SCALABLE APPROXIMATION AND SOLVERS FOR IONIC

ELECTRODIFFUSION IN CELLULAR GEOMETRIES

PIETRO BENEDUSI∗ , ADA J. ELLINGSRUD∗ , HALVOR HERLYNG∗ , AND MARIE
E. ROGNES∗

Abstract. The activity and dynamics of excitable cells are fundamentally regulated and moder-
ated by extracellular and intracellular ion concentrations and their electric potentials. The increasing
availability of dense reconstructions of excitable tissue at extreme geometric detail pose a new and
arXiv:2403.04491v1 [math.NA] 7 Mar 2024

clear scientific computing challenge for computational modelling of ion dynamics and transport. In
this paper, we design, develop and evaluate a scalable numerical algorithm for solving the time-
dependent and nonlinear KNP-EMI equations describing ionic electrodiffusion for excitable cells
with an explicit geometric representation of intracellular and extracellular compartments and in-
terior interfaces. We also introduce and specify a set of model scenarios of increasing complexity
suitable for benchmarking. Our solution strategy is based on an implicit-explicit discretization and
linearization in time, a mixed finite element discretization of ion concentrations and electric potentials
in intracellular and extracellular domains, and an algebraic multigrid-based, inexact block-diagonal
preconditioner for GMRES. Numerical experiments with up to 108 unknowns per time step and up to
256 cores demonstrate that this solution strategy is robust and scalable with respect to the problem
size, time discretization and number of cores.

Key words. Electrodiffusion, electroneutrality, KNP-EMI, finite elements, preconditioning

MSC codes. 65F10, 65N30, 65N55, 68U20, 68W10, 92-08, 92C20

1. Introduction. Our brains are composed of intertangled tissue consisting of

neurons, glial cells and interstitial space, penetrated by blood vessels. Ions, such as
potassium (K+ ), sodium (Na+ ), calcium (Ca2+ ) and chloride (Cl− ), move within and
between the intracellular and extracellular compartments in a highly regulated man-
ner. These ions, their movement and concentration differences, are fundamental for
the function and well-being of the brain [52, 45]. For instance, the brain’s electrical
signals fundamentally rely on action potentials induced by rapid neuronal influx of
Na+ and efflux of K+ [52]. Moreover, ion concentrations and importantly ion concen-
tration gradients may modulate brain signalling, regulate brain volume, and control
brain states [41, 4, 45]. Notably, these physiological processes are only partially un-
derstood and currently attract substantial interest from the neurosciences [35, 45,
60, 6, 56, 43, 20]. High-fidelity in-silico studies would offer an innovative avenue of
investigation with significant scientific potential.
Electrical and chemical activity in excitable tissue are naturally modelled via cou-
pled partial differential equations (PDEs) describing ionic electrodiffusion [39, 42, 38,
21, 22]. These models go beyond the classical electrophysiology models, such as the
monodomain or bidomain equations [53] and their cellular-level counterparts [3, 54],
by describing spatial and temporal variations of ion concentrations in addition to the
electric potentials. Within this context, we here consider explicit geometric repre-
sentations of the intracellular and extracellular compartments and their joint inter-
face representing the cell membrane [39, 61, 22]; i.e., the so-called KNP-EMI model
(Kirchhoff-Nernst-Planck, extracellular-membrane-intracellular). In their seminal re-
search [39], Mori and Peskin introduced and analyzed a finite volume-based spatial
discretization of these cell-based ionic electrodiffusion equations, with numerical ex-
periments in idealized two- and three-dimensional geometries. Similar equations and

∗ Department
for Numerical Analysis and Scientific Computing, Simula Research Laboratory, Oslo,
Norway ([email protected], [email protected], [email protected], [email protected])
1
2 P. BENEDUSI ET AL

