLING ELEMENT FOR PERIODIC B.pnt. src. % finally add cpl to fem. % TRANSIENT SOLUTION OVER 2 PERIODS USING FEMTIME % solve nonlinear electrical problem fem.var = {’phi_cpl’ ’u_cpl’ ’v_cpl’ ’p_cpl’}.map = {{’1’} {’1’} {’1’} {’1’}}.php clear cpl src dst map % extrusion coupling variables acting in geometry one cpl. ’fx_test*(u-us)+fy_test*(v-vs)’ ’u_test*fx+v_test*fy’} {} {} {}}. map. % linear transformation *from* destination *to* intermediate mesh map.pnt. fem. cpl. clear cpl src dst map % BUILD EXTENDED MESH fem.dg = ’1’.’Solcomp’.{’phi’ ’q’}).dv = {’1’.’2’}. % solve linear flow problem with constant system matrix t = fem.pnt. cpl. cpl.’12’}. {’phi-phi_cpl’ ’u-u_cpl’ ’v-v_cpl’ ’p-p_cpl’}}..sol = femlin(fem.ind = {{’8’}}.120 FEMLAB code {’phi_test*qt’ ’q_test*(Vext-phi-zeta+q*delta)’ .comsol. .tlist. % source expressions and trivial map {0} *from* source *to* intermediate mesh src..

tt = 2*pi*(0:nt-1)/nt.Q. % compute time average flow rate <Q> Qavg = postint(fem.sol.’Keep’.’U’.{’K’ ’N’}).’r:’) 3.[0 4*pi].5 1 0.1.5 0 0 2 4 6 8 10 12 Time t Figure A.mean(fem. Result from Femlab’s built-in timestepping algorithm femtime (solid line) and time average from custom written pseudospectral algorithm femfreq (dashed).tt.sol = femlin(fem.’Plist’.’t’.’b. % solve nonlinear electrical problem fem.’Edim’.’Symmetry’. 121 % COMPUTE PERIODIC SOLUTION USING CUSTOM WRITTEN ALGORITHM FEMFREQ fem0 = fem.1: Flowrate Q as a function of time t from weakly nonlinear Femlab code example. .’t’. while the code example computes a stationary solution at each point in time corresponding to the limit εSc ≫ 1. nt = 16.sol.’Keep’. for εSc ≪ 1 we ought to solve a stationary Stokes ﬂow problem driven by the time average ˜ s¯ slip ζ∇ φ ..1.’Solcomp’.1.{’phi’ ’q’}.1.’u’.sol = femfreq(fem. .’Plist’.’Dl’.’Dl’.1:length(t)).t. plot(t.A.tt.fem.’Solnum’.5 2 1.’u’. because the problem is linear we get the same result for the time average ﬂow rate Q in any case.u.1 Weakly nonlinear ﬁnite element model ’Pname’. % solve linear flow problem fem. However. ’Pname’.fem.’U’. Strictly.{’K’ ’N’}).’Edim’.5 3 Flowrate Q 2.sol.’Solcomp’.’odd’).’U’.{’u’ ’v’ ’p’ ’fx’ ’fy’}.’Tlist’..2)).-’. % compute volume flow rate Q(t) [large Schmidt number limit] Q = postint(fem.[Qavg Qavg].

20) electrodes . for all test functions C for c. electrodes ¯ for all test functions ϕ for the potential φ at all times.19) s for all test functions ̟ for w. s (A.8. c ¯ c (A.122 FEMLAB code A. and performing partial integration only on the ∇2 c term.2.18) ¯ using ∇ · u = 0. it sometimes gives more stable ¯ ¯ numerical results to rewrite ∇ · Fa in non-conservative form ¯ ¯¯ ¯ ∇ · Fa = ∇ · (−Da ∇¯ + Pe u c) = −Da ∇2 c + Pe u · ∇¯. Alternatively. The weak form of the bulk electrical problem ¯ ∇ · J becomes 0=− = Ω Ω ¯ ϕ∇ · J dr = Ω ¯ ∇ϕ · J dr + ¯ ∂Ω ¯ ϕ n · J ds (A.16) for all test functions ̺ for q .14) ¯ ∇ϕ · J dr + ϕ Jn ds. c (A. which eﬀectively determines the normal ﬂux ˜ ˇn . where we used partial integration to move the ˜ derivative ∇ onto ̺ and assumed ˜ = 0 at the electrode edges. The conservation law for the accumulated charge q in the Debye layer takes the form ˜ 0= ̺ ∂t q + ˜ electrodes ¯ Jn + ε∇ · ˜ ds = s ̺ ∂t q + Jn ) − ˜ electrodes ¯ ε∇ ̺ · ˜ ds.2 Strongly nonlinear ﬁnite element model In this section we give an example of our implementation of the strongly nonlinear model from Sec. where we used the ¯n to eliminate the normal current n· J in the boundary ¯ Lagrange multiplier J integral. (A. (A. and the boundary condition for the potential drop across the double layer is imposed by requiring 0= Jn (Vext electrodes ¯ ˜ ˜ − φ − ζ + q δ) ds. For the bulk s salt concentration c we write ¯ 0=− = Ω Ω C∇ ¯ · Fa dr = Ω ¯ ∇C · Fa dr + ˇ ¯ C n · Fa ds ∂Ω (A.15) ¯ for all test functions Jn for Jn .17) ¯ ∇C · Fa dr + C Fn − electrodes ¯ ∆D Jn ds. ¯ The conservation law for the surface excess accumulated salt w becomes ˜ 0= ̟ (∂t w ˜ electrodes ˇ ˜ + Fn + ε∇ · Γ ds = s ̟ (∂t w ˜ electrodes ˇ ˜ + Fn ) − ε∇ ̟ · Γ ds. 2. whereas the excess amount of salt is determined by F 0= Fn w − ˜ q 2 + 4ˇ + ˜ c √ 4ˇ ds.

In the special case of equal diﬀusivities where ∆D = 0. Otherwise it is necessary to rewrite the body force as ¯ = ∇2 φ ∇φ = ∇ · T .24) With this formulation we can use Gauss’ law to move the derivative from ¯ T onto the test function u for the ﬂuid velocity.A. with the second derivative undeﬁned on element boundaries. p.1. For simplicity we assume an open system and impose unit concentration c = 1 and no ¯ ¯ = 0 boundary conditions far above the electrode array. This allows ¯ ¯ ¯ the computational cost of solving the full coupled problem to be reduced . u.22) ¯ where T is the Maxwell stress tensor ¯ cf. and γ have ˜ ˇ ˇ ¯ even symmetry.25) for all test functions u.2 Strongly nonlinear ﬁnite element model 123 ˇ for all test functions Fn for Fn . (A. and the f c problem does not arise. whereas c. and us = ζ ˇ s ¯ in the Debye and diﬀusion layers. Fn .23) ¯ ¯ ¯ ¯ ¯ Tij = ∂i φ ∂j φ − δij ∂k φ ∂k φ/2.2. ¯ Eq.22) expresses ¯ in terms of the gradients ∇¯ and ∇φ only. and Jn have odd symmetry in time. 4. (A. Finally. Eq. it is not necessary to subtract Fn − ∆DJn ˇ ′ ) is blind to any time average component by because the kernel G(t − t construction. and µ are constant in time. and the weak form of the Stokes ﬂow problem becomes 0= Ω ¯ ¯ ¯ ∇u : (σ + T ) − ξ ∇ · u dr + electrodes ¯ u ˜ u · µ + f · (¯ − us ) ds. ξ. In the code example below we consider the simple equal diﬀusivity case with ∆D = 0 for the symmetric geometry from Sec. (2. For shear force n · σ ¯ ˜ ¯ ∆D = 0 both φ. q . (A. ¯ ¯ f c ¯ (A. w. the surface excess concentration γ = c − c at y = 0 is determined by the (excess) ambipolar ﬂux in and out ˇ ˇ ¯ ˇ of the diﬀusion layer 0= Y γ− ˇ electrodes ε 1 ωDa T T 0 ˇ ¯ G(t − t′ ) Fn − ∆DJn dt′ ds. and subject to the no-slip condition on the ¯ ¯ ¯ channel walls.21) ˇ ¯ for all test functions Y for γ . f . (A. Here µ = n · (σ + T ) is the total reaction force from the ˜+ ς ∇ φ is the eﬀective slip velocity induced ˜ electrode surface. In the strongly nonlinear model we need to include the bulk electrical body force ¯ into the Stokes ﬂow problem. The dependence on ∇2 φ or ∇2 c is problematic because ¯ with standard ﬁnite element approximations the solution is only continuous and piecewise diﬀerentiable. ¯ ¯ ¯ f (A.60). where f 2 ¯ c ¯ ¯ ¯ = ∇2 φ ∇φ = − ∇¯ · ∇φ + ∆D∇ c ∇φ.

dk/mifts % AC ELECTOKINETIC FLOW BY STRONGLY NONLINEAR MODEL clear fem % G W H L R GEOMETRY = 1. and excessr ˇ ∞ computes the integral ω/Da 0 1 − 1 dˇ given γ (t) and c by numerical y ˇ ¯ c ˇ c ¯ quadrature.Pe = 0.equ.delta = 1.fillet(rect2(-W/2.H) .V0 = 50.expr = {’rho’ ’(cx*phix+cy*phiy)/c’ . ’ux_test*(-p+2*ux)+vy_test*(-p+2*vy)+(uy_test+vx_test)*(uy+vx)’ . % MESH fem.sdim = {’x’ ’y’}. ’Jx’ ’-c*phix’ ’Jy’ ’-c*phiy’ ’Fx’ ’-cx+Pe*u*c’ ’Fy’ ’-cy+Pe*v*c’}...0.0. 0. fem.const.124 FEMLAB code by a factor of two as compared to the case ∆D = 0.1. % % % % % gap between electrodes width of electrodes height of channel period of electrode array electrode corner radius fem.shape = {[1:5]}.weak = {{’phix_test*Jx+phiy_test*Jy’ .-1.75. % GOVERNING EQUATIONS fem. fem. fem. Code also available on http://www.shape = [2 2 2 2 1 2 2 2 2 2 2 2 2]. % % % % % driving frequency driving voltage surface capacitance ratio Debye layer thickness Peclet number % VARIABLES fem...[2.const.init = {{0 1 0 0 0 0 0 0 0 0 0 0 0}}. The convection from ﬂuid motion induced in the diﬀusion layer is neglected when solving the coupled electrohydrodynamical problem.’Hmaxedg’.’Hmax’. ’-p_test*(ux+vy)’ ’u_text*rho*phix+v_test*rho*phiy’}}. fem.dim = {’phi’ ’c’ ’u’ ’v’ ’p’ ’q’ ’w’ ’Jn’ ’Fn’ ’gamma’ ’Zeta’ ’fx’ ’fy’}.R).epsilon = 0.const.dtu.mic.R).equ.equ. The auxiliary function fkernel evaluates the kernel G(t − t′ ) at y = 0. fem.const.25. % BOUNDARY CONDITIONS . = 2.001. and ﬂow induced in the diﬀusion layer.L/2. = 5..W/2.. ’-cx_test*cx-cy_test*cy-c_test*Pe*(u*cx+v*cy)’ . fem.1]).mesh = meshinit(fem. bulk electroconvection.’Radii’. fem. % PARAMETERS fem. fem. = 0.omega = 1.const.geom = rect2(0. but in a postprocessing step we single out the components of the net ﬂuid motion due to electroosmotic slip..equ.form = ’weak’.. fem. fem. = W+G.

