Multiscale Modeling of Interface Phenomena in Biologyricsac/NotesCMElBioMath.pdf · Multiscale...
Transcript of Multiscale Modeling of Interface Phenomena in Biologyricsac/NotesCMElBioMath.pdf · Multiscale...
Notes of the Course
Multiscale Modeling of InterfacePhenomena in Biology
Ph.D. School in Mathematical Models and Methods in Engineering
Dipartimento di Matematica “F. Brioschi” Politecnico di MilanoMobility Project “Athens” - Politecnico di Milano
Prof. Riccardo Sacco
Dipartimento di Matematica “F. Brioschi” Politecnico di Milano
Piazza Leonardo da Vinci 32 20133 Milano, Italy
E-mail: [email protected]
Home Page: http://www1.mate.polimi.it/~ricsac/
Contents
1. Structure of the Course 3
I. Cellular Interfaces: Structure and Transport Phenomena 5
2. Introduction to Cellular Interfaces and Ion Electrodiffusion 7
2.1. Cells: structure, membrane and ion transport . . . . . . . . . . . . . . . . 7
2.1.1. The cell membrane . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2. Ionic channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2. Transport of charged particles . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1. Units and conventions . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2. The Nernst-Planck equation . . . . . . . . . . . . . . . . . . . . . . 11
3. ODE-Based Modeling of Transmembrane Ion Flow in Cellular Electrophysiol-
ogy 13
3.1. Membrane electrophysiology . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2. General form of transmembrane ionic current densities . . . . . . . . . . . 14
3.3. The ODE model of ion transport . . . . . . . . . . . . . . . . . . . . . . . 15
3.4. Transmembrane current models . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.1. The linear resistor model . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.2. The Goldman-Hodgkin-Katz model . . . . . . . . . . . . . . . . . . 16
3.4.3. The Hodgkin-Huxley model . . . . . . . . . . . . . . . . . . . . . . 19
3.5. Thermal equilibrium of a system of monovalent ions . . . . . . . . . . . . 20
II. Multiscale Mathematical Models of Transmembrane Ion Transport 23
4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology 25
4.1. Electrochemical description of ion flux . . . . . . . . . . . . . . . . . . . . 25
4.1.1. The electric field . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.2. Ion mass flux density . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.1.3. Ion electrical current density . . . . . . . . . . . . . . . . . . . . . 27
4.2. Microscale model of cellular ion flow . . . . . . . . . . . . . . . . . . . . . 27
4.3. Multiscale reduction of cellular ion flow . . . . . . . . . . . . . . . . . . . 30
4.4. Macroscale model of ion flow . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.4.1. Electrostatic model of the membrane . . . . . . . . . . . . . . . . . 31
4.4.2. Electrodynamical model of the membrane . . . . . . . . . . . . . . 33
II Indice
4.5. The PNP system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.6. The Cable Equation model . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5. Mathematical Analysis of the Poisson-Nernst-Planck Model 41
5.1. Summary of model equations . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1.1. Scaling of the PNP system . . . . . . . . . . . . . . . . . . . . . . 42
5.2. The scaled PNP model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3. The electroneutral PNP model (EN-PNP) . . . . . . . . . . . . . . . . . . 45
III. Functional Techniques and Discretization Methods 47
6. Solution Map for the PNP system 49
6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.2. PNP solution map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.2.1. The Nonlinear Poisson equation . . . . . . . . . . . . . . . . . . . 51
6.2.2. The continuity equations . . . . . . . . . . . . . . . . . . . . . . . 53
7. Unified Framework and Well-Posedness Analysis 55
7.1. Unified framework for the PNP solution map . . . . . . . . . . . . . . . . 55
7.2. Weak formulation and well posedness . . . . . . . . . . . . . . . . . . . . . 57
7.2.1. Multi-domain functional setting . . . . . . . . . . . . . . . . . . . . 57
7.2.2. DAR problem reformulation . . . . . . . . . . . . . . . . . . . . . . 57
7.2.3. Weak formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.2.4. Well posedness analysis and stability estimate . . . . . . . . . . . . 59
8. Finite Element Approximation of the DAR Problem 61
8.1. Motivation to the use of a DMH method . . . . . . . . . . . . . . . . . . . 61
8.2. Geometric Discretization and Finite Element Spaces . . . . . . . . . . . . 63
8.3. A Mixed–Hybridized Method with Numerical Quadrature . . . . . . . . . 66
8.4. Implementation and Post-Processing of the DMH-FV Method . . . . . . . 74
8.4.1. Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
8.4.2. Post-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
IV. Simulation Results, Applications and Advanced Topics 77
9. Numerical Validation of the DMH-FV Method 79
9.1. Static condensation CAMBIA TITOLO . . . . . . . . . . . . . . . . . . . 79
9.2. A one-dimensional heterogeneous domain . . . . . . . . . . . . . . . . . . 82
9.3. Stationary profile of a binary electrolyte at a boundary . . . . . . . . . . . 83
9.4. Simulation of a neuro-chip . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
9.5. Action potential propagation in an axon . . . . . . . . . . . . . . . . . . . 87
10.Advanced Topics, A Look at the Future and Concluding Remarks 91
10.1. Applications of the future . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
10.2. Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Indice III
Acknowledgments
I wish to gratefully acknowledge the fundamental contribution given by Matteo Porro,
PhD student in ”Mathematical Models and Methods in Engineering” at Politecnico di
Milano, in the development and computer implementation of the various multiscale mod-
els illustrated in the course, for his personal assistance during the hands-on sessions on
the computer and for carefully reading (and proof-checking!!) these Lecture Notes.
I wish also to thank Emanuela Abbate, MSc student in Mathematical Engineering,
Politecnico di Milano, for her precious assistance in developing multiscale models and
algorithms for neural interfaces.
Last but certainly not least, my grateful thanks go also to Prof. Joseph W. Jerome,
Dr. Bice Chini, Dr. Yoichiro Mori and Dr. Paola Causin for all the enlightning and
productive discussions and years of common work in the area of mathematica modeling
and simulation of Ionic Transport in Cellular Biology.
Finally, I wish also to thank Marco Brera and Fabio Manganini, two former MSc bril-
liant students under my supervision at Politecnico di Milano, that helped me developing
and implementing a substantial part of models and methods discussed in these Notes.
Indice 1
Introduction
These notes represent the supporting material for the course entitled
Multiscale Modeling of Interface Phenomena in Biology
to be held within the PhD programme ”Mathematical Models and Methods in Engineer-
ing” and within the Mobility Project ”Athens”. For further information on these two
initiatives, refer to:
- http://www.mate.polimi.it/dottorato/index.php?lg=it
- www.athensprogramme.com (course catalogue POLI13)
The course has a duration of 27 hours, distributed over a period of one week and
organized into lectures and hands-on sessions on the computer. The scope of the course
is to introduce the mathematical modeling aspects of biological interfaces and the com-
putational techniques that can be used for their successful simulation.
These notes are divided into four distinct parts:
• Part I gives a short introduction to cellular biology and electrophysiology, providing
also a first example of mathematical model of ion transport across a cellular mem-
brane. The proposed model is based on an equivalent electrical representation of
the cell and is constituted by a system of Ordinary Differential Equations (ODEs)
derived from the application of Kirchhoff’s current law.
• Part II illustrates the mathematical models of ion transport based on systems
of Partial Differential Equations (PDEs) that generalize to the multi-dimensional
setting the basic ideas developed in Part I.
• Part III describes the functional techniques used to decouple the equation systems
and the corresponding numerical methods used for their temporal and spatial ap-
proximate solution.
• Part IV concludes the course presentation with the simulation of several test studies
of interest in biological applications and with a short overview of a few selected
advanced topics including:
1. advanced bio-hybrid interfaces (example: prototype of artificial retina);
2. cellular mechanotransduction processes (example: calcium sparks induced by
chemical and mechanical stimuli).
1. Structure of the Course
Interfaces are ubiquitous in Biology. They have the function to separate two neighbouring
regions maintaining a condition of local equilibrium through a dynamical control of the
electro-chemical and mechanical processes occurring thereby. The ratio between the
interface thickness and the characteristic size of the system is usually a very small quantity
- of the order of 1/1000 in the case of cellular membrane - so that a hierarchical multiscale
vision of the coupling between interface and surrounding environment seems to be the
right option for an efficient and accurate modeling of the problem.
MATHEMATICAL MODELS Motivated by the above consideration, in this course
we aim to provide an introduction to Computational Modeling of interface phenomena
in biological systems by illustrating in detail a hierarchy of ODE/PDE equation-based
models for the study of cellular electrical activity (CEA). The members of the hierarchy,
in increasing order of complexity, include:
(CEA1) 0D equivalent electrical lumped parameter models;
(CEA2) the Cable Equation (CE) model;
(CEA3) the Poisson-Nernst-Planck (PNP) and the electroneutral PNP (EN-PNP) models
for electrodiffusion of Mion ionic species in an electrolyte.
In all CEA formulations, the dynamical response of the interface to externally applied
stimuli is represented by equivalent electrical models including capacitors and nonlinear
resistors to account for charge accumulation and trans-interface currents (TIC). TICs
are treated mathematically by adopting hierarchical models of increasing complexity
that comprise:
(TIC1) the linear resistor model;
(TIC2) the Goldman-Hodgkin-Katz nonlinear model;
(TIC3) the Hodgkin-Huxley system of ODEs.
LINEARIZATION AND NUMERICAL APPROXIMATION Combining together the
CEA/TIC model hierarchies, a highly nonlinear system of ODE/PDEs is obtained. CEA1
is numerically treated by resorting to ODE solvers for stiff problems, while in CEA2 and
CEA3 temporal semi-discretization is treated using the Backward Euler method. System
linearization is dealt with using a Gummel-type fixed-point iteration. This approach
leads to the successive solution of an heterogeneous-domain advection-diffusion-reaction
4 1. Structure of the Course
(HD-ADR) boundary value model problem in conservative form with appropriate flux
transmission conditions across the interface. The numerical approximation of the HD-
ADR problem is dealt with mixed-hybridized finite element methods (MH-FEM) char-
acterized by strong flux conservation across interelement boundaries and robustness in
treating sharp fronts and/or discontinuities.
MATHEMATICAL AND NUMERICAL ISSUES
a) well-posedness of the linearized boundary-value problems;
b) existence and uniqueness of a fixed point;
c) convergence of the iterations;
d) asymptotic analysis with respect to singular perturbation parameters;
e) stability of the MH-FEM scheme (discrete maximum principle);
f) convergence analysis with respect to mesh size;
g) algebraic structure of the MH method and efficient implementation through the use
of static condensation;
h) derivation of a robust and efficient finite volume MH scheme by lumping of the local
mass flux matrix.
SIMULATION RESULTS AND MODEL VALIDATION
A) study with the PNP model of the stationary profile of a binary electrolyte (Mion = 2)
at a boundary (comparison with exact solution);
B) study with the PNP model of the time-dependent electrodiffusion of a ternary elec-
trolyte (Mion = 3) across a membrane (critical analysis of the electroneutrality
assumption);
C) study with PNP and EN-PNP models of the propagation of an action potential along
a neuronal axon;
D) simulation with PNP and EN-PNP models of cellular interfacing with an electronic
substrate of inorganic/organic type (neurochip).
ADVANCED TOPICS AND FURTHER READING The concluding remarks of the
Course are devoted to a short overview of a few selected advanced topics (listed be-
low), including bibliographical references, with suggestions for further development and
application of ideas and methods addressed in the notes:
1. study and simulation of advanced bio-hybrid interfaces (example: prototype of
artificial retina);
2. study and simulation of cellular mechanotransduction processes (example: calcium
sparks induced by chemical and mechanical stimuli).
Part I.
Cellular Interfaces: Structure and
Transport Phenomena
2. Introduction to Cellular Interfaces and
Ion Electrodiffusion
In this chapter, we give a short introduction to the biological setting object of the course.
In particular, we illustrate the structure of a cell and of its membrane. Then, we introduce
the ionic channels and we discuss their role in the regulation of the cellular electrical
activity. For more detailed information, we refer to [Hil01, GH06, BF01].
2.1. Cells: structure, membrane and ion transport
The basic living unit of the body is the cell : each organ is an aggregate of many different
cells held together by intercellular supporting structures. The entire body contains about
100 trillion cells (∼ 1014 cells). Although the many cells of the body often differ markedly
from one another (since each type of cell is specially adapted to perform one or a few
particular functions), all of them have certain basic characteristics that are alike: a
typical cell, as seen by the light microscope, is shown in Fig. 2.1. Its two major parts
are the nucleus and the cytoplasm. The nucleus is separated from the cytoplasm by a
nuclear membrane, and the cytoplasm is separated from the surrounding fluids by a cell
membrane. The principal fluid medium of the cell is water, which is present in most
cells, except for fat cells, in a concentration of 70 to 85 %. Many cellular chemicals are
dissolved in the water. Others are suspended in the water as solid particulates. Chemical
reactions take place among the dissolved chemicals or at the surfaces of the suspended
particles or membranes.
2.1.1. The cell membrane
The cell membrane (also called the plasma membrane), which envelops the cell, is a
thin, pliable, elastic structure with a thickness tm of only 7.5 to 10 nanometers. The
diameter dc of cells varies from 7.5 to 150 micrometers . Thus, the ratio tm/dc varies
in the range 50 × 10−6 ÷ 1.3 × 10−3. Despite of its relative small dimension, the cell
membrane plays some fundamental roles for the living of the cell: it preserves the cell
integrity separating the intracellular fluid from the extracellular fluid and it regulates
the passage of substances from the outside to the inside of the cell, and viceversa.
The cell membrane consists almost entirely of a lipid bilayer, but it also contains large
numbers of protein molecules in the lipid, many of which penetrate all the way through
the membrane, as shown in Fig. 2.2. Most of these penetrating proteins constitute a
pathway through the cell membrane. Some proteins, called channel proteins, allow free
movement of water as well as selected ions or molecules. The channel proteins are usually
8 2. Introduction to Cellular Interfaces and Ion Electrodiffusion
Figure 2.1.: Structure of the cell as seen with the light microscope.
highly selective with respect to the types of molecules or ions that are allowed to cross
the membrane.
Figure 2.2.: Cross-section of a cell membrane.
Ions provide inorganic chemicals for cellular reactions. Also, they are necessary for
operation of some of the cellular control mechanisms. For instance, ions acting at the
cell membrane are required for transmission of electrochemical impulses in nerve and
muscle fibers. The most important ions in the cell are potassium K+, magnesium Mg++,
phosphate, sulfate SO−−4 , bicarbonate HCO−3 , and smaller quantities of sodium Na+,
chloride Cl−, and calcium Ca++. An ion is called cation if positively charged, anion if
negatively charged.
2.1. Cells: structure, membrane and ion transport 9
2.1.2. Ionic channels
Certain cells, commonly called excitable cells, are unique because of their ability to gen-
erate electrical signals. Some examples are neuron cells, muscle cells, and touch receptor
cells. Like all cells, an excitable cell maintains a different concentration of ions in its cy-
toplasm than in its extracellular environment. Together, these concentration differences
create a small electrical potential across the plasma membrane. Then, when the potential
difference reaches a threshold value, typically 55 millivolts, specialized channels in the
plasma membrane, called ion channels, open and allow rapid ion movement into or out
of the cell, and this movement creates an electrical signal. All of these processes charac-
terize the so-called cellular electrical activity (CEA) that represents the way ion channels
can generate an electrical current flowing between a cell and another, thus providing a
fundamental aspect in the life of every biological system.
Ionic channels are large proteins that reside in the membrane of cells (Fig. 2.3) and
conduct ions through a narrow tunnel of fixed charge formed by the amino acid residues
of the protein.
Channels are ideally placed across the membrane in series with the intracellular en-
vironment to control the cellular biological function [SBT02]. Ion channels should be
viewed as natural nanotubes that relate the electrolyte solutions in and outside the cell
to the electric field that is established across the cell membrane.
Figure 2.3.: Ion channels function as pores to permit the flux of ions down their electro-
chemical potential gradient.
Channels are responsible for signaling in the nervous system, coordination of muscle
contraction, and transport in all tissues. Channels are obvious targets for drugs and
disease [Sch96]: as a matter of fact, many of the drugs used in clinical medicine act
directly or indirectly through channels.
10 2. Introduction to Cellular Interfaces and Ion Electrodiffusion
2.2. Transport of charged particles
In biological channels, ions move under the controlled action of gradients of concentration
and electric potential [Hil01]. The (opposite) of the gradient of the electric potential,
which is the electric field, forces a particle to move accordingly to its sign (Coulomb’s
law). This mechanism is called drift (see Fig. 2.4(a)). The gradient of concentration
moves particles from regions at higher concentration to regions at lower concentration
(Fick’s law). This mechanism is called diffusion (see Fig. 2.4(b)). The flow of ions
through the channel generates an electrical current denoted henceforth by the symbol
I. From a practical point of view, we are mainly interested in computing the current
density J , which is the charge that crosses a generic cross-sectional area S in a unit of
time, in such a way that (in terms of modules) we can write
J =I
S. (2.1)
Using linear superposition of the drift and diffusion processes, the current density takes
the following form
J = Jdrift + Jdiffusion. (2.2)
Notice that Jdrift and Jdiffusion might act in opposite directions, as shown in Fig. 2.4.
(a) Drift (b) Diffusion
Figure 2.4.: Drift and diffusion currents. The electric field is denoted by E.
2.2.1. Units and conventions
To characterize in mathematical terms the drift-diffusion model for the current density
we need the following fundamental quantities:
• c the concentration of the ionic species;
• z the valence of the ion (dimensionless);
• D the diffusion coefficient of the ionic species (units: m2 s−1);
• E = −∂ϕ∂x
electric field (units: V m−1), where x denotes the spatial coordinate of
ion flow (units: m) and ϕ is the electric potential (units: V).
2.2. Transport of charged particles 11
The definition of the units for the ion concentration c is not unique, according to the
physical meaning expressed by the variable. In fact, if c is a molar density its appropriate
units are mol m−3, while if c is a number density the right units are #ions m−3 (more
shortly, m−3). The first choice is typical of bio-electrochemical applications while the
second choice is the preferred one in electronics applications. We adopt henceforth the
following convention: in numerical computations and/or input/output tables of data,
the concentration will be expressed in mol m−3, while in all the mathematical modeling
treatment the concentration will be expressed in m−3. Thus, possible change of units
will have to be done only twice, one time before computations, the other as a simple
post-processing.
2.2.2. The Nernst-Planck equation
The number flux density of ions through the membrane along the direction x that is driven
by the simultaneous effect of concentration gradients and electric field is expressed by
the relation
f = −D ∂c
∂x︸ ︷︷ ︸Fick diffusion
− zq
KTDc
∂ϕ
∂x︸ ︷︷ ︸Drift
(2.3)
where KB is Boltzmann’s constant (units: J K−1) and q is the electron charge (units: C).
The units of the flux are
[f ] = [D] ·m−4 = m2 s−1 ·m−4 = m−2 s−1.
Multiplying this latter expression by qz yields
J = qz × f = −qzD(∂c
∂x+
zq
KTc∂ϕ
∂x
). (2.4)
The units of J are
[J ] = [q]× [f ] = Cm−2 s−1 = A m−2
from which we conclude that J is a current density. Equation (2.4) is known as Nernst-
Planck relation and is the most widely used modeling tool in Electrophysiology to describe
the motion of ions under the effect of diffusive and drift forces.
We conclude this section with some concepts and definitions that will play an impor-
tant role in the discussion to follow.
Remark 2.2.1 (Thermal equilibrium). The special (and important) case where Jdrift =
−Jdiffusion corresponds to J = 0 and is commonly referred to as thermal equilibrium
(see Sect. 3.2). In this situation, the net current flow across a section is null although
the separate drift and diffusion contributions are non-zero. Thermal equilibrium is a dy-
namical state and allows to introduce the so-called Nernst potential that is a fundamental
quantity in the analysis of the electrical activity of every excitable cell.
Definition 2.2.1 (Nernst potential). The Nernst potential (units: V) associated with
the ionic species c is defined as
Ec =KBT
zqln
(c(out)
c(in)
)=RT
zFln
(c(out)
c(in)
)(2.5)
12 2. Introduction to Cellular Interfaces and Ion Electrodiffusion
where R is the universal gas constant (units: J mol−1 K−1) and F is Faraday’s constant
(units: C mol−1).
Definition 2.2.2 (Thermal equilibrium).
Thermal equilibrium ⇔ J = 0 ⇔ ϕm = Ec. (2.6)
3. ODE-Based Modeling of Transmembrane
Ion Flow in Cellular Electrophysiology
As anticipated in Chapter 2, the cell membrane is a biological interface that separates
the interior of every cell from the outside environment. The cell membrane is selectively
permeable to ions and organic molecules and controls the movement of substances in and
out of the cell. The basic function of cell membrane is to protect the cell from its sur-
roundings, but it is also involved in a variety of cellular processes such as cell adhesion,
ion conductivity and cell signaling. In this chapter, focus on the membrane electrical
activity developing a biophysically sound description of the ionic currents through the
membrane. The resulting mathematical model is based on an equivalent electrical rep-
resentation of the cell and is constituted by a system of Ordinary Differential Equations
(ODEs) derived from the application of Kirchhoff’s current law. The system must be
solved at each spatial point of the membrane surface and at each time level of the tempo-
ral evolution of the biophysical problem under given initial conditions for the membrane
potential and the ionic concentrations in the intra- and extracellular sites.
3.1. Membrane electrophysiology
In the simplest picture, the electrical properties of cell membrane are represented in terms
of the electrical equivalent circuit illustrated in Fig. 3.1. Several electrical components
can be identified in the scheme. Resistors (of both linear and nonlinear types) are used
to model the various types of ion channels embedded in the membrane. Voltage supply
generators (batteries) are used to represent the electrochemical potentials induced by
different values of intra- and extracellular ion concentrations (Nernst potentials, also
called reversal potentials). Capacitors are used to model the charge storage capacity of
the cell membrane during transient phenomena.
In the equivalent circuit, the current across the membrane has two major components,
one associated with the membrane capacitance and the other associated with the flow of
ions through resistive membrane channels. The behavior of the electrical circuit shown
in Fig. 3.1 can be described by the following Ordinary Differential Equation (ODE)
Cmdϕmdt
+ Itot = 0 (3.1)
where:
• ϕm := ϕin − ϕout is the membrane potential, given by the difference between the
value of the electric potential ϕ in the intracellular space and that in the extracel-
lular space;
143. ODE-Based Modeling of Transmembrane Ion Flow in Cellular
Electrophysiology
Figure 3.1.: Electrical equivalent circuit for cell membrane. The capacitor represents
the capacitance of the cell membrane; the three variable resistors repre-
sent voltage-dependent conductances, the fixed resistor represents a voltage-
independent conductance and the batteries represent reversal potentials for
the corresponding conductances.
• Cm is the cell membrane capacitance (units: F);
• Itot is the sum of the transmembrane ionic currents flowing through the linear and
nonlinear resistors in Fig. 3.1 (units: A = C s−1).
Equation (3.1) is the fundamental Kirchhoff current law relating the time rate of change
of the membrane potential to the currents flowing across the membrane.
Definition 3.1.1 (0D - Lumped parameter model). A model like that schematically rep-
resented in Fig. 3.1 will be henceforth referred to as a 0D model, because the dependence
of the unknown ϕm on the spatial variable is neglected in Eq. (3.1). Using the language
of electrical engineering, the 0D model is also called lumped parameter model, because
the complex bio-physical behaviour of the cell membrane with respect to ion transport
is synthetically represented by the lumped electrical parameters (membrane capacitance,
conductances, batteries and current generators) appearing in Fig. 3.1. More general
mathematical formulations based on the use of a distributed parameter approach will be
considered in the remainder of these notes.
At this point, the main challenge is to characterize the form of Itot on the basis of
definition (2.1) and of the general drift-diffusion relation (2.2).