approximation challenges are also encountered in connection with ionic transport

in lithium-ion batteries [32]. In our previous work [22], we introduced a mixed-
dimensional mortar finite element method for these equations, again with numerical
experiments limited to idealized geometries, and with little attention to computational
performance.
Dense reconstructions and segmentations of brain tissue are increasingly becom-
ing openly available [40, 62, 17, 36]. These imaging-based representations reveal
an extraordinary geometric complexity that is very different from the more struc-
tured layers of cardiac tissue [14, 30]. This complexity places clear demands on the
scalability of numerical solution algorithms. Solvers for the EMI (or cell-by-cell)
electrophysiology equations, which can be viewed as a subsystem of the ionic elec-
trodiffusion equations, have been studied quite intensively in recent years [3, 54, 11],
including finite difference based methods [31], mixed finite element methods with and
without Lagrange multipliers [54], boundary element methods [46], cut finite element
methods [13], and finite volume methods [58]. Another key topic is scalable precon-
ditioning [29, 11], and we highlight the scalable solution strategy for electrodiffusion
(without interior interfaces) using a discontinuous Galerkin discretization proposed
by Roy et al [47]. Parallel scalability may also be approached via domain decomposi-
tion techniques [29, 31], though the transition to complex geometries pose additional
challenges. In fact, most studies consider highly idealized settings [39, 22], structured
cell patterns as encountered in cardiac excitable tissue [54, 30], or simplified neuronal
geometries [57, 59, 50, 15, 13, 24].
In this paper, we address the challenge of how to design scalable solution al-
gorithms for the cell-based ionic electrodiffusion (KNP-EMI) equations with high
geometric complexity and physiological membrane mechanisms. To this end, we pres-
ent a lightweight single-dimensional finite element formulation with a scalable and
efficient monolithic preconditioning strategy, vetted by a series of numerical experi-
ments with realistic model parameters and geometries. The approximation scheme
and solution strategy yield accurate solutions in a bounded (low) number of Krylov
iterations. Near-optimal parallel scalability allows for rapidly obtaining solutions to
large KNP-EMI problems using high performance computing systems, thus bridging
the technology gap between dense tissue reconstructions and ionic electrodiffusion
simulations.
This manuscript is organized as follows. In Section 2, we present the KNP-
EMI equations and describe active and passive membrane dynamics involved in the
interface coupling. In Section 3, we introduce a weak formulation of the continuous
problem together with discretizations in time and space. In Section 4, we examine
the block-structure and properties of the resulting discrete systems, and present a
tailored solution strategy. In Section 5 we present four model scenarios in idealized
and image-based geometries of increasing complexity defining a benchmark suite for
KNP-EMI solvers. Section 6 reports on numerical experiments for the four scenarios,
investigating robustness, efficiency, and parallel performance of the proposed strategy.
Finally, Section 7 provides some concluding remarks and outlook.

2. A model for ionic electrodiffusion in cellular spaces. Following [22],

we consider the domain Ω = Ωi ∪ Ωe ∪ Γ ⊂ Rd for d ∈ {2, 3} with Ωi and Ωe
representing the intra- and extracellular domains (ICS, ECS), respectively, and Γ =
∂Ωi the cellular membrane(s). In general, the intracellular domain can be composed of
SNcell
Ncell ∈ N disjoint domains Ωi = j=1 Ωi,j , with Ωi,j representing the jth (biological)
cell with membrane Γj = ∂Ωi,j . We consider a set K of ionic species, for example
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 3

K = {Na+ , K+ , Cl− }, with cardinality |K|. For each ionic species k ∈ K and physical
region r ∈ {i, e}, we model the evolution of ionic concentrations [k]r : Ωr × [0, T ] → R,
as well as electric potentials ϕr : Ωr × [0, T ] → R.
2.1. Conservation equations. The assumption of conservation of ions for each
region r ∈ {i, e} yields the following time-dependent partial differential equation for
each ionic species k ∈ K and for t ∈ (0, T ) and x ∈ Ωr :

(2.1) ∂t [k]r (x, t) + ∇ · Jkr (x, t) = frk (x, t),

where Jkr : Ωr × [0, T ] → Rd is the ion flux density. We consider a Nernst–Planck

relation describing diffusion and drift in the electric field induced by the potential ϕr :
Drk zk
(2.2) Jkr = Jkr,diff + Jkr,drift = −Drk ∇[k]r − [k]r ∇ϕr ,
ψ
where Drk ∈ R+ is the effective diffusion coefficient of ionic species k in region r, zk is
the valence of ion k, and ψ = RT F −1 is composed of the gas constant R, the absolute
temperature T , and the Faraday constant F . Finally, frk : Ωr × [0, T ] → R represent
source terms. For consistency with the Kirchhoff-Nernst-Planck (KNP) assumption
below, frk should maintain electroneutrality in the sense that
X
zk frk (x, t) = 0 for r ∈ {i, e}.
k∈K