.’Out’. fem. ’jx’ ’-(1+2*Pe)*w*phiTx’ ’jy’ ’-(1+2*Pe)*w*phiTy’ ... fem. iJn = iJn(iZeta).pnt.omega)*Jn.init = {{0 1 0 0 0 0 eps 0 0 0 0 0 0}}.(1:ndofs)’. ’Solcomp’. % GET INDICES OF [Zeta gamma Jn c] IN GLOBAL SOLUTION VECTOR iZeta = find(asseminit(fem..bnd.’tlist’. fem.epsilon/fem. ’c_test*Fn’ ’Fn_test*(w-sqrt(q^2+4*(c+gamma))+sqrt(4*(c+gamma)))’ .’Init’.:) = Zeta.sol.. R = excessr(gamma..’U’). .sol = femfreq(fem. iJn = asseminit(fem..’Kernel’. igamma = asseminit(fem.’U’).*R.shape = {[1:13] [1:13]}.u.’Tlist’.bnd.shape = {[1:13] [1:5] [1:5] [1:5] [1:5]}. .tlist. % COMPUTE EXCESS POTENTIAL DROP ACROSS THE DIFFUSION LAYER gamma = fem. fem...sol.. fem.{’phi’ ’c’ ’q’ ’w’ ’Jn’ ’Fn’ ’gamma’ ’u’ ’v’ ’p’ ’fx’ ’fy’}. % BUILD EXTENDED MESH fem..’U’)).tlist).const. ’fx_test*(u-us)+fy_test*(v-vs)+u_test*fx+v_test*fy’ . Zeta = sqrt(fem..bnd. ic = ic(iZeta).ind = {[1] [4]}.weak = {{’phi_test*Jn’ ’Jn_test*(Vext-phi-zeta+q*delta)’ .{’Zeta’ ’Jn’}.bnd.ind = {[2 6] [4] [1] [5] [3]}.’U’).sol. u0 = fem. . igamma = igamma(iZeta)..’Init’. ’w_test*(wt+Fn)-epsilon*(wTx_test*Gammax+wTy_test*Gammay)’ .u(igamma. % POINT CONSTRAINTS % clamp weak constraint force where overruled by ideal constraints fem.’leapfrog’.fkernel(0..sol.constr = {{’fx’} {’fx’ ’fy’}}.pnt.:).A.const..u(iJn. c = fem.bnd...2 Strongly nonlinear ﬁnite element model % group [1] electrode % group [2] substrate % group [3] even symmetry % group [4] odd symmetry % group [5] "infinity" fem.nt.4)).bnd.:).’Init’.{’Zeta’ ’gamma’}. Jn = fem. 125 % COMPUTE PERIODIC SOLUTION USING FEMFREQ nt = 32. ndofs = flngdof(fem). ’const’ ’const’ ’const’ ’const’ ’const’}. ’Rule’.’U’.c).expr = {’Vext’ {’V0*sin(omega*t)’ {} {} {} {}} .. fem. .’Out’. fem.xmesh = meshextend(fem).’U’.{’Zeta’ ’c’}.pnt.. ’zeta’ ’-2*asinh(q/(2*sqrt(c+gamma)))’ ..u(ic. ’q_test*(qt+Jn)-epsilon*(qTx_test*jx+qTy_test*jy)’ ... fem. u0(iZeta.’Out’.constr = {{} {’u’ ’v’} {’u’} {’u’ ’phi’} {’c-1’}}.’Out’..sol = femsol(u0.1). ’Gammax’ ’-(1+2*Pe)*q*phiTx’ ’Gammay’ ’-(1+2*Pe)*q*phiTy’ .’Init’.. ’us’ ’(zeta+Zeta)*phiTx’ ’vs’ ’(zeta+Zeta)*phiTy’}.{’odd’ ’const’ ’odd’ ’even’ ’odd’ ’even’ ’even’ . fem.’U’..(1:ndofs)’. ’Symmetry’. ’gamma_test*(gamma-Fntt*sqrt(epsilon/omega))’} {} {} {} {}}. tlist = 2*pi*(0:nt-1)/nt.{’Zeta’ 1}.(1:ndofs)’. ic = asseminit(fem..

’Keep’. ’Const’.{’u’ ’v’ ’p’ ’fx’ ’fy’}. ’Pname’.2)).sol.{’K’ ’N’}).2)).’Plist’.’U’..’Out’. .{’u’ ’v’ ’p’ ’fx’ ’fy’}. . ’Const’. % flow due to slip induced in the diffusion layer uec = femlin(fem.’x*v-y*u’.’U’.. . % COMPUTE ANGULAR MOMENTUM FROM FLUID VORTEX Leo = postint(fem.’Keep’.fem.fem. .’t’.’Solcomp’.’Plist’..’U’.’U’.tlist.’U’.2)). ’Pname’.{’K’ ’N’}).’U’.fem. Lb = postint(fem.’Out’..’x*v-y*u’. Ltot = Leo + Lec + Lb.’Keep’. .’t’.’Out’.’t’. ’Pname’. ..126 FEMLAB code % COMPUTE INDIVIDUAL FLOW COMPONENTS % flow due to slip induced in the Debye layer ueo = femlin(fem.mean(ub.. ’Const’.{’zeta’ 0 ’Zeta’ 0}.sol..’U’.’Solcomp’.’Plist’.’U’. Lec = postint(fem..{’u’ ’v’ ’p’ ’fx’ ’fy’}.{’zeta’ 0 ’rho’ 0}.’Solcomp’..tlist.’U’.tlist..mean(uec.{’Zeta’ 0 ’rho’ 0}. .mean(ueo..{’K’ ’N’}). % flow due to bulk electroconvection ub = femlin(fem.’x*v-y*u’..sol.

2. . t) ∂t U0 − 1 T T 0 1 T T 0 G(t. we wish to use a relaxation method to compute the periodic solution instead of the standard time-steppig algorithm. M(U. t) the residual vector. G(t.31) where the coeﬃcients 1 ˆ Um = N N U(tn )e−imω tn n=1 (A. . (A. and N = −∂M/∂U the matrix that couples the Lagrange multiplier into the problem.3 Periodic integro-diﬀerential equations in time We wish to solve an integro-diﬀerential equation in time of the form D ∂t U + subject to the constraint 0 = M. this may cause problems in the Newton itertions if those are strongly nonlinear functions. this is identical to the generic form of a time-dependent problem after discretization in Femlab. t′ ) δU(t′ ) dt′ = L0 − K δU − NT Λ. i.29) The derivative and integral terms of the trial solution are absorbed into the residual vector L0 = L(U0 . t) is the mass matrix. L(U. t) the constraint vector. Λ(t) a Lagrange multiplier for that constraint.28) G(t. (A. (A. (A. U(t) = U(t + T ). . Linearizing around a given trial solution U0 (t) we get D ∂t δU + subject to 0 = M0 − N δU. t′ )U(t′ ) dt′ = L − NT Λ.26) 1 T T 0 G(t. t and K = −∂L/∂U. t′ ) the kernel matrix in the convolution. N then it can be written as a discrete Fourier transform N U(tn ) = m=1 ˆ Um eimω tn .32) .A. Apart from the convolution operator.27) where the solution U(t) is periodic in time.3 Periodic integro-diﬀerential equations in time 127 A. However. t) − D(U0 . (A. If we represent the temporal solution for U by N equispaced points in time tn = n T /N for n = 1. t′ )U0 (t′ ) dt′ . We neglect the dependence of D and N on U when linearizing. Here D(U.30) and further M0 = M U0 .. .e. (A.

n . [68]. e. 2.38) for n = 1. By construction Um = Um+N .33) where the prime indicates that the terms m = ±N/2 are counted half only. and for a nice introduction to spectral methods in general. . t′ )U(t′ ) dt′ ≈ 1 N N G(tn . N tan[ω(t − t′ )/2] (A. The function in the curly braces is the periodic sinc function SN (t − t′ ) 1 SN (t − t ) = N ′ N/2 ′ m=−N/2 eimω(t−t ) = ′ 1 sin[N ω(t − t′ )/2] . the cost of one step forward in time by a standard timestepping algoritm is O(1) which is much more memory eﬃcient but sometimes requires integrating for a long time before a periodic solution is reached. (A.34) which allows the time detivative of U(t) to be evaluated as N ∂t U(tn ) = m=1 ′ SN (tn − tm )U(tm ). which would reduce the computational cost for a diﬀerential equation from O(N 2 ) to O(N ) at the expense of the spectral accuracy. For comparison.g. (A. .. m=1 (A. . (A. If the problem cannot be solved accurately with just a few Fourier components.n . we could replace the spectral derivative with a low order ﬁnitediﬀerence scheme.37) Nn δUn = M0. tm )U(tm ). .128 FEMLAB code ˆ ˆ can be computed by a fast Fourier transform. i. For an explanation why this is “appropriate”.36) Substituting into the linearized problem we get N m=1 ′ Dn SN (tn − tm ) + 1 Gnm δUm + Kn δUn + NT Λn = L0.35) The convolution integral can be approximated as 1 T T 0 G(tn . n N (A.. the point is that for harmonically driven weakly nonlinear systems the number of Fourier components required to represent the solution is often fairly small. Unless N is small this problem will make most linear solvers choke.e. a simultaneous matrix problem for all the solution components δUm . see Ref. If the problem . ∗ ∗ ˆ ˆ ˆ and since U is real we also have Um = U−m = UN −m . The appropriate way of interpreting U(t) for t = tn is U(t) = 1 N N/2 ′ m=−N/2 N ˆ Um eimωt = n=1 U(tn ) 1 N N/2 ′ m=−N/2 eimω(t−tn ) . N . the leapfrog scheme [∂t U]n ≈ [Un+1 − Un−1 ]/2∆t.

dk/mifts. a pseudo-timestepping algorithm designed speciﬁcally for rapid convergence towards the periodic solution might provide a fairly inexpensive solution to the problem. We also support odd.A. (A.3 Periodic integro-diﬀerential equations in time 129 involves convolution there is no obvious way to work around the nonlocal coupling in time. where G(t−t′ ) is a some scalar function supplied by the user via the property ’Kernel’.dtu.39) Substituting this and the truncated series Eq. and it is originally deﬁned in terms of an inﬁnite Fourier series G(t − t′ ) = ∞ n=−∞ ′ ˆ Gn einω(t−t ) . (A. even. . (A. we have not had time to look further into this subject. However. i. t′ ) that could depend explicitly on both time and spatial position. but it serves our purpose. This approach is signiﬁcantly less ﬂexible than a general G(t. the convolution kernel does not depend explicitly on time but only on the time delay. namely. 0}.42) Implementation: FEMFREQ We have implemented the relaxation scheme in a function femfreq. We do not include the code here but it is available at http://www.41) k=−N/2 This essentially tells us how to truncate the inﬁnite series Eq.e. it is equivalent to replacing SN (t − t′ ) by a piecewise linear interpolant. If the spectral time derivative is replaced by the leapfrog rule. In that case it is more consistent to truncate the convolution kernel as GN (t − t′ ) = N T 0 G(t − t′′ )PN (t′′ − t′ ) dt′′ . For problems involving convolution we abuse the syntax for second order problems by T 2 interpreting the term E ∂t U instead as E 0 G(t−t′ )U(t′ ) dt′ . For our particular application where the nonlocal coupling in time comes about from our semi-analytical model for the dynamics in the diﬀusion layer. the “hat” function PN (t − t′ ) = max{1 − N |t − t′ |. (A.33) for U(t) into the convolution integral we obtain 1 T where GN (t − t ) = ′ T 0 G(tn − t′ )U(t′ ) dt′ = 1 N N m=1 N/2 ′ ′ ˆ Gk eikω(t−t ) . G = G(t − t′ ). It computes periodic solutions to simple ﬁrst order problems in time with a syntax similar to Femlab’s standard time-stepping algorithm femtime.mic..39) in a consistent way. GN (tn − tm )U(tm ). and constant symmetries in the solution which allows the computational cost to be reduced.40) (A. (A.

130 FEMLAB code .

Appendix B Paper published in Phys. Henrik Bruus. and nonlinear surface capacitance Authors Laurits Højgaard Olesen. Faradaic current injection. Rev. 056313 (2006) (16 pages) . E Title AC electrokinetic micropumps: the eﬀect of geometrical conﬁnement. and Armand Ajdari Reference Phys. E 73. Rev.