3.2. General form of transmembrane ionic current densities
Transmembrane ionic currents are currents that flow through ion channels, transporters,
or pumps that are located within the cell membrane. To define in mathematical terms
these currents we adopt the formalism of Hodgkin and Huxley [HH52], generalized here
to allow for nonlinear instantaneous current-voltage relations and ion concentration ef-
fects [Mor06]. In the remainder of these notes we always assume, otherwise differently
stated, that in the considered electrolyte solution a number of Mion ≥ 1 ionic species is
flowing. Each ion has a concentration ci and a ionic valence zi in such a way that the
3.3. The ODE model of ion transport 15
amount of charge carried by the ionic species per unite volume is qzici, q being the ele-
mentary charge of the electron equal to 1.602 · 10−19 Coulomb. In accordance with (2.1),
we introduce the quantity S that represents an arbitrarily chosen cross-sectional area of
the membrane surface across which the considered transmembrane current Ii is flowing.
Then, the transmembrane current density (units: A m−2) associated with the i-th ionic
species has the following expression
Ji = Ji
(x,y(x, t), ϕm(x, t), c(in)(x, t), c(out)(x, t)
)i = 1, . . . ,Mion. (3.2)
In the above relation, x denotes the spatial position vector along the membrane, t is the
time variable while the other parameters are defined as follows:
• y(x, t) = (y1, . . . , yNg) is a vector of gating variables where Ng is the total number of
gating variables in all of the channel types that arise in our system. The individual
components of y are dimensionless variables in the range [0, 1] and describe the
time-dependent activation or inactivation profile of the channel.
• ϕm(x, t) is the transmembrane potential. Keeping the other parameters fixed in
Ji, and letting only ϕm vary, we get the instantaneous current-voltage relationship
for the i-th ion from the extracellular space to the intracellular space at point x at
time t.
• c(in)(x, t) = (c(in)1 , . . . , c
(in)Mion
) (and similarly c(out)) is the vector of ion concentra-
tions in the intracellular (respectively extracellular) space. By including the whole
vector of ion concentrations, we allow for the possibility that the current density
carried by the i-th species of ion is influenced by the concentrations of other ionic
species on the two sides of the membrane.
The functional relation (3.2) expresses the bio-physical fact that the current density
of the i-th ion may be influenced by the transmembrane current density of the other
channels (possibly of more than one type) that carry the i-th species of ion across the
membrane separating the intracellular and the extracellular space. Finally, the explicit
dependence of Ji on x reflects the possible inhomogeneity of the membrane, because the
density of channels may vary from one location to another.
3.3. The ODE model of ion transport
Replacing the general expression of the transmembrane current density (3.2) into (3.1)
we obtain the following ODE to be solved at each time level t > 0 in correspondance of
each spatial position x on the membrane:
cmdϕmdt
(x, t) = −Mion∑i=1
Ji
(x,y(x, t), ϕm(x, t), c(in)(x, t), c(out)(x, t)
)(3.3a)
ϕ(x, 0) = ϕ0(x). (3.3b)
In (3.3), the quantity cm is the specific membrane capacitance (units: F m−2) related
to Cm through the relation cm = Cm/Sm, Sm being the area of the membrane, while
ϕ0(x) is the initial value of the membrane potential at each point of the membrane.
163. ODE-Based Modeling of Transmembrane Ion Flow in Cellular
Electrophysiology
The concentrations c(in)(x, t) and c(out)(x, t) of the Mion ionic species are assumed to
be (biophysically suitable) given functions, and the same holds for the gating variables
y(x, t). We shall see that this is not the case with the more general PDE-based models
treated in the remainder of these notes. Under these assumptions, (3.3) is a system of
Cauchy problems for the membrane potential ϕm to be solved (in principle) at each spatial
position x of the membrane. In practice, this cannot be done and the ODE system (3.3)
is solved only at a finite number of points xk suitable selected over the membrane surface.
Examples of this approach will be discussed in Part IV of these notes.
3.4. Transmembrane current models
In this section we present the models for the transmembrane current densities that are
most commonly used in the theoretical and computational description of cellular electrical
activity. For a fully detailed treatment of this complex subject, we refer to [KS98, Hil01,
ET10].
3.4.1. The linear resistor model
This is the simplest current-voltage relationship because the ionic current density Ji of
the i-th ion can be expressed as
Ji = gi (ϕm − Ei) = gi
(ϕin − ϕout −
KBT
zqln
(c
(out)i
c(in)i
)), (3.4)
gi being the specific conductance associated with the ionic species ci. The units of gi are
[gi] = A V−1 m−2 = S m−2.
By inspection, we see that (3.4) is consistent with the thermal equilibrium condition (2.2.2).
The graphical representation of the current-voltage relationship (3.4) in the (ϕm, J) plane
is a straight line whose slope is equal to gi. Despite its simplicity, the linear resistor model
proves to be quite accurate in many cases and it is used for instance in [Fro03].
3.4.2. The Goldman-Hodgkin-Katz model
The Goldman-Hodgkin-Katz (GHK) model is a first, significant, example of improvement
of the linear resistor formulation of the previous section. To derive a realistic model of
fluxes and currents that flow across the cellular membrane it is necessary to make some
simplifying assumptions (see also [KS98, Rub90]):
1. equation (2.4) holds across the membrane;
2. ions move independently;
3. the electric field constant across the membrane;
4. the current density is constant across the membrane.
3.4. Transmembrane current models 17
Figure 3.2.: Schematics of ion electrodiffusion across the cell membrane.
Referring to Fig. 3.4.2 for the notation and indicating by x the spatial coordinate par-
allel to the channel (black arrow in the figure) and by l the membrane thickness, the
application of assumption 3. yields
E = −∂ϕ∂x
= constant = −ϕout − ϕinl
=ϕml.
Replacing this relation in (2.4) we get
Ji = −qziDi
(∂ci∂x− ziq
KTciϕml
). (3.5)
Then, applying assumption 4. to (3.5) we obtain
Ji = constant =⇒ ∂Ji∂x
= 0 =⇒ ∂2ci∂x2
− ziq
KT· ϕml· ∂ci∂x
= 0.
The solution of the above differential equation is given by
ci(x) = A+B exp( ziqKT
ϕmx
l
)(3.6)
where A and B are arbitrary constants that can be found by imposing the following
boundary conditions at the two sides of the channel:
ci(0) = c(in)i
ci(l) = c(out)i
=⇒
A+B = c
(in)i
A+Beziq
KTϕm = c
(out)i
=⇒
A = c
(in)i −
c(out)i − c(in)
i
eziq
KTϕm − 1
B =c
(out)i − c(in)
i
eziq
KTϕm − 1
.
For notational simplicity, we introduce the dimensionless variable
X := ziϕmKT/q
which has the physical meaning of a normalized electric potential. Then, to compute the
constant current density throughout the channel we replace (3.6) into (3.5) to obtain
Ji = −FziDi
(BX
lexp
(Xx
l
)− X
lA−BX
lexp
(Xx
l
)).
183. ODE-Based Modeling of Transmembrane Ion Flow in Cellular
Electrophysiology
The first and third term in the braces at the right-hand side mutually cancel out, and
we are left with the constant current density
Ji = FziDiX
lA = −FziDi
1
l
[X
eX − 1c
(out)i − XeX
eX − 1c
(in)i
].
It is useful to introduce the inverse of the Bernoulli function
B(X) :=X
eX − 1. (3.7a)
A plot of B(X) and B(−X) is reported in Fig. 3.3.
Figure 3.3.: Plot of B(X) (solid line) and of B(−X) (dashed line).
This function has several properties:
B(X) > 0 ∀X ∈ R; (3.7b)
B(0) = 1; (3.7c)
eXB(X) = B(−X) = X + B(X); (3.7d)
limX→+∞
B(X) = 0+, limX→−∞
B(X) = −X; (3.7e)
limX→+∞
B(−X) = X, limX→−∞
B(X) = 0+. (3.7f)
Using the definition (3.7) and property (3.7d), the constant current density can be
written as
Ji = −qziDi1
l
[B(ziϕmKT/q
)c
(out)i − B
(− ziϕmKT/q
)c
(in)i
]. (3.8)
This is the celebrated Goldman-Hodgkin-Katz (GHK) equation for the current density
associated with the i-th ion.
Let us check that (3.8) satisfies the thermal equilibrium condition (2.2.2). Using
properties (3.7b) and (3.7d), we have
Ji = 0 =⇒ c(out)i − c(in)
i exp(ziqϕmKT
)= 0
from which we get
ϕm =KT
ziqln
(c
(out)i
c(in)i
)≡ Ei
3.4. Transmembrane current models 19
that is, the membrane potential coincides with the Nernst potential as required at thermal
equilibrium.
The GHK current density enjoys other interesting properties. Assume that ϕm = 0
(i.e., the intra- and extracellular potentials have the same value). Then, using prop-
erty (3.7c) the GHK current density degenerates in
Ji = −qziDic
(out)i − c(in)
i
l.
This formula corresponds to a pure diffusion ion flow across the membrane in agreement
with the Nernst-Planck relation (3.5) in absence of electric field.
Conversely, assume that c(out)i = c
(in)i = ci (i.e., the intra- and extracellular ion
concentrations have the same value ci). Then, using property (3.7d) the GHK current
density degenerates in
Ji = −qziDicil
(−zi
ϕmRT/q
).
This formula corresponds to a pure drift ion flow across the membrane in agreement with
the Nernst-Planck relation (3.5) in absence of a concentration gradient.
The above analysis shows that the GHK expression of the ion current density au-
tomatically adapts itself to all possible transport regimes. This makes it amenable to
numerical computations and gives the reason of its wide success and implementation in
contemporary simulation tools.
3.4.3. The Hodgkin-Huxley model
The Hodgkin-Huxley (HH) model [HH52] is a further extension of the simple linear
resistor theory and accounts for voltage-gating mechanism of the channels, which in turn
permits the simulation of the propagation of an action potential. Four ionic species, Na+,
K+ and Cl− are included in the mathematical description because they are responsible
for the majority of the ionic current in a cellular action potential [KS98, Hil01, ET10].
The HH model requires solving a nonlinear, stiff ODE system, that is:
Jm = Cmdϕmdt
+ JK + JNa + JL
JK = n4 gK(ϕm − EK)
JNa = hm3 gNa(ϕm − ENa)
JL = gL(ϕm − EL).
(3.9)
The membrane current Jm is the sum of two ionic current fluxes, associated with
potassium and sodium, a membrane capacitance contribution Cmdϕmdt
and a leakage
current JL. For this latter term, Hodgkin and Huxley do not specify what ion is carrying
it [HH52]. However, since the contribution of the leakage current is quite small, it may
be neglected or arbitrarily associated with both K+ and Cl−. gK , gNa and gL are the
constant specific conductances, while EK , ENa and EL are the reversal potentials of
each ion. The variables m, n and h are called gating variables, because they describe the
203. ODE-Based Modeling of Transmembrane Ion Flow in Cellular
Electrophysiology
opening state of the channel. These variables typically vary between 0 and 1, and each
of them are governed by the following ordinary differential equations at each point of the
membrane∂s
∂t= αs(ϕm)(1− s)− βs(ϕm)s, s = m,n, h, (3.10)
where αs and βs are experimentally determined functions. Hodgkin and Huxley [HH52]
used the following functions:
αm = B((25− ϕm)/10)
βm = 4 exp(−ϕm/18)
αh = 0.07 exp(−ϕm/20)
βh = C((30− ϕm)/10)
αn = 0.1B((10− ϕm)/10)
βn = 0.125 exp(−ϕm/80)
where ϕm is expressed in mV and C(t) := 1/(et + 1). A complete analysis of this ODE
system can be found in [KS98]. It is worth noting that accounting for each considered
type of ion channel would lead to a system of increased complexity compared to the linear
model (3.4) and the GHK model (3.8) in terms of the number of state variables. Since
each of these state variables is a local property of the membrane interface, the overall
complexity introduced by the HH model is significant.
3.5. Thermal equilibrium of a system of monovalent ions
In Sect. 2.2.2 we have addressed the important issue of thermal equilibrium for the
electrodiffusive flow of a single ionic species flowing across the cell membrane. In this
section we consider the more general case of a system of Mion monovalent ions (i.e.,
zi = ±1) that are moving in a biological fluid environment, in such a way that
Mion = M+ +M−
where M+ is the number of ions with zi = +1 and M− the number of ions with zi = −1.
To extend Defns. 2.2.1 and 2.2.2 to this (more realistic) situation we can profitably use
the GHK theory of Sect. 3.4.2 and write the following generalization of Def. 2.2.2 as
Definition 3.5.1 (Thermal equilibrium for a system of Mion ≥ 1 monovalent ions).
Thermal equilibrium ⇔ Jtot,Mion = 0 ⇔ ϕm = Ec,Mion (3.11a)
where
Jtot,Mion =
Mtot∑i=1
Ji (3.11b)
3.5. Thermal equilibrium of a system of monovalent ions 21
Ji being given by (3.8), and
Ec,Mion =RT
Fln
M+∑i=1
P+i c
(+,out)i +
M−∑i=1
P−i c(−,out)i
M+∑i=1
P+i c
(+,in)i +
M−∑i=1
P−i c(−,in)i
. (3.11c)
Formula (3.11c) is the so-called Goldman equation and the quantity P±i := D±i /l is the
membrane permeability with respect to the i-th ion, i = 1, . . . ,Mion. The quantity Ec,Mion
is the reversal (or equilibrium) potential of the whole system of ions and is also called
Goldman potential.
Remark 3.5.1. It is important to notice that Def. 3.5.1 does not require the single ion
current density to be equal to zero in thermal equilibrium, but only the (weaker) condition
that the total current density sums up to zero. It is clear that in the case where the cell is
much more permeable to a specific ion than to all the others (highly selective membrane),
the Goldman relation tends to the classical Nernst potential (2.5) and thermal equilibrium
of the whole system is regulated by the most permeant species.
Part II.
Multiscale Mathematical Models of
Transmembrane Ion Transport
4. PDE-Based Multiscale Modeling of Ion
Flow in Cellular Biology
In this chapter, we propose and investigate a hierarchy of mathematical formulations,
based on Partial Differential Equations (PDEs), for the description of ion flow in a cellular
system. In doing this, we extend to the multi-dimensional setting the basic Nernst-Planck
ion transport model (2.4). We start from a microscale view of the system, where the
membrane is described in detail as well as the extra-and intracellular sites. Then, we
perform an upscaling procedure that allows us to eliminate the geometrical description of
the membrane through the introduction of suitable transmission conditions that regulate
ion flow and charge accumulation at the two sides of the cellular environment. The
resulting formulation is the so-called macroscale model. This latter is characterized by
a considerable reduction of computational complexity compared to the microscale model
without a significant loss of bio-physical accuracy, and for this reason it is the basic
starting choice in numerical simulations.
4.1. Electrochemical description of ion flux
In this section, we give a brief summary of the basic laws describing the flow of a chemical
species in an ionic solution (denoted henceforth as the environment or medium). To this
purpose, let c = c(x, t) denote the concentration of a given ionized species having valence
z at point x and time t. We assume that the units of x, t and c are m, s and m−3,
respectively. We also denote by D (units: m2 s−1) and µ (units: V m−2 s−1) the diffusion
coefficient and mobility of ion c, respectively. Assuming the validity of Einstein’s relation
(cf. [KS98], Chpt. 2), we have
D =µVth|z|
(4.1)
where Vth := KBT/q = RT/F is the thermal voltage (units: V), T being the absolute
temperature of the environment, KB being the Boltzmann constant, q the electron charge,
R the gas constant (Rydberg constant) and F the Faraday constant, respectively.
4.1.1. The electric field
Charged ions in motion throughout the medium experience, at a point x and at time t,
an electric force proportional to the local value of the electric field E = E(x, t) (units:
V m−1). According to the quasi-static assumption in Electrodynamics [Jac99], which is
by far satisfied in the present context, the electric field can be expressed as the gradient
26 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
of the electric potential ϕ = ϕ(x, t) as
E = −∇ϕ. (4.2)
The electric field cannot, in general, be considered a given function, rather it is self-
consistently determined by the solution of the Poisson equation
divD = ρ (4.3)
where D is the electric displacement vector defined as
D = εE (4.4)
ε being the dielectric constant of the medium in which ion flow takes place, and ρ is the
electrical charge density at each point x of the environment and at each time t, defined
as
ρ = ρ0 + q
Mion∑i=1
zici (4.5)
where Mion is the number of ion species that are simultaneously flowing in the medium
and ρ0 is a given fixed charge density accounting, for instance, for the presence of im-
mobile charged ions localized throughout the environment. The units of D, ρ and ε are
C m−2, C m−3 and F m−1, respectively. In the remainder of these notes we take
ε = εw = ε0εwr = 80ε0 (4.6)
where ε0 is the dielectric constant of vacuum and εwr is the relative dielectric constant of
water. The above formula amounts to assuming that the liquid constituting the biological
environment in which ions flow (intracellular site, extracellular site, membrane) is made
entirely of water.
4.1.2. Ion mass flux density
Flux balance for the ion species c is expressed by the following PDE in conservative form
∂ c
∂ t+ divf = P (4.7)
where P is a production term and the flux of ion c is given by the following DD relation,
that is well known as the Nernst-Planck (NP) equation [Rub90, KS98, Hil01]
f = µz
|z|cE︸ ︷︷ ︸
drift flux
−D∇ c︸ ︷︷ ︸diffusion flux
. (4.8)
Notice that (4.8) is the multi-dimensional counterpart of (2.4) upon the use of Einstein
relation (4.1) and of the definition (4.2). Using (4.1) into (4.8), we see that the DD ion
flux can be written in the following equivalent purely advective form
f = cvc (4.9)
where vc is the advective (or drift) velocity of ion c, defined as
vc = −µ z
|z|∇ϕec. (4.10)
4.2. Microscale model of cellular ion flow 27
The scalar function ϕec is the electrochemical potential, defined as
ϕec = ϕ+Vthz
ln
(c
cref
)(4.11)
where cref is a reference concentration. The units of ϕec are V. From (4.9), we see that
the units of f are m−3 m s−1 = m−2 s−1, consistently with the fact that f is a number
flux density, that is, a time rate of number of ions per unit area.
4.1.3. Ion electrical current density
Since ions are electrically charged, it is natural to associate with the ion flux f the ion
current density
J := qzf . (4.12)
Again, we notice that (4.12) is the multi-dimensional counterpart of the first relation
in (2.4). The units of J are C m−2 s−1 = A m−2, consistently with the fact that J repre-
sents a time rate of charge per unit area. Replacing (4.9) and (4.10) into (4.12), we see
that the ion current density can be expressed in the form of generalized Ohm’s law
Jc = σEc (4.13)
where
σ := q |z|µ c, Ec := −∇ϕc (4.14)
represent the ionic conductivity and the electrochemical field acting on ion c, respectively.
The units of σ are S m−1. We notice also that if the ion concentration is constant, then
Ec = E and the ion current density reduces to the classical Ohm’s law
J = σE
expressing current flow in an electrolyte with electrical conductivity σ and subject to an
electric field E. However, if a concentration gradient if also present, then ion current
flow is driven by the electrochemical field Ec = −∇ϕc so that the generalized Ohm’s
law (4.13) must be used. The relative weight of diffusion and drift forces thus determines
the flow of ion c in the solution environment.
Having introduced the notion of flux current density, the balance equation (4.7) can
also be consistently modified into the following current continuity equation for the ionic
species c
q z∂ c
∂ t+ divJ = q z P. (4.15)
The generalized Ohm’s form (4.13) of ion current flux and the current continuity equa-
tion (4.15) will be useful for interpreting ion exchange between cell compartments in
terms of equivalent electrical circuit parameters, as discussed in Sects. 4.2 and 4.4.
4.2. Microscale model of cellular ion flow
In this section, we introduce the microscale model for electrochemical ion flow in a cellular
system. To this purpose, we consider the simplified geometrical setting shown in Fig. 4.1.
28 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
The computational domain Ω is the union of three open disjoint bounded subsets of Rd,d = 1, 2, 3, denoted Ω1, Ω2 and Ωm. The two subdomains Ω′k, k = 1, 2, represent two
cell compartments while Ωm is a thin membrane of thickness tm that separates Ω1 from
Ω2. The external boundary of Ω, denoted ∂Ω, is the union of the boundary portions
Γ′1, Γ′2 and Γm. The outward unit normal vector on ∂Ω is n. On the internal boundary
Γ1 = ∂Ω′1 ∩ ∂Ωm, the outward unit normal vector from Ω′1 into Ωm is n1 while the unit
outward normal vector from Ωm into Ω′1 is nm = −n1. Similarly, on Γ2 = ∂Ω′2 ∩ ∂Ωm,
the outward unit normal vectors are n2 (from Ω′2 into Ωm) and nm (from Ωm into Ω′2),
respectively.
The main feature of the membrane subdomain Ωm is that the thickness tm, according
to biophysical evidence, is much smaller than the characteristic size of the domain ` :=
diam(Ω). This feature will be exploited to construct a reduced-order model of ion flow
in the cellular system.
Figure 4.1.: Two cellular compartments Ω′1 and Ω′2 separated by a membrane Ωm. The
membrane thickness tm is much smaller than ` := diam(Ω).
Unless otherwise specified, we assume from now on that the production term P is
equal to zero and that d = 3 (three-dimensional case). We also assume that Mion ≥ 1
chemical species are present in the medium, each one being characterized by a concen-
tration ci and a valence zi, i = 1, . . . ,Mion. The flow of the medium is neglected in this
analysis, as the electrolyte fluid velocity with respect an inertial system is typically much
smaller compared to the velocity of each ion species. Should this assumption fail to hold,
the ion flow model would have to be completed by adding the Navier-Stokes system for
fluid velocity (see [Rub90]).
The microscale description of ion motion consists of the following nonlinearly coupled,
incompletely parabolic system of PDEs, for the dependent variables ci, i = 1, . . . ,Mion,
4.2. Microscale model of cellular ion flow 29
and ϕ, to be solved in the space-time cylinder QT := Ω× (0, Tend) [Rub90]:
q zi∂ ci∂ t
+ divJ i = 0 (4.16a)
J i = qµi|zi|ciE − qziDi∇ ci (4.16b)
divD = ρ0 +
Mion∑i=1
qzici (4.16c)
D = εE = −ε∇ϕ, (4.16d)
supplied with the following set of initial and boundary conditions:
ci(x, 0) = c0i (x) in Ω. (4.17a)
−κiJ i · n+ αici = βi on ∂Ω (4.17b)
−κϕD · n+ αϕϕ = βϕ on ∂Ω. (4.17c)
The equation system (4.16) is well known as the Poisson-Nernst-Planck (PNP) model for
ion flow in cellular biology [Rub90]. The diffusion coefficients Di, proportional to the ion
mobilities µi through Einstein’s relation (4.1), are piecewise constant and positive over
the domain Ω, not necessarily assuming the same value in each subdomain Ω1, Ω2 and Ωm.
The same assumption applies to the dielectric permittivity ε. The initial concentrations
c0i are nonnegative given functions belonging to L∞(Ω). The dimensionless coefficients
κi and κϕ may take only the values 0 (corresponding to a Robin boundary condition)
and 1 (corresponding to a Dirichlet boundary condition). The coefficients αi and βi are
nonnegative and belong to L∞(∂Ω), while αϕ and βϕ are given functions in L∞(∂Ω). The
units of αi and βi are Coul m s−1 and Coul C m s−1, respectively, while the units of αϕ
and βϕ are F m−2 and Coul m−2, respectively. It can be verified that any weak solution
of (4.16)- (4.17) satisfies the following transmission properties on each internal boundary
Γk, k = 1, 2:
[[ci]]Γk= 0 (4.18a)
[[J i]]Γk= 0 (4.18b)
[[ϕ]]Γk= 0 (4.18c)
[[D]]Γk= 0. (4.18d)
In (4.18), for any scalar function f : Ω→ R, we let
[[f ]]Γkm:= fknk + fmnm k = 1, 2 (4.19)
be the jump of f across the internal boundaries Γ1 and Γ2, fk and fm being the traces
of f along Γk from the subdomains Ωk and Ωm, respectively. In the case of a vector
function V : Ω→ Rd, we let
[[V ]]Γk:= V k · nk + V m · nm k = 1, 2, (4.20)
be the jump of V across Γ1 and Γ2. Notice that the jump of a scalar is a vector, while the
jump of a vector is a scalar. The transmission conditions (4.18) essentially state that the
concentrations and the related fluxes, the electric potential and the related displacement
vector are continuous across the internal boundaries of the cellular domain.