Equation (2.1) can be interpreted as an advection-diffusion equation with advection

velocity given by the electric field ∇ϕr . Since ϕr is itself an unknown, it is not trivial
to classify (2.1) as a diffusion- or advection-dominated problem a-priori. We note
that (2.1) ignores convective effects due to movement of the domain or medium; thus
assuming a medium at rest (see e.g. [21, 49] for contrast).
We introduce the KNP assumption of bulk electroneutrality for each region r ∈
{i, e}, for t ∈ [0, T ] and x ∈ Ωr :
X
zk [k]r (x, t) = 0,
k∈K

which, differentiated with respect to time and combined with (2.1), yields
X
(2.3) zk ∇ · Jkr (x, t) = 0.
k∈K

The KNP assumption can be compared to the Poisson-Nernst-Planck (PNP) assump-

tion. The latter would explicitly account for also the (rapid) charge relaxation dy-
namics, more relevant at the nanoscale [19, 51]. Note that the ion fluxes nonlinearly
couple the ion concentrations [k]r via the potential ϕr . Given the |K| + 1 unknown
scalar fields in each of the two regions, the KNP-EMI model thus consists of two
systems of |K| + 1 partial differential equations (|K| parabolic, one elliptic). These
two systems are coupled via EMI-type interface conditions.
2.2. Interface conditions. In this section, we introduce interface conditions
describing currents and dynamics across the cellular membrane(s) Γ. We first define
the membrane potential ϕM : Γ × (0, T ] → R as the jump in the electric potential:

(2.4) ϕM (x, t) = ϕi (x, t) − ϕe (x, t) for x ∈ Γ, t > 0.

4 P. BENEDUSI ET AL

We next assume that the total ionic current density IM : Γ × (0, T ] → R is continuous
across Γ such that for t > 0 and x ∈ Γ,
X X
(2.5) IM (x, t) = F zk Jki (x, t) · ni (x) = −F zk Jke (x, t) · ne (x),
k∈K k∈K

where nr denotes the normal on the interface pointing out from Ωr . We further
assume that IM consists of two components: a total ionic channel current Ich and
a capacitive current Icap . Both of these currents have contributions from each ionic
species:
X X
k k
(2.6) IM = Ich + Icap = Ich + Icap .
k∈K k∈K
k
The ionic channel currents Ich
will be subject to modelling, as described in Section 2.4,
while we consider the capacitor equation for the total capacitive current
(2.7) Icap = Cm ∂t ϕM ,
where Cm ∈ R+ is the membrane capacitance.
As in [22, Section 2.1.3], we derive an expression for Jkr · nr that will be of useful
in the weak formulation of the KNP-EMI equations. For k ∈ K and r ∈ {i, e}, we
introduce the relation
k
(2.8) Icap = αrk Icap = αrk Cm ∂t ϕM ,
and define the ratio
Drk zk2 [k]r X
(2.9) αrk = P ℓ 2
∈ [0, 1], with αrk = 1.
ℓ∈K Dr zℓ [ℓ]r k∈K

Combining (2.5), (2.6) and (2.8), we can express the normal fluxes corresponding to
a specific ion k ∈ K, as a function of the ionic currents and the membrane potential:
k k
Ich + Icap I k + αrk Cm ∂t ϕM
(2.10) Jkr · nr = ±r = ±r ch ,
F zk F zk
for r ∈ {i, e}, ±i = 1, and ±e = −1.
2.3. Initial and boundary conditions. To close the KNP-EMI system, we
impose initial conditions for k ∈ K and r ∈ {i, e}:
(2.11) [k]r (x, 0) = [k]0r (x), for x ∈ Ωr ,
(2.12) ϕM (x, 0) = ϕ0M (x), for x ∈ Γ,
with [k]0r and ϕ0M representing initial ion concentration and an initial membrane stim-
ulus, respectively. We also assume that the initial ion concentrations are prescribed
such that bulk electroneutrality is satisfied in each region:
X
(2.13) zk [k]0r (x) = 0.
k∈K

We impose homogeneous Neumann boundary conditions on the exterior boundary ∂Ω

(2.14) Jkr (x, t) · nr (x) = 0 for x ∈ ∂Ω.
The electric potentials ϕr in (2.2) are only determined up to a (common) additive
constant. Therefore, an additional constraint must be considered to enforce unique-
ness. Previous works have for example considered a zero-mean condition on ϕe [22].
We return to this point in Remark 4.1.
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 5