132 Paper published in Phys. Rev. E .

The latter fall into two major categories: One category comprises mechanical actuation and deﬂectable membranes to create pumps and valves. Often large external pumps. ©2006 The American Physical Society 056313-1 . DOI: 10. and fast pumping velocities mm/ s with low driving voltage of a few volts only.73.65. which was soon after demonstrated experimentally by Brown et al. when a dc current is drawn from electrodes integrated in a microﬂuidic system there are problems with bubble formation due to electrolysis. Based on general symmetry arguments. y. These have proven to be versatile and simple to fabricate and operate. the theory disagrees with experimental observations in several ways: e. III we describe our model for the electrokinetic system. The second category involves no moving parts but uses electric ﬁelds to induce electrokinetic pumping. VII we conclude the paper. Technical University of Denmark.2 Henrik Bruus. F-75231 Paris Cedex 05. a pumping velocity scaling as the square of the driving voltage. the above analysis is strictly valid only at low driving voltage V0 kBT / e 25 mV. Studer et al. In Sec. such as syringe pumps. 10 rue Vauquelin. As a ﬁrst step we here address a few generalizations of the existing linear response theory to the weakly nonlinear regime by taking into account.056313 PACS number s : 47. Ramos et al.32. and iii nonlinear surface capacitance of the Debye layer. This provides a ﬁrm starting point when in Sec. and we investigate the effect of Faradaic current injection in a linearized scheme.90. We report here that these effects indeed affect the pump performance in a way that we can rationalize by physical arguments. whereas experiments are usually performed with at least a few volts. We therefore extend the latter theories to account for three experimentally relevant effects: i vertical conﬁnement of the pumping channel. except that in the original paper Ajdari predicted pumping in the “reverse” direction at low frequency when Faradaic electrode reactions were included in the model 5 . and ﬁnally. 0 but looks more like a linear scaling 10. ii Faradaic currents from electrochemical reactions at the electrodes. Department of Micro and Nanotechnology. France Received 9 October 2005.1. d. in Sec. within the thin Debye layer approximation. found a reversal of the pumping direction at driving frequencies well 1539-3755/2006/73 5 /056313 16 above that for maximal pumping in the “forward” direction 11 and Ramos and co-workers observed reversal of the pumping direction on a traveling-wave device for driving voltages above 2 V at 1 kHz driving frequency 13. but there are some difﬁculties with further downscaling. several groups have observed bubble formation and electrode degradation at low frequency and high voltage which indicates that electrode reactions are actually taking place 8–11.g. performed a more detailed linear response investigation of the pumping mechanism and found a speciﬁc pumping direction. and nonlinear surface capacitance MIC. The existing theory does not give many clues as to the mechanism for this reversal. but for portable systems a number of on-chip micropumps have been developed over the last decade. As a consequence. However. the effect of Faradaic currents both in a linearized scheme and using the full nonlinear Butler-Volmer reaction kinetics. are used to deliver the necessary pressures.14 . 6 and later by several other groups obtaining pumping velocities mm/ s with driving voltage of a few volts 7–14 . However. ac electro-osmosis has recently been observed to induce ﬂuid motion over pairs of microelectrodes 2–4 .. We also include the nonlinear surface capacitance of the Debye layer as described in Gouy-Chapman theory. IV we extend the linear analysis for low driving voltage to study in more detail the effects of the device geometry and vertical conﬁnement of the system. local. In Sec. their experiments were performed with driving frequencies in the MHz range which is of the order of the Debye frequency for the electrolyte. they also displayed many quantitative and qualitative discrepancies with existing theories. V we study the nonlinear model both with and without Faradaic current. These discrepancies between experimental observations and existing theory demonstrate the need for a more complete theoretical understanding. 056313 2006 ac electrokinetic micropumps: The effect of geometrical conﬁnement.11 or tends to saturate at large voltage 13 . classical dc electroosmosis requires a relatively large voltage and ﬁeld strength. DTU bldg. and a well-deﬁned frequency for maximal pumping 15 . published 31 May 2006 Recent experiments have demonstrated that ac electrokinetic micropumps permit integrable. Theoretically. Moreover. UMR CNRS-ESPCI 7083. experimentally the pumping velocity does not always compare well with the predicted V2 scaling. Faradaic current injection. II we ﬁx the device geometry and in Sec. However. However. Moreover. which was attributed to strong Faradaic current injection 16 .14 . DK-2800 Kongens Lyngby. VI we summarize our results and compare to experiments reported in the literature. whereas ac electro-osmosis generally occurs around the inverse RC time of the device. h 1 Laurits Højgaard Olesen. observed ﬂuid motion in the reverse direction of that usually expected for ac electro-osmosis. 85.1 and Armand Ajdari2 I.PHYSICAL REVIEW E 73.1103/PhysRevE. Indeed. Ajdari predicted that the same mechanism would generate a net ﬂow over an asymmetric array of electrodes 5 . 345 east. INTRODUCTION Lab-on-a-chip systems require micropumps and valves to manipulate small volumes of a liquid sample 1 . The paper is organized as follows: In Sec. 47. which is undesirable. Lastochkin et al. Denmark 2 Laboratoire de Physico-Chimie Théorique.

OLESEN, BRUUS, AND AJDARI

PHYSICAL REVIEW E 73, 056313 2006

FIG. 1. Sketch of the device geometry. The interdigitated electrode array is biased with an ac voltage that induces a buildup of a Debye screening layer and electro-osmotic ﬂuid motion. II. DEVICE GEOMETRY

We consider a two-dimensional 2D geometry like that shown in a side view in Fig. 1, similar to devices used experimentally 6–11,14 . It consists of a substrate on which an asymmetric array of interdigitated microelectrodes is deposited. Each electrode pair in the array consists of a narrow electrode of width W1 and a wide electrode of width W2 separated by a narrow and a wide gap G1 and G2, respectively, and the full period is L = W1 + G2 + W2 + G2. Notice that in order to break the left-right symmetry of the array it is necessary that both W1 W2 and G1 G2. On top of the electrode array a microﬂuidic channel of height H is placed and we assume that the device extends sufﬁciently far into the third dimension that a 2D description in the xy plane is appropriate. The channel is ﬁlled with an electrolyte, and the narrow and wide electrodes are biased with an ac voltage Vext = ± V0 cos t , respectively; i.e., the applied voltage difference is 2V0 cos t . This induces the formation of a Debye layer on the electrodes that acts to partially screen the electric ﬁeld. The nonuniform partial screening gives rise to tangential electric ﬁeld in the Debye layer and electro-osmotic ﬂow. This ﬂow takes the form of ﬂuid rolls above the electrode edges as indicated in Fig. 1.

III. ELECTROKINETIC MODEL

troosmotic ﬂow induced in the Debye layer. v Finally, we also assume that the bulk concentration of the reactants in the Faradaic electrode reaction is constant; i.e., we neglect the effect of mass transfer. Upon these approximations, the only dynamical variables we are left to consider are the potential distribution x , y , t in the bulk and the local surface charge density q x , t accumulated in the Debye layer on the electrodes. The instantaneous value of q determines the potential drop from the electrode to the electrolyte immediately outside the Debye layer, whereas the time evolution of q is determined by the balance between the Ohmic current from the bulk and Faradaic current from electrochemical reactions on the electrodes. We emphasize that this simple electrokinetic model is appropriate only in the weakly nonlinear regime of not-toohigh driving voltage 125 mV . Beyond this, the assumption of uniform bulk electrolyte concentration breaks down, the effect of mass transfer on the Faradaic reaction kinetics cannot be neglected, and the Debye layer may be driven out of local equilibrium. This is discussed in more detail in Sec. VI.

A. Bulk electrolyte

In the bulk electrolyte we assume charge neutrality such that the charge continuity equation reduces to · J = 0. 1

The electric current J is described simply as the Ohmic current J = − where is the conductivity of the electrolyte. Since we assume uniform salt concentration and conductivity throughout the bulk, we then simply end up with a Laplace problem for the potential r , t .

B. Debye layer

The Debye layer is assumed to be in local equilibrium with the electrolyte immediately outside the layer. This implies that the driving frequency should be well below the Debye frequency D = / of the electrolyte and that the Faradaic current injection cannot be too strong. The total charge accumulated in the Debye layer can then be directly related to the potential drop x , t across it 18 q x,t =

Debye layer

x,y,t dy

2

Our model for the electrohydrodynamics of the system is similar to that of Ramos et al. 15 and is based on the following classical approximations 17–20 : i The bulk electrolyte is assumed to be charge neutral with uniform salt concentration, such that the ionic transport can be described as Ohmic current. ii The Debye layer is assumed to be in local equilibrium with the electrolyte immediately outside the layer, such that the charge distribution and potential variation in the Debye layer can be described by GouyChapman theory. iii Moreover, we assume that the thickness of the Debye layer is much smaller than the size of the electrodes, and we neglect surface diffusion and migration of charge. iv The bulk ﬂuid motion is described by Stokes ﬂow with a slip condition on the electrodes set by the elec-

=− sgn

2 k BT

n

c* e−zne n

/kBT

−1 ,

3

where is the charge density, the permittivity of the solvent, kB Boltzmann’s constant, T the temperature, c* the bulk n concentration number density of the nth ionic species, and kBT / ze, Eq. 3 zn its valence. In the Debye-Hückel limit can be linearized to q=−

D

,

4

where D = kBT / nc*z2e2 is the Debye length and / D is n n the Debye layer capacitance per unit area at low voltage.

056313-2

ac ELECTROKINETIC MICROPUMPS: THE EFFECT…

PHYSICAL REVIEW E 73, 056313 2006

For larger potentials we focus on the simple case of a monovalent symmetric 1:1 electrolyte where Eq. 3 reduces to q=− 2kBT e sinh . e 2kBT D 5

O+e

R,

11

where zO = zR + 1 to ensure charge conservation. Activatedcomplex theory predicts an electric current density jext from this process given by 18

s jext = e kR cRe 1−

e

s/kBT

s − k O c Oe −

e

s/kBT

,

12

As in the classical Gouy-Chapman-Stern model we assume that the compact Stern layer on the electrode simply gives rise to an additional surface capacitance Cs in series with the Debye layer 18 . Alternatively Cs could model an oxide layer grown intentionally on the electrodes to inhibit electrochemical reactions 4,12,20 . In either case we shall for simplicity assume that Cs is independent of potential. The total potential drop across the double layer Debye and Stern is Vext − = − 1 q, Cs 6

where kO and kR are the forward and backward standard rate s s constants for the reaction, cO and cR are the concentrations of the oxidized and reduced species directly at the electrode Stern layer surface, is the transfer coefﬁcient, and s = −q / Cs is the potential drop across the Stern layer. Assuming local equilibrium, the concentrations directly at the surface are related to those immediately outside the Debye layer by a Boltzmann factor, cs = cne−zne /kBT. Then Eq. 12 can be n reexpressed as the current-overpotential equation

where Vext is the electrode potential and is the potential in the electrolyte immediately outside the Debye layer. The overall capacitance per unit area of the double layer is deﬁned as C =− q Vext −

−1

jext = j0e− zO−

e /kBT

cR 1− *e cR

e /kBT

−

cO − *e cO

e /kBT

, 13

.

7

**In the Debye-Hückel limit this reduces to C0 =
**

D

+

1 Cs

=

1 1+

,

D

8

where we introduced the parameter = / DCs for the surface capacitance ratio. At larger potential the Debye layer capacitance becomes very large and C is dominated by the always satisﬁes C0 Stern layer only. Notice that C C s. C The charging of the Debye layer is due to both Ohmic current from the bulk and Faradaic current from electrochemical reactions at the electrode:

tq

where j0 = ek is the exchange current, k is the standard rate constant, cO and cR are the concentrations im* * mediately outside the Debye layer, and cO and cR are the corresponding bulk values. = Vext − − Vext − eq is the difference between the actual potential drop across the double layer and the thermal equilibrium for the given redox process and bulk concentration; is also termed the overpotential. Appreciating that cs in Eq. 12 differs from cn by a n Boltzmann factor is known as the Frumkin correction to the standard Butler-Volmer equation. Linearization for small and for cO and cR close to their bulk values yields jext = j0e− zO−

e eq/kBT

* cR

* cO 1−

cR cO e , * − * + c R c O k BT

14

= − n · J + jext ,

9

where n is a unit normal pointing out of the electrode, −n · J is the Ohmic current running into the Debye layer, and jext is the Faradaic current from the electrode reaction. We neglect any surface diffusion and migration of charge in the Debye layer.

C. Insulating walls

where eq is the intrinsic equilibrium potential drop across the Debye layer. For simplicity we shall take eq = 0 such that = Vext − in Eqs. 13 and 14 , set the transfer coefﬁcient equal to 1 / 2, and assume that cO and cR are virtually at their bulk values; i.e., we neglect the effect of mass transfer. Moreover, for surface potentials outside the Debye-Hückel limit we shall focus on a monovalent symmetric 1:1 electrolyte with zO = zR + 1 = 1. Then Eqs. 13 and 14 reduce to jext = j0e−e and jext = j0 e Vext − k BT = Vext − Rct , 16

/2kBT

2 sinh

On the channel substrate between the electrodes and on the channel lid we assume no normal current: n · J = 0. 10

e Vext − 2kBT

15

This is justiﬁed because the amount of charge required to screen any ﬁeld normal to the channel wall is both very small and builds up on the Debye time scale D = / .