30 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
Remark 4.2.1 (Why microscale model?). The name “microscale model” is motivated
by the fact that system (4.16) is to be solved in the partitioned domain Ω = Ω′1 ∪ Ω′2 ∪Ωm where the membrane is distinguished from the cellular compartments. In view of a
numerical solution of the problem, this means that both the small-size membrane region
and the large-size cellular compartments must be geometrically discretized, with an obvious
increase of meshing effort and memory storage.
Remark 4.2.2 (The Poisson equation). Nonlinear coupling in the equation system (4.16)
is due to the presence of the drift term in the flux constitutive equations (4.16b). As a
matter of fact, at each time level the Poisson equation (4.16c) must be self-consistently
solved to update the electric field inside the domain. This update step represents an
elliptic constraint on the remaining parabolic part of the problem (the Mion continuity
equations for the species ci) and may significantly increase the computational effort of
the solution algorithm, especially in the case of sharp transients and/or long-time system
dynamics.
4.3. Multiscale reduction of cellular ion flow
The principal difficulty in the numerical solution of the microscale model (4.16) is the geo-
metrical discretization of the membrane region, which may in turn give rise to a huge num-
ber of degrees of freedom of the numerical method. To reduce computational complexity,
in this section we propose a two-level procedure that combines the micro-to-macro scale
transition method used in [dFPSV12] in the study of nanostructured organic solar cells
with the effective membrane model proposed and investigated in [Mor06, MJP07, MP09]
in the three-dimensional study of cellular electrical activity. These two approaches share
some similarities with those used in [MJR05] for porous media with thin fractures or
in [LS11] for reaction problems with sharp moving reaction fronts, and, more recently,
in [CSV12] for Tissue Engineering applications in regenerative medicine.
The proposed approach consists of a geometrical level and a modeling level. The
geometrical level is based on the following steps:
(G1) introduce the d − 1-dimensional manifold Γ corresponding to the middle cross-
section of the membrane volume Ωm;
(G2) partition the membrane Ωm into the union of the disjoint subregions Ωm1, Ωm2
and Γ, where Ωm1 and Ωm2 are the two open portions of Ωm in contact with Ω′1and Ω′2, respectively;
(G3) define the two “extended” subdomains Ω1 := Ω′1 ∪ Ωm1 and Ω2 := Ω′2 ∪ Ωm2, in
such a way that Ω = Ω1 ∪ Ω2 ∪ Γ.
The new geometrical partition of the cell structure is shown in Fig. 4.2.
The modeling level is based on the following steps:
(M1) replace equations (4.16) in the membrane subdomain Ωm with suitable transmis-
sion conditions across the d− 1-dimensional manifold Γ;
4.4. Macroscale model of ion flow 31
Figure 4.2.: Two cellular compartments Ω1 and Ω2 separated by an interface Γ with zero
thickness. The original boundaries Γ1 and Γ2 between the membrane and
the cell compartments are drawn in dashed lines, together with the thickness
of the membrane.
(M2) solve equations (4.16) into the new partitioned domain Ω \ Γ = Ω1 ∪ Ω2, subject
to the initial/boundary conditions (4.17) and to the interface conditions as in step
(M1).
Remark 4.3.1 (Intra- and extracellular sites). By convention, we assume throughout
that quantities labeled with the subscript ”1” refer to the intracellular site while those
labeled with subscript ”2” refer to the extracellular site. In such an event, we let nΓ := n1
to indicate a prescribed orientation for ion flow across the interface separating the two
biological environments. According to this choice, a current flowing from Ω1 into Ω2 is
positive otherwise it is negative.
4.4. Macroscale model of ion flow
In this section, we apply the general multiscale modeling reduction procedure illustrated
in Sect. 4.3 to the study of ion flow in a cellular system. This requires, in principle, to
characterize in detail the extracellular and intracellular signals that ultimately determine
the biophysical behaviour of the cellular membrane (see, for instance, [WT06, CL08a]).
As the complexity of such phenomena goes far beyond the scope of this article, here, we
merely focus on the basic electrochemical effects that characterize the dynamical response
of the membrane and that have been discussed in Chapt. 3. The membrane modeling
reduction methodology presented here is based on the approach proposed in [Mor06] to
which we refer for more details on the asymptotical analysis of the formulation.
4.4.1. Electrostatic model of the membrane
Let us consider the schematical picture of the cellular region near the interface shown in
Fig. 4.3.
The principal assumption of the modeling reduction of the membrane is that the electric
potential varies linearly inside Ωm. This assumption agrees with the fact that tm `
32 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
Figure 4.3.: Microscale model of the membrane: 1D cross-section of the cellular domain.
The electric potential ϕ is assumed to be a linear function of position in-
side the membrane. In the example depicted in the figure, the extracellular
potential is higher than the intracellular potential so that the membrane
potential ϕm is negative.
and, replaced into the transmission condition (4.18d) at Γ1, yields
D1 · n1
∣∣∣Γ1
= −Dm · nm∣∣∣Γ1
= −Dm · (−n1)∣∣∣Γ1
= −εm∇ϕ · n1
∣∣∣Γ1
= −εmϕ′2 − ϕ′1tm
≡ Cm(ϕ′1 − ϕ′2)
(4.21)
where ϕ′1 and ϕ′2 are the traces of ϕ at Γ1 and Γ2, respectively, while
cm :=εmtm
(4.22)
is the intrinsic membrane capacitance and has units Fm−2. In an analogous manner, the
transmission condition (4.18d) at Γ2 becomes
D2 · n2
∣∣∣Γ2
= −Dm · nm∣∣∣Γ2
= εm∇ϕ · nm∣∣∣Γ2
= εmϕ′2 − ϕ′1tm
≡ cm(ϕ′2 − ϕ′1). (4.23)
At this stage of the procedure, the two transmission conditions (4.21) and (4.23) are
still localized at the internal interfaces Γ1 and Γ2 that were introduced in the microscale
description of the cellular geometry. To construct the electrostatic reduced-order model
of the membrane, we perform steps (G1)-(G3), and enforce conditions (4.21) and (4.23)
directly at the interface Γ, by replacing the trace ϕ1 along Γ with the value ϕ′1 along Γ1,
and the trace ϕ2 along Γ with the value ϕ′2 along Γ2. This is equivalent to replacing the
linear variation of ϕ in the membrane with a jump ϕ′2−ϕ′1 and corresponds to performing
step (M1). This step provides the appropriate electrostatic transmission conditions to
be satisfied by the displacement vector in the electrostatic reduced-order model of the
membrane:
−D1 · n1 + cm(ϕ1 − ϕ2) = 0 on Γ (4.24a)
−D2 · n2 + cm(ϕ2 − ϕ1) = 0 on Γ (4.24b)
4.4. Macroscale model of ion flow 33
where the superscript (·)′ has been omitted for sake of clarity. The electrical equivalent
representation of the reduced model of the membrane is shown in Fig. 4.4.
ϕ2
ϕ1
Ω1 Ω2
_____
+++++
Γ
cm
E, D
Figure 4.4.: Macroscale equivalent electrostatic representation of the membrane. The two
sheets of surface charge (positive and negative) are clearly visible on the two
sides of the specific capacitance cm.
In conclusion, the function pair (ϕ, D) satisfies the following conditions at the inter-
face Γ in the macroscale model:
[[D]]Γ = 0 (4.25a)
[[ϕ]]Γ · n1 = ϕ1 − ϕ2 6= 0. (4.25b)
Relation (4.25a) is an immediate consequence of taking the sum of the two relations (4.24),
and shows that in the macroscale model the electric displacement vector D satisfies the
same compatibility condition as in the macroscale model (cf. (4.18d)). Relation (4.25b),
instead, shows that the a-posteriori effect of upscaling the membrane thickness is that
ϕ in the macroscale model is no longer continuous across Γ (as it was in the microscale
model across Γk, k = 1, 2), but it experiences a finite jump discontinuity whose strength
is equal to |ϕ1 − ϕ2|.
4.4.2. Electrodynamical model of the membrane
Let us consider the schematical picture of the membrane region shown in Fig. 4.5.
For each ionic species i = 1, . . . ,Mion, the following contributions to current flow across
the membrane can be distinguished:
• J1,i, J2,i: current densities flowing into the membrane region Ωm from subdomain
Ω′1 and Ω′2, resp.;
• Jm1,i, Jm2,i: transmembrane current densities from Ω′1 into Ω′2 and viceversa, resp.;
• Jσ1,i, Jσ2,i: surface currents on Γ1 and Γ2, respectively.
The transmembrane current densities Jm1,i and Jm2,i represent the biophysical fact that ion
charge flows throughout ion channels, transporters and pumps that are located within
the membrane region. A biophysical characterization of these currents is far from trivial,
and for a more detailed treatment of this issue we refer to [KS98, Hil01] and to [Mor06].
34 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
Γ1 Γ
2
nmn1
tm
nm n2
J1,i
J2,i
Jσ
1.i
mJ
1,i
Jm
2,i
Jσ
2,i
Figure 4.5.: Microscale model of the membrane: 3D view of the membrane and current
flow contributions. Each red and green arrow represents a current density
injected into the membrane from each respective compartment.
Examples of models for Jm1,i, Jm2,i that are practically usable in numerical computations
have been given in Sect. 3.2.
The transmembrane current densities satisfy the following conservation conditions
Jm1,i + Jm2,i = 0 i = 1, . . . ,Mion. (4.26)
As a consequence, also the total transmembrane current densities
Jm1 :=
Mion∑i=1
Jm1,i, Jm2 :=
Mion∑i=1
Jm2,i (4.27)
are such that
Jm1 + Jm2 = 0, (4.28)
so that the total ion flux traversing the membrane thickness through the aid of ion
channels, transporters and pumps is conserved.
The current densities Jσ1,i and Jσ2,i represent the biophysical fact that some of the ion
charge density qzici flowing across the membrane accumulates on the separating surfaces
Γ1 and Γ2. To define such currents, it is convenient to introduce first the total surface
charge densities on Γ1 and Γ2
σm1 := cm(ϕ′1 − ϕ′2), σm2 = −σm1 . (4.29)
Relations (4.29) state that the membrane behaves like a linear capacitor and that the
total charge accumulated on Γ1 is instantaneously counterbalanced by the total charge
accumulated on Γ2, Then, the fraction of the total surface charge densities associated
with ion ci over Γ1 and Γ2 is given by
σm1,i := λ′1,iσm1 , σm2,i := λ′2,iσ
m2 i = 1, . . . ,Mion, (4.30)
where
λ′1,i :=z2i (c′1,i)
2∑Mionj=1 z2
j (c′1,j)2, λ′2,i :=
z2i (c′2,i)
2∑Mionj=1 z2
j (c′2,j)2, (4.31)
4.4. Macroscale model of ion flow 35
(uk)′ denoting again the trace of a function u over Γk, k = 1, 2.
The dimensionless parameters λ′1,i ∈ (0, 1] and λ′2,i ∈ (0, 1] have the biophysical
meaning of membrane charge fraction of the i-th ionic charge density with respect to the
total membrane ion charge. By construction, these parameters satisfy the property of
partition of unityMion∑i=1
λ′1,i = 1,
Mion∑i=1
λ′2,i = 1. (4.32)
The flux transmission condition (4.18b) at Γ1, for i = 1, . . . ,Mion, yields
J1,i · n1 = J1,m · n1 := Jm1,i + Jσ1,i = Jm1,i +∂
∂t
(λ′1,icm(ϕ′1 − ϕ′2)
), (4.33)
while the flux transmission condition (4.18b) at Γ2, for i = 1, . . . ,Mion, becomes
J2,i · n2 = J2,m · n2 := Jm2,i + Jσ2,i = Jm2,i +∂
∂t
(λ′2,icm(ϕ′2 − ϕ′1)
). (4.34)
Let
J :=
Mion∑i=1
J i (4.35)
be the total ion flux, and denote by J1 and J2 the traces of J on Γ1 (from Ω′1) and on Γ2
(from Ω′2), respectively. Then, taking the sum of (4.33) and (4.34) over i = 1, . . . ,Mion,
summing the resulting currents, upon using (4.28) and (4.32), yields
J1 · n1 + J2 · n2 = 0 (4.36)
so that the total ion flux traversing the membrane is conserved.
To construct the electrodynamical reduced-order model of the membrane, we perform
steps (G1)-(G3), and enforce conditions (4.33) and (4.34) directly at the interface Γ, by
replacing the quantities ϕ′1, ϕ′2 with ϕ1, ϕ2, the quantities c′1,i, c′2,i with c1,i, c2,i, and the
quantities λ′1,i, λ′2,i with
λ1,i :=z2i c
21,i∑Mion
j=1 z2j c
21,j
, λ2,i :=z2i c
22,i∑Mion
j=1 z2j c
22,j
, (4.37)
respectively. Notice that the dimensionless parameters λ1,i and λ2,i still satisfy the parti-
tion of unity property (4.32). The above procedure corresponds to performing step (M1),
and provides the appropriate electrodynamical transmission conditions to be satisfied by
the current density of the i-th ion species in the reduced-order model, for i = 1, . . . ,Mion:
−J1,i · n1 + Jm1,i +∂
∂t(λ1,icm(ϕ1 − ϕ2)) = 0 on Γ
−J2,i · n2 + Jm2,i +∂
∂t(λ2,icm(ϕ2 − ϕ1)) = 0 on Γ.
(4.38)
In conclusion, the function pairs (ci, J i), i = 1, . . . ,Mion, satisfy the following condi-
tions at the interface Γ in the macroscale model:
[[J i]]Γ = cm∂
∂t((λ1,i − λ2,i)(ϕ1 − ϕ2)) 6= 0 (4.39a)
[[ci]]Γ · n1 = c1,i − c2,i 6= 0. (4.39b)
36 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
ϕ1,c
1ϕ2,c
2
Ω1 Ω2
n1
λ1cm
J1
J2
Γ
gm
Figure 4.6.: Equivalent electrical representation of cellular ion flux and membrane. λ1cm
is the intrinsic (nonlinear) membrane capacitance (in F m−2). gm is the
intrinsic (generally nonlinear) membrane conductance (in Ω−1 m−2) associ-
ated with transmembrane current densities. In the example represented in
the figure, J1 and J2 are ion current densities flowing across the membrane
into the intracellular compartment so that the corresponding currents are
negative according to Remark 4.3.1.
Relation (4.39b) is the analogue of (4.25b), and shows that the a-posteriori effect of up-
scaling the membrane thickness is that ci in the macroscale model is no longer continuous
across Γ (as it was in the microscale model across Γk, k = 1, 2), but it experiences a finite
jump discontinuity whose strength is equal to |c1,i − c2,i|.
Consequently, relation (4.39a), that is the result of taking the sum of the two rela-
tions (4.38), shows that in the macroscale model the current flux of the i-th ion expe-
riences a jump in crossing the membrane. This jump is determined by the unbalanced
charge accumulation associated with the flow of ion charge qzici occurring at the two
sides of the interface Γ. Such unbalanced accumulation of charge actually cancels out
when summing over all ionic species flowing across the membrane. As a matter of fact,
taking the sum over i = 1, . . . ,Mion in (4.39a), upon using (4.28) and (4.32), we obtain
[[J ]]Γ = 0 (4.40)
so that the total ion flux traversing the membrane is conserved, in agreement with (4.36).
The electrical equivalent representation of the electrodynamical behaviour of the mem-
brane in the macroscale model is shown in Fig. 4.6.
4.5. The PNP system
Collecting the results of Sects. 4.4.1 and 4.4.2, the macroscale description of ion motion
consists of the following PNP system to be solved in the space-time cylinder QT :=
4.6. The Cable Equation model 37
(Ω \ Γ)× (0, Tend):
q zi∂ ci∂ t
+ divJ i = 0 (4.41a)
J i = qµi|zi|ciE − qziDi∇ ci (4.41b)
divD = ρ0 +
Mion∑i=1
qzici (4.41c)
D = εE = −ε∇ϕ, (4.41d)
supplied with the following set of initial and boundary conditions:
ci(x, 0) = c0i (x) in Ω. (4.42a)
−κiJ i · n+ αici = βi on ∂Ω (4.42b)
−κϕD · n+ αϕϕ = βϕ on ∂Ω, (4.42c)
and with the following set of interface conditions:
−D1 · n1 + cm(ϕ1 − ϕ2) = 0 on Γ (4.43a)
−D2 · n2 + cm(ϕ2 − ϕ1) = 0 on Γ (4.43b)
−J1,i · n1 + Jm1,i +∂
∂t(λ1,icm(ϕ1 − ϕ2)) = 0 on Γ (4.43c)
−J2,i · n2 + Jm2,i +∂
∂t(λ2,icm(ϕ2 − ϕ1)) = 0 on Γ (4.43d)
where the dimensionless parameters λk,i are defined in (4.37).
Remark 4.5.1 (Why macroscale model?). In a similar manner as in Rem. 4.2.1, sys-
tem (4.41) is called “macroscale model” because it does no longer require the detailed
geometrical representation of the volumetric membrane region and the corresponding so-
lution of the equation system (4.16) into that region, but only to characterize the func-
tional behaviour of the membrane through an equivalent model in terms of the trans-
membrane fluxes jmk,i and of the transmembrane intrinsic capacitances λk,icm, k = 1, 2
and i = 1, . . . ,Mion. Both these two sets of parameters are nonlinear functions of the
interface potentials ϕ1, ϕ2 and of the ion concentrations c1,i, c2,i, i = 1, . . . ,Mion. This
makes the communication between the two cellular compartments Ω1 and Ω2 a strongly
nonlinear interface coupling process, that requires the adoption of appropriate iterative
solution maps. For this latter issue, we refer to [MJP07] and [Bre09], and, in these
notes, to Chapt. 6.
4.6. The Cable Equation model
In this section, we consistently derive, starting from the PNP model introduced in
Sect. 4.5, a simplified description of intra-extracellular ion flow that is well-known as
the Cable Model. This approach is a PDE-based formulation that is characterized by a
significantly reduced computational effort than the PNP model and turns out to be quite
appropriate for studying passive electrical flow in neuronal networks. For a biologically
complete treatment of this subject, we refer to [KS98], Chapt. 8 and to [Hil01].
38 4. PDE-Based Multiscale Modeling of Ion Flow in Cellular Biology
The principal assumption in the cable equation (CE) model is that the ionic concen-
trations in the intra- and extracellular sites are constant in both space and time. As
a consequence, from the constitutive relation (4.41b) it turns out that current flow is
purely ohmic, i.e., we have
J i = qµi|zi|ciE i = 1, . . . ,Mion, (4.44a)
where ci, i = 1, . . . ,Mion are the values of ion concentrations equal to c(in)i and c
(out)i
in the intra- and extracellular sites, respectively. Let us introduce the total electrical
conductivity of the solution (units: S m−1)
σtot :=
Mion∑i=1
qµi|zi|ci. (4.44b)
Then, summing (4.41a) over i = 1, . . . ,Mion and using (4.44b), we obtain the following
CE system of model equations for the electric potential ϕ to be solved in the space-time
cylinder QT :
divJ tot = 0 (4.45a)
J tot = σtotE = −σtot∇ϕ (4.45b)
supplied with the following set of initial and boundary conditions:
ϕ(x, 0) = ϕ0(x) in Ω. (4.46a)
−κϕε
σtotJ tot · n+ αϕϕ = βϕ on ∂Ω, (4.46b)
and with the following set of interface conditions:
−J tot,1 · n1 + Jmtot,1 + cm∂
∂t(ϕ1 − ϕ2) = 0 on Γ (4.47a)
−J tot,2 · n2 + Jmtot,2 + cm∂
∂t(ϕ2 − ϕ1) = 0 on Γ. (4.47b)
The initial datum ϕ0 is the solution of the linear Poisson equation (4.41c) at t = 0, while
Jmtot,1 and Jmtot,2 are the total transmembrane current densities injected from side 1 and
side 2 of the membrane, respectively.
Remark 4.6.1 (Time dependence in the CE model). The current continuity equa-
tion (4.45a) is an elliptic constraint to be satisfied at each time level t > 0. Time
evolution of the Cable Equation model is governed by the interface conditions at the
membrane (4.47). Integrating these latter equations across the membrane surface, from
each cellular compartment and using the convention on positive currents introduced in
Remark 4.3.1, we immediately get the following Kirchhoff current laws (KCLs):
Itot,1 = Imtot,1 + Cm∂ϕm∂t
from intracellular site (4.47c)
−Itot,2 = −Imtot,2 − Cm∂ϕm∂t
into extracellular site (4.47d)
4.6. The Cable Equation model 39
where, as usual, ϕm = ϕ1−ϕ2 denotes the membrane potential. Summing the two KCLs,
we get
(Itot,1 − Itot,2) =(Imtot,1 − Imtot,2
). (4.47e)
This latter equation has a clear physical meaning: the net (total) electrodiffusive current
crossing the membrane interface is exactly equal to the net (total) transmembrane current.
In the event where no charge is accumulated at the two membrane sites or trapped inside
the lipid layer then we can conclude that the right-hand side of (4.47e) is equal to zero
and the (total) electrodiffusive current is conserved.
For further details on the CE model, its applications in cellular electrophysiology and
its comparison with the full PNP formulation, we refer to [QS89] and to [Mor06].
5. Mathematical Analysis of the
Poisson-Nernst-Planck Model
In the present chapter we focus our attention on the Poisson-Nernst-Planck (PNP) model
for ion electrodiffusion introduced in Chapter 4. In particular, we consider in detail some
mathematical issues that play a significant role in view of i) the analysis of the properties
of the solutions of the model equations and ii) of their numerical approximation. With
this purpose, we first proceed with the adimensionalization of the PNP system; then,
based on the singularly perturbed character of the Poisson equation, we derive a further
electroneutral (EN) reduction of the system, denoted EN-PNP model, that represents a
very good choice for rapid and sufficiently accurate simulations of ion transport problems.
For both PNP and EN-PNP models in adimensional form, we provide suitable sets of
boundary, interface and initial conditions.
5.1. Summary of model equations
For convenience of exposition, we write below the Poisson-Nernst-Planck equation system
for ion electrodiffusion:
q zi∂ci∂t
+ divJi = 0 i = 1, . . . ,Mion (5.1a)
divD = ρ0 + q
Mion∑i=1
zi ci (5.1b)
Ji = −qziDi
(ziVth
ci∇ϕ+ ∇ci
)i = 1, . . . ,Mion (5.1c)
D = εwE = −εw∇ϕ. (5.1d)
To derive relation (5.1c), we use (4.2) and Einstein’s relation (4.1) into Eq. (4.41b) and
obtain
J i = −qµi|zi|ci∇ϕ− qziDi∇ ci
= −qziDi
(µi|zi|Dizi
ci∇ϕ+ ∇ ci
)= −qziDi
(ziVth
ci∇ϕ+ ∇ ci
).
Boundary, interface and initial conditions for (5.1) have been specified in (4.42) and (4.43).
42 5. Mathematical Analysis of the Poisson-Nernst-Planck Model
5.1.1. Scaling of the PNP system
A closed form solution for system (5.1) is, in general, impossible to obtain and an ap-
proximate solution is therefore required. The first step towards a numerically stable
approximation consists of reformulating the system in order to obtain a scaled set of
equations, where variables are adimensional and normalized. This operation, called scal-
ing, is very useful in view of numerical computations, since the variables of the problem
may have physical values of very different magnitudes. The scaling procedure leads to a
set of PDEs where dimensioneless variables have comparable orders of magnitude.
For any quantity w we set
w = w · w (5.2)
where w is the scaled (dimensionless) variable and w is the corresponding scaling factor.