2.4. Ionic channels. In terms of ionic species, we focus on sodium, potassium,

and chloride: K = {Na+ , K+ , Cl− }, and two modelling scenarios (active and passive
k
dynamics) for the ionic channel currents Ich .
2.4.1. Active dynamics: Hodgkin-Huxley model. To model the membrane
dynamics of axons, we consider the Hodgkin-Huxley (HH) ionic model [27] with the
following expressions for the ionic currents:
Na
Ich (ϕM , [Na+ ]i , [Na+ ]e , w) = (gstim
Na Na
+ gleak + ḡ Na m3 h)(ϕM − ENa ),
K
(2.15) Ich (ϕM , [K+ ]i , [K+ ]e , w) = (gleak
K
+ ḡ K n4 )(ϕM − EK ),
Cl
Ich (ϕM , [Cl− ]i , [Cl− ]e ) = gleak
Cl
(ϕM − ECl ),
given the ion-specific reversal potentials
RT [k]e
(2.16) Ek = ln .
F zk [k]i
In (2.15), ḡ k , gleak
k
∈ R+ are the maximum and leak conductivities, respectively. The
Hodgkin-Huxley model also includes a set of time-dependent gating variables w(t) =
(n(t), m(t), h(t)) ∈ [0, 1]3 governed by the initial value problem
∂w
(2.17) = αw (ϕM )(1 − w) − βw (ϕM )w,
∂t
(2.18) w(0) = w0 = (n0 , m0 , h0 ).
The coefficients αw , βw depend on the difference ϕM − ϕrest , given a resting potential
Na
ϕrest , see [27]. The stimulus contribution gstim : Γ×[0, T ] → R+ triggers the activation
of the membrane dynamics.
2.4.2. Passive dynamics: Kir–Na/K. To model astrocyte and dendrite mem-
brane dynamics, we consider the following passive model (labeled Kir–Na/K) , which
includes leak channels, an inward-rectifying K channel, and a Na/K pump [25]:
Na
Ich (ϕM , [Na+ ]i , [Na+ ]e ) = (gstim
Na Na
+ gleak )(ϕM − ENa ) + 3F zNa · jpump ,
K
Ich (ϕM , [K+ ]i , [K+ ]e ) = gleak
K
(ϕM − EK )f Kir − 2F zK · jpump ,
Cl
Ich (ϕM , [Cl− ]i , [Cl− ]e ) = gleak
Cl
(ϕM − ECl ),
where the Kir-function fKir , controlling the inward-rectifying K current, is given by:
s
+ AB [K+ ]e
fKir ([K ]e , ∆ϕK , ϕM ) = ,
CD [K+ ]0e
where
0
A = 1 + exp(0.433), B = 1 + exp(−(0.1186 + EK )/0.0441),
C = 1 + exp((∆ϕK + 0.0185)/0.0425), D = 1 + exp(−(0.1186 + ϕM )/0.0441).
0
Here, ∆ϕK = ϕM − EK , and EK is the K-reversal potential at t = 0. Finally, the
pump flux density is given by:
1.5
!
[Na+ ]i [K+ ]e

jpump = ρpump 1.5 ,
[Na+ ]i 1.5
+ PNa [K+ ]e + PK
where ρpump , Pk ∈ R+ are, respectively, the maximum pump rate and a threshold for
ion k. All physical parameters are given in Table 2 (Section 5).
6 P. BENEDUSI ET AL

3. Variational formulation and discretization.

3.1. Single-dimensional formulation of the KNP-EMI problem. We con-
sider a so-called single-dimensional formulation of the KNP-EMI problem, following
the naming convention introduced in [34, 55]. In this formulation, the membrane cur-
rent variable IM is eliminated and the problem is no longer mixed-dimensional, since
all the unknowns are defined over d-dimensional domains. For the EMI problem, us-
ing the single-dimensional formulation results in a more compact algebraic structure
with favorable properties (symmetry and positive definiteness) [11]. In the case of
KNP-EMI, only positive definiteness is maintained.
Let Vrk = Vrk (Ω) and Wr = Wr (Ω) be Hilbert spaces of sufficiently regular func-
tions. For r ∈ {i, e}, assuming the solutions [k]r and ϕr to be sufficiently regular over
Ωr , we multiply the PDE in (2.1) by test functions vrk ∈ Vrk and integrate over Ωr .
After integration by parts and using the normal flux definition from (2.10), we obtain
for k ∈ K and r ∈ {r, e}:
Z Z Z
1
∂t [k]r vrk − Jkr · ∇vrk dx ±r k
+ αrk Cm ∂t ϕM vrk ds = frk vrk dx