D. Electrochemistry

**respectively, deﬁning the area speciﬁc charge-transfer resistance Rct = kBT / j0e of units m2 .
**

E. Fluid dynamics

For the Faradaic electrode reaction we consider a simple one-step, one-electron redox process of the form

Since we are considering a microsystem where the Reynolds number is usually very low—i.e., where viscosity

056313-3

OLESEN, BRUUS, AND AJDARI

PHYSICAL REVIEW E 73, 056313 2006

**dominates over inertia—the ﬂuid motion in the electrolyte is described by the Stokes equation
**

m tu

C0

=− p+

2

u,

17

R0 Rct

FIG. 2. Equivalent circuit diagram for the electrokinetic system: R0 is the bulk electrolyte Ohmic resistance, C0 is the double-layer capacitance, and Rct is the charge-transfer resistance for the electrode reaction. The characteristic time for charging the Debye layer through the electrolyte is 0 = R0C0, and the time scale for decharging the layer through Faradaic electrode reaction is ct = RctC0. IV. LINEAR ANALYSIS

together with the incompressibility constraint · u = 0. 18

Here m is the ﬂuid mass density, u the velocity, p the pressure, and the dynamic viscosity. On the insulating walls the no-slip boundary condition applies, whereas on the electrodes we impose a tangential slip condition based on the Helmholtz-Smoluchowski slip velocity: us = − Et , 19

where Et = − x is the tangential ﬁeld. This classical result for the slip at a ﬂat surface with an externally applied tangential ﬁeld also holds in general for thin Debye layers in quasiequilibrium at a metal-electrolyte interface 21 . The ﬂuid ﬂow pattern is complex with rolls above the electrode edges as sketched in Fig. 1. However, we are mostly concerned with the net pumping Q = H ux dy. A 0 Fourier analysis shows that on average in time and space the ﬂow is a simple Couette ﬂow driven by the average slip velocity U on the bottom wall 15 U= 1 L

x2 x4

us dx +

x1 x3

us dx ,

20

in terms of which Q = HU / 2—that is, assuming zero backpressure on the device. Here xi denote the positions of the electrode edges cf. Fig. 1 and the insulating walls do not contribute due to the no-slip condition. Notice that us and hence U are determined entirely by the solution to the electrical problem. Therefore we can study pumping without resolving any details of the actual ﬂow pattern above the electrodes—i.e., without solving Eq. 17 —which is very convenient.

F. Characteristic dimensions

We ﬁrst study the problem in the Debye-Hückel limit kBT / ze. Here the electrokinetic problem can be linearized which simpliﬁes the analysis signiﬁcantly. Indeed the entire problem is characterized by only two dimensionless groups measuring the ratios of the two relaxation times 0 and ct and the period of the driving voltage. Of course, the solution also depends on the geometrical parameters W1, W2, G1, G2, and H, of which four can be varied independently once a characteristic length scale is ﬁxed. We do not vary all these parameters simultaneously, though. In the simplest case with negligible conﬁnement H L and no Faradaic electrode reaction we recover the earlier results of Ramos et al. We also determine the optimal parameters for the geometry of the electrode array. Then with the relative size of the electrodes ﬁxed we consider how vertical conﬁnement of the ﬂuidic channel to H L affects the system and we investigate the behavior when electrode reactions occur. Finally, we discuss how part of the linear analysis can be extended to cover systems with arbitrary intrinsic zeta potential eq on the electrodes but low driving voltage V0 1 + kBT / ze.

A. Dimensionless form for linear analysis

Before going into a detailed analysis it is convenient to discuss the overall properties of the electrokinetic system. Very roughly speaking the system is equivalent to the simple RC circuit shown in Fig. 2. The characteristic length scale 0 of the device is the narrow electrode gap G1, which determines the magnitude of the electric ﬁeld strength. The characteristic time 0 is the RC time for charging the Debye layer through the bulk electrolyte 0 = R0C0, where R0 is the area speciﬁc bulk electrolyte resistance. However, at high voltage the Stern layer dominates the capacitance of the double = R 0C s = 0 1 layer and the relaxation time changes to + −1 . The characteristic time for de charging the Debye layer through the Faradaic reaction is ct = RctC0. When the electrode reaction is very facile and Rct R0 this can be signiﬁcantly faster than the Ohmic charging, acting effectively as a “short-circuit” on the Debye layer. Finally, the characV0 / 1 + and Et teristic ﬂuid velocity obtained with V0 / 0 is u0 = V2 / 0 1 + . 0

In preparation of our numerical analysis of the problem we rescale the variables based on the characteristic dimensions deﬁned in Table I, denoting dimensionless variables by a tilde: r= r 0˜ , t = 0˜, t q = C0V0˜ , q =

−1 0 ˜

, 21

= V0 ˜ ,

u = u0˜ . u

In terms of these, the linearized relations 4 , 6 , 9 , and 16 for the charging of the Debye layer on the electrodes reduce to

˜˜ tq

= n · ˜ ˜ + K ˜ ext − ˜ , V ˜ − ˜ = −˜, q Vext

22 23

where K = R0 / Rct = 0 / ct is a measure of the facility of the electrode reaction—i.e., a Faradaic conductance. The har-

056313-4

ac ELECTROKINETIC MICROPUMPS: THE EFFECT… TABLE I. Characteristic dimensions of the electrokinetic system and typical experimental values for a 10−4 M KCl working electrolyte, assuming = 0.1. Geometric length Debye length Ohmic relaxation time at large voltage Faradaic relax. time Debye relaxation time Double-layer cpt. Capacitance ratio Bulk resistance Charge-transfer rst. Thermal voltage Fluid velocity

0 D 0

**PHYSICAL REVIEW E 73, 056313 2006
**

0.15 0.1 0.05

G1 k BT / nc *z 2e 2 n n R 0C 0 R 0C s RctC0 / / D 1+ / DC s 0/ k BT / j 0e k BT / e V2 / 0 1 + 0

us ˜

5 m 30 nm 70 s 770 s —s 0.5 s 20 mF m−2 0.1 — 3.3 m m2 — m2 25 mV — ms−1

0 −0.05 −0.1 −0.15 0

ct D

C0 R0 Rct — u0

2

4

6

8

10

12

14

x ˜

FIG. 3. Time average slip velocity ˜ s across the electrodes u obtained for ˜ = 1 and K = 0 no Faradaic current in a geometry ˜ ˜ ˜ ˜ L with W1 = 1.5, W2 = 7, G2 = 5, and H ˜ .

FEMLAB. We use second-order Lagrange elements to approximate the solution and a nonuniform ﬁnite-element mesh with increased resolution at the electrode edges. In the simple case without any Faradaic current, the Debye layer is charged by Ohmic current through the bulk only. Figure 3 shows the time average slip velocity ˜ s across the u ˜ electrodes for ˜ = 1 and K = 0 in a geometry with W1 = 1.5, ˜ = 7, G = 5, and H ˜ . The slip velocity is strongest close ˜ ˜ L W2 2 to the edges of the electrodes and is always directed from the edges towards the center of the electrodes. To understand this behavior, notice that close to the electrode edges the electric ﬁeld and hence the Ohmic current are stronger, which makes the charging of the Debye layer faster and the screening of ˜ more efﬁcient. Then Eq. 28 clearly shows that ˜ is us Vext always directed from regions where the screening is good ˆ and Vext − ˆ 2 is large towards regions where the screening is ˆ less efﬁcient and Vext − ˆ 2 is smaller—i.e., from the electrode edges towards the center. The slip velocity on the electrodes gives rise to bulk ﬂuid motion with rolls above the electrode edges as sketched in Fig. 1. However, the net pumping is determined only by the ˜ average slip U which is shown in Fig. 4 as a function of ˜ . When ˜ 1 the screening is almost complete at frequency all times. Hence, there is no tangential ﬁeld and no electroosmotic ﬂow. Conversely, when ˜ 1 the driving is too fast for any signiﬁcant screening to occur; i.e., there is no charge in the Debye layer and no ﬂow. As a consequence, in Fig. 4 we ˜ see that the pumping velocity U is maximized around ˜ 2 1, whereas it falls off as ˜ at low frequency and slightly faster than ˜ −2 at high frequency. In the low-frequency limit ˜ 1 we can verify our numerical results analytically by expanding the problem in powers of ˜ 22 . It then turns out that the leading-order time average slip ˜ s scales as ˜ 2 with a corresponding pumping u ˜ shown with a dashed line in Fig. 4. It is interestvelocity U ing to note that the overall magnitude of ˜ s falls off only u linearly with ˜ at low frequency. This results from the uniform surface charge density to zeroth order in ˜ interacting

**monic time dependence is most conveniently dealt with by introducing complex variables
**

t ˜ ˜ ,t = 1 ˆ ˜ ei ˜˜ + c.c. r˜ 2 r

24

q and similarly for ˜ ext and ˜ . Then ˜ in Eq. 22 can be elimiV q nated using Eq. 23 to obtain ˆ n · ˜ ˆ = i ˜ + K ˆ − Vext 25

for the charge balance in the Debye layer. On the insulating walls, n · ˜ ˆ = 0, 26

whereas in the bulk Eq. 1 reduces to a Laplace problem − ˜ 2 ˆ = 0. 27

**The time average electro-osmotic slip velocity can be expressed as ˜ s = ˜ ext − ˜ u V
**

˜ x

˜ =− 1 2

˜ x

˜ −˜ Vext

2

=−

1 4

˜ x

ˆ Vext − ˆ 2 , 28

**where we used that ˜ ˜ ext = 0 across the electrodes. The spaxV tial average is then 15 ˜ 1 U= ˜ L =−
**

˜2 x ˜4 x

˜ s dx + u ˜

˜1 x ˜3 x 2 ˜2 x

˜ s dx u ˜ ˆ − Vext − ˆ

2 ˜4 x 2 ˜1 x

1 ˆ Vext − ˆ ˜ 4L

2 ˜3 x

ˆ + Vext − ˆ

ˆ − Vext − ˆ

.

29

B. Validation of the numerical scheme

We solve Eqs. 25 – 27 numerically with the ﬁniteelement method, using the commercial software package

056313-5

**OLESEN, BRUUS, AND AJDARI
**

10

−2

PHYSICAL REVIEW E 73, 056313 2006

0.15

10

−3

0.1

0.05

10

us ˜

−4

˜ U

0

10

−5

−0.05

−0.1

10

−6

−0.15 0 0.2 0.4 0.6 0.8 1

10

−2

10

−1

10

0

10

1

10

2

ω ˜

˜ FIG. 4. Pumping velocity U as a function of frequency ˜ . ˜ Dashed line: analytical result at low frequency where U ˜ 2.

x/Wi

FIG. 5. Time average slip velocity ˜ s for both electrodes u against the relative position on the electrode x / Wi. Symbols: results of Ramos et al. 15 for the narrow and wide electrodes, respectively. Solid lines: our numerical solution.

with the ﬁrst-order electric ﬁeld, which gives rise to oscillating ﬂow but no time average. A similar expansion in the high-frequency limit is more difﬁcult. Still, our numerical results show that the magnitude of ˜ s falls off as ˜ −2 at high u frequency for ﬁxed ˜ , but also that it remains ﬁnite in a x narrow region of width ˜ −1 from the electrode edges. The latter appears to be an artifact due to our simpliﬁed geometry model with inﬁnitely thin electrodes: For electrodes of ﬁnite thickness and radius of curvature ˜ at the edges, the magniR tude of ˜ s falls of as ˜ −2 everywhere on the electrodes for u ˜ ˜ −1; see also 23 . R One particular question that one might ask from Fig. 4 is ˜ why the maximum net pumping velocity Umax 0.003 is so low: If the chosen velocity scale u0 is appropriate, should we ˜ not expect Umax 1? Starting from Fig. 3 we note that already the time average slip ˜ s is less than O 1 because the u ˆ double-layer potential drop Vext − ˆ in Eq. 28 is only ˆ around half the total applied voltage Vext at the relaxation frequency. The spatial average over the full period of the array includes both positive and negative values of ˜ s and u passive sections of insulating channel wall, which reduces ˜ the value of U further compared to unity. Interestingly, the average slip across only the narrow electrode, 1 ˜ U1 = ˜ L

˜2 x

˜ ˜ the same graph for ˜ = 1.24 in a geometry with W1 = 1, W2 ˜ ˜ L = G2 = 10/ 3, and H ˜ . The results from the two different numerical schemes are hardly distinguishable, except close to the electrode edges where our ﬁnite-element solution displays a slightly larger slip velocity 24 . We also checked our results for convergence with mesh reﬁnement. The potential ˆ r is everywhere smooth, but the tangential ﬁeld is singular at the electrode edges due to the abrupt change in boundary condition between Eqs. 25 and 26 25 . Of course, our ﬁnite-element solution cannot represent an inﬁnite slope, but by choosing a strongly nonuniform mesh with very ﬁne resolution near the edges we obtain fairly good quality of the solution even close to the singularity. Moreover, we are mostly concerned with the net pump˜ ing U which is an integral quantity and therefore less sensitive to the details at the electrode edges 26 .