We introduce the following independent scaling factors: x, c, ϕ and µ for x, ci, ϕ and
µi, and we set:
x = diam(Ω) (5.3a)
c = maxi=1,...,Mion
c
(in)i,eq , c
(out)i,eq
(5.3b)
ϕ = Vth (5.3c)
D = maxi=1,...,Mion
Di (5.3d)
where c(in)i,eq and c
(out)i,eq are the intra- and extracellular values of the ion concentrations
in thermal equilibrium conditions. Characteristic values of the scaling constants are
reported in Tab. 5.1.
Scaling factor Value Units
x 10−6 m
c 150 ·NAV = 150 · 6.023 · 1023 #ions m−3
D 1.87 · 10−9 m2V−1s−1
ϕ = Vth = Kb T / q ∼= 27 · 10−3 V
t = x/ v 5.67 · 10−6 s
λ2 :=εw ϕ
q x2 c2 · 10−4 adimensional
J = q c v 1.7 · 106 Am−2
v = D / x 1.87 · 10−3 ms−1
E = ϕ / x ∼= 2.5 · 104 Vm−1
Table 5.1.: Scaling factors and relevant parameters.
We start by scaling Eq. (5.1a) and obtain
q zi∂(c ci)
∂(t t)+
1
xdiv J i J = 0,
5.1. Summary of model equations 43
where t and J are the scaling factors (yet unspecifed) for the time variable t and the
current density J i, respectively. Defining
J :=q c x
t≡ q c v
where v has the units of a velocity ms−1, we obtain the following scaled form of the
continuity equation
zi∂ ci
∂ t+ div J i = 0 i = 1, . . . ,Mion. (5.3e)
We fix the scaling factor J by using the notion of drift velocity (see (4.10))
v = µE =µϕ
x
so that
J = q c v = q cµ x
ϕ(5.3f)
and, as a consequence, we get
t =x2
µϕ. (5.3g)
The current density reads
J i = −qziDic
x
(zici∇ ϕ+ ∇ ci
).
Introducing the scaling factor D for the diffusivity Di, we get
J i = −qDcxziDi
(zici∇ ϕ+ ∇ ci
):= J J i
from which, using (5.3f) and (5.3g), the scaling factor D turns out to be given by
D =x2
t= µϕ. (5.3h)
and the scaled form of the current density finally reads
J i = −qziDi
(zici∇ ϕ+ ∇ ci
)i = 1, . . . ,Mion. (5.3i)
The scaling of Eq. (4.41c) gives
−εwx2 div
(∇(ϕ ϕ)
)= q(ρ0 c) + q
Mion∑i=1
zi(ci c),
from which, upon introducing the parameter
λ2 :=εw ϕ
q x2 c, (5.3j)
we obtain the following scaled form of Poisson’s equation
− λ2 div ∇ϕ = ρ0 +
Mion∑i=1
zici. (5.3k)
44 5. Mathematical Analysis of the Poisson-Nernst-Planck Model
Remark 5.1.1 (The scaled Debye length). The parameter λ is the so-called scaled Debye
length and is defined as
λ =
(εw ϕ
q c
)1/2 1
x≡ λD
x,
where λD ≈ 1nm is the standard definition of the Debye length. This latter quantity gives
a measure of the free mean path that a charged ion can travel inside the electrolyte fluid
before experiencing the screening effect of the electric field produced by the presence of the
surrounding charged ions. It is relevant to observe that if λ2 << 1, then the PNP system
exhibits a singularly perturbed character (see [Mar86] and, for a general presentation,
[RST96]), and the corresponding solutions may exhibit internal and/or boundary layers.
In the applications considered in these notes, a typical value is λ2 = O(10−4), so that
we can conclude that the PNP system is markedly singularly perturbed. This feature of
the problem we are dealing with has several important consequences, both in terms of
mathematical model reduction and of numerical discretization. These critical issues will
be addressed in the remainder of this chapter and in Chapt. 8.
5.2. The scaled PNP model
Continuing to denote each scaled variable with the same symbol as in the unscaled case,
the PNP model in scaled form reads:
zi∂ci∂t
+ divJi = 0 i = 1, . . . ,Mion (5.4a)
λ2divE = ρ0 +
Mion∑i=1
zi ci (5.4b)
Ji = −ziDi (zici∇ϕ+ ∇ci) i = 1, . . . ,Mion (5.4c)
E = −∇ϕ. (5.4d)
Applying the same scaling procedure also to (4.42) and (4.43), we end up with the
following set of initial and boundary conditions:
ci(x, 0) = c0i (x) in Ω. (5.5a)
−κiJ i · n+ αici = βi on ∂Ω (5.5b)
−κϕE · n+ αϕϕ = βϕ on ∂Ω, (5.5c)
and with the following set of interface conditions:
−E1 · n1 + cm(ϕ1 − ϕ2) = 0 on Γ (5.6a)
−E2 · n2 + cm(ϕ2 − ϕ1) = 0 on Γ (5.6b)
−J1,i · n1 + Jm1,i + λ2 ∂
∂t(λ1,icm(ϕ1 − ϕ2)) = 0 on Γ (5.6c)
−J2,i · n2 + Jm2,i + λ2 ∂
∂t(λ2,icm(ϕ2 − ϕ1)) = 0 on Γ (5.6d)
where the dimensionless parameters λk,i are defined in (4.37).
Remark 5.2.1 (Analysis of the PNP system). The analysis of the PNP system (5.4) in
steady-state conditions and in the one-dimensional case can be found in [PJ97, BCEJ97],
5.3. The electroneutral PNP model (EN-PNP) 45
while an extensive discussion of the mathematical properties of solutions of the model
(with special emphasis on their asymptotical behaviour with respect to small parameters)
can be found in [Rub90]. Well-posedness and global estimates in time of the nonlinearly
time-dependent coupled system constituted by the PNP and Navier-Stokes equations have
been thoroughly investigated in [Jer02, JS09] and in [Sch09].
5.3. The electroneutral PNP model (EN-PNP)
The very small value of the singular perturbation parameter λ2 = O(10−8) suggests that
the Poisson equation (5.4b) can be substituted, without appreciable loss of accuracy, by
the following relation
ρ(x, t) = ρ0(x) +
Mion∑i=1
zi ci(x, t) = 0 ∀x ∈ Ω, ∀t ∈ (0, Tend). (5.7)
Relation (5.7) expresses the physical property that the fluid solution containing ions in
motion and fixed charges is electroneutral. For this reason, the reduced model emanating
from the PNP formulation under the previous simplification is known as the Electroneu-
tral PNP system (EN-PNP), and reads:
zi∂ci∂t
+ divJi = 0 i = 1, . . . ,Mion (5.8a)
ρ0 +
Mion∑i=1
zi ci = 0 (5.8b)
Ji = −ziDi (zici∇ϕ+ ∇ci) i = 1, . . . ,Mion (5.8c)
E = −∇ϕ (5.8d)
with the following set of initial and boundary conditions:
ci(x, 0) = c0i (x) in Ω. (5.9a)
−κiJ i · n+ αici = βi on ∂Ω (5.9b)
−κϕE · n+ αϕϕ = βϕ on ∂Ω, (5.9c)
and with the following set of interface conditions:
−E1 · n1 + cm(ϕ1 − ϕ2) = 0 on Γ (5.10a)
−E2 · n2 + cm(ϕ2 − ϕ1) = 0 on Γ (5.10b)
−J1,i · n1 + Jm1,i + λ2 ∂
∂t(λ1,icm(ϕ1 − ϕ2)) = 0 on Γ (5.10c)
−J2,i · n2 + Jm2,i + λ2 ∂
∂t(λ2,icm(ϕ2 − ϕ1)) = 0 on Γ. (5.10d)
The EN-PNP model illustrated above was presented and mathematically investigated
in [MJP07] and recently extended in [MP09]. Compared with the PNP system, the EN-
PNP formulation has the mathematical advantage to neglect a-priori the occurrence of
interior layers in the electric potential spatial distribution at the interface Γ. Such layers
would require, as a matter of fact, heavy mesh refinement in the vicinity of Γ in order to
capture the discontinuity of the potential ϕ2 − ϕ1. Instead, using the model (5.8) such
46 5. Mathematical Analysis of the Poisson-Nernst-Planck Model
a problem is removed and the associated computational effort is considerably reduced.
Numerical examples demonstrating the quality of the EN-PNP model compared with
that of the full PNP approach can be found in [Mor06]. Applications of the EN-PNP
system will be considered in Part IV of these notes.
Part III.
Functional Techniques and
Discretization Methods
6. Solution Map for the PNP system
In this chapter, we describe in detail the solution map to treat the PNP model equation
system that has been illustrated in Chapter 5. Similar approaches can be used, with
proper modifications to handle the electroneutrality condition (5.7), for the EN-PNP
model. Time advancing is carried out using the Backward Euler method. Then, a
functional iteration approach that is widely used in the decoupled solution of the Drift-
Diffusion semiconductor device equations, well-known as Gummel map, is applied to
the case of the PNP model. Special care is spent in the linearization of the membrane
transmission conditions within the modular structure of the fixed-point algorithm. It is
shown that such a treatment leads to solving a sequence of linear advection-diffusion-
reaction boundary-value problems in a multi-domain fashion.
6.1. Introduction
The system of PDEs, with the associated initial and boundary conditions, illustrated in
Chap. 5 are in general impossible to solve in closed form, hence an approximate numerical
solution must be found. Toward this end, the simultaneous occurrence of the following
issues must be taken care of:
• time dependence adds a dimension to the solution space and complicates the treat-
ment of the equations;
• non-linearity forces the use of an iterative procedure, where a sequence of linearized
problems must be solved until convergence;
• the interface embodied by the membrane has to be adequately treated by the nu-
merical method.
The solution maps, based on functional iterations, that are used to specifically deal
with the PNP and the EN-PNP models share some common approaches, which will be
discussed in the next subsections. Before entering into the details of the algorithms, let
us shortly address the general treatment of each of the issues listed above.
Time dependence is managed for both the PNP and EN-PNP systems and boundary
and interface conditions by introducing a simple temporal semi-discretization. Consider-
ing the problem having a time-span of [0, Tend], Tend being the final time, we indicate with
tm = m∆t the m-th time level, where m = 0, . . . ,MT − 1, MT ≥ 1, and ∆t = Tend/MT ,
and we use a Backward-Euler (BE) method to approximate all time derivatives in such
a way that
∂ci∂t
(x, tm) 'cm+1i (x)− cmi (x)
∆tm = 0, . . . ,MT − 1, i = 1, . . . ,Mion
50 6. Solution Map for the PNP system
where (·)m indicates that the quantity (·) is evaluated at the time level tm. It is well known
that the BE method is unconditionally stable, is easy to implement and introduces a time
discretization error of order ∆t, which is of the same order as the spatial discretization
error introduced by the finite element scheme discussed in Chapt. 8. At each time level
tm a non-linear system of equations must now be solved.
The non-linearity of the system of PDEs (5.4) is due to the coupling between potential
ϕ and concentrations ci through the drift term in the ionic flux constitutive equation
in (5.4c) and (5.8c)
−z2i Di ci∇ϕ = |zi|µi ciE, i = 1, . . . ,Mion
where µi is the scaled mobility associated to the i-th ionic species.
The nonlinear nature of the equations describing the ionic membrane currents, as
shown in Sect.3.4, is also a problem to be dealt with. Functional iterations provide
a well known approach to translating the original nonlinear system into a sequence of
linear problems [OR70], the solution of which should converge to a corresponding, non-
necessarily unique solution of the original problem. The most relevant example of func-
tional iteration is the so-called fixed-point method, an instance of which is the ubiquitous
Newton method [OR70, Jer96]. While this method has the very appealing property of
being quadratically convergent, some essential drawbacks must be pointed out:
• a “good” initial guess must be provided in order to fully enjoy the property of
quadratic convergence, and in some cases to even simply assure that convergence
can be reached;
• the algebraic system associated with the discrete solution of the linearized problem
may be very large in size, as the system must be solved for all variables (potential
and the Mion ions) simultaneously. This large size often proves to be computa-
tionally intensive, as well as causing the associated coefficient matrix to be often
ill-conditioned.
As illustrated in Sect.6.2, the PNP model will be solved with a staggered algorithm,
where the equations defining the potential and each ion will be solved separately, instead
of using a monolithic algorithm as the Newton method. This approach simplifies the
treatment for the following reasons:
• the decoupling of the potential from the concentration variables renders the conti-
nuity equation for the concentrations linear, hence easily solvable;
• the size of the corresponding algebric system is reduced, hopefully decreasing the
global time required to reach a solution, as well as improving the numerical condi-
tioning of the coefficient matrix.
6.2. PNP solution map
As noted in Chapt. 5, system (5.4) is formally identical to the well-known drift-diffusion
(DD) model for semiconductor devices [Jer96]. This similarity can be profitably exploited
6.2. PNP solution map 51
for both mathematical and computational purposes. Precisely, it is convenient to use the
definition of electrochemical potential given in (4.11) to express the ion concentration ci
(in scaled form) as a nonlinear function of the electric and electrochemical potentials as
ci = exp (zi(ϕec,i − ϕ)) i = 1, . . . ,Mion (6.1)
where ϕec,i is the (scaled) electrochemical potential associated with the ionic concentra-
tion ci. The use of (6.1) into (5.4b) has the effect of transforming the linear Poisson
equation into a nonlinear (more properly, semilinear) equation. This approach is the
same as done in the iterative solution of the DD semiconductor model with the so-called
Gummel map, originally introduced in [Gum64] in one spatial dimension and thoroughly
analyzed in multiple spatial dimensions [Jer96]. Thus, because of the mathematical equiv-
alence between the PNP and DD models, Gummel’s iteration appears to be a natural
choice also for the treatment of the PNP system.
The Gummel map is a decoupled algorithm where each dependent variable of the
problem and its corresponding equation are treated separately, as shown in the flow
chart of Fig.6.1. At each time level tm, the iterations start with a guess for the electric
and electro-chemical potential, ϕ(0) and ϕ(0)eci (or equivalently, with a concentration guess
c(0)i ). By indicating with (k) the generic k-th iteration, a single step of the Gummel
process consists of:
• the solution of a nonlinear Poisson equation (NLP), in order to obtain an updated
potential ϕ(k+1);
• the solution of a linear continuity equation for each electro-chemical potential
ϕ(k+1)eci , i = 1, . . . ,Mion, given the known updated potential ϕ(k+1);
• a check of convergence of the iteration, carried out by controlling whether the
maximum absolute difference between two consecutive iterates (k) and (k+1) is less
than a prescibed tolerance, for each i = 1, . . . ,Mion:
‖ϕ(k+1) − ϕ(k)‖L∞(Ω) < εϕ, ‖ϕ(k+1)eci − ϕ(k)
eci‖L∞(Ω) < εϕeci,
where, for any measurable function f , we set (cf. [QV97], Chapt. 1)
‖f‖L∞(Ω) := infM ≥ 0 | |f(x)| ≤ M almost everywhere in Ω.
A complete analysis of the convergence of the Gummel map is carried out in [Jer96]
in the case Mion = 2. The main result is that as k →∞, the map converges to a unique
solution ϕ∗ and c∗i , i = 1, 2, provided that suitable constraints are enforced on boundary
data and problem coefficients. Minor modifications are expected to extend the previous
result to the case of the PNP model (5.4). For more details, see [Rub90].
6.2.1. The Nonlinear Poisson equation
Each step of the Gummel Map for the PNP system requires solving a NLP equation for
the updated potential ϕ(k+1). The electrochemical potentials ϕ(k)ec,i are given functions and
remain unchanged in the solution process of the NLP equation. Dropping the superscript
52 6. Solution Map for the PNP system
Figure 6.1.: PNP solution map.
(k) for notation simplicity and using (6.1), we can write the NLP as the problem of finding
the zero of the abstract operator
F (ϕ) = −λ2div∇ϕ− ρ0 −Mion∑i=1
zi exp(zi(ϕeci − ϕ)), (6.2)
i.e., we solve the nonlinear PDE
F (ϕ) = 0 (6.3)
supplied by suitable boundary and interface conditions. A solution of (6.3) can be ob-
tained by applying the damped Newton method [OR70, Sel84]. This latter method is a
fixed-point iteration that amounts to solving at each step j, j ≥ 0, until convergence the
following homogeneous linear boundary value problem for the Newton update U (j):
F ′(ϕ(j))U (j) = −F (ϕ(j)) in Ω \ Γ (6.4a)
− κϕE(U (j)) · n+ αϕU(j) = 0 on ∂Ω (6.4b)
−E1(U(j)1 ) · n1 + cm(U
(j)1 − U (j)
2 ) = 0 on Γ (6.4c)
−E2(U(j)2 ) · n2 + cm(U
(j)2 − U (j)
1 ) = 0 on Γ (6.4d)
where
F ′(ϕ(j))U (j) = −λ2div∇U (j) +
[Mion∑i=1
z2i exp(zi(ϕeci − ϕ(j)))
]U (j). (6.4e)
6.2. PNP solution map 53
Once problem (6.4) is solved, the new iterate is found by the correction step
ϕ(j+1) = ϕ(j) + τ (j) U (j), (6.4f)
where τ (j) ∈ (0, 1] is a damping parameter to be properly chosen.
Remark 6.2.1 (Boundary conditions). The Newton increment U (j) satisfies the homo-
geneous boundary conditions (6.4b) in such a way that the updated potential ϕ(j+1) satis-
fies, because of (6.4f), the corresponding non-homogeneous conditions (5.5c) (assuming,
of course, that this also holds for ϕ(j) for all j ≥ 0).
Remark 6.2.2 (Frechet derivative). Relation (6.4e) expresses the action of the Frechet
derivative F ′(ϕ), evaluated at ϕ(j), on the increment funtion U (j).
Remark 6.2.3 (Damping). The damping parameter τ (j) is chosen in order to enforce a
norm-reduced residual at each iteration (see [Sel84], Chapt. 7), namely
‖F (ϕ(j+1))‖L∞(Ω) < ‖F (ϕ(j))‖L∞(Ω). (6.4g)
The value of τ (j) is found by a simple iteration loop to equation (6.4f), in which τ (j) is
initially set equal to one and then reduced (e.g., divided by two) until condition (6.4g) is
satisfied.
6.2.2. The continuity equations
Once the updated electric potential ϕ(k+1) has been computed by iteratively solving the
NLP equation as illustrated in Sect. 6.2.1, the next step of the Gummel map consists of the
successive solution of Mion linear continuity equations supplied, however, by (possibly)
nonlinear boundary and interface conditions.
Setting for brevity Ci := c(k+1)i (x) (unknown functions), V := ϕ(k+1)(x) (known
function) and Q := λ2cm(V1 − V2) (known function), the boundary value problems to be
solved to determine the updated ion concentrations Ci (and consequently, the updated
electrochemical potentials ϕ(k+1)ec,i ) for each i = 1, . . . ,Mion, read as follows:
ziCi∆t
+ divJi(Ci;V ) = zicmi∆t
in Ω \ Γ (6.5a)
Ji(Ci;V ) = −ziDi (ziCi∇V + ∇Ci) (6.5b)
− κiJ i(Ci;V ) · n+ αiCi = βi on ∂Ω (6.5c)
− J1,i(Ci;V ) · n1 + J1,mi(Ci;V ) + λ1,iQm+1 −Qm
∆t= 0 on Γ (6.5d)
− J2,i(Ci;V ) · n2 + J2,mi(Ci;V )− λ2,iQm+1 −Qm
∆t= 0 on Γ. (6.5e)
The differential equation (6.5a) and boundary conditions (6.5c) are of linear type. A
possible non-linearity is hidden, instead, in the interface conditions (6.5d)-(6.5e) because
of the definition of the ionic membrane currents Jmi as a function of Ci. Referring
to Sect.3.4, we see that:
54 6. Solution Map for the PNP system
• the linear resistor model (3.4), reported below for clarity
Jmi = gi
(Vin − Vout −
1
ziln
(C
(out)i
C(in)i
)),
is of nonlinear type with respect to the dependent variable Ci and therefore requires
an iterative treatment for solution;
• the GHK current equation (3.8) is of linear type with respect to the dependent
variable Ci, hence requires no further processing;
• the HH ionic current equations (3.9)2−4 are similar to that of the linear resistor
model so that they require an iterative treatment for solution.
In view of the numerical implementation of a nonlinear interface condition, a further
iteration scheme has to be devised. The solution of the continuity equation for each
i = 1, . . . ,Mion is then divided into a sequence of (p) steps, each one consisting in:
• linearization of the membrane ionic current Jmi evaluated in correspondance of a
known concentration C(p)i . Indicating with an overline the values of the concentra-
tion C(p)i and with no overline the concentrations for the next iteration (p+1), this
linearization can be expressed as
Jmi |ci ' gi
ϕm − 1
ziln
C(out)i
C(in)i
+ gi
C(in)i
C(in)i
−C
(out)i
C(out)i
.
• solution of the continuity equation for each ion, to obtain the updated concentra-
tions C(p+1)i ;
• convergence check between two consecutive iterates (p) and (p+1)
‖C(p+1)i − C(p)
i ‖L∞(Ω) < εci i = 1, . . . ,Mion.
Remark 6.2.4 (Convergence of inner iteration loop for continuity equations). While
no general mathematical proof is given regarding the convergence of this algorithm, ex-
tensive computational tests show that a fast convergence is obtained in every performed
simulation.
Remark 6.2.5 (Linearized interface conditions). Naming u the dependent variable and
J its corresponding flux, the general form of a linear interface condition can be expressed
as: J · n1 = αu1 − β u2 + σ1
J · n2 = β u2 − αu1 − σ2
(6.6)
where u1 and u2 denote the traces of u on either sides of the membrane, n1 = −n2 is
the outward unit normal to the membrane pointing from side (1) (intracellular, symbol:
(·)(in)) to side (2) (extracellular, symbol: (·)(out)) of the membrane, while α, β, σ1 and σ2
are given coefficients. We see that the linearized forms of the interface conditions (6.5d)
and (6.5e) fit the general form (6.6) by setting:
α =gi
C(in)i
, β =gi
C(out)i
, σ1 = −σ2 = gi
ϕm − 1
ziln
C(out)i
C(in)i
.
7. Unified Framework and Well-Posedness
Analysis
In this chapter, we construct a unified PDE model that includes as special cases all the
linearized differential subproblems that we need to solve at each step of the Gummel
iterative procedure illustrated in Chapt. 6. The model is a linear diffusion-advection-
reaction (DAR) PDE to be solved in a bounded open domain Ω with Lipschitz boundary
Γ := ∂Ω, composed by the union of two open subdomains, Ωi, i = 1, 2, and by one
interface Γm that separates them and that plays the role of the membrane between the
intra- and extracellular environments (see Fig. 7.1).
Here and in the next chapter, we assume Ω to be a bounded open domain in R2 despite
many of the ideas, results and methodologies can be extended to the three dimensional
case. The model problem is complemented by mixed type boundary conditions on the
external boundary Γ and by transmission conditions of the form (6.6) on Γm. Then, upon
specifying suitable functional spaces for the model coefficients and the solution, a weak
formulation of the DAR problem is obtained and well posedness of the corresponding
weak solution is found in a suitable subset of H1(Ω). The numerical discretization of the
DAR equation is the object of Chapt. 8.
7.1. Unified framework for the PNP solution map
All the equations that must be solved in the iteration steps described in Sect.6.2 can be
cast into the Diffusion-Advection-Reaction (DAR) model problem, with gradient advec-
tive field, of finding u : Ω = (Ω1 ∪ Ω2)→ R such that:
divJ(u) + c u = f in Ω (7.1a)
J(u) = −D (∇u+ cϕ u∇ϕ) in Ω (7.1b)
u = uD on ΓD (7.1c)
J(u) · n = γ u+ jR on ΓR (7.1d)
J(u) · n1 = αu1 − β u2 + σ1 on Γm,1 (7.1e)
J(u) · n2 = β u2 − αu1 − σ2 on Γm,2. (7.1f)
In view of the multi-domain formulation of problem (7.1), we denote by Γi the boundaries
of Ωi, i = 1, 2, in such a way that ΓD,i = Γi ∩ ΓD, ΓR,i = Γi ∩ Γi and Γm,i = Γi ∩ Γm
for i = 1, 2. Consistently, as represented in Fig. 7.1, on the external portions of the
subdomain boundaries Γi∩Γ, the outward unit normal vector ni coincides with n, while
on the interface Γm we have two distinct outward unit normal vectors n1 and n2, such
that n2(x) = −n1(x) for all x ∈ Γm.