(3.1) Ich
Ωr F zk Γ Ωr

Similarly, multiplying (2.3) by test functions wr ∈ Wr and integrating by parts, we

obtain
X Z 1
Z
(3.2) − zk Jkr · ∇wr dx ±r (Ich + Cm ∂t ϕM ) wr ds = 0.
Ωr F Γ
k∈K

k
We recall that Ich linearly depends on ϕM = ϕi − ϕe (cf. Section 2.4) and non-linearly
on [k]r . We thus define the weak solution to the KNP-EMI problem as {[k]r , ϕr } for
r ∈ {i, e} and k ∈ K satisfying (3.1)–(3.2) for all test functions vrk ∈ Vrk and wr ∈ Wr .
3.2. Time discretization. We consider an implicit-explicit time discretization
scheme, leading to a convenient linearization. We partition the time axes uniformly
into Nt intervals:

tn = n∆t, with n = 0, . . . , Nt , ∆t = T /Nt .

Given the initial conditions [k]0r , ϕ0M , w0 defined in Sections 2.3–2.4, we consider the
following first-order, implicit approximations at times tn for n = 1, . . . , Nt :

(3.3) ∂t [k]r ≈ ([k]nr − [k]n−1

r )/∆t, ∂t ϕM ≈ (ϕnM − ϕn−1
M )/∆t,

with [k]nr = [k]r (x, tn ), ϕnr = ϕr (x, tn ), and frk,n = frk (x, tn ). The ion flux densities
are approximated with an implicit-explicit scheme, while the coefficients αrk and ionic
k
currents Ich are treated explicitly:

Drk zk n−1 n
Jkr (x, tn ) ≈ Jk,n
r = −Drk ∇[k]nr − [k]r ∇ϕr ,
ψ

Drk zk2 [k]n−1

r
αrk (x, tn ) ≈ αrk,n−1 = P ℓ z 2 [ℓ]n−1
.
ℓ∈K D r ℓ r

If gating variables w are present, we first solve the ODEs (2.17) numerically with
Node steps of the Rush-Larsen method with step size ∆t/Node to obtain wn ≈ w(tn ),
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 7

setting ϕM = ϕn−1
M in (2.17), given the initial condition wn−1 . We then approximate
the ionic currents by
k,n−1
(3.4) k
Ich (x, tn ) ≈ Ich k
= Ich (ϕn−1 n−1
M , [k]i , [k]n−1
e , wn ).

Inserting the previous expressions into (3.1), we obtain the following linearized weak
formulation:
Z Z
(3.5) [k]nr vrk − ∆t Jk,n
r · ∇v k
r dx ± C
r r
k,n−1
ϕnM vrk ds
Ωr Γ
Z Z
= ([k]r + frk,n )vrk dx − grk,n−1 vrk ds,
n−1
Ωr Γ

for k ∈ K and n = 1, . . . , Nt , having introduced the short-hand

αrk,n Cm 1 k,n
(3.6) Crk,n = and grk,n = ±r (∆tIch − αrk,n Cm ϕnM ).
F zk F zk

Similarly, for (3.2), we obtain

X Z Cm
Z Z
(3.7) − ∆t zk Jk,n
r · ∇wr dx ±r ϕ n
w r ds = − hn−1 wr ds,
Ωr F Γ M Γ
r
k∈K