C. Optimal device geometry

˜ s dx , u ˜

˜1 x

30

˜ changes sign for ˜ 1. At low frequency, U1 and the corre˜ sponding average U2 across the wide electrode are both posi˜ ˜ tive with U1 0.1U2 ˜ 2. Above the relaxation frequency −1 ˜ ˜ ˜ ˜ ˜ , although their sum U = U1 + U2 falls we have −U1 U2 −2 off as or slightly faster than ˜ . In order to further validate our numerical scheme we compare to the results obtained by Ramos et al. using a boundary element method based on the Greens function for a unit charge line source 15 . This is shown in Fig. 5, where the time average slip ˜ s from both electrodes is shown on u

An important question in relation to actual device design is what electrode geometry would maximize the pumping velocity. It is immediately clear that since the characteristic velocity u0 is inversely proportional to the characteristic length scale 0, the pumping is increased by downscaling the overall size of the electrodes. We focus again on the limit H L, where the results are independent of H. Still, there remains four different length scales describing the electrode geometry—namely, W1, W2, G1, and G2—of which three can be varied independently once a characteristic length scale is ﬁxed. Ramos et al. also investigated how the pumping velocity depends on the electrode geometry 15 . They rescaled the problem using the width of the wide electrode W2 as their characteristic length scale, but found that the pumping velocity would increase without bounds as G1 / W2 → 0. Therefore they ﬁxed the ratio G1 / W2 at 0.1751, as in the experiment by Brown et al. 6 , and found that the pumping velocity was then maximized when W1 / W2 = 0.24 and G2 / W2 = 0.8.

056313-6

ac ELECTROKINETIC MICROPUMPS: THE EFFECT…

PHYSICAL REVIEW E 73, 056313 2006

TABLE II. List of geometries used experimentally and our linear model prediction for their dimensionless maximum pumping velocity and optimal driving frequency. Of course, the actual pumping velocity depends on the absolute size of the device through u0. ˜ W1 Brown et al. Studer et al. Mpholo et al. Debesset et al. Ramos et al. Global optimum 0.93 0.67 1.0 0.8 1.37 1.51 ˜ W2 5.7 4.33 5.0 5.2 5.71 6.55 ˜ G2 3.5 2.67 3.0 3.2 4.57 4.74 ˜ Umax 0.00331 0.00293 0.00332 0.00317 0.00357 0.00359 ˜ max 1.12 1.47 1.12 1.26 0.88 0.80 Refs. 6,7,11 8 9 10 15 theory Present work

However, the choice G1 / W2 = 0.1751 is arbitrary. It is more appropriate to identify the narrow electrode gap G1 as the characteristic length scale and look for an optimal device ˜ ˜ ˜ geometry as a function of W1, W2, and G2 in the parameter ˜ ˜ ˜ and 1 G2 . Notice that when ranges 0 W1 W2 ˜ ˜ ˜ W2 and G2 attain their lower limits, the net pumping U vanishes due to left-right symmetry; i.e., the optimum cannot be on the parameter range boundaries. Therefore we can use a simple unconstrained optimization algorithm such as the MATLAB function FMINSEARCH to determine the global optimum: ˜ Umax = max

˜ ˜ ˜ ˜ ,W1,W2,G2

˜ ˜ ˜ ˜ U ˜ ,K = 0,W1,W2,G2 .

31

In this way we ﬁnd that the pumping velocity is maximized ˜ ˜ ˜ ˜ for W1 = 1.51, W2 = 6.55, and G2 = 4.74, where Umax = 0.003 59 at ˜ max = 0.80. Given that we have tracked down the global optimum, it is natural to ask how much better this geometry performs as compared to that suggested by Ramos et al. and those used experimentally. This is summarized in Table II: Almost all experimental geometries come within 10% of the maximal pumping velocity, and the theoretical result of Ramos et al. is less than 1% off. Yet the frequency ˜ max at which pumping is maximized for the various geometries differs by almost a factor of 2. The overall conclusion to be drawn is that the pumping velocity is fairly insensitive to the particular choice of electrode geometry. Henceforth, we shall therefore continue to ˜ ˜ use a geometry with rounded numbers W1 = 1.5, W2 = 7, and ˜ = 5 which is still close to optimal; see also Fig. 6. G2

D. Conﬁned geometry

pumping velocity increases by roughly a factor of 2 and the ˜ optimal driving frequency ˜ max becomes proportional to H for L / H 1. The latter is easily understood from our simple circuit model of the system in Fig. 2: While the double-layer capacitance C0 is largely independent of conﬁnement, the ˜ resistance of the bulk electrolyte changes from R0 to R0 / H because the cross section of the conducting channel is decreased. The relaxation frequency therefore becomes ˜ ˜ H / R0C0—i.e., proportional to H. Our ﬁndings here are important in relation to actual device design because they indicate that working with a strongly conﬁned geometry does not, as one might have worried, mask the asymmetry and destroy the ability to pump. Thus the channel height H and the period of the electrode array L can be chosen independently. It is important to realize that while the pumping velocity U does not depend much on conﬁnement, the maximal ﬂow rate Qmax = HU / 2 that the pump can deliver per unit width of the channel and the maximal backpressure pmax = 6 NLU / H2 it can sustain where N is the total number of electrode pairs certainly do.

20 18 16 14 12

˜ W2

10 8 6 4 2 0 0 0.5 1

0.0 03 5

0.0

03

0.0

025 .002 0

In several experimental studies the height of the pumping channel H has been comparable to the period of the electrode array L 8,10,11 . It is not immediately clear how this would affect the properties of the device compared to the simple case of no conﬁnement H L. ˜ Figure 7 shows the pumping velocity U as a function of frequency ˜ and degree of conﬁnement L / H. Our previous results correspond to the base line L / H 1, whereas the conﬁnement becomes signiﬁcant for L / H 2. In particular, the

etry Symm

1.5 2 2.5 3 3.5 4 4.5 5

˜ W1

FIG. 6. Contour plot of the maximal pumping velocity over ˜ ˜ ˜ ˜ frequency max˜ U ˜ , . . . as a function of W1 and W2 for G2 = 5. ˜ ˜ When W1 = W2 the net pumping vanishes due to symmetry. There is ˜ ˜ ˜ a unique optimum close to W1 = 1.5 and W2 = 7 with Umax = 0.00358 and ˜ max = 0.8.

056313-7

**OLESEN, BRUUS, AND AJDARI
**

10

3

PHYSICAL REVIEW E 73, 056313 2006

−0.004

−0.002

10

2

10

2

a

b

0.0

L/H

08 0.0

10

1

10

1

07 0.0 06 0.0 05 0.0 0

0.00 3

K2

10

0

c

d

0.012 0.010 0.008 0.006 0.004

0.002

4

10

−1

0.002

10

0

10

−2

0.

0

0.001

10

−2

10

−1

10

0

10

1

10

−3

ω ˜

˜ FIG. 7. Contour plot of pumping velocity U as a function of ˜ and degree of conﬁnement L / H. frequency E. Faradaic current injection

10

−3

10

−2

10

−1

10

0

10

1

10

2

10

3

K1

˜ FIG. 9. Contour plot of U in the zero-frequency limit ˜ 0 with asymmetric current injection: K1 and K2 denote the Faradaic conductance on the narrow and wide electrodes, respectively. The slip velocity ˜ s at the points a – d is shown in Fig. 10. u

When we include electrode reactions and Faradaic current in our model, one of the most important new features is that the complete screening of the electrodes at low frequency is unbalanced. When the reaction is slow and the Faradaic conductance K = R0 / Rct ˜ , the effect is small and only perturb the result from purely Ohmic charging slightly. Conversely, when K ˜ the reaction is fast enough to establish a steady state where the Ohmic and Faradaic currents are balancing. For either K 1 and/or ˜ 1 the Debye layer is effectively “short-circuited;” cf. Fig. 2. On the basis of these observations we would expect to see pumping even in the lowfrequency limit when the Faradaic reaction speed is moderate and K 1. ˜ Figure 8 shows the pumping velocity U as a function of ˜ and K. For K ˜ we recover our previous results for pure

10

2

10

1

10

0

10

−1

10

−2

10

−3

10

−2

10

−1

10

0

10

1

10

2

ω ˜

˜ FIG. 8. Contour plot of the pumping velocity U as a function of frequency ˜ and Faradaic conductance K.

Ohmic charging. However, for K ˜ we observe no pumping at all, not even for K 1. Figure 9 shows the result of a more detailed investigation in the zero-frequency limit ˜ 0, where we consider a device with asymmetric surface properties on the electrodes, such that the Faradaic conductance takes the values K1 and K2 on the narrow and wide electrodes, respectively. The ﬁgure shows that whenever K1 K2 the device does actually pump: When K1 K2 it pumps in the forward direction—i.e., in the same direction as observed for purely Ohmic charging—but for K2 K1 the pumping direction is reversed. This reversal of the pumping direction can be understood qualitatively by inspecting the form of the time average slip ˜ s for the four different points marked a to d in Fig. 9, u the result of which is shown in Fig. 10. When K1 = K2 = 1 c the slip velocity is clearly nonzero on both electrodes, but the average across the narrow electrode exactly cancels that across the wide electrode. When K1 K2 = 1 d the Debye layer on the narrow electrode is essentially short-circuited by the fast electrode reaction, leaving no charge and no electroosmotic ﬂow there except from in a narrow region of width K−1 from the electrode edges. Therefore the pump1 ˜ ing velocity U is dominated by the contribution from the wide electrode. Similarly, when K2 K1 = 1 a the Debye layer on the wide electrode is essentially short-circuited and the pumping velocity is dominated by the contribution from the narrow electrode—which tends to be negative. In conclusion, the absence of pumping in the lowfrequency limit when K1 = K2 is a special case due to high symmetry; whenever K1 K2 there is some pumping and this falls off linearly with Ki for Ki 1. More generally we can prove that for a device with spatially varying capacitance ratio ˜ and Faradaic conductance K ˜ , there can be no x x pumping in the low-frequency limit when the product 1 x x + ˜ K ˜ is a constant on all electrodes. The details of the proof are outlined in Appendix A.

K

0.001

0.002

0.003

056313-8

**ac ELECTROKINETIC MICROPUMPS: THE EFFECT…
**

0.15 0.1 0.05

PHYSICAL REVIEW E 73, 056313 2006

(a)

˜ U <0

(b)

˜ U =0

1 djext = Rct d Vext −

=

j 0e − z − e O k BT

e eq/kBT ,

36

us ˜

0

−0.05 −0.1 −0.15

and the Faradaic conductance is redeﬁned as K = R0 / Rct eq accordingly. The interaction of the linear response tangential ﬁeld with the equilibrium Debye layer charge gives rise to a linear response slip velocity

(c)

˜ U =0

0.15 0.1 0.05

(d)

˜ U >0

eq

us =

x

.

37

us ˜

0

−0.05 −0.1 −0.15 2

4

6

8

10

12

2

4

6

8

10

12

x ˜

x ˜

However, because is harmonic in time, this does not give rise to any net pumping. The leading-order time average slip is to second order in V0, where the interaction of the induced charge to ﬁrst order with the tangential ﬁeld gives rise to a slip velocity of us =

x

FIG. 10. Time average slip velocity ˜ s on the electrodes for the u parameter values marked in Fig. 9: a K1 = 1, K2 = 102; b K1 = K2 = 102; c K1 = K2 = 1; d K1 = 102, K2 = 1. F. Linear response at nonzero intrinsic zeta potential

=−

d 4 d Vext −

x

ˆ Vext − ˆ

2

. 38

Many electrode-electrolyte systems spontaneously form a Debye layer and possess an intrinsic zeta potential typically of the order eq 100 mV which is well beyond the DebyekBT / ze 25 mV. Hence, our foregoing linHückel limit ear analysis cannot be applied immediately to such systems. But then it is standard procedure in electrochemical impedance spectroscopy to study the linear response of the electrokinetic system to a small applied voltage 18 . Consider a system where the electrolyte forms a Debye layer on the electrodes with an intrinsic zeta potential eq. In equilibrium there is a uniform surface charge density qeq given by Eq. 3 on the electrodes and a constant potential eq = − eq + qeq / Cs in the bulk. If the electrode array is biased with a small-amplitude ac voltage Vext = ± V0 cos t , we can 1 linearize around equilibrium and writing = 2 eq + ˆ ei t + c.c. and similarly for q and we recover the form of Eqs. 26 and 27 for the linear response ˆ . Only now the characteristic time used for the rescaling in Eq. 21 becomes

0

This is fully equivalent to Eq. 28 , and the net pumping produced can be obtained directly from the analysis presented earlier in the text. Only now the characteristic ﬂuid velocity becomes u0 = V2 0 d d Vext − 0 , 39

where, for a symmetric z : z electrolyte, d d Vext − = 1 1 + cosh ze . 40

eq

/2kBT

= R 0C d ,

32

Notice that u0 eq falls off exponentially with eq because becomes negligible. The linear response dominates over the higher-order components only at low applied voltage. Speciﬁcally, we must and V0 require that both V0 Cd / dCd / d Vext − , and a sufﬁcient condition for this is Rct / dRct / d Vext − V0 1 + kBT / ze. However, because the time average slip velocity vanishes to ﬁrst order in V0, one cannot neglect the second-order tangential ﬁeld − x , which gives rise to an additional component ¯ s to the slip velocity given by u

eq

where Cd layer:

eq

is the differential capacitance of the double ¯s = u dq Cd = − d Vext − . 33

x

.