56 7. Unified Framework and Well-Posedness Analysis
Figure 7.1.: Computational domain for the DAR model problem.
The subscripts in (7.1e)- (7.1f) indicate the restrictions on Γm of functions and quan-
tities defined in the subdomains Ω1 and Ω2, respectively. This means, in particular, that
the function u is not in general continuous across the interface Γm.
The partition of Γ into two mutually disjoint subsets is in accordance with the nu-
merical examples illustrated in Part IV of these notes. In particular, on ΓD the Dirichlet
boundary condition (7.1c) is enforced, on ΓR the Robin boundary condition (7.1d) is en-
forced while on the internal interface Γm the transmission conditions (7.1e) and (7.1f) are
enforced. Clearly, conditions (7.1c)-(7.1f) are special cases of the conditions (6.4b)- (6.4d)
(for the NLP equation) and of (6.5c)- (6.5e) (for the linearized continuity equations).
The diffusion coefficient D is a given function in L∞(Ω) such that
0 < Dmin ≤ D(x) ≤ Dmax a.e. in Ω.
The reaction coefficient c is a given function in L∞(Ω) such that
c(x) > 0 a.e. in Ω.
The potential ϕ is a given piecewise smooth function on Ω, such that ϕi ∈ (H1(Ωi) ∩L∞(Ωi)) for i = 1, 2, with a possible jump discontinuity at Γm. We denote by ϕmin and
ϕmax the mininum and maximum of ϕ over Ω. The quantity cϕ is a constant parameter
over Ω (equal to 0 in the case of the linearization of the NLP equation and equal to zi
in the case of the linearization of the continuity equation for the i-th ionic species). The
production term f is a given function in L2(Ω).
Boundary and interface data uD, γ, jR, α, β, σ1 and σ2 are given functions with
suitable regularity on their respective domains of definition, such that
α(x) > 0, β(x) > 0 a.e. on Γm
and
γ(x) ≥ 0 a.e. on ΓR.
For the definition and properties of the Hilbert function spaces used in the present
chapter, see [QV97] and the references cited therein.
7.2. Weak formulation and well posedness 57
7.2. Weak formulation and well posedness
In this section we construct the weak formulation of the DAR model system (7.1) and
prove that it admits a unique solution depending with continuity on problem data. To
ease the presentation, and without any loss of generality, we assume uD = 0 (to avoid
introducing lifting of boundary data). We also set cϕ = +1 as is the case with the
continuity equation (6.5a) for the sodium ion Na+. The construction consists of two
preliminary steps: 1) introduction of an appropriate functional setting; 2) reformulation
of (7.1) as an equivalent boundary value problem of diffusion-reaction type through the
application of a suitable change of the dependent variable.
7.2.1. Multi-domain functional setting
Since the function u is in general discontinuous across the interface Γm we can not
take H1(Ω) (more precisely, a proper subspace accounting for homogeneous boundary
conditions on ΓD) as ambient space for the weak formulation of the DAR equation because
this choice would contradict Prop. 3.2.1 of [QV97]. Therefore, we adopt a more flexible
approach suggested in a natural manner by the multi-domain structure of the problem
at hand. We start introducing the spaces
Vi =v ∈ H1(Ωi) | v = 0 on ΓD,i i = 1, 2
. (7.2a)
The spaces Vi are Hilbert spaces endowed with the norm
‖v‖Vi = ‖∇v‖L2(Ωi) ∀v ∈ Vi (7.2b)
because of Poincare’s inequality. Then, associated with the spaces Vi, we introduce the
product space
V = V1 × V2 (7.2c)
that is a Hilbert function space endowed with the graph norm
‖q‖V =(‖q1‖2V1 + ‖q2‖2V2
)1/2 ∀q ∈ V (7.2d)
where q1 and q2 are the restrictions to Ω1 and Ω2 of the generic function q : Ω → R,
respectively.
The space V is the appropriate ambient space of the weak formulation of the DAR
boundary value problem (7.1) as we are going to see in the remainder of this section.
7.2.2. DAR problem reformulation
The presence of the drift term
−Du∇ϕ
in the flux density J(u) (7.1b) prevents us from a straightforward application of the
Lax-Milgram Lemma because of the difficulty of proving coerciveness of the bilinear
form emanating from the weak formulation of (7.1). To exit this deadlock we exploit
relation (6.1) and set
u = U exp (−ϕ) (7.3)
58 7. Unified Framework and Well-Posedness Analysis
where U := exp (ϕec). Replacing (7.3) into system (7.1) we can rewrite this latter in the
equivalent form of finding U : Ω = (Ω1 ∪ Ω2)→ R such that:
divJ(U) + cU e−ϕ = f in Ω (7.4a)
J(u) = −De−ϕ∇U in Ω (7.4b)
U = UD on ΓD (7.4c)
J(U) · n = γ U e−ϕ + jR on ΓR (7.4d)
J(U) · n1 = αU1 e−ϕ1 − β U2 e
−ϕ2 + σ1 on Γm,1 (7.4e)
J(U) · n2 = β U2 e−ϕ2 − αU1 e
−ϕ1 − σ2 on Γm,2. (7.4f)
Clearly, problems (7.1) and (7.4) are completely equivalent so that, should the solution
of (7.4) exist and be unique the same property would immediately hold for that of (7.1)
in virtue of the change of variable (7.3).
Remark 7.2.1 (The exponential change of variable). The change of variable (7.3) is
known as Cole-Hopf transformation [Hop50] and strongly relies on the assumptions that
a) Einstein’s relation holds; and b) the drift term in the transport model is a gradient
field. The same formula was also proposed by Jan Slotboom in [Slo73] and from that time
on it is referred to as Slotboom change of variable. The exponential transformation is
widely used in all analytical and numerical treatments of the DD semiconductor model
(see [Mar86, MRS90, Jer96, Sel84]).
Comparing the two equation systems, we see that the application of the Slotboom
transformation has had the following consequences:
1. the DAR problem has become a DR problem;
2. the drift term is disappeared;
3. the diffusion coeddicient D is now modified in De−ϕ;
4. the same modifications apply to the reaction term c, to the Robin coefficient γ and
to the interface coefficients α and β;
5. all of the above modifications share the property that originally bounded coefficients
remain bounded after the application of (7.3) because ϕ is essentially bounded from
above and from below. This avoids blow-up of the solution of (7.4).
7.2.3. Weak formulation
From now on in this chapter, we denote by v = (v1, v2) any function of V whose restric-
tions on Ω1 and Ω2 are v1 and v2, respectively. Then we proceed in the standard manner,
that is, we multiply (7.4a) by v, integrate over Ω and integrate by parts the terms∫Ωi
vi divJ(Ui) dΩi
7.2. Weak formulation and well posedness 59
to obtain the following integral identity∫Γm,1
v1 J(U1) · n1 ds+
∫Γm,2
v2 J(U2) · n2 ds−∫Ω
J(U) ·∇v dΩ
+
∫Ω
cU e−ϕ v dΩ =
∫Ω
f v dΩ ∀v ∈ V.
Using the fact that vi = 0 in the sense of traces on ΓD,i, the boundary condition (7.4d)
and the interface conditions (7.4e)- (7.4f), the above identity translates into the weak
formulation of problem (7.4):
find U ∈ V such that
B(U, v) = F (v) ∀v ∈ V (7.5a)
where:
B(U, v) =
∫Ω
De−ϕ∇U ·∇v dΩ +
∫Ω
cU e−ϕ v dΩ +
∫ΓR
γ U e−ϕ v ds
+
∫Γm,1
(αU1e−ϕ1 − βU2e
−ϕ2) v1 ds+
∫Γm,2
(βU2e−ϕ2 − αU1e
−ϕ1) v2 ds U, v ∈ V
(7.5b)
F (v) =
∫Ω
f v dΩ−∫
ΓR
jR v ds+
∫Γm,2
σ2 v2 ds−∫
Γm,1
σ2 v2 ds v ∈ V.
(7.5c)
7.2.4. Well posedness analysis and stability estimate
To the sole purpose of simplifying the estimates, we assume that the coefficients α and
β, γ are positive constant quantities, with A := max α, β. We also denote by CP the
largest between the two Poincare’s constants of Ω1 and Ω2 and by C∗ the largest between
the two trace constants of Ω1 and Ω2.
To verify that problem (7.5) admits a unique solution in the distributional sense, we
apply the Lax-Milgram Lemma (see, e.g., [QV97], Chapt. 5), and try to prove that there
exist three positive constants M, K and Λ such that:
|B(U, v)| ≤ M‖U‖V ‖v‖V ∀U, v ∈ V (7.6a)
B(U,U) ≥ K‖U‖2V ∀U ∈ V (7.6b)
|F (v)| ≤ Λ‖v‖V ∀v ∈ V. (7.6c)
Continuity of B Let us check condition (7.6a) that expresses the continuity of B(·, ·).Using the upper bound for e−ϕ and Cauchy-Schwarz, Poincare and trace inequalities we
get (7.6a) with
M = e−ϕmin[Dmax + 2C2
P ‖c‖L∞(Ω) + (2‖γ‖L∞(ΓR) + 4A)(C∗)2].
60 7. Unified Framework and Well-Posedness Analysis
Coercivity of B Let us check condition (7.6b) that expresses the coercivity of B(·, ·).Using the lower bound for e−ϕ we get
B(U,U) ≥ e−ϕmaxDmin‖U‖2V +
∫Γm
(αU2
1 e−ϕ1 + βU2
2 e−ϕ2 − (αe−ϕ1 + βe−ϕ2)U1U2
)ds.
We need to investigate the sign of the term in the integral at the right-hand side. Set
A := αe−ϕ1 , B := βe−ϕ2 , X :=√AU1 and Y :=
√BU2. Then, requiring the integrand
to be a non-negative quantity amounts to solving the following inequality
X 2 + Y2 −
(√A
B+
√B
A
)XY ≥ 0 ∀X ,Y
A sufficient condition for the above condition to hold is clearly
A = B (7.6d)
because in such a case the integrand becomes
X 2 + Y2 − 2XY = (X − Y)2 ≥ 0 ∀X ,Y.
Condition (7.6d) amounts to requiring that
αe−ϕ1 = βe−ϕ2 . (7.6e)
Remark 7.2.2. It can be checked, using (3.7d), that condition (7.6e) is automatically
satisfied if the GHK model (3.8) is used for the transmembrane current.
Under the (sufficient) condition (7.6d), we immediately obtain (7.6b) with
K = Dmine−ϕmax .
Continuity of F Let us check condition (7.6c) that expresses the continuity of F (·).Using Cauchy-Schwarz, Poincare and trace inequalities we get (7.6c) with
Λ = 2CP ‖f‖L2(Ω) + 2C∗‖jR‖L2(ΓR) + 2C∗Σ
where
Σ := max‖σ1‖L2(Γm), ‖σ2‖L2(Γm)
.
The above analysis yields the following concluding result.
Theorem 7.2.1 (Well posedness and stability). Assume that (7.6e) holds. Then, prob-
lem (7.5) admits a unique solution U ∈ V such that
‖U‖V ≤Λ
K≤ 2
eϕmax
Dmin
[CP ‖f‖L2(Ω) + C∗‖jR‖L2(ΓR) + C∗Σ
]. (7.6f)
As a consequence, in virtue of the equivalence between problem (7.4) and problem (7.1),
also this latter problem admits a unique weak solution u ∈ V .
8. Finite Element Approximation of the
DAR Problem
In this chapter, we discuss the numerical approximation of the linear advection-diffusion-
reaction model problem introduced in Chapt. 7 using a dual mixed–hybridized finite el-
ement (volume) method (DMH-FV) with numerical quadrature of the mass flux matrix.
The resulting method is a conservative finite volume scheme over triangular grids, for
which a discrete maximum principle is proved under the assumption that the mesh is of
Delaunay type in the interior of the domain and of weakly acute type along the domain ex-
ternal boundary and internal interface. Compared to a standard displacement-based ap-
proach, the proposed DMH-FV method has the advantage of being flux-conservative (as
standard FV schemes) and self-equilibrated (as standard mixed methods). Moreover, the
fact that the novel DMH-FV formulation incorporates the classical Scharfetter-Gummel
(SG) exponentially fitted scheme allows to compute sharp fronts without spurious oscilla-
tions even in the presence of dominating convective phenomena. The stability, accuracy
and robustness of the proposed method are validated in Chapt. 9 on several numerical
examples motivated by applications in Biology, Electrophysiology and Neuroelectronics.
8.1. Motivation to the use of a DMH method
For convenience of the presentation, we write below the Diffusion-Advection-Reaction
(DAR) model problem (7.1) with gradient advective field introduced and analyzed in
Chapt. 7, assuming cϕ = +1 in (7.1b). Find u : Ω = (Ω1 ∪ Ω2)→ R such that:
divJ(u) + c u = f in Ω (8.1a)
J(u) = −D (∇u+ u∇ϕ) in Ω (8.1b)
u = uD on ΓD (8.1c)
J(u) · n = γ u+ jR on ΓR (8.1d)
J(u) · n1 = αu1 − β u2 + σ1 on Γm,1 (8.1e)
J(u) · n2 = β u2 − αu1 − σ2 on Γm,2. (8.1f)
The geometrical description of the problem and the assumptions on model coefficients
and boundary data have been already discussed in detail in Chapt. 7 where the well
posedness of the DAR system has been proved upon introducing the exponential change
of variable (7.3). This approach will be also exploited in the numerical approximation
of (8.1) in Sect. 8.3.
The model problem (8.1) is representative of several important applications, rang-
ing from electrokinetic flows in nanofluidics [Rub90, KBA05] to cell biology [KS98,
62 8. Finite Element Approximation of the DAR Problem
FMWT02]. A common feature of these applications is the presence of active inter-
faces (membranes) whose selective behavior controls mass transport from a subdomain
to the neighbouring one according to the difference between the values of the electro-
static potential ϕ across the membrane. In this chapter, ϕ is assumed to be a given
function, but in realistic situations, for example in the study of current flux across ionic
channels using the so-called Poisson-Nernst-Planck (PNP) model [Rub90], the potential
is itself an unknown of the problem and is dynamically determined by the solution of
Gauss’ law in differential form in the domain Ω as described in detail in Chapt. 6 and
in [Jer96, Jer02, MJP07, MJC+07, JCLS08, JS09].
Equation (8.1a) is a conservation law expressing the balance between the flux of the
advective–diffusive vector field J(u) across an arbitrary control volume B ⊆ Ω and the
production term f − c u within the volume itself. In particular, the jump of the nor-
mal component of J(u) is equal to zero across each segment belonging to the interior
of Ω1 and Ω2, respectively, while it is equal to σ1 − σ2 across the membrane Γm, as it
can be checked by summing (8.1e) and (8.1f). Using the terminology of Computational
Mechanics, where u has the meaning of displacement and J(u) is the stress field, it is
well-known that standard displacement-based finite element methods for the numerical
approximation of (8.1) generally fail at satisfying the above properties, despite the op-
timal convergence of the approximate solution uh to u in the H1-norm (see [QV97]).
An effective alternative is represented by dual mixed (DM) methods, where two inde-
pendent discrete solutions uh and Jh are simultaneously sought for, leading to a linear
system in saddle-point form. DM methods satisfy both local self-equilibrium and con-
servation, and optimal error estimates hold for the pair (uh, Jh) in appropriate function
space norms (see [BF91]). However, there are several drawbacks that make them not
so amenable to realistic computations, namely, the increased computational cost, the
indefinite algebraic character of the system, and a possible failure at satisfying the dis-
crete maximum principle (DMP) for uh in the case of a nonvanishing reaction term c
(see [BMM+05] and references cited therein). A considerable improvement consists of
resorting to the hybridization of the DM formulation (see, [AB85] and [BF91], Chapt.
V; for more recent development in the framework of Discontinuous Gakerkin methods,
see also [CDG+09]). The hybridization procedure is based on the introduction of a
Lagrange multiplier denoted by λh (hybrid variable), which is an approximation of u
along mesh edges and allows one to enforce the interelement continuity of the normal
component of Jh. The local elimination of the variables uh and Jh as functions of λh
(static condensation) leads to a dual mixed–hybridized (DMH) finite element scheme of
displacement–based type, acting on the sole λh, which is completely equivalent to the
original DM approximation but at a much reduced computational effort. Moreover, it
can be shown that the hybrid variable enjoys superconvergence properties. However,
the question of ensuring a numerically stable computed solution in the presence of dom-
inating convection and/or reaction terms still remains an open issue, and appropriate
stabilization techniques must be used (see [BMM+05, BMM+06] and the more recent
work [CDG+09]). To this end, we propose in this chapter a finite volume variation of
the standard DMH method, denoted DMH-FV method, based on the introduction of a
quadrature formula for the diagonalization of the local flux mass matrix. This approach,
8.2. Geometric Discretization and Finite Element Spaces 63
that extends to the heterogeneous transport problem (8.1) previously introduced MFV
formulations [ABMO95, BMO96, MSS01, BMM+05], has three important advantages.
The first is that the resulting numerical scheme has a simple and very compact finite
volume structure where for each element of the grid, the computational stencil consists
of the element itself and, at most, its three neighbours. The second is that the treatment
of the exponentially varying diffusion coefficient across inter-element edges allows, under
mild geometric conditions, a DMP for the computed discrete solution. The third is that
the novel method enjoys the same convergence properties as the standard DMH scheme,
including superconvergence in the L2-norm of the post-processed solution obtained from
λh (see [AB85, dFS11]).
A brief outline of the chapter is as follows. In Sect. 8.2, we introduce the geometric
entities and finite element spaces; in Sect. 8.3, we use the change of variable (7.3) to
write problem (8.1) in symmetric form. Then, we describe the DMH-FV method, while
in Sect. 8.4 we illustrate the computer implementation of the scheme and related post-
processing.
For a thorough validation of the numerical performance of the DMH-FV scheme, we
refer to Chapt. 9.
8.2. Geometric Discretization and Finite Element Spaces
Let Th be a regular family of given partitions of the domain Ω into open triangles K
satisfying the usual admissibility condition (see [QV97], Sect. 3.1 and Def. 3.4.1). For a
given Th, we denote by NT and Ne the total number of triangles and edges, respectively,
by |K| and hK the area and the diameter of K, respectively, and we set h = maxTh hK .
Let x = (x, y)T be the position vector in Ω; then, for each K ∈ Th, we denote (see
Fig.8.1) by xq, q = 1, 2, 3, the three vertices of K ordered according to a counterclockwise
orientation, by eq the edge of ∂K which is opposite to xq, by θKq the angle opposite to
eq and by CK the circumcenter of K. We denote by |eq| the length of eq and by nq
the outward unit normal vector along eq. Moreover, we define sKq as the signed distance
between CK and the midpoint Mq of eq. If θKq < π/2 then sKq > 0, while if K is
Figure 8.1.: Geometrical notation on triangle K.
obtuse in θKq then sKq < 0, and CK falls outside K. Notice also that if θKq = π/2 then
64 8. Finite Element Approximation of the DAR Problem
sKq = 0, and CK coincides with Mq. We denote by Eh the set of edges of Th, and by
Eh,int and Eh,Γ those belonging to the interior of Ω and to the boundary Γ, respectively.
For each e ∈ Eh,int, we indicate by K1e and K2
e the pair of elements of Th such that
e = ∂K1e ∩ ∂K2
e . Finally, we let se = sK1
ee + s
K2e
e denote the signed distance between CK1e
and CK2e
(see Fig.8.2). If θK1
ee + θ
K2e
e < π for all e ∈ Eh,int, then se > 0, and Th is called
Figure 8.2.: Two neighbouring elements.
a Delaunay triangulation [Del34]. If the inequality is replaced by an equality, for some
e ∈ Eh,int, we call Th a degenerate Delaunay triangulation. For such an edge, se = 0 and
the two circumcenters CK1e, CK2
ecollapse into the midpoint of e. The Delaunay condition
prevents the occurrence of pairs of obtuse neighbouring elements in Th, still allowing the
possibility of having single obtuse triangles in the computational grid (see [FG08] for
algorithmic details). From now on, we assume that Th is a Delaunay triangulation.
For k ≥ 0 and a given set S, we denote by Pk(S) the space of polynomials of degree
≤ k defined over S. We also denote by RT0(K) := (P0(K))2 ⊕ P0(K)x the Raviart–
Thomas (RT) finite element space of lowest degree [RT77], and by P0 the L2-projection
over constant functions. Then, for g ∈ L2(ΓD), we introduce the following finite element
spaces:
Vh := v ∈ (L2(Ω))2 |v|K ∈ RT0(K) ∀K ∈ Th
Wh := w ∈ L2(Ω) |w|K ∈ P0(K) ∀K ∈ Th
Mh,g := m ∈ L2(Eh) |m|∂K ∈ R0(∂K)∀K ∈ Th,
mK1e |e = mK2
e |e ∀e ∈ Eh,int, m|e = P0g|e,∀e ∈ ΓD,
(8.2)
whereR0(∂K) := v ∈ L2(∂K)|v|e ∈ P0(e)∀e ∈ ∂K, and mK1e , mK2
e are the restrictions
of the generic function m ∈ Mh,g on K1e and K2
e , respectively. For each K ∈ Th, the
basis functions of RT0(K) are τ j(x) = (x − xj)/(2|K|), j = 1, 2, 3, and are such that
div τ j = 1/|K| and τ j ·nj = 1/|ej | for each ej ∈ ∂K, which implies that∫ejτ i · nj dς =
δij , i, j = 1, 2, 3, δij being the Kronecker symbol. Functions belonging to Mh,g are single-
valued on Eh,int∪ΓD∪ΓR, while they admit two distinct values on each edge e ∈ Γm. This
8.2. Geometric Discretization and Finite Element Spaces 65
latter, special, situation reproduces, on the discrete level, the selectivity characteristic of
the membrane Γm, and allows accounting for the occurrence of finite jump discontinuities
across Γm. The local degrees of freedom associated with the finite element spaces (8.2)
are depicted in Fig.8.3. It is also useful to introduce the following global finite element
Figure 8.3.: Degrees of freedom of the finite element spaces (8.2) over the element K.
The black circle is associated with Wh, the arrows with Vh and the black
squares with Mh,g.
space
Λh = vh ∈ L2(Ω) | vh ∈ P1(K) ∀K ∈ Th,
vh(MK1
ee ) = vh(M
K2e
e ) ∀e ∈ Eh,int = span ωee∈Eh ,
where the basis functions ωe are the non-conforming elements of Crouzeix-Raviart [CR73]
(see Fig.8.4). Functions in Λh are piecewise linear over Th, continuous at the midpoint
Figure 8.4.: Basis function ωe. The support of the function is the neighbouring pair K1e ,
K2e .
of each edge e ∈ Eh,int and possibly admitting a finite jump discontinuity at each edge
e ∈ Γm.
66 8. Finite Element Approximation of the DAR Problem
8.3. A Mixed–Hybridized Method with Numerical Quadrature
With the aim of constructing a finite element approximation of the DAR model problem
(8.1), we use the change of dependent variable (7.3) is such a way that the original
advection–diffusion–reaction system (8.1) is transformed into the equivalent problem of
finding the solution ρ : Ω → R of the following linear diffusion-reaction model problem
in conservative form:
Lρ = div J(ρ) + c ρ e−ϕ = f in Ω (8.3a)
J(ρ) = −De−ϕ∇ρ in Ω (8.3b)
ρ = ρD on ΓD (8.3c)
J(ρ) · n = γ ρ e−ϕ + jR on ΓR (8.3d)
J(ρ) · n1 = αρ1 e−ϕ1 − β ρ2 e
−ϕ2 + σ1 on Γm,1 (8.3e)
J(ρ) · n2 = β ρ2 e−ϕ2 − αρ1 e
−ϕ1 − σ2 on Γm,2, (8.3f)
where ρD := uD eϕD and ϕD := ϕ|ΓD
. Comparing (8.3) with (8.1), we see that the
use of relation (7.3) has transformed the original DAR problem into a new equivalent
diffusion-reaction problem with an exponentially varying diffusion coefficient De−ϕ and
a new dependent variable ρ. From now on, we assume that D, c and f are piecewise
constant given functions over Th, and that ρD, γ and jR are piecewise constant boundary
data over Eh,Γ, with the same assumption for the transmission coefficients α, β, σ1 and
σ2, and that ϕ ∈ Λh. Moreover, given a function η, we denote by ηK and ηe the constant
values of η over each element K ∈ Th and each edge e ∈ Eh, respectively. Finally, we set
for brevity a := De−ϕ and A := a−1.