with hnr = ±r (∆tIch

n
+ Cm ϕnM )/F .
In summary, the semi-discrete KNP-EMI problem reads as: for n = 1, . . . , Nt , for
r ∈ {i, e} and k ∈ K, find [k]nr , ϕnr such that (3.5) and (3.7) hold for all vrk ∈ Vrk and
wr ∈ Wr .
3.3. Spatial discretization. We discretize Ωi and Ωe by conforming simplicial
meshes Ti and Te , respectively, such that T = Ti ∪ Te forms a conforming simplicial
tessellation of Ω; in particular, that Ti and Te match at Γ. For Tr and r ∈ {i, e}, we
approximate the Hilbert spaces Vrk and Wr by finite element spaces Vr,h k
and Wr,h of
continuous piecewise polynomials of degree pr,k and pr , for [k]r and ϕr respectively.
The (fully) discrete KNP-EMI problem to be solved at each time step n thus reads:
find [k]nr,h ∈ Vr,h
k k
and ϕnr,h ∈ Wr,h such that (3.5) and (3.7) hold for all vrk ∈ Vr,h and
for all wr ∈ Wr,h .
In practice, we use the same polynomial order p = pr = pr,k for all regions r and
unknowns, therefore using a single finite element space Vr,h of dimension (number
k
of degrees of freedom) Nr with Vr,h = Vr,h = Wr,h for all k ∈ K. We remark that
extending the content of the current and subsequent sections to the general case,
with multiple finite element spaces, is straightforward, but cumbersome in terms of
notation. We use Lagrangian nodal basis functions {φr,j }N j=1 of order p (omitting the
r

order in the notation) so that

 
Vr,h = span {φr,j }Nj=1 ,
r

and the real coefficients {[k]nr,j }N n Nr

j=1 and {ϕr,j }j=1 define the discrete approximations
r

(3.8)
Nr
X Nr
X
[k]nr (x) ≈ [k]nr,h (x) = [k]nr,j φr,j (x) and ϕnr (x) ≈ ϕnr,h (x) = ϕnr,j φr,j (x),
j j
8 P. BENEDUSI ET AL

for x ∈ Ωr and n = 0, . . . , Nt .
The potential jump ϕM and gating variables w are approximated with the same
Lagrangian elements restricted to the membrane. To illustrate, we have

X
ϕnM,h (x) = ϕni,h (x) − ϕne,h (x) = ϕnM,j φr,j (x) for x ∈ Γ,
j∈JΓ

for a fixed but arbitrary r, and JΓ = {1 ≤ j ≤ Nr : supp(φr,j ) ∩ Γ ̸= ∅} being the

set of indices corresponding to Γ. When applicable (if gating variables are present),
we solve (2.17) point-wise in space at the coordinates associated with the Lagrange
degrees of freedom and using the coefficient values ϕM,j .

4. Discrete algebraic structure and solver strategy. We now turn to ex-

amine the structure and properties of linear operators associated with the discrete
KNP-EMI equations, before presenting a tailored iterative and preconditioned linear
solution strategy.

4.1. Algebraic structure and matrix assembly. We begin by defining mass

M and (weighted) stiffness matrices A for the concentrations and potentials in each
of the bulk regions Ωr :

Z Nr
Mr = φr,j (x)φr,l (x) dx ∈ RNr ×Nr ,
Ωr j,l=1
Z Nr
Ar = ∇φr,j (x) · ∇φr,l (x) dx ∈ RNr ×Nr ,
Ωr j,l=1
Z Nr
Ak,n
r = [k]nr (x)∇φr,j (x) · ∇φr,l (x) dx ∈ RNr ×Nr ,
Ωr j,l=1

We next define the membrane (mass) matrices:

Z Nr
Mr,Γ = φr,j (x)φr,l (x) ds ∈ RNr ×Nr ,
Γ j,l=1
Z (Nr ,Nq )
Mrq,Γ = φr,j (x)φq,l (x) ds ∈ RNr ×Nq ,
Γ j,l=1

with q = {i, e}/r. Note that the membrane operators, denoted by the subscript Γ,
have non-zero elements only for (j, l) entries corresponding to basis functions with Γ
in their support. For example,

Γ ∩ supp(φr,j ) ∩ supp(φr,l ) ̸= ∅ ⇒ [Mr,Γ ]j,l ̸= 0.

Therefore, membrane operators can be seen as low-rank perturbations of bulk ones.

To recast the discrete KNP-EMI equations (3.5)–(3.7) as a linear system, we define
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 9

the non-symmetric matrix1 corresponding to Ωr , given τrk = ∆tDrk and τ̃rk = τrk zk ψ −1 :

Mr + τr1 Ar 0 ··· 0 Cr1,n Mr,Γ + τ̃r1 A1,n

 
r
..
Mr + τr2 Ar Cr2,n Mr,Γ + τ̃r2 A2,n
 
 0 0 . r 
.. .. ..
 
n
Ar =  . ,
 
. 0 0 .
|K| |K|,n |K| |K|,n
 

 0 ··· 0 Mr + τr Ar Cr Mr,Γ + τ̃r Ar 

|K|
z1 τr1 Ar z2 τr2 Ar