41

**For a symmetric z : z electrolyte this can be written as Cd =
**

D

+ sech ze

eq

/2kBT

.

34

In the absence of Faradaic currents, the time average potential vanishes to second and all higher orders within our simple electrokinetic model, as is easily seen by taking the is determined as a time average of Eq. 9 . Otherwise, solution to the Laplace equation with the boundary condition 0= n· − 1 Rct + 1 d2 jext 4 d Vext − ˆ Vext − ˆ

2

**The relaxation time for the Faradaic electrode reaction becomes
**

ct =

2

RctCd ,

35

where Rct given by

eq

42 at the electrodes, where

is the differential charge-transfer resistance

056313-9

BRUUS. and we therefore focus on the case of a monovalent symmetric 1:1 electrolyte with no intrinsic zeta potential formed on the electrodes—i. The transition between the two rewhen 1 and for ˜ 0 −2 log V gimes occurs for ˜ 0 V when 1. corresponding to V = 25 V. Moreover. j 1+ k BT e 2 ˜V U˜ 2 . the data are scaled with ˜ 2 to recover the voltage dependence of the physical V0 pumping velocity: U= 0 = n · ˜ ˜ + ˜ext . reduce to ˜ − ˜ =˜ −˜ / 1 + q Vext ˜˜ tq .. where we represent the periodic solution by a set of equispaced points over one period in time. Then Eqs. and calculate the time derivative in Eq. at higher voltage when the doublelayer capacitance is dominated by the Stern layer. 0 52 where ˜ext is the dimensionless Faradaic current given by j ˜ext = 2Ke−˜ /2 sinh ˜ ext − ˜ /2 . Figure 11 also shows that the dimensionless pumping ˜ velocity U falls off at high voltage. 9 . Eqs. the actual computation of the time average ﬂuid ﬂow and net pumping for ﬁnite intrinsic zeta potential on the electrodes and/or the insulating channel walls requires a separate and more elaborate study. even with the simple nonlinear model we consider here. However. 45 46 At low voltage this reduces to ˜ V0 / 1 + whereas at high V voltage ˜ 2 log ˜ 0 / . This is conﬁrmed by Fig. cf. 48 We must emphasize at this point that the upper voltage limit ˜ = 103 in Figs. 12. and 15 . see Appendix B for details. Due to the nonlinearity. 51 Apart from this. 46 using the leapfrog ﬁnite-difference scheme. Beyond this voltage our results can therefore only provide some characteristic features of the system and should not be regarded as an accurate description. At the base line for ˜ V0 1 voltage V0 ˜ we recover the results from the linear analysis with U peaking for ˜ 1. 45 – 50 . is solved numerically with the ﬁnite-element method using FEMLAB. Instead we employ a relaxation method.1 and K = 0. and where the zeta potential is directly related to the Debye layer charge by ˜ ˜ = − 2 sinh−1 ˜ /2 1 + q q On the insulating walls we have again . which we have to leave open despite its experimental relevance. When the driving frequency is not too large. We therefore introduce a new rescaling for the electrical part of the problem: k BT ˜ = . 056313-10 . VI we discuss a number of strongly nonlinear phenomena that become signiﬁcant for V0 125 mV in a typical experiment. A. the smallest characteristic voltage of the electrokinetic system is the thermal voltage kBT / e 25 mV and not the driving voltage V0. it is not possible to solve the problem with a single complex variable for the base frequency component. for ˜ 0 V − 2 log we expect us to scale as V0 log V0 rather than V2. Hence. is far V0 0 outside the range of validity of our simple electrokinetic model: In Sec. 49 This demonstrates that when Faradaic electrode reactions occur. where ˜ = 0 / = / 1 + is the relaxation frequency at high voltage. However. 2 43 n·˜˜ =0 and. Dimensionless form for nonlinear analysis The problem. 056313 2006 d2 jext d Vext − 2 =− e 2 kBT Rct zO − d d Vext − + − 1 . in the bulk. V. 6 . B. most experimental work has been done with driving voltages of a few volts in order to obtain an appreciable ﬂuid motion. we retain u0 = V2 / 0 1 + for the characteristic 0 ﬂuid velocity. for the Faradaic electrode reaction we take zO = zR + 1 = 1 and assume a symmetric transfer coefﬁcient = 1 / 2. 5 . the fre˜ quency at which U is maximized drops to ˜ ˜ . − ˜ 2 ˜ = 0. e k BT ˜ V 0. Eqs. 21 . This is simply due to the fact that the physical electro-osmotic slip velocity us 2 x no longer scales as V0. Figure 11 shows the pumping ˜ velocity U as a function of the driving frequency ˜ and ˜ for = 0. governing the charging of the Debye layer on the electrodes. We ﬁrst investigate the system behavior without the presence of the electrode reaction so that the only nonlinearity in our model arises from the nonlinear capacitance of the Debye layer. AND AJDARI PHYSICAL REVIEW E 73. the scaling remains as in Eq. the parameter space is large. Then we turn on the electrode reaction and study the full nonlinear model. NONLINEAR ANALYSIS 50 While theoretically it is convenient to work in the lowvoltage limit where the system can be characterized by relatively few dimensionless parameters. Table I. q = C0 q e 44 In the absence of Faradaic electrode reactions the only nonlinearity in the model arises from the nonlinear surface capacitance in the Debye layer. eq = 0.OLESEN. V0 = e k BT ˜. 0 ˜ V which shows Umax as a function of ˜ 0. Nonlinear Debye layer capacitance Beyond the Debye-Hückel limit. 45 and 48 allow us to estimate the zeta potential from ˜ V0 ˜ + 2 sinh ˜ /2 .e. in particular. 11 and 12. V j 47 with K = R0 / Rct as before.

1. Eq. V0 V0 V0 ˜ with a peak value of U less than 0.ac ELECTROKINETIC MICROPUMPS: THE EFFECT… 10 3 PHYSICAL REVIEW E 73.0 03 0.1 and Faradaic conductance K = 0.001 0. Effectively. Nonlinear Faradaic current injection ω ˜ ˜ FIG.. C. yielding an overall V0 voltage we ﬁnd that the pumping levels off to become almost independent of ˜ 0. such that the exponential increase in the Faradaic current cf. reﬂecting the larger relaxation time when the double-layer capacitance is dominated by the Stern layer.001 10 −1 10 −2 10 −1 10 0 10 1 10 −1 10 −2 10 −1 10 0 10 1 ω ˜ ˜ FIG. 12. Notice the ˜ pumping in the low-frequency limit for V0 1. While the nonlinearity induced by the Debye layer capacitance manifests itself relatively slowly as the voltage is increased. This behavior can be partially understood V by considering the scaling of ˜ and the tangential ﬁeld in this limit: When the frequency is low enough that a steady state is established with the Ohmic and Faradaic currents balancing. but for ˜ 0 10 it V bends sharply up towards higher frequency.003. Contour plot of the pumping velocity U as a function ˜ and voltage ˜ 0 with capacitance ratio V of driving frequency = 0. 11. this makes the Faradaic conductance larger on the narrow electrode than on the wide one. Correspondingly. Qualitatively it is not so difﬁcult to see how the pumping at low frequency arises: At low voltage the magnitude of the charge density in the Debye layer on the narrow electrode is a factor of W2 / W1 larger than on the wide electrode to ensure overall charge conservation. Apart from the intrinsic net pumping grows rapidly with V0 ˜ 2 scaling.002 10 0 10 0. The dashed line shows the V optimum driving frequency ˜ max as a function of ˜ 0.003 0 0. However. At larger as well. the time average asymmetry in effective Faradaic V0 conductance between the two electrodes grows roughly as ˜ 2 V0 ˜ 4 voltage dependence. For V0 comparison. 11. At the base line for ˜ 0 1 we recover the results from V the linear analysis in Fig. we can write the applied voltage ˜ 0 as the sum of the V potential drop across the double layer ˜ ext − ˜ and the potenV ˜ ˜ Umax V02 056313-11 . For ˜ 0 1 the ˜ . Figure 14 shows in more detail the voltage dependence of V the pumping velocity at low frequency ˜ 1. we expect a more dramatic effect from the Faradaic current due to the exponential voltage dependence in Eq. for ˜ 0 V ˜ 1 we do observe pumping at low frequency. as V evaluated along the white dashed line in Fig. ˜ 47 . whereas U falls off rapidly for all ˜ when ˜ 0 10.e. Contour plot of the pumping velocity U as a function ˜ and driving voltage ˜ 0 with capacitance ratio V of frequency = 0. The white dashed line shows ˜ V the frequency ˜ max at which U is maximized as a function of ˜ 0. 9 from our linear analysis this should indeed give rise to pumping in the forward direction at low frequency. 11. 056313 2006 10 3 10 2 10 2 ˜ V0 10 1 ˜ V0 10 1 0. 10 2 ˜ V0 10 0 10 −2 ˜2 V0 ˜ ˜ V0 logV0 10 −4 10 −1 10 0 10 1 10 2 10 3 ˜ V0 ˜ FIG. while the dash-dotted line is a ﬁt to ˜ log ˜ . also the double-layer potential drop ˜ ext − ˜ is larger on the narV row electrode. 8 for K = 0. The data are scaled with ˜ 2 to recover the physical dependence on driving voltage.002 0. the dashed lines show linear and quadratic scaling with ˜ .1. K = 0. 13. and according to Fig.1 and K of driving frequency ˜ and voltage V0 = 0. no pumping in the low-frequency limit and maximal pumping for ˜ 1.1—i. The ﬁgure also V shows that at low voltage the optimal driving frequency ˜ max follows the same pattern as in Fig. Maximal pumping velocity Umax as a function of ˜ 0. 47 starts off earlier there. Figure 13 shows the pumping velocity U as a function ˜ for = 0. At large driving voltage ˜ max shifts from 1 to ˜ .1 and no Faradaic current.

the scaling of the pumping velocity with driving voltage changes from V2 0 to V0 log V0 because the zeta potential entering the Helmholtz-Smoluchowski slip velocity becomes only a small fraction of the overall potential drop across the double layer. the maximal ﬂow rate Qmax = HU / 2 that the pump can deliver and the maximal backpressure pmax U / H2 it can sustain certainly do. Therefore.5. and hence also the optimal frequency for pumping. Studer et al. the characteristic behavior from Fig. The electrode material in those experimental studies was uncoated gold and platinum. Brown et al. where = / DCs is = 1 / R0C0 to the surface capacitance ratio.6 mM KCl solution: Naively.15 mM and a 0. we ﬁnd that as the driving voltage V0 is increased well beyond the thermal voltage kBT / e. and we would ex− / 1 + go as log V0 pect the physical electro-osmotic slip velocity us x to grow as slowly as log V0 2 with driving voltage. It is remarkable that this value was obtained for both a 0. changing to ˜ ext − ˜ V ˜ / K at higher voltage. reported that the optimal driving frequency f max dropped from 3 kHz at 0. Nonlinear double-layer capacitance tial drop across the bulk electrolyte when an Ohmic current equal to ˜ext is passed through it—i. we comment on the extent to which our results agree or disagree with the experiments and discuss a number of nonlinear effects not included in our model that could account for some of the more exotic experimental observations. 14 with the pumping velocity leveling off dominates even for ˜ 1 when ˜ 0 102. mapped out the entire ﬂuid velocity ﬁeld over a single pair of gold electrodes with titanium coating 4 . drops from 0 = 1 / R0Cs = 0 / 1 + . The observed downshift in optimal frequency therefore ﬁts well with our model predictions if we assume a 0. The observed velocity magnitude and frequency dependence at 500 mV driving voltage was matched with linear theory predictions by assuming a capacitance ratio of = 3. consistent with a large Stern layer capacitance and hence small . For their particular electrolyte and electrode geometry the relaxation frequency at low voltage was expected to be f 0 = 0 / 2 1+ 2 kHz. Finally. whereas our results indicate that conﬁnement becomes signiﬁcant only for L / H 2. that while the pumping velocity U does not depend much on conﬁnement. where the sole nonlinearity in our model is due to the nonlinear capacitance of the Debye layer. The optimal choice of channel height therefore depends strongly on whether the pump is intended to work as a pressure or current generator 27 . In effect. found the capacitance ratio of optimal driving frequency around 1 kHz for driving voltages larger than 1 Vrms with similar electrolyte and electrode geometry 11 . Pumping velocity U as a function of ˜ 0 in the lowV V0 frequency limit ˜ 1. DISCUSSION Our numerical study of the simple electrokinetic model has brought about a number of interesting theoretical predictions that we brieﬂy summarize below and compare to experimental observations reported in the literature. 056313 2006 A. The data are scaled with ˜ 2 to recover the physical dependence on driving voltage. as one might have worried.e. 14. though. AND AJDARI PHYSICAL REVIEW E 73. The titanium spontaneously forms a thin oxide layer which was intended to inhibit Faradaic electrode reactions in the experiment. And when the optimal driving frequency in V Fig. we ﬁnd immediately that ˜ ext − ˜ ˜ 0 / 1 + K at V V low voltage ˜ 0 K − 2 log K. the “low-frequency” limit where the Faradaic and Ohmic currents balance extends higher and higher in frequency as the driving voltage is increased. At the same time. both the tangential 2 log V0 ﬁeld ˜ ˜ = − ˜ ˜ ext − ˜ and the zeta potential ˜ ˜ ext V x x V ˜ at large voltage. Green et al. 13 shifts towards higher frequency for ˜ 0 10. j ˜ V0 ˜ − ˜ +˜ Vext jext ˜ − ˜ + 2Ke−˜ /2 Vext 53 sinh ˜ ext − ˜ /2 . Experimentally.OLESEN.2 Vrms to 1 kHz at 1. it is V essentially a trace of the point at which the Faradaic current starts to dominate. In particular. V If we neglect for simplicity the ˜ -dependent exponential V prefactor. Cahill et al. we note that since the electrode reaction is very fast at large voltage. Likewise. Our analysis has shown that vertical conﬁnement of the pumping channel does not mask the asymmetry of the device and destroy the ability to pump. BRUUS. VI.2 Vrms 6 . respectively. but instead enhances the pumping velocity by roughly a factor of 2 and causes the optimal driving frequency to shift down inversely proportional to L / H. B. the conﬁnement has been relatively weak with L / H 2. assuming a constant value for the Stern layer capacitance Cs one would expect = / DCs to depend on the electrolyte concentration through D 28 . observed ac electroosmotic ﬂuid motion due to traveling-wave electric ﬁelds induced on a four-phase in- 056313-12 . In all experimental studies to date. the relaxation frequency for the electrokinetic system. Conﬁnement 10 0 ˜4 V0 ˜ ˜ U V02 10 −2 10 −4 10 −6 10 −1 10 −8 10 0 10 1 10 2 10 3 ˜ V0 ˜ FIG.. In the absence of Faradaic electrode reactions. It is important to keep in mind.

And Lastochkin et al. and the characteristic distance that the nth species can diffuse over one cycle of the driving voltage is n = Dn / .14 . it is difﬁcult to see how the potential drop across and hence the charge in the double layer could change sign. as well as a similar reversal on a traveling-wave array for driving voltages above 2 V and frequencies around 1 kHz 13. However. IV F indicates that when Faradaic electrode reactions are present. We note that the reversal of the pumping direction that we predict at low frequency is essentially due to an imbalance between net pumping contributions of opposite signs from the two electrodes. whereas it is in the reverse direction otherwise. Moreover. also found reversal of the pumping direction on an asymmetric electrode array. When the oxidized and reduced species O and R are consumed or produced at the electrode by the Faradaic reaction. Based on the simple circuit model for the electrokinetic system from Fig. 056313 2006 terdigitated electrode array and found an accurate V2 scaling 0 for 100− 500 mV driving voltage. In the symmetric special case when Rct is equal on both electrodes there is no net pumping at low frequency. And the reversal of the time average slip direction observed by Lastochkin et al. However. at even larger voltage our nonlinear analysis shows that the pumping velocity levels off to a constant value. and vice versa for cR. the pumping at low frequency is in the forward direction. 4 and 16 are no longer valid. we predict that the direction of the pumping in the low-frequency limit depends on the relative magnitude of Rct on the narrow and wide electrodes. Hence. C. Finally. where Rct is the charge-transfer resistance characterizing the electrode reaction and R0 is the bulk electrolyte Ohmic resistance. When Rct R0 the Faradaic current injection merely imbalances the otherwise perfect screening at low frequency and therefore induces electro-osmotic ﬂuid motion. D. the diffusion process is essentially one dimensional and the area speciﬁc Warburg impedance takes the form 056313-13 . Studer et al. observed that the pumping direction was reversed when the driving frequency was increased above 20 kHz and the voltage above 3 Vrms 11 .11 . the lower cO drops at the electrode. respectively: If Rct is smaller on the narrow electrodes than on the wide ones. the experimentally observed pumping velocities tend to be smaller than the linear theory prediction and sometimes resemble more an afﬁne dependence than a quadratic scaling with V0 10. is certainly a different phenomenon. although the pumping velocity peak position and magnitude were smaller than the linear theory predictions 12 . If this distance is small compared to the characteristic dimension of the electrodes 0. could be due to imbalance between the net pumping contributions from the two electrodes: Although our nonlinear analysis predicts pumping in the forward direction. in series with the charge-transfer resistance Rct in Fig. This is at least in qualitative agreement with our nonlinear model results because it is difﬁcult to distinguish V0 from V0 log V0 over the limited voltage range used experimentally.1 as in Fig. Ramos et al. the Debye layer cannot be assumed to be in local equilibrium and Eqs.0 Vrms. Effect of mass transfer on reaction kinetics One important effect that we have neglected in our model is the inﬂuence of mass transfer on the electrode reaction kinetics. yielding a capacitance ratio of = 75– 150 depending on the electrolyte concentration.ac ELECTROKINETIC MICROPUMPS: THE EFFECT… PHYSICAL REVIEW E 73. Their optimal driving frequency around 2 kHz was matched with the linear theory by assuming = 1. When the driving voltage is increased. we recall that for simplicity we have been focusing here on the case with no intrinsic zeta potential eq on the electrodes and/or the insulating channel walls. In their experiments the electrodes were coated with a 50-nm-thick Teﬂon-like insulating layer.5– 4. 14. The transport of O and R from the bulk to the surface is by diffusion. 2 18 . We never ﬁnd reversal of the time average slip direction within our simple electrokinetic model. Generally. which is of the order of the Debye frequency for the electrolyte. even if the intrinsic surface properties are symmetric. reported that the direction of the time average slip velocity on an array of T-shaped electrodes was reversed as compared to the direction usually expected for ac electroosmosis 16 . 2. the pumping is strongly suppressed for Rct R0 due to “short-circuiting” of the double layer. depending on the driving frequency and electrode material 8–11. Our analysis in Sec. since the driving frequency in their experiments was in the MHz range.14 . their local concentrations cO and cR become different * * from the bulk concentrations cO and cR. the magnitude of the pumping velocity at this plateau is roughly U 100 m / s. Now. This can be modeled as an additional impedance Z . Faradaic current injection Our linear analysis at low driving voltage shows that when Faradaic electrode reactions occur. respectively. called the Warburg impedance. argued that strong Faradaic current injection in their system would reverse the polarity of the charge in the Debye layer and hence the direction of the slip. where Dn is the diffusion constant. they remained in the linear regime over the entire voltage range applied. Moreover. Using the typical parameter values from Table I and R0 / Rct = 0. the behavior for nonzero eq needs to be studied more carefully. Ramos et al. regardless of how small Rct becomes. Ramos et al. Moreover.5. Lastochkin et al. especially in the nonlinear regime. while the maximum pumping velocity at low frequency was an order of magnitude below the prediction from linear theory. observed pumping reversal on a traveling-wave array where all the electrodes are identical—this rules out any imbalance between the contributions from individual electrodes in their setup. the nonlinearity in the Faradaic current injection induces an effective asymmetry in the charge-transfer resistance and we obtain pumping in the forward direction at low frequency. It is possible that the pumping reversal observed by Studer et al. The observation of electrode degradation and bubble formation from electrolysis is experimental evidence that Faradaic reactions do occur above 1. the more difﬁcult it becomes to run the reaction forward. performed a similar study using titanium-coated gold electrodes 13 . there could be other nonlinear effects leading to imbalance into the reverse direction.

We do not anticipate that this alone could be the mechanism responsible for reversal of the time average slip velocity or the direction of net pumping. E. Surface conduction in the Debye layer V0 / 1 + and using typical parameter values Taking from Table I. when the amount of salt absorbed in the Debye available in the layer approaches the total amount c* diffusion zone. For a symmetric monovalent 1:1 electrolyte the Dukhin number can be expressed as 17 Du = 1 + k BT e 2 increases at low frequency because the reNotice that Z actants need to diffuse longer. This is quantiﬁed by the Dukhin number Du= s / 0. the Faradaic current can become so strong that the concentrations of the reactants at the electrodes differ signiﬁcantly from their equilibrium values. as one of them approaches zero.1 mM. the diffusion layer should remain charge neutral and the conductivity largely unaffected.. we do expect the prediction of the pumping direction based on Fig. 056313-14 .OLESEN. at very low freD / 2—the diffusion layer quency when 0—i.e. Assuming the frequency is low enough that the Ohmic and Faradaic currents V0 / 0—we ﬁnd that the limiting are balancing—i. F. the reaction stagnates. 4kBT 56 We note that Z0 R0 = 0 / . 59 At large voltage the Debye layer can accumulate enough charge that the mean ion density. Taking again V0 / 1 + and using typical parameter values from Table I. For the typical parameter values from Table I and 0. we get local depletion of salt and our assumption of uniform electrolyte concentration throughout the bulk breaks down. if there is an excess of supporting electrolyte that does not participate in the electrode reaction. is the bulk conductivity. BRUUS. This underlines the potential role of the speciﬁcs of the electrolyte and electrode material. because it is the larger of Rct and Z0 that dominates. Local salt depletion at electrodes Another issue relating to the massive accumulation of ions in the Debye layer at large voltage is discussed by Bazant et al. we expect surface currents to smear out the charge distribution across the electrodes. we ﬁnd that surface conductance becomes signiﬁcant and Du 1 for V0 250 mV. 056313 2006 Z = k BT 2e2 * cO 1 DO + 1 * cR 1−i DR . The current at which this occurs is termed the limiting current 18 . In this case one cannot neglect lateral surface currents in Eq. Ultimately. Generally.e. 0 extends much longer than the separation between the electrodes and the Warburg impedance levels off to a constant Z0 k BT 0 0 + * . it is not possible to short-circuit the double layer for 2 D / 0.. where the electrolyte conductivity is given by = nDic*z2e2 / kBT. the effect of mass transfer is to make the short-circuiting of the double layer discussed in Secs. even if Rct n n R0. More generally. incorporating both migration and electro-osmotic convection 2 where we wrote D = D / D. Among other things. We are currently studying in more detail the role of this nonlinear mechanism for inducing ﬂuid motion. and cO 0. and hence the conductivity. IV E and V C less dramatic. In this case the system will be dominated by capacitive charging and behave as if no Faradaic reaction is taking place. 19 —namely.33 . it interacts with the ﬂuid in a place more remotely from the wall and is therefore more efﬁcient in setting the ﬂuid in motion globally.t = 1 2 c+ + c− − 2c* dy Debye layer 57 =c* D4 sinh2 e . reducing the tangential ﬁeld and hence the electroosmotic slip and pumping velocity. and 0 is the characteristic geometrical length scale. and it is given by the maximal rate at which the reactants can be transported from the bulk to the electrodes by diffusion. known to play a decisive role in a few electrically generated phenomena 32. jext * current is reached around V0 eDOcO 0 / O . 9 to hold for ˜ 0 and Rct R0. for the forward reaction the * lim limiting current is jO eDOcO / O . However. the concentration cO tends to zero at the electrode and the Faradaic current levels off at the limiting value. this makes the conductivity in the diffusion layer differ signiﬁcantly from the bulk value and the Debye layer is driven out 1/3 2/3 of local equilibrium and expands in width to 0 D 30. this would change the local Debye length as seen from the electrode surface and give rise to space charge like y y log in the diffusion layer 32 . However. becomes signiﬁcantly larger than in the bulk electrolyte. Therefore. For a symmetric monovalent 1:1 electrolyte we deﬁne the excess amount of neutral salt absorbed in the Debye layer as w x. Beyond this voltage. Hence. this corresponds to V0 100 mV. This occurs for k BT 2 log e / D = k BT log e D/ . where s is the surface conductivity in the Debye layer.31 . AND AJDARI PHYSICAL REVIEW E 73. While the amount of space charge in the diffusion layer would typically be much smaller than in the Debye layer. we ﬁnd that salt depletion around the relaxation frequency 0 becomes an issue for V0 125 mV. * e 2 c OD O c RD R 55 2 D D 0 4 sinh2 e . 4kBT 58 This salt is periodically taken up and released again between = D/ . There are several interesting phenomena arising when the approximation of uniform bulk electrolyte concentration and conductivity breaks down. 54 of charge. 9 . the Debye layer and a diffusion zone of width However. Still. that of where those ions come from. As the driving voltage is increased. If there is no supporting electrolyte. D O * 10−9 m2 / s.

Our results compare well with experiments in many aspects but there still remains unexplained phenomena—e. we ﬁnd x Nq and similarly for ˜ on n ˜ n=1. 25 describing the charge balance in the Debye layer generalizes to 1+ 0 n · ˜ ˆ = i˜ +K ˜ x 1+ ˜ x ˆ ˆ − Vext . we expect our results to hold with some generality for other similar electrokinetic systems. the reversal of the net pumping or the time average slip direction. . 056313-15 . As these arguments are more general than the speciﬁc model. for ˜ to be a weak solution we require that ˜ V ˜ · ˜ dr + electrodes ˜˜ tq − ˜ext ˜ dx = 0 j q ˜ B1 Combining these we can write the time average slip as ˜s = u 1+ 1+ 1 ˆ Re Vext − ˆ ˜ 2 x 0 ˜ x ˜ x ˆ* for all and at all times. the net pumping U goes to zero in the low-frequency limit ˜ → 0. A8 clearly shows that regardless ˜ of .g. In the high-symmetry x x case where ˜ is a constant independent of ˜ . over one period. This points out the need for further studies of other nonlinear phenomena to fully understand the complex electrohydrodynamic system. A6 follows from ˜ 2 ˆ = 0. The weak problem is discretized using the Galerkin method. and ﬁnally. 2 . we expand ˜ on a ﬁnite set of basis functions n ˜ n=1 as r N N 1 ˜ 2L − Re V n·˜ ˆ i˜ + ˜ x ˆ * ds ˜. A7 follows from simple complex arithmetics. M. . CONCLUSION PHYSICAL REVIEW E 73. the solution is periodic in time. We therefore believe that our work will be useful for design of electrokinetic micropumps. The constraint.ac ELECTROKINETIC MICROPUMPS: THE EFFECT… VII. . For our particular model system of an asymmetric electrode array micropump we have obtained a number of results that we subsequently have interpreted using simple physical arguments. and current injection from Faradaic electrode reactions in both a linear and a nonlinear scheme. . B1 . APPENDIX A 1 1 ˜ U= Re ˜ ˜+ i 2L = 1 1 Re ˜ i˜ + 2L 1 2 V ˜ x ˜· ˜ˆ V ˜ x ˆ * dr ˜ ˆ * dr ˜ ˆ* ˜ dr A5 ˜ˆ ·˜ V ˜ x A6 1 = ˜ 2L = 1 ˜ 2L ˜ ˆ · ˜ ˆ * + ˜ Im ˜ ˆ · ˜ 2 ˜ x + ˜2 ˜ x A7 ˜ 2 + ˜2 Im V ˜ˆ ·˜ ˆ * dr . so we discretize on a set of M t equispaced points ˜m = m ˜. m = 1 .e.. ˜ A8 We give here a proof that there can be no net pumping at low frequency when the electrode surface properties are symmetric and we stay within the Debye-Hückel limit. and we use those same q basis sets as test functions in Eqs. integrating r over the computational domain V. x The electro-osmotic slip velocity becomes ˜s = u 1+ 1+ 0 Here Eq. The apparent singular behavior for = 0 is only apparent because the screening is then complete and ˜ ˆ zero in the bulk. our analysis rules out any net pumping in the low-frequency limit of our simple model when the electrode surface properties are symmetric in the sense of makx ing the parameter independent of ˜ . Eq. 056313 2006 We have extended existing theory for ac electrokinetic pumping to account for vertical conﬁnement of the pumping channel.. The ﬁnal result in Eq. nonlinear surface capacitance of the double layer. and using Gauss’ law with Eqs. Eq. Hence. A2 The nonlinear problem is converted to weak form by multiplying a test function ˜ for ˜ on Eq. Conx sider a device where the capacitance ratio ˜ and inverse charge-transfer resistance K ˜ vary across the electrodes. 45 . B1 and B2 . 46 and 49 to eliminate n · ˜ ˜ . APPENDIX B ˜ x ˜ −˜ Vext ˜ x ˜. Further. Eq. A1 where 0 is the capacitance ratio used in the rescaling of the problem. the net pump˜ ing velocity U can be manipulated as follows: ˜ 1 U= ˜ L ˜ s dx = u ˜ electrodes for all and at all times. on the potential drop across the double layer is satisﬁed by multiplying a test function ˜ for ˜ and requiring that x q ˜ − ˜ −˜ ˜ +˜ / 1 + Vext q q electrodes =− n·˜ ˆ 1 Re 2 i˜ + ˜ x ˆ* . t and use the leapfrog ﬁnite-difference scheme ˜˜ n ˜m tq t ˜ n ˜m+1 − ˜ n ˜m−1 / 2 ˜ to evaluate the time derivative in q t q t t Eq. A8 holds because the integral of the term ˜ ˜ ˆ · ˆ * cancels out due to x periodicity. and Eq. i. Hence. 50 .t r˜ n=1 ˜n˜ t n ˜ r B3 where V denotes the boundary of the computational domain and we used the fact that n · ˜ ˆ vanishes on insulating walls and cancels over periodic boundaries. Then. A4 ˜ ˜ . A3 ˜ dx = 0 B2 x x x where ˜ = K ˜ 1 + ˜ / 1 + 0 . using Gauss’ law and recalling that n is a unit normal pointing out of the electrodes and into V.

97. 021506 2004 . Cahill. Green. whereas the maximal relative V error on ˜ is about twice as large. 539 2004 . Phys. Argoul. 084906 2005 . Y. E 61. H. H. Z. 915 2002 . J. Smith. F. and A. A. Fluid Mech. Hayden. Mpholo. 1 H. E 70. C. Hunter. Ramos. Stone. Rev. 6 A. Ramos. and H. 129. Ajdari. A. R. and H. 27 A. 11 V. Gonzalez. Ajdari. C. G. 15 A. Manz. Chen. G. 26 At high frequency ˜ s is nonzero only in a region of width ˜ −1 u from the electrode edges. J. Rev. Figures 11 and 12 show the q results for M = 64. 4019 2000 . G. 22 L. Langmuir 21. Eijkel. 4 N. Heyderman. Electrochim. 387–399 2002 . H. used the width W2 of the wide electrode as their characteristic length scale rather than G1. Morgan. E 61. A. Olesen. Foundations of Colloid Science Oxford University Press. Ybert. and H. 32 P. Bazant. Some authors therefore ﬁnd it efﬁcient to use the empirical so-called constant-phase element to model the double layer. Ben. 33 J. C. and A. and A. using the commercial ﬁnite-element software FEMLAB to deﬁne the basis functions n ˜ and n ˜ and to evaluate the Jacobian mar x trix of the discretized problem. New York.OLESEN. 2 N. U. Here r is the distance from the edge and the polar angle. D. G. Phys. K. Ajdari. Here the tangential ﬁeld diverges as x Li2 z = n=1zn / n2 is the dilogarithm function. Annu. Chang. T. 13 A. Faulkner. R. A. Rev. 207211 2000 . A. 1485 2005 . 31 K. which makes the solution more demanding: Comparing results for M = 64 and M = 128 we ﬁnd that the ˜ maximal relative error on U is less than 0. the charging and decharging of the Debye layer becomes very rapid at high voltage. it is often found even for perfectly polarizable solid electrodes that the double-layer impedance does not match that of a perfect capacitor 29 . and A. C. Chu and M. and D. Morgan. Thornton. D. C. at the lowest frequency. Rev. Green. J. Pajkossy. E 70. Lab Chip 4. A. Phys. Phys. P. J. 20 N. Faraday Soc. E 9. Chen. Rennie. 11 and 13 there are ﬁve data points per decade for both frequency and voltage. 1730 2004 . 381 2004 . Rev. G. Ramos. we obtain fairly accurate results with a coarse time resolution of M = 16: When compared to a ﬁne resolution ˜ result at M = 64. A. Y. Ramos. Stroock. Kerner and T. R. H. Brown. following an approach initially suggested in Y. 207 1967 . C. 3 A. 4011 2000 . J. Castellanos. Ben unpublished . and A. Ajdari. Gonzalez. Zaltzman. Morgan. Nadal. Smyrl and J. J. Phys. J. 28 Experimentally. N. Phys. Rev. Microelectron. Analyst Cambridge. Appl. A. 19 M. Ajdari. and H. VI D–VI F. L. Dalton. the potential r1/2 cos / 2 . 29 Z. Lastochkin. Math. Hence in 15 both the dimensionless frequency and slip velocity are larger than ours by a factor of W2 / G1. Rubinstein and B.K. 944 2004 . G. one should keep in mind that at high voltage our weakly nonlinear model also becomes physically inadequate for reasons discussed in Secs. 2745 2003 . A. L. Appl. Acta 46. Ajdari. Actuators B 92. Debesset. Newman. Studer.1% 1. Castellanos. Eur. N. Prieve. Gonzalez. 056313 2006 The full coupled problem for all M time steps is solved numerically by damped Newton iterations. Ajdari. A. 8 V. Studer. Pepin. E 71. Eijkel. 24 Ramos et al. AND AJDARI PHYSICAL REVIEW E 73. Appl. A. Phys. C. Stemmer. Ramos. A. R45 2000 . 056302 2003 . Rev. Manz unpublished . 65. Langmuir 19. Bard and L. and H. Sens. Castellanos. However. but that the corresponding error on ˜ is as large as 0. 016305 2000 . E 67. G. 056306 2005 . J. Rev. 11. 10 S. Morgan. Pepin. Phys. and A. 5. Gobrecht. 18 A. 263 2001 .5% for 1 3 ˜ q V0 10 10 . Green. 026305 2002 . 23 F. 25 and therefore cannot be singular. Wang.5% 20% . 16 D. Pouligny. Sci. However. Morgan unpublished . Dalton. Sides. In particular. Garcia. Appl. Eng. and A. Phys. Figures 13 and 14 show results for M = 128. 21 I. Rev. Green. 96. C. Brown. J. Therefore our numerical solution becomes inaccurate for ˜ 103 and likewise for K 103 in the case of Faradaic current injection. Z. Bruus. Rev. Fagan. 9 M. we use the solution from lower voltage as initial guess when stepping to higher voltage. J. E 61. New York. Green.-C. and A. Gonzalez. 30 W. 61-62. Mortensen. Y. Castellanos. 036305 2004 . C.. and A. N. Hayden. With Faradaic current injection. 396 2004 . Belmon. 1991 . In the frequency-voltage maps in Figs. 262 2003 . 1980 . Models Meth. D. E 66. Smith. A. The convergence of the Newton iterations is accelerated by a continuation scheme where the converged solution at one frequency is used as initial guess for the next solution at higher frequency and the same voltage. A. 36. Gonzalez. L. the maximal relative error on U is less than 1 3 1% 3% for ˜ 0 10 10 . H. C. close to the sharp electrode edge goes like and the electric ﬁeld diverges as r−1/2. N. 7 S. Electrochemical Methods Wiley. Math. T. where the screening is negligible. 056313-16 . C. 5 A. A. the tangential ﬁeld diverges logarithmically. Ramos. At ﬁnite frequency the normal ﬁeld is linked to the potential drop across the double layer through Eq. A. T. Debesset. B. 25 A well-known instance of this singular behavior is that at high frequency. P. Morgan. 63. Phys. B. at low frequency it can be shown analytically that goes like the potential to ﬁrst order in Im Li2 e−2 i x−x j +y /L close to the electrode edge 22 and −log x − x j . B. Zhou. 14 P. and A. Bazant. Castellanos. J. Phys. Krestener. 12 B. E 63. Phys. P. Green. SIAM J. Sides. BRUUS. P. and A. and A. With nonlinear Debye layer capacitance but no Faradaic currents. 1784 2005 . Olesen unpublished . G. H. but we prefer not to do so due to the lack of structural justiﬁcation for it. 17 R. J. Trans. Phys. and A.

Meth. Num. Meth. Title A high-level programming-language implementation of topology optimization applied to steady-state Navier–Stokes ﬂow Authors Laurits Højgaard Olesen. 65. J. Numer. J. Eng. and Henrik Bruus Reference Int. Eng. 975-1001 (2006) (27 pages) .Appendix C Paper published in Int. Fridolin Okkels.

150 Paper published in Int. J. Numer. Meth. . Eng.

DTU Bldg. Numer. † E-mail: lho@mic. Recently Borrvall and Petersson introduced the method for ﬂuids in Stokes ﬂow [6]. 345 East. Ltd. MIC—Department of Micro and Nanotechnology. a direction pioneered by the work of Sigmund and Gersborg-Hansen [7–9].com). Ltd. We analyse the physical aspects of the solutions and how they are affected by different parameters of the optimization algorithm. contract/grant number: 26-03-0037 ∗ Correspondence Copyright 2005 John Wiley & Sons. 41:77–107).dk Contract/grant sponsor: The Danish Technical Research Council. Navier–Stokes ﬂow. J. Denmark SUMMARY We present a versatile high-level programming-language implementation of non-linear topology optimization.1002/nme.wiley. it is desirable to extend the method to ﬂuids de