The DMH Galerkin approximation of problem (8.3) consists of finding (Jh, ρh, λh) ∈(Vh ×Wh ×Mh,ρD) such that:
(AJh, τ h)Th − (ρh, div τ h)Th + 〈λh, τ h · n〉Eh = 0 ∀τ h ∈ Vh (8.4a)
(divJh + c e−ϕ ρh, qh)Th = (f, qh)Th ∀qh ∈Wh (8.4b)
〈Jh · n, µh〉Eh = 〈γ λh e−ϕ, µh〉ΓR+ 〈jR, µh〉ΓR
+ 〈αλh e−ϕ1 , µh〉Γm,1 − 〈β λh e−ϕ2 , µh〉Γm,1
+ 〈σ1, µh〉Γm,1 + 〈β λh e−ϕ2 , µh〉Γm,2
− 〈αλh e−ϕ1 , µh〉Γm,2 − 〈σ2, µh〉Γm,1 ∀µh ∈Mh,0, (8.4c)
where (·, ·)Th and 〈·, ·〉S denote the elementwise L2 inner products over Th and over any
subset S ⊆ Eh, respectively. The equations in (8.4) have the following interpretation:
(8.4a) expresses the approximate local constitutive law; (8.4b) expresses the approxi-
mate local balance between net flux across K and net production of mass inside K;
(8.4c) expresses the approximate continuity of J · n across each interelement edge, the
Robin boundary condition and the interface transmission condition. The approximate
interelement continuity of ρ and the Dirichlet boundary condition are automatically ex-
pressed by the fact that λh is a single–valued function over Eh,int ∪ ΓD ∪ ΓR. Using the
static condensation procedure allows one to eliminate uh and Jh in favor of the sole
8.3. A Mixed–Hybridized Method with Numerical Quadrature 67
hybrid variable λh and leads to solving a linear algebraic system whose size is of the or-
der of Ne, which makes the DMH formulation a generalized displacement-based method.
Once λh is available, the variables uh and Jh can be recovered by post-processing over
each mesh element. The DMH formulation was originally proposed and theoretically
analyzed in [AB85] in the study of an elliptic model problem with Dirichlet boundary
conditions. Further analysis and extensions can be found in [BF91, RT79, RT91]. Re-
lated approaches in the framework of Discontinuous Galerkin methods have been recently
proposed and analyzed, in the general case of higher order polynomials, in the series of
papers [CG04, CG05, CDG08, CGL09, CGW09, CDG+09].
To construct a DMH scheme with reduced computational effort, we proceed as follows.
For each K ∈ Th, we set
JKh (x) =3∑j=1
ΦKj τ j(x) x ∈ K, (8.5)
where the degree of freedom ΦKj =
∫ej
JKh ·njdζ is the flux of JKh across edge ej , j = 1, 2, 3.
Then, we consider the following quadrature formula∫K
AJKh · τ idK =
3∑j=1
ΦKj
∫K
A τ j · τ idK
' 1
2ΦKj A
Ki cot(θKi )δij = ΦK
j AKi
sKi|ei|
δij i, j = 1, 2, 3,
(8.6)
where AKi :=
∫Mi
CKAK(ζ)dζ/|sKi |. Using the fact that ϕ ∈ P1(K), we have
AKi =
∫Mi
CKD−1(ζ)eϕ(ζ)dζ
|sKi |=
1
DiK
eϕi
Be (ϕK − ϕi)(8.7)
where Be(t) := t/(et−1) is the inverse of the Bernoulli function, and DiK
is the constant
value of the diffusion coefficient along the segment CKMi defined as:
DiK
=
DK if sKi ≥ 0
DKi if sKi < 0.(8.8)
The above definition is consistent with physical intuition, because in the case where
sKi < 0 (i.e., θKi > π/2) the path of the integral in (8.7) lies completely in Ki, so that the
diffusion coefficient that must be used to compute the average AKi is that associated with
triangle Ki (opposite to K with respect to edge ei). Using (8.8) makes the average AKi
always a strictly positive quantity. Moreover, it can be shown that the diagonalization
formula (8.6) is affected by the following quadrature error∣∣∣∫K
A τ j · τ idK −AKi
sKi|ei|
δij
∣∣∣ ≤ ChK‖τ i‖H(div ;K)‖τ j‖H(div ;K), (8.9)
where H(div ;K) = τ ∈ (L2(K))2 |div τ ∈ L2(K)K ∈ Th and C is a positive constant
depending on A and on the mesh regularity (see [MSS01, BMM+05, dFS11] for a proof).
Using (8.6) into (8.4a), we obtain the following discrete equations for the DMH
method with diagonalized local mass flux matrix.
68 8. Finite Element Approximation of the DAR Problem
• Equation (8.4a):
AKi ΦK
i
sKi|ei|− ρK + λKi = 0 ∀K ∈ Th i = 1, 2, 3. (8.10)
• Equation (8.4b):
3∑i=1
ΦKi + cKe−ϕ
KρK |K| = fK |K| ∀K ∈ Th. (8.11)
• Equation (8.4c):
ΦK1e
e =
−ΦK2
ee e ∈ Eh,int
(γe λe e−ϕe + jRe) |e| e ∈ ΓR
(αeλe,1 e−ϕe,1 − βeλe,2 e−ϕe,2 + σe,1) |e| e ∈ Γm,1
(βeλe,2 e−ϕe,2 − αeλe,1 e−ϕe,1 − σe,2) |e| e ∈ Γm,2.
(8.12)
Equation (8.11) is already in genuine FV form, so that, to construct a finite volume
approximation starting from system (8.10)–(8.12), we need to express the flux ΦKi as a
function of ρK and ρKi , for each K ∈ Th and i = 1, 2, 3, proceeding as follows.
(Step 1). Consider equation (8.10) and assume that θKi 6= π/2. Then, for each K ∈ Thwe obtain the explicit relation
ΦKi = −(A
Ki )−1 λ
Ki − ρK
sKi|ei| i = 1, 2, 3. (8.13)
In the special case where θKi = π/2, then sKi = 0 and equation (8.10) yields
ρK = λKi irrespective of the (undetermined) value of ΦKi . Such a value can be
recovered by post-processing the computed solution ρh by using (8.10) as
ΦKi = (fK − cKe−ϕK
ρK) |K| −3∑
j=1,j 6=iΦKj . (8.14)
(Step 2). For each e ∈ Eh,int we replace (8.13) into (8.12)1, obtaining the explicit relation
λe =(A
K1e
i sK1
ei )−1ρK
1e + (A
K2e
i sK2
ei )−1ρK
2e
(AK1
ei s
K1e
i )−1 + (AK2
ei s
K2e
i )−1. (8.15)
Let Le be the “lumping region” connecting CK1e, CK2
eand the two endpoints of e
(the shaded area in Fig. 8.5). Then, introducing the harmonic average of a over
Le, defined as
He(a) :=
(∫sea−1(ζ) dζ
se
)−1
=se
AK1
ei s
K1e
i +AK2
ei s
K2e
i
,
we can write (8.15) in a more expressive manner as
λe = He(a)
(AK2
ei
sK2
ei
seρK
1e +A
K1e
i
sK1
ei
seρK
2e
)≡ C1
e ρK1
e + C2e ρ
K2e . (8.16)
8.3. A Mixed–Hybridized Method with Numerical Quadrature 69
K1
e
K2
e
Figure 8.5.: Lumping region Le.
The two constants C1e and C2
e are such that C1e + C2
e = 1. This ensures that the
average (8.16) is consistent, i.e., if we set ρK1e = ρK
2e = ρ, then we get λe = ρ, as
should be expected. Using (8.7) over K1e and K2
e , we have
He(a) = e−ϕese
ζ1e + ζ2
e
, (8.17)
where ζre = sKr
ei /
(Di
Kre Be(∆ϕK
re ))
and ∆ϕKre := (ϕK
re − ϕe), r = 1, 2. The har-
monic average (8.17) is a positive quantity, because se > 0 and(ζ1e + ζ2
e
)> 0 due to
the fact that Th is a Delaunay triangulation and (8.8), respectively. For a discussion
of the use of the harmonic average in the finite element approximation of elliptic
problems, and of its impact on the computational performance of the method, we
refer to [BO83, AB85].
(Step 3). Substituting back (8.16) into (8.13) yields for each K ∈ Th the explicit relation
ΦKi = −(A
Ki )−1 CKi
i
ρKi − ρK
sKi|ei| = −(A
Ki )−1 A
Ki s
Ki
siHei(a)
ρKi − ρK
sKi|ei|
= −Hei(a)ρKi − ρK
si|ei| ∀ei ∈ ∂K ∩ Eh,int.
(8.18)
The finite volume nature of the formulation proposed in the present chapter can be
clearly recognized by comparison of the approximate flux ΦKi with the exact flux∫
ei
− a∇ρ · ni ds, which shows that the effect of the quadrature formula (8.6) is to
replace the term −(∇ρ ·ni)|Lei with the incremental ratio −(ρKi − ρK)/si and the
diffusion coefficient a|Lei with its harmonic average Hei(a). This result extends to
the dual mixed finite element setting the approach proposed in [MW94b] for the
Petrov-Galerkin finite element discretization of the convection-diffusion-reaction
equation.
(Step 4). Equation (8.18) already relates the unknown ρK to the neighbouring unknowns
ρKi , then to complete the derivation of the finite volume scheme, we need to consider
the case where ei ∈ Γ. We have:
70 8. Finite Element Approximation of the DAR Problem
ei ∈ ΓD: in this case, combining (8.13) and (8.18) immediately yields
ΦKi = −Hei(a)
ρD − ρK
sKi|ei|. (8.19)
ei ∈ ΓR: in this case, equating (8.13) with (8.12)2 and eliminating the hybrid vari-
able λei , yields
ΦKi =
γi e−ϕi ρK + jRi
AKi γi e
−ϕi sKi + 1|ei|. (8.20)
ei ∈ Γm: in this case, combining relations (8.10) and (8.12)3,4 and eliminating the
hybrid variables λe,1 and λe,2, we get:
ΦK1
eii =
αi e−ϕi,1 ρK
1ei − βi e−ϕi,2 ρK
2ei + σi,1 +A
K2ei
i sK2
eii βie
−ϕi,2(σi,1 − σi,2)
1 +AK1
eii s
K1ei
i αie−ϕi,1 +AK2
eii s
K2ei
i βie−ϕi,2
|ei|
ΦK2
eii =
βi e−ϕi,2 ρK
2ei − αi e−ϕi,1 ρK
1ei − σi,2 +A
K1ei
i sK1
eii αie
−ϕi,1(σi,1 − σi,2)
1 +AK1
eii s
K1ei
i αie−ϕi,1 +AK2
eii s
K2ei
i βie−ϕi,2
|ei|.
(8.21)
Replacing the expression of the flux ΦKi into the equilibrium equation (8.11), we ob-
tain the following linear system of algebraic equations that characterize the DMH-FV
approximation of problem (8.3)
Aρ ρ = f (8.22)
where Aρ ∈ RNE×NE is the stiffness matrix, ρ ∈ RNE is the unknown vector and f ∈ RNE
is the load vector, accounting for the contribution of the source function f and of the
boundary and interface data. To write down the entries of Aρ and f , we indicate by IK
and JKi the global indices of element K and Ki, i = 1, 2, 3. Moreover, for each K ∈ Th,
we introduce the non-negative quantities NKD , NK
R and NKm representing the number of
edges of K which belong to ΓD, ΓR and Γm, respectively. Clearly, these quantities are
all equal to zero if ∂K ∩ Γ = ∅. Then, the diagonal entries of Aρ read:
AρIKIK=
3∑i=1
ξKi + cKe−ϕK |K|
ξKi =
Hei(a)|ei|si
ei ∈ Eh,int
Hei(a)|ei|sKi
ei ∈ ΓD
γi e−ϕi
1 +AKi γi e
−ϕi sKi
|ei| ei ∈ ΓR
αi e−ϕi,1
1 +AK1
eii s
K1ei
i αie−ϕi,1 +AK2
eii s
K2ei
i βie−ϕi,2
|ei| ei ∈ Γm,1
βi e−ϕi,2
1 +AK1
eii s
K1ei
i αie−ϕi,1 +AK2
eii s
K2ei
i βie−ϕi,2
|ei| ei ∈ Γm,2,
(8.23)
8.3. A Mixed–Hybridized Method with Numerical Quadrature 71
the off-diagonal entries of Aρ are:
AρIKJKi=
−Hei(a)|ei|si
ei ∈ Eh,int
− βi e−ϕi,2
1 +AK1
eii s
K1ei
i αie−ϕi,1 +AK2
eii s
K2ei
i βie−ϕi,2
|ei| ei ∈ Γm,1
− αi e−ϕi,1
1 +AK1
eii s
K1ei
i αie−ϕi,1 +AK2
eii s
K2ei
i βie−ϕi,2
|ei| ei ∈ Γm,2,
(8.24)
and the load vector entries are:
fuIK = fK |K|+NK
D∑i=1
ηK,Di +
NKR∑
i=1
ηK,Ri +
NKm∑
i=1
ηK,mi
ηK,Di =uDi
ζKi|ei|
ηK,Ri =jRi
1 + γi ζKi|ei|
η1,mi = −σi,1 + βi ζ
2i (σi,1 − σi,2)
1 + αi ζ1i + βi ζ2
i
|ei|
η2,mi =
σi,2 + αi ζ1i (σi,2 − σi,1)
1 + αi ζ1i + βi ζ2
i
|ei|.
(8.25)
Some remarks about the properties of the numerical formulation illustrated in this
section are in order.
The first remark concerns the algebraic properties of the DMH-FV method. Matrix
Aρ has, at most, four non–zero entries on each row, and is structurally symmetric, i.e.,
if Aρij 6= 0 then also Aρ
ji 6= 0. In particular, denoting for each e ∈ Eh,int by I and J
the indices of the two triangles such that e = ∂KI ∩ ∂KJ, we have that AρIJ = Aρ
JI if
e ∈ Eh,int \ Γm while AρIJ 6= Aρ
JI if e ∈ Γm. The lack of symmetry numerically translates
the nonsymmetric action of the transmission conditions (8.3e) and (8.3f) with respect to
the neighbouring subdomains Ω1 and Ω2. To make this issue more precise, we associate
with each edge ei ∈ Γm the following “transmission” matrix Tρi ∈ RNE×NE
I J
Tρi =
|ei|∆i
. . . 0αie−ϕi,1 −βie−ϕi,2
. . .
−αie−ϕi,1 βie−ϕi,2
0 . . .
,
I
J
(8.26)
where ∆i := 1 + AK1
eii s
K1ei
i αie−ϕi,1 + A
K2ei
i sK2
eii βie
−ϕi,2 . By construction, the non-zero
entries of Tρi are the contributions ξKI
i , ξKJ
i to the diagonal entries of Aρ and the off-
diagonal entries AρIJ, AρJI, from which we see that Tρ
i is a nonsymmetric singular matrix
72 8. Finite Element Approximation of the DAR Problem
with zero column sum. The global stiffness matrix Aρ can therefore be partitioned
into the sum of a symmetric positive definite part AρS (associated with all the triangles
belonging to the interior of Ω1 and Ω2) and of a non-symmetric part AρNS =
∑ei∈Γm
Tρi .
By suitably renumbering the mesh elements, we see that the non-zero portion of matrix
AρNS has a block diagonal structure, where each block of 2 × 2 size corresponds to the
triangle pair sharing an edge on Γm (for example, KI, KJ or KP, KQ in Fig.8.6).
Figure 8.6.: Neighbouring triangles across the menbrane.
Having characterized the structure and basic properties of the stiffness matrix Aρ,
let us now investigate the numerical stability of the DMH-FV scheme. In this respect, an
important issue in heterogeneous flow transport problems is that the adopted numerical
scheme is monotone or, equivalently, it satisfies a Discrete Maximum Principle (DMP).
This property is the discrete counterpart of the continuous maximum principle associated
with problem (8.3), and is quite desirable because it prevents ρh from being affected by
spurious oscillations and ensures that each component of ρ is positive if each component
of the load vector f is > 0.
The need of devising a monotone approximation of problem (8.3) (typically studied
under more standard Dirichlet-Neumann boundary conditions, i.e., without the pres-
ence of an internal interface) has driven a considerable interest towards the develop-
ment of a special class of finite element schemes, known as exponentially fitted schemes
(see [RST96] for a detailed analysis and references). Such schemes are based on the so–
called Scharfetter–Gummel (SG) finite difference scheme [SG69], also known as Allen–
Southwell method [dGAS55]. The SG method is an optimal upwind difference scheme, it
is nodally exact in the case of constant problem coefficients [BH82] and satisfies a DMP
irrespective of the relative weight between diffusive and convective terms. Extending
the SG scheme to the two and three–dimensional setting, on triangular and tetrahe-
dral decompositions of the computational domain, has been the object of several works:
mixed-hybrid formulations [BMP87, BMP89b, BMP89a, SS97], Petrov–Galerkin formu-
lations [MW94b, MW94a, MW94c, Ker96], and Galerkin formulations with averaging of
the model coefficients along the element edges [BBFS90, GMS98, BJC98, XZ99, LZ05].
These methods share some common features: (i) they recover the SG approximation if
applied to one-dimensional problems; (ii) they satisfy a DMP under proper assumptions
on the angles of the triangulation Th; (iii) they ensure flux conservation across suitably
defined control volumes. Moreover, as a general trend, the schemes exhibit a common
ability in capturing sharp fronts without spurious oscillations, at the price of introducing
a certain amount of crosswind dissipation if the grid is not favorably aligned with the
8.3. A Mixed–Hybridized Method with Numerical Quadrature 73
advection field (cf. the numerical experiments in [BMP89a, SS97] and [GMS98]). More-
over, in some cases (as in the mixed formulation proposed in [BMP89b, BMP89a]), the
presence of a reaction term in the differential model introduces a difficulty in proving
the DMP for any value of the coefficient and of the mesh size, and requires a suitable
modification of the finite element space to reinforce the desired property [MP90]. Con-
ceptually similar approaches (based on the use of a proper lumping quadrature formula)
are adopted in the case of nodal-based formulations [BRF83, BR87, Ker96]
The following result provides sufficient conditions for the DMH-FV method to satisfy
a DMP.
Proposition 8.3.1. Let Th be a Delaunay triangulation such that for each edge e ∈ Γm
we have θK1
ee ≤ π/2, θ
K2e
e ≤ π/2, and for each edge e ∈ ΓD ∪ ΓR we have θKee ≤ π/2.
Then, Aρ is an irreducible M-matrix with strictly positive inverse [Var62], so that ρ > 0
if f ≥ 0 (DMP).
Proof. Under the above geometric assumptions on Th and the properties of AρS and Aρ
NS ,
it turns out that the stiffness matrix Aρ has zero column sums, with strictly positive
diagonal entries and nonpositive off diagonal entries. Moreover, for each element K with
an edge on ΓD, the matrix is diagonally dominant on the column corresponding to K.
The result then immediately follows by application of Theorem 3.1, p.202 of [RST96].
Prop. 8.3.1 provides a characterization of the numerical stability of the DMH-FV
scheme under proper assumptions on the geometrical discretization. It is important
to notice that the monotonicity of the proposed numerical method does not depend
on the value of the reaction coefficient ce−ϕ in (8.3), as is the case with the standard
dual-mixed method of [BMP89b, BMP89a], because in the FV structure of the scheme
such a term introduces a diagonal non-negative contribution to the stiffness matrix which
increases its diagonal dominance. The requirement of weak acuteness of Th on the domain
external boundary is standard and not restrictive for implementation (see [XZ99] and
the references cited therein). The requirement of weak acuteness of Th along the internal
interface is not strictly necessary, as a sufficient (more general) condition for Prop. 8.3.1
to hold is that ∆i > 0 for each edge ei ∈ Γm. In all the numerical experiments reported
in Chapt. 9 the finite element triangulation is chosen to be weakly acute along Γm and
ΓD ∪ ΓR.
The second remark concerns the relation between the proposed DMH-FV method
and other classical methods for the numerical solution of (8.3). Each row of (8.22) is
the finite volume discretization of the restriction to each element K ∈ Th of the mass
balance equation system (8.3a)- (8.3b). Using Euler’s theorem, we have that Ne→ 3NE/2
as the mesh size is refined, so that we can conclude that the computational effort of the
DMH-FV method is substantially lower than that of the standard DMH formulation.
Comparing the DMH-FV scheme to standard displacement-based methods, we see from
relations (8.19) and (8.20) that in the former approach both Dirichlet and Robin boundary
conditions are accounted for in an essential manner, unlike in the latter where Robin
conditions are accounted for in a weak manner. This indicates the robustness of the DMH-
FV method in treating boundary conditions on the flux variable, which are typically the
most important in the applications we are focusing on in the present chapter.
74 8. Finite Element Approximation of the DAR Problem
8.4. Implementation and Post-Processing of the DMH-FV
Method
In this section, we discuss how to implement the DMH-FV method in a numerically stable
manner and how to use the computed discrete solution to obtain a further approximation
of the exact solution u of (8.1) that enjoys a better convergence behavior.
8.4.1. Implementation
The solution of system (8.22) is not convenient from the numerical standpoint because
of the dynamic range of the function e−ϕ. This requires one to go back to the original
variable u using the inverse of (7.3) on each element K ∈ Th (cf. [BMP89b, BMP89a,
BMM+05])
ρK = uK eϕK ∀K ∈ Th. (8.27)
The action of (8.27) is a right diagonal scaling of Aρ which transforms (8.22) into the
equivalent algebraic linear system
Au u = f , (8.28)
where Au = AρDϕ ∈ RNE×NE is the new stiffness matrix and u ∈ RNE is the new unknown
vector, Dϕ being a diagonal matrix such that DϕIKIK
= eϕK
, K ∈ Th.
Proposition 8.4.1. Under the same assumptions as in Prop. 8.3.1, we have that Au is
an M-matrix with strictly positive inverse. This implies that u > 0 if f ≥ 0.
8.4.2. Post-Processing
The approximate flux density Jh can be recovered from the computed solution of (8.28)
by using (8.5) over each element K ∈ Th. With this aim, we need the expression of the
flux ΦKi across each edge ei ∈ ∂K, i = 1, 2, 3, such that ei ∈ Eh,int. A similar treatment
holds for the edges belonging to ΓD, ΓR or Γm. Using (8.27) and (8.17) in (8.18) yields
ΦKi = −e
∆ϕKii uKi − e∆ϕK
i uK
ζKi + ζKii
|ei| ei ∈ ∂K ∩ Eh,int. (8.29)
Proposition 8.4.2. Let ei ∈ ∂K ∩ Eh,int and assume that DK = DKi ≡ D and that
ϕ ∈ C1([CK , CKi ]). Then, the flux approximation (8.29) coincides with the classical
Scharfetter-Gummel (SG) exponentially fitted difference formula [SG69]
ΦKi = −D uKi Be(∆ϕi)− uK Be(−∆ϕi)
si|ei|, ∆ϕi := ϕK − ϕKi . (8.30)
Proof. We have to prove that:
e∆ϕKi
sKiBe(∆ϕ1
i )+
sKii
Be(∆ϕ2i )
=Be(−∆ϕi)
si
e∆ϕKii
sKii
Be(∆ϕ1i )
+sKii
Be(∆ϕ2i )
=Be(∆ϕi)
si.
(8.31)
8.4. Implementation and Post-Processing of the DMH-FV Method 75
Let us consider (8.31)1. Noting that ∆ϕi = ∆ϕKi −∆ϕKii , we have
Be(−∆ϕi)
si=
−∆ϕisi(e−∆ϕi − 1)
=−∆ϕi e
∆ϕKi
si(e∆ϕ
Kii − e∆ϕK
i )=
e∆ϕKi
e∆ϕKi − 1
∆ϕi/si− e∆ϕ
Kii − 1
∆ϕi/si
,
which coincides with the left-hand side of (8.31)1 because ∆ϕKi = ∆ϕi(sKi /si) and
∆ϕKii = −∆ϕi(s
Kii /si). In the same manner, we prove (8.31)2.
Proposition 8.4.2 shows that (8.29) is the consistent generalization of the SG method
to the case where both diffusivity coefficient and advective field are piecewise constant
quantities over the interval se, with a possible finite jump discontinuity in correspondance
of the midpoint Me of the inter-element edge e. This connection between the DMH-FV
formulation and the SG discretization is relevant in view of the analysis of the numerical
performance of the former scheme in the presence of dominating convection, as thoroughly
addressed in Chapt. 9.
Let λh ∈ Λh be the hybrid variable representing the approximation of u over Eh.
To recover λh from the computed solution of (8.28) we need to use (8.16) and then
apply (8.17), (8.7), (8.27) and (7.3) to obtain
λe =ζ2e e
∆ϕK1e uK
1e + ζ1
e e∆ϕK2
e uK2e
ζ1e + ζ2
e
∀e ∈ Eh,int. (8.32)
A similar treatment holds for the edges belonging to ΓD, ΓR and Γm, to yield:
λe =
P0(uD,e) e ∈ ΓD
e∆ϕKuK − ζe jRe
1 + γe ζKee ∈ ΓR,
(8.33)
while on Γm we have:
λe,1 =e∆ϕK1
e(1 + βe ζ
2e
)uK
1e + βe ζ
1e e
∆ϕK2e uK
2e − ζ1
e
(σe,1 + βe ζ
2e (σe,1 − σe,2)
)1 + αe ζ1
e + βe ζ2e
λe,2 =e∆ϕK2
e(1 + αe ζ
1e
)uK
2e + αe ζ
2e e
∆ϕK1e uK
1e + ζ2
e
(σe,2 + αe ζ
1e (σe,2 − σe,1)
)1 + αe ζ1
e + βe ζ2e
.
(8.34)
The above expressions of the degrees of freedom of λh over Eh can be used to construct
the following approximation of the exact solution u of (8.1)
u∗h(x) =∑e∈Eh
λe ωe(x), x ∈ Ω. (8.35)
The function u∗h ∈ Λh is the non-conforming piecewise linear interpolate of λh over
the computational grid Th. A thorough experimental analysis illustrated in Chapt. 9
demonstrates that u∗h satisfies the following convergence result
‖u− u∗h‖L2(Ω) ≤ Ch2, (8.36)
C being a positive constant depending on u and J but independent of the mesh size h. For
a proof of (8.36) in the case of homogeneous Dirichlet boundary conditions, see [dFS11].
76 8. Finite Element Approximation of the DAR Problem
Since the expected order of accuracy of uh in the L2-norm is O(h), we conclude
that (8.36) represents the superconvergence of the non-conforming approximation u∗h to
the exact solution u of (7.1), indicating, at least experimentally, that the DMH-FV
method, applied to the heterogeneous transport model, enjoys the same convergence
behavior proved in [AB85] for the standard DMH formulation in the case of the elliptic
model problem with Dirichlet boundary conditions.
Part IV.
Simulation Results, Applications and
Advanced Topics
9. Numerical Validation of the DMH-FV
Method
In this chapter, we perform a thorough numerical validation of the DMH-FV method in
the study of several test problems that represent significant examples of realistic appli-
cations in Biology and Electrophysiology. In particular, we include:
1. an experimental convergence analysis of the DMH-FV formulation in the solution
of problem (8.1) on a single domain;
2. a boundary-layer example of a two-ion PNP electrodiffusive transport;
3. the simulation of a prototype of a neuro-chip device; and
4. the simulation of the propagation of the action potential in an axon.
Geometry, parameters and data used in all of the reported test cases are as in [Bre09,
BJMS10].
9.1. Static condensation CAMBIA TITOLO
In Sect. 8.3 problem (8.4) was reformulated in the FV form (8.21), in which the only
unknown left is ρK and the quantities Φ and λK can be obtained with post-processing
operations. In this section instead, we show an alternative approach in which (8.4) is
reduced to a problem having the trace values λK as the only unknowns. For clarity we
report once again the considered problem: find u : Ω = (Ω1 ∪ Ω2)→ R such that:
divJ(u) + c u = f in Ω (9.1a)
J(u) = −D (∇u+ u∇ϕ) in Ω (9.1b)
u = uD on ΓD (9.1c)
J(u) · n = γ u+ jR on ΓR (9.1d)
J(u) · n1 = αu1 − β u2 + σ1 on Γm,1 (9.1e)
J(u) · n2 = β u2 − αu1 − σ2 on Γm,2. (9.1f)
80 9. Numerical Validation of the DMH-FV Method
If we choose the finite element spaces as in (8.2) the DMH Galerkin approximation of (9.1)
consists in finding (Jh, uh, λh) ∈ (Vh ×Wh ×Mh,uD) such that:
(D−1 Jh, τ h)Th + (D−1 ∇ϕuh, τ h)Th
− (uh, div τ h)Th + 〈λh · n, τ h〉Eh = 0 ∀τ h ∈ Vh (9.2a)
(divJh, qh)Th + (c uh, qh)Th = (f, qh)Th ∀qh ∈Wh (9.2b)
〈Jh · n, µh〉Eh = 〈γ λh µh〉ΓR+ 〈jR, µh〉ΓR
+ 〈αλh µh〉Γm,1 − 〈β λh µh〉Γm,1 + 〈σ1, µh〉Γm,1
+ 〈β λh µh〉Γm,2 − 〈αλh µh〉Γm,2 − 〈σ2, µh〉Γm,1 ∀µh ∈Mh,0. (9.2c)
If we restrict our analysis to one element K at a time and define the matrices
(AK)ij = (D−1τ j , τ i)K i, j = 1, . . . , d+ 1 (9.3a)
(BK)ij = −(qi, div τ j)K i = 1, j = 1, . . . , d+ 1 (9.3b)
(CK)ij = (D−1∇qj , τ i)K i = 1, . . . , d+ 1, j = 1 (9.3c)
(DK)ij = 〈µi · n, τ j〉∂K i, j = 1, . . . , d+ 1 (9.3d)
(EK)ij = −(c qj , qi)K i, j = 1 (9.3e)
(fK)i = −(f, qi)K i = 1, (9.3f)
where with d we denote the dimensionality of the problem (Ω ⊂ Rd), Eqs. (9.2a)
and (9.2b) read
AKJK + CKuK + (BK)TuK + (DK)Tλ∂K = 0 (9.4a)
−BKJK −EKuK = −fK . (9.4b)
Since AK is symmetric positive definite, we can obtain an explicit form for JK from (9.4a)
JK = −(AK)−1[(CK + (BK)T )uK + (DK)Tλ∂K
](9.5)
and being AK a small matrix (AK ∈ R(d+1)×(d+1)) its inversion is a quite numerically
cheap operation. If we insert (9.5) into (9.4b) we get[−BK(AK)−1(CK + (BK)T ) + EK
]︸ ︷︷ ︸UK
uK −BK(AK)−1(DK)Tλ∂K = fK (9.6)
and again it is possible to obtain an explicit expression for uK
uK = (UK)−1[fK + BK(AK)−1(DK)Tλ∂K
](9.7)
which depends only on the values of the traces λ∂K . Substituting (9.7) into the previously
obtained expression for JK (9.5) we get
JK =− (AK)−1(CK + (BK)T )(UK)−1BK(AK)−1(DK)Tλ∂K
− (AK)−1(CK + (BK)T )(UK)−1fK − (AK)−1(DK)Tλ∂K
=−[(AK)−1(CK + (BK)T )(UK)−1BK + I
](AK)−1(DK)Tλ∂K
− (AK)−1(CK + (BK)T )(UK)−1fK
=MKλ∂K + bK (9.8)
9.1. Static condensation CAMBIA TITOLO 81
which again depends solely on λ∂K . If we enforce the continuity of the normal component
of the flux across the edges, i.e.
[[Ji · n]]ei = 0 ∀ei ∈ Eh
we obtain ∑K
〈JK · n, µh〉∂K =∑K
∑j
〈µi · n, JKj τ j〉∂K = 0. (9.9)
In (9.9) we recognize (9.3d) so we can finally retrieve the desired algebraic problem in
the sole unknown λ using (9.8) ∑K
DKJK = 0
∑K
DK(MKλ∂K + bK
)= 0∑
K
−DKMKλ∂K =∑K
DKbK∑K
ΛKλ∂K =∑K
BK . (9.10)
Since the obtained problem is obtained with a sum over all the elements of the triangu-
lation Th, it is possible to construct the corresponding matrix and right hand side vector
block by block, summing up all the local contributions.
As for boundary conditions, their implementation does not present particular com-
plexities. Consider the case a Robin (Neumann) conditions has to be enforced on one of
the edges of the element K
κJK · n = γu− δ on ei ⊂ ∂K. (9.11)
By its definition it holds λ|ei = u|ei and in the (9.9) corresponding to the element K the
contribution of ei is ⟨JK · n, µi
⟩ei
=⟨γκλeiµi, µi
⟩ei−⟨δ
κ, µi
⟩ei
.
In order to include the boundary condition in the algebraic problem (9.10) it is enough
to modify the line corresponding to the edge ei
(ΛK)ii → (ΛK)ii +γ
κ(GK)ii and (BK)i → (BK)i +
(δK)iκ
where we defined GK and δK as
(GK)ij = 〈µi, µj〉∂K i, j = 1, . . . , d+ 1
(δK)i = 〈δ, µi〉∂K i = 1, . . . , d+ 1.
The transmission condition (9.2c) can be implemented using the same procedure and for
this reason we do not report the details here.
In addition, should the flux show jumps in its normal component across the inter-element
interfaces according to
[[Ji · n]]ei = σ
82 9. Numerical Validation of the DMH-FV Method
the numerical scheme can be easily modified to take into account that by modifying the
right hand side term corresponding to the elements K1 and K2 that share the edge ei
(K1 ∪K2 = ei)
(BKj )i → (BKj )i +(σKj )i
2for j = 1, 2,
where
(σK)i = 〈σ, µi〉∂K i = 1, . . . , d+ 1.
9.2. A one-dimensional heterogeneous domain
In this section, we consider problem (8.1) in the case where Ω = (0, 1) × (−0.5, 0.5)
and a membrane Γm is located at x = 0.5 to separate the left subdomain Ω1 from the
right subdomain Ω2. We set f = 0 and ∇ϕ = [−5, 0]T , while having two different
constant values in each subdomain for the diffusion constant D1 = 50, D2 = 0.5. The
Dirichlet data are uD = 0 at x = 0, y ∈ [−0.5, 0.5] and uD = 1 at x = 1, y ∈ [−0.5, 0.5],
while homogeneous Neumann conditions are enforced along y = 0.5 and y = −0.5, and
σ1 = σ2 = 0 on the interface. These data correspond to a one-dimensional transmembrane
flow along the x-direction. The following three sets of input data are considered:
1) c1 = c2 = 0 and α = β with α→ +∞;
2) c1 = c2 = 0 and α = β = 10;
3) c1 = 0.1, c2 = 10 while again α = β = 10.
Notice that case 1) corresponds to enforcing that u and J ·n are continuous across Γm.
The computed solutions for cases 1) and 2) are depicted in Fig. 9.1 (left), representing a
section at y = 0 of the post-processed quantity u∗h. The problem is diffusion–dominated
in Ω1, and is advection-dominated in Ω2, with an exact solution u almost linear over Ω1
and exponential over Ω2. In case 2), the solution has a finite jump across Γm because
of the selective behaviour of the membrane, while in case 1), the solution is continuous,
because the membrane is completely transparent to the flow of transported mass since
the interface condition is reduced to u1 = u2 on Γm, which is equivalent to eliminating
the membrane and treating the edges on Γm as belonging to Eh,int. The numerical
implementation of problem (8.1) is achieved using the transmission matrices defined
in (8.26). In all simulated cases, the DMH-FV method captures the solution layer without
introducing spurious oscillations, and it can be checked that the post-processed solution
u∗h is nodally exact up to machine precision.
In case 3), because of the fact that c 6= 0, the variable u∗h is no longer nodally exact;
however, the experimental convergence analysis reported in Fig.9.1 (right) indicates that
u∗h exhibits second order accuracy according to the error estimate (8.36). Fig. 9.2 shows
a three-dimensional plot of u∗h. The finite jump across Γm and the non-conforming
interpolation properties of the finite element space Λh are clearly visible.
9.3. Stationary profile of a binary electrolyte at a boundary 83
Figure 9.1.: Left: u∗h. Solid line: α = β = 10, dash-dotted line: α = β → ∞. Right:
‖u− u∗h‖L2(Ω) as a function of the mesh size h.
Figure 9.2.: Left: post-processed solution u∗h in the case c1 = 0.1, c2 = 10 and α = β =
10. Right: zoom of the solution on Ω2.
9.3. Stationary profile of a binary electrolyte at a boundary
In this section, we apply the DMH-FV to numerically study the stationary Poisson-
Nernst-Planck (PNP) system of partial differential equations introduced in Chapt. 5 (see
also [Rub90]). In stationary conditions (∂ci/∂t = 0), the PNP nonlinear differential
system is treated using the decoupled functional iteration described in Chapt. 6 (see
also [Bre09]).
This approach leads to the successive solution of linearized differential subproblems of
the form (8.1). In the considered case, we study a binary electrolyte (i.e., Mion = 2) with
zi = ±1, and the boundary value problem (5.4) is to be solved in the semi-infinite domain
x ∈ [0,+∞) with an applied external voltage drop ∆Vext = ϕ(0) − ϕ(+∞) = 100mV ,
with ϕ(+∞) = 0 and a surface at x = 0 impermeable to the ions. An analytical solution
84 9. Numerical Validation of the DMH-FV Method
of this problem for ϕ and u1,2 can be written as [BB93]:
ϕ = 2VT log
(1 +K exp(−
√2x/LD)
1−K exp(−√
2x/LD)
)ui = N0 exp (−zi ϕ/VT ),
where K = tanh(∆Vext/(4VT )), LD =√ε VT /(q N0) is the Debye length, VT ≈ 25mV is
the thermal voltage (having assumed T = 300K), q is the unit charge, ε ≈ 7·10−10 Fm−1
is the dielectric constant of the medium (water in this case) and N0 = 1mM is the bulk
concentration of both ions. The performed simulation is actually carried out on the two-
dimensional domain Ω = (0, L)2, with L = 5LD. The Debye length gives a measure of
the screening effect of a space charge layer, so that the choice of truncating the semi-
infinite domain to a finite length equal to a positive multiple of LD is a very good
approximation of the decaying behavior of ionic densities far away from the layer. The
boundary conditions for ϕ and u1,2 on y = 0 and y = 5LD are of homogeneous Neumann
type in order to obtain a solution dependent upon x solely, while on x = 0 and x = 5LD
the boundary conditions are obtained from the analytical solution. Fig. 9.3 illustrates a
slice along the x-axis of the computed ion concentrations ci and the discretization error as
a function of h. No spurious oscillations affect the results, and, again, superconvergence
as predicted by (8.36) can be observed for both u1 and u2.
Figure 9.3.: Computed ion concentrations (left) and ‖ui − u∗i,h‖L2(Ω), i = 1, 2 (right).
9.4. Simulation of a neuro-chip
In this section, we carry out a validation of the numerical accuracy and robustness of
the DMH-FV formulation in the simulation, using the PNP differential model, of a basic
configuration of a neuro-chip for neuroscience applications [ZF01, SMF01, Fro03, BF05].
The bio-hybrid device is the EOSFET (Electrolyte Oxide-Semiconductor Field Effect
Transistor) schematically depicted in Fig. 9.4 (left).
The aim of the device is to interface a biological component (a neuronal cell) to an
electrical component (solid-state substrate), in order (i) to transduce a chemical signal
generated by the biological component into an electronically readable signal, or, viceversa,
9.4. Simulation of a neuro-chip 85
(ii) to activate the biological component by the application of an electronic signal. In
operation mode (i), the EOSFET is working as a bio-sensor, while in operation mode (ii)
the EOSFET is working as a neuro-prosthetic device, i.e., playing the role of a neuronal
connection or even of a full neuronal network, thus opening the view for future use of
the neuro-chip as a cure for neuro-degenerative deseases like Alzheimer or Parkinson
[BYTV+05].
Figure 9.4.: Left: schematics of a neuro-chip (by courtesy reprinted from: E. Neher,
Molecular biology meets microelectronics, Nature Biotechnology, 19, 114
(2001)). Right: computational domain for stationary neuro-chip simulation.
The computational domain Ω is depicted in Fig. 9.4 (right), where we can distinguish
a portion Ωcell of the cell cytoplasm, the interstitial electrolyte cleft Ωbath separating cell
from substrate, the cell membrane Γm, the cell-to-chip contacting interface Γel and two
reference contacts Γcell and Γref . Dirichlet boundary conditions as in (8.1c) are enforced
on Γcell and Γref , a Robin boundary condition as in (8.1d) is enforced on Γel, while
interface boundary conditions as in (8.1e)- (8.1f) are enforced on Γm. On the remaining
portions of the domain boundary, ΓA, ΓN and the left vertical side of Ω, a homoge-
neous Neumann condition is enforced (γ = jR = 0 in (8.1d)). The geometrical data
used in computations are L = 0.8µm, H = 0.3µm, δcell = 0.25, µm, rcell = 0.5µm and
δcleft = 50nm. Ionic charge flow includes three species, K+, Na+ and Cl−, whose refer-
ence values are kept fixed respectively at (139, 12, 151)mM on Γcell and (4, 145, 149)mM
on Γref . As for the boundary condition for the electrostatic potential ϕ, we set ϕ = 0V
on Γref and ϕ ∈ [−100,+60]mV on Γcell. On the membrane Γm, the interface condi-
tions (6.4c)- (6.4d) are enforced for the potential while the interface conditions (6.5d)-
(6.5e) are enforced for the concentrations. In this latter case, the Goldman-Hodgkin-Katz
model (3.8) is used to describe the flow of ionic concentrations through the membrane.
On Γel, an homogeneous Neumann condition is enforced for the concentrations ci, while
the following compatibility condition is enforced for the polarization vector D
−εw∇ϕ · nΓel= Cel(ϕ) (ϕ− Vel) ,
where Vel is a fixed external potential and Cel(ϕ) is a MOS (Metal Oxide Semiconductor)
capacitance nonlinearly depending upon ϕ as described, e.g., in [TN98].
Fig. 9.5 (left) shows the computed static current-voltage characteristics, which de-
scribes the behaviour of the average value of Ji · n|Γref(positive if current flows out of
86 9. Numerical Validation of the DMH-FV Method
Figure 9.5.: Left: static I-V characteristics. Right: computed Na+ current density.
Γref , negative otherwise) as a function of ϕ|Γcell. The accuracy of the results is demon-
strated by the very good agreement between the estimated reverse potentials and the
physiologically correct values (2.5) computed using values for the intra-and extracellu-
lar ion concentrations typical of electrophysiology measurements [KS98, Hil01]. To help
interpret the results shown in Fig. 9.5 (left), we recall that the Nernst potential of each
ionic species is the value of ϕ|Γcellat which the corresponding ionic current density is
equal to zero, in accordance with Def. 2.2.2.
Figure 9.6.: Variations with respect to reference values of potential (left) and Na+ con-
centration (right).
Fig. 9.5 (right) shows the distribution of sodium current density over the computa-
tional domain. We can notice the higher current density in the cleft region between the
cell membrane and the electrical substrate. Such higher current density in turn causes
the rise of the potential in the cleft region, which can be measured by the field-effect
transistor in the substrate. Computed current value is again in quite a good agreement
with measured data [BF05]. We conclude this discussion by showing in Fig. 9.6 the com-
puted variations over Th of the electric potential ϕ and of the Na+ concentration with
respect to their corrsponding reference values (enforced at the Dirichlet boundary). The
results give an idea of the steep boundary layer effects occurring across the membrane
separating the intracellular region from the electrolyte cleft and at the interface with the
electronic substrate, and demonstrates the effectiveness of the DMH-FV formulation in
9.5. Action potential propagation in an axon 87
capturing the essential phenomena without introducing spurious oscillations that would
otherwise make the simulation prediction completely unreliable and inaccurate.
9.5. Action potential propagation in an axon
In this section, we test the performance of the DMH-FV method on a case of physiological
interest, namely, the Hodgkin-Huxley axon [HH52]. To this purpose, we consider the
same problem numerically investigated in [MP09], that is, the propagation of an action
potential in an unmyelinated neuronal axon. Ion transport in the intra and-extracellular
space is described by the EN-PNP model (5.8), while the full Hodgkin-Huxley system
of ordinary differential equations [HH52] is self-consistently solved at each time level of
signal propagation, to account for voltage gating mechanisms occurring along the axon
membrane.
Figure 9.7.: Computed action potential at two different time levels.
A full simulation of an action potential has been performed over the time interval
[0, 5ms], for the potential and three ionic species, namely K+, Na+ and Cl−. An
artificial increase at time t = 0 s of the Cl− conductance at the center of the axon triggers
the action potential, which propagates along the simulated axon of length 4000µm and
diameter 1µm.
Time snapshots of the electric potential spatial distribution along the axon, at t =
3ms and t = 5ms, shown in Fig. 9.7 clearly demonstrate the spreading of the action po-
tential towards the two ends of the axon. A similar trend can be observed in Figs. 9.8, 9.9
and 9.10, where the variations of each ionic concentration with respect to a reference value
are shown.
Fig. 9.11 depicts the spatial distribution of the gating variables at the two considered
time levels.
88 9. Numerical Validation of the DMH-FV Method
Figure 9.8.: Computed potassium concentration variation at two different time levels.
Figure 9.9.: Computed sodium concentration variation at two different time levels.
Figure 9.10.: Computed chlorine concentration variation at two different time levels.
9.5. Action potential propagation in an axon 89
Figure 9.11.: Computed gating variables at two different time levels.
10. Advanced Topics, A Look at the Future
and Concluding Remarks
In this concluding chapter, we provide a short overview of a few selected advanced topics,
with related bibliographical references, giving some suggestions for further development
and application of ideas and methods addressed in the notes.
10.1. Applications of the future
Among the wide variety of frontier applications of biological interfaces existing nowadays,
we mention:
1. the realization of advanced bio-hybrid interfaces, such as the prototype of an ar-
tificial retina for artificial vision. For this subject, we suggest to look at the
works [MDD+09, GADM+11, DRR+13] and all references cited therein. Exam-
ples of modeling and simulation methods based on the same material covered in
these notes can be found in [GPR+12] and [PLdF+14];
2. the study of the cellular mechanotransduction processes that control muscle con-
traction (for example, calcium sparks induced by chemical and mechanical stim-
uli that overview heart beat and pumping) and autoregulation phenomena (for
example, but not limited to, in retinal and brain blood circulation and nutrient
delivery/uptake).
For calcium sparks, we refer to the fundamental review work [CL08b] and, for
specific applications in neuroscience and renal smooth muscle cells, [Fri95, FC08,
BAL+07].
For autoregulation mechanisms in retinal microcirculation, we refer to [AHS+13,
GHC+14, HGA+13] and to all references cited therein, while for further important
aspects of electro-chemo-mechanical effects in the brain and other organs in animals
(and also human body), we refer to [YCBR03a, YCBR03b, YCBR05, WCS+11].
10.2. Concluding remarks
Along these Course, we have overviewed a hierarchy of mathematical models for the
study and simulation of charge transport processes occurring at the interfaces separating
two neighbouring biological environments, such as the intra- and extracellular sites of
a cell. These phenomena concern primarily with the electrodiffusive motion of charged
92 10. Advanced Topics, A Look at the Future and Concluding Remarks
ionic species in a biological solution under the action of an external applied force (of
electro-chemical and/or mechanical type) and are very well represented by the use of
mathematical descriptions based on nonlinearly coupled systems of ODEs and PDEs.
We have considered the mainly adopted approaches, namely, the GHK and Hodgkin-
Huxley current relations for the ODE part, and the PNP system for the PDE part,
and we have also discussed how to iteratively solve the problem and how to numerically
treat the linearized resulting equations using the finite element method. Finally, we have
illustrated in several examples the mathematical and biological accuracy of the proposed
models and methods on a set of significant case studies.
In the study of all of the above referenced issues and scientific articles, the mathe-
matical ODE-based and PDE-based models illustrated and analyzed in these notes, are
extensively adopted for describing and simulating model problems as well as realistic sit-
uations that aim to better interpret and, possibly, predict the occurrence of a pathology
and the way(s) for curing it, or, at least, maintaining under control its dangerous effect
and development.
Mathematics, in this sense, can provide a fundamental framework for constructing
a virtual laboratory in which the scientist can easily verify assumptions, claims and hy-
potheses on the functional behavior of a cellular system, a tissue, an organ or a synthetic
prosthesis of this latter, in order to improve knowledge and, ultimately, the health and
life of millions of people. This is a mission for which all of us should invest time and
resources for getting inside the fascinating world of Computational Biology, and, all in
all, represents the main scope of this Course and of these notes.
Bibliography
[AB85] D.N Arnold and F. Brezzi. Mixed and nonconforming finite element meth-
ods: implementation, postprocessing and error estimates. Math. Modeling
and Numer. Anal., 19(1):7–32, 1985.
[ABMO95] A. Agouzal, J. Baranger, J.F. Maitre, and F. Oudin. Connection between
finite volume and mixed finite element methods for a diffusion problem
with nonconstant coefficients. application to a convection-diffusion problem.
East-West J. Numer. Math., 34:237–254, 1995.
[AHS+13] J. Arciero, A. Harris, B. Siesky, A. Amireskandari, V. Gershuny, A. Pickrell,
and G. Guidoboni. Theoretical analysis of vascular regulatory mechanisms
contributing to retinal blood flow autoregulation. Invest. Ophthalmol. Vis.
Sci., 54(8):5584–5593, 2013.
[BAL+07] L. Balasubramanian, A. Ahmed, C. M. Lo, J.S.K. Sham, and K.P. Yip.
Integrin-mediated mechanotransduction in renal vascular smooth muscle
cells: activation of calcium sparks. Am J Physiol Regul Integr Comp Phys-
iol, 293:R1586–R1594, 2007.
[BB93] C.M.A. Brett and A.M. Oliveira Brett. Electrochemistry: Principles, Meth-
ods, and Applications. Oxford University Press, USA, 7 1993.
[BBFS90] R.E. Bank, J.F. Burgler, W. Fichtner, and R.K. Smith. Some upwinding
techniques for finite element approximations of convection-diffusion equa-
tions. Numer. Math., 58:185–202, 1990.
[BCEJ97] V. Barcilon, D. Chen, R.S. Eisenberg, and J.W. Jerome. Qualitative proper-
ties of steady-state Poisson-Nernst-Planck systems: perturbation and sim-
ulation study. SIAM J. Appl. Math., 57(3):631–648, 1997.
[BF91] F. Brezzi and M. Fortin. Mixed and Hybrid Finite Element Methods.
Springer Verlag, New York, 1991.
[BF01] A.J. Bard and L.K. Faulkner. Electrochemical Methods. John Wiley &
Sons, 2 edition, 2001.
[BF05] M. Brittinger and P. Fromherz. Field-effect transistor with recombinant
potassium channels: fast and slow response by electrical and chemical in-
teractions. Appl. Phys. A, 81:439–447, 2005.
94 Bibliography
[BH82] A.N. Brooks and T.J.R. Hughes. Streamline upwind Petrov-Galerkin for-
mulations for convection dominated flows with particular emphasis on the
incompressible Navier-Stokes equations. Comput. Meths. Appl. Mech. and
Engr., 32:199–259, 1982.
[BJC98] R.E. Bank, W.M. Coughran Jr., and L.C. Cowsar. The finite volume
Scharfetter–Gummel method for steady convection–diffusion equations.
Comput. Visual. Sci., 1(3):123–136, 1998.
[BJMS10] M. Brera, J.W. Jerome, Y. Mori, and R. Sacco. A conservative and mono-
tone mixed-hybridized finite element approximation of transport problems
in heterogeneous domains. Computer Methods in Applied Mechanics and
Engineering, 199(41-44):2709 – 2720, 2010.
[BMM+05] F. Brezzi, L.D. Marini, S. Micheletti, P. Pietra, R. Sacco, and S. Wang.
Discretization of semiconductor device problems (i). In E.J.W. ter Maten
P.G. Ciarlet, W.H.A. Schilders, editor, Lecture Notes in Computational Sci-
ence and Engineering, volume XIII, pages 317–441. Elsevier North-Holland,
Amsterdam, 2005.
[BMM+06] F. Brezzi, L.D. Marini, S. Micheletti, P. Pietra, and R. Sacco. Stability
and error analysis of mixed finite volume methods for advective-diffusive
problems. Comput. Math. Appl., 51:681–696, 2006.
[BMO96] J. Baranger, J.F. Maitre, and F. Oudin. Connection between finite volume
and mixed finite element methods. M2AN, 30:445–465, 1996.
[BMP87] F. Brezzi, L.D. Marini, and P. Pietra. Methodes d’elements finis mixtes et
schema de Scharfetter–Gummel. C.R. Acad. Sci. Paris, 305, Serie 1:599–
604, 1987.
[BMP89a] F. Brezzi, L.D. Marini, and P. Pietra. Numerical simulation of semicon-
ductor devices. Comput. Meths. Appl. Mech. and Engr., 75:493–514, 1989.
[BMP89b] F. Brezzi, L.D. Marini, and P. Pietra. Two-dimensional exponential fitting
and applications to drift-diffusion models. SIAM J. Numer. Anal., 26:1342–
1355, 1989.
[BO83] I. Babuska and J. Osborn. Generalized finite element methods, their per-
formance and their relation to mixed methods. SIAM J. Numer. Anal.,
20:510–536, 1983.
[BR87] R.E. Bank and D.J. Rose. Some error estimates for the box method. SIAM
J. Numer. Anal., 24(4):777–787, 1987.
[Bre09] M. Brera. Multiphysics/multiscale computational modeling in neuroelec-
tronics, 2009. MSc. Thesis, Politecnico di Milano, http://www1.mate.
polimi.it/biblioteca/tesvview.php?id=66&L=i.
Bibliography 95
[BRF83] R.E. Bank, D.J. Rose, and W. Fichtner. Numerical methods for semi-
conductor device simulation. SIAM J. Sci. and Stat. Comput., 4:416–435,
1983.
[BYTV+05] M. Bani-Yaghoub, R. Tremblay, R. Voicu, G. Mealing, R. Monette, C. Py,
K. Faid, and M. Sikorska. Neurogenesis and neuronal communication on
micropatterned neurochips. Biotechnology and Bioengineering, 92 (3):336–
345, 2005.
[CDG08] B. Cockburn, B. Dong, and J. Guzman. A superconvergent LDG-
hybridizable Galerkin method for second-order elliptic problems. Math.
Comp., 77:1887–1916, 2008.
[CDG+09] B. Cockburn, B. Dong, J. Guzman, M. Restelli, and R. Sacco. An hybridiz-
able discontinuous galerkin method for steady-state convection-diffusion-
reaction problems. SIAM J. Sci. Comp., 31 (5):3827–3846, 2009.
[CG04] B. Cockburn and J. Gopalakrishnan. A characterization of hybridized
mixed methods for second order elliptic problems. SIAM J. Numer. Anal.,
42:283–301, 2004.
[CG05] B. Cockburn and J. Gopalakrishnan. Error analysis of variable degree mixed
methods for elliptic problems via hybridization. Math. Comp., 74:1653–
1677, 2005.
[CGL09] B. Cockburn, J. Gopalakrishnan, and R. Lazarov. Unified hybridization of
discontinuous Galerkin, mixed and continuous Galerkin methods for second
order elliptic problems. SIAM J. Numer. Anal., 47:1319–1365, 2009.
[CGW09] B. Cockburn, J. Guzman, and H. Wang. Superconvergent discontinuous
Galerkin methods for second-order elliptic problems. Math. Comp., 78:1–
24, 2009.
[CL08a] H. Cheng and W. J. Lederer. Calcium sparks. Physiol Rev, 88(4):1491–
1545, 2008.
[CL08b] H. Cheng and W.J. Lederer. Calcium sparks. Physiol Rev, 88:1491–1545,
2008.
[CR73] M. Crouzeix and P.A. Raviart. Conforming and non-conforming finite el-
ement methods for solving the stationary Stokes equations. R.A.I.R.O.,
R-3:33–76, 1973.
[CSV12] P. Causin, R. Sacco, and M. Verri. A multiscale approach in the computa-
tional modeling of the biophysical environment in artificial cartilage tissue
regeneration. Biomechanics and Modeling in Mechanobiology, 2012.
[Del34] B. Delaunay. Sur la sphere vide. Izv. Akad. Nauk. SSSR., Math. and Nat.
Sci. Div., 6:793–800, 1934.
96 Bibliography
[dFPSV12] C. de Falco, M. Porro, R. Sacco, and M. Verri. Multiscale modeling and
simulation of organic solar cells. Computer Methods in Applied Mechanics
and Engineering, 245-246:102 – 116, 2012.
[dFS11] C. de Falco and R. Sacco. Error estimates for a mixed hybridized finite vol-
ume method for 2nd order elliptic problems. In Carmelo Clavero, Jos Luis
Gracia, and Francisco J. Lisbona, editors, BAIL 2010 - Boundary and Inte-
rior Layers, Computational and Asymptotic Methods, volume 81 of Lecture
Notes in Computational Science and Engineering, pages 109–117. Springer
Berlin Heidelberg, 2011.
[dGAS55] D.N. de G. Allen and R.V. Southwell. Relaxation methods applied to deter-
mine the motion, in two dimensions, of a viscous fluid past a fixed cylinder.
Quart. J. Mech. Appl. Math., 8:129–145, 1955.
[DRR+13] Ghezzi D., Antognazza M. R., Maccarone R., Bellani S., Lanzarini E.,
Martino N., Mete M., Pertile G., Bisti S., Lanzani G., and Benfenati F.
A polymer optoelectronic interface restores light sensitivity in blind rat
retinas. Nature Photonics, 7:400–406, 2013.
[ET10] G. Bard Ermentrout and David H. Terman. Mathematical Foundations of
Neuroscience. Springer, 1st edition. edition, 2010.
[FC08] D. D. Friel and H. H. Chiel. Calcium dynamics: analyzing the Ca2+ regu-
latory network in intact cells. Trends in Neuroscience, 31(1):8–19, 2008.
[FG08] P.J. Frey and P.-L. George. Mesh Generation. Application to Finite Ele-
ments (2nd Ed.). Wiley, London, 2008.
[FMWT02] C.P. Fall, E.S. Marland, J.M. Wagner, and J.J. Tyson, editors. Compu-
tational Cell Biology, volume 20 of Interdisciplinary Applied Mathematics.
Springer, New York, 2002. 1st ed 2002. Corr. 3rd printing, 2004.
[Fri95] D. D. Friel. [ca2+] oscillations in sympathetic neurons: An experimental
test of a theoretical model. Biophysical Journal, 68:1752–1766, 1995.
[Fro03] P. Fromherz. Neuroelectronics interfacing: Semiconductor chips with ion
channels, nerve cells and brain. In R. Weise, editor, Nanoelectronics and
Information Technology, pages 781–810. Wiley-VCH, Berlin, 2003.
[GADM+11] Diego Ghezzi, Maria R. Antognazza, Marco Dal Maschio, Erica Lanzarini,
Fabio Benfenati, and Guglielmo Lanzani. A hybrid bioorganic interface for
neuronal photoactivation. Nature Communications, 2:166+, January 2011.
[GH06] A.C. Guyton and J.E. Hall. Textbook of Medical Physiology. Elsevier, 11
edition, 2006.
[GHC+14] G. Guidoboni, A. Harris, L. Carichino, Y. Arieli, and B. A. Siesky. Effect
of intraocular pressure on the hemodynamics of the central retinal artery:
a mathematical model. Math. Biosci. Eng., 11(3):523–546, 2014.
Bibliography 97
[GMS98] E. Gatti, S. Micheletti, and R. Sacco. A new Galerkin framework for the
drift-diffusion equation in semiconductors. East-West J. Numer. Math.,
6(2):101–135, 1998.
[GPR+12] M. Garbugli, M. Porro, V. Roiati, A. Rizzo, G. Gigli, A. Petrozza, and
G. Lanzani. Light energy harvesting with nano-dipoles. Nanoscale, 4:1728–
1733, 2012.
[Gum64] H.K. Gummel. A self-consistent iterative scheme for one-dimensional steady
state transistor calculations. IEEE Trans. Electron Devices, ED-11:455–
465, 1964.
[HGA+13] A. Harris, G. Guidoboni, J. C. Arciero, A. Amireskandari, L. A. Tobe, and
B. A. Siesky. Ocular hemodynamics and glaucoma: the role of mathematical
modeling. Eur. J. Ophthalmol., 23(2):139–146, 2013.
[HH52] A.L. Hodgkin and A.F. Huxley. Currents carried by sodium and potas-
sium ions through the membrane of the giant axon of loligo. Journal of
Physiology, 116:449–472, 1952.
[Hil01] B. Hille. Ionic Channels of Excitable Membranes. Sinauer Associates, Inc.,
Sunderland, MA, 2001.
[Hop50] E. Hopf. The partial differential equation ut + uux = uxx. Comm. Pure
Appl. Math., 3:201–230, 1950.
[Jac99] J. D. Jackson. Classical Electrodynamics (3rd Ed.). Wiley, New York, 1999.
[JCLS08] J.W. Jerome, B. Chini, M. Longaretti, and R. Sacco. Computational mod-
eling and simulation of complex systems in bio-electronics. Journal of Com-
putational Electronics, 7(1):10–13, 2008.
[Jer96] J.W. Jerome. Analysis of charge transport: a mathematical study of semi-
conductor devices. Springer-Verlag, 1996.
[Jer02] J.W. Jerome. Analytical approaches to charge transport in a moving
medium. Transport Theory and Statistical Physics, 31:333–366, 2002.
[JS09] J.W. Jerome and R. Sacco. Global weak solutions for an incompressible
charged fluid with multi-scale couplings: Initial-boundary value problem.
Nonlinear Analysis, 71:e2487–e2497, 2009.
[KBA05] G.E. Karniadakis, A. Beskok, and N. Aluru, editors. Microflows and
Nanoflows. Fundamentals and Simulation, volume 29 of Interdisciplinary
Applied Mathematics. Springer, New York, 2005.
[Ker96] T. Kerkhoven. Piecewise linear Petrov–Galerkin error estimates for the box
method. SIAM J. Numer. Anal., 33(5):1864–1884, 1996.
[KS98] J. P. Keener and J. Sneyd. Mathematical Physiology. Springer, New York,
1998.
98 Bibliography
[LS11] S. Lee and V. Sundararaghavan. Multi-scale modeling of moving interface
problems with flux and field jumps: Application to oxidative degradation of
ceramic matrix composites. International Journal for Numerical Methods
in Engineering, 85(6):784–804, 2011.
[LZ05] R.D. Lazarov and L.T. Zikatanov. An exponential fitting scheme for
general convection–diffusion equations on tetrahedral meshes. Comput.
Appl. Math., (Obchysljuval’na ta prykladna matematyka, Kiev), 1(92):60–
69, 2005.
[Mar86] P.A. Markowich. The Stationary Semiconductor Device Equations.
Springer-Verlag, Wien-New York, 1986.
[MDD+09] Antognazza M.R., Ghezzi D., Musitelli D., Garbugli M., and Lanzani G.
A hybrid solid-liquid polymer photodiode for the bioenvironment. Applied
Physics Letters, 94:243501–243503, 2009.
[MJC+07] G. Marino, J. W. Jerome, B. Chini, M. Longaretti, and R. Sacco. Com-
putational models in nano-bio-electronics: simulation of ionic transport in
voltage operated channels. Journal of Nanoscience and Nanotechnology,
8(7):1–9, 2007.
[MJP07] Y. Mori, J .W. Jerome, and C. S. Peskin. A three-dimensional model of cel-
lular electrical activity. Bulletin of the Institute of Mathematics, Academia
Sinica, 2 (2):367–390, 2007.
[MJR05] V. Martin, J. Jaffre, and J. E. Roberts. Modeling fractures and barriers
as interfaces for flow in porous media. SIAM J. Sci. Comp., 26:1667–1691,
2005.
[Mor06] Y. Mori. A Three-Dimensional Model of Cellular Electrical Activity. PhD
thesis, New York University, 2006.
[MP90] L.D. Marini and P. Pietra. An abstract theory for mixed approximations
of second order elliptic problems. Math. Aplic. Comp., 8:219–239, 1990.
[MP09] Y. Mori and C. S. Peskin. A numerical method for cellular electrophysiol-
ogy based on the electrodiffusion equations with internal boundary condi-
tions at internal membranes. Communications in Applied Mathematics and
Computational Science, 4 (1):85–134, 2009.
[MRS90] P.A. Markowich, C.A. Ringhofer, and C. Schmeiser. Semiconductor Equa-
tions. Springer-Verlag, 1990.
[MSS01] S. Micheletti, R. Sacco, and F. Saleri. On some mixed finite element meth-
ods with numerical integration. SIAM J. Sci. Comput., 23-1:245–270, 2001.
[MW94a] J.J. Miller and S. Wang. An analysis of the Scharfetter-Gummel box
method for the stationary semiconductor device equations. RAIRO M2AN,
28(2):123–140, 1994.
Bibliography 99
[MW94b] J.J. Miller and S. Wang. A new non–conforming Petrov–Galerkin finite ele-
ment method with triangular elements for an advection–diffusion problem.
IMA J. Numer. Anal., 14:257–276, 1994.
[MW94c] J.J. Miller and S. Wang. A tetrahedral mixed finite element method for the
stationary semiconductor continuity equations. SIAM J. Numer. Anal., 31
(1):196–216, 1994.
[OR70] J.M. Ortega and W.C. Rheinboldt. Iterative Solution of Nonlinear Equa-
tions in Several Variables. Academic Press, New York, 1970.
[PJ97] J.H. Park and J.W. Jerome. Qualitative properties of steady-state Poisson-
Nernst-Planck systems: mathematical study. SIAM J. Appl. Math.,
57(3):609–630, 1997.
[PLdF+14] M. Porro, G. Lanzani, C. de Falco, R. Sacco, and M. Verri. Multiscale
simulation of organic heterojunction light harvesting devices. COMPEL:
The International Journal for Computation and Mathematics in Electrical
and Electronic Engineering, 33, 2014. to appear.
[QS89] N. Qian and T.J. Sejnowski. An electro-diffusion model for computing mem-
brane potentials and ionic concentrations in branching dendrites, spines and
axons. Biol. Cybern., 62:1–15, 1989.
[QV97] A. Quarteroni and A. Valli. Numerical Approximation of Partial Differen-
tial Equations. Springer-Verlag, 1997.
[RST96] H. G. Roos, M. Stynes, and L. Tobiska. Numerical methods for singularly
perturbed differential equations. Springer-Verlag, Berlin Heidelberg, 1996.
[RT77] P.A. Raviart and J.M. Thomas. A mixed finite element method for second
order elliptic problems. In I. Galligani and E. Magenes, editors, Mathemat-
ical Aspects of Finite Element Methods,I. Springer-Verlag, Berlin, 1977.
[RT79] P.A. Raviart and J.M. Thomas. Dual finite element methods for second
order elliptic problems. In R. Glowinski, E.Y. Rodin, and O.C. Zienkiewicz,
editors, Energy methods in finite element analysis. John Wiley and Sons
Ltd., New-York, 1979.
[RT91] J.E. Roberts and J.M. Thomas. Mixed and hybrid methods. In P.G. Ciarlet
and J.L. Lions, editors, Finite Element Methods, Part I. North-Holland,
Amsterdam, 1991. Vol.2.
[Rub90] I. Rubinstein. Electrodiffusion of Ions. SIAM, Philadelphia, PA, 1990.
[SBT02] L. Stryer, J.M. Berg, and J.L. Tymoczko. Biochemistry. New York:
W.H.Freeman., 2002.
[Sch96] S.G. Schultz. Molecular Biology of Membrane Transport Disorders.
Springer, 1996.
100 Bibliography
[Sch09] M. Schmuck. Analysis of the Navier-Stokes-Nernst-Planck-Poisson system.
Mathematical Models and Methods in Applied Sciences, 19(6):993–1015,
2009.
[Sel84] S. Selberherr. Analysis and Simulation of Semiconductor Devices. Springer-
Verlag, 1984.
[SG69] D.L. Scharfetter and H.K. Gummel. Large signal analysis of a silicon Read
diode oscillator. IEEE Trans. Electron Devices, ED-16(1):64–77, 1969.
[Slo73] J. W. Slotboom. Computer-aided two-dimensional analysis of bipolar tran-
sistors. IEEE Trans. El. Dev., 20(8):669–679, 1973.
[SMF01] B. Straub, E. Meyer, and P. Fromherz. Recombinant maxi-K channels on
transistor, a prototype of iono-electronic interfacing. Nature Biotechnology,
19:121–124, 2001.
[SS97] R. Sacco and F. Saleri. Stabilized mixed finite volume methods for
convection-diffusion problems. East West J. Numer. Math., 5 (4):291–311,
1997.
[TN98] Y. Taur and T.H. Ning. Fundamentals of modern VLSI devices. Cambridge
University Press, New York, NY, USA, 1998.
[Var62] R.S. Varga. Matrix Iterative Analysis. Englewood Cliffs, New Jersey, 1962.
[WCS+11] Tong WC, Choi CY, Kharche S, Holden AV, Zhang H, and Taggart MJ.
A computational model of the ionic currents, ca2+ dynamics and action
potentials underlying contraction of isolated uterine smooth muscle. PLoS
One, 6:e18685, 2011. doi: 10.1371/journal.pone.0018685.
[WT06] J. H.-C. Wang and B. P. Thampatty. An introductory review of cell
mechanobiology. Biomechanics and Modeling in Mechanobiology, 5:1–16,
2006. 10.1007/s10237-005-0012-z.
[XZ99] J. Xu and L. Zikatanov. A monotone finite element scheme for convection–
diffusion equations. Math. Comp., 68(228):1429–1446, 1999.
[YCBR03a] Jin Yang, John W. Clark, Robert M. Bryan, and Claudia S. Robertson.
The myogenic response in isolated rat cerebrovascular arteries: smooth
muscle cell model. Medical Engineering & Physics, 25:691–709, 2003.
10.1016/S1350-4533(03)00100-0.
[YCBR03b] Jin Yang, John W. Clark, Robert M. Bryan, and Claudia S. Robert-
son. The myogenic response in isolated rat cerebrovascular arteries: vessel
model. Medical Engineering & physics, 25:711–717, 2003. 10.1016/S1350-
4533(03)00101-2.
[YCBR05] Jin Yang, John W. Clark, Robert M. Bryan, and Claudia S. Robertson.
Mathematical modeling of the nitric oxide/cgmp pathway in the vascular
Bibliography 101
smooth muscle cell. American Journal of Physiology. Heart and Circulatory
Physiology, 289, 2005. DOI:10.1152/ajpheart.00216.2004.
[ZF01] G. Zeck and P. Fromherz. Noninvasive neuroelectronic interfacing with
synaptically connected snail neurons on a semiconductor chip. PNAS,
98:10457–10462, 2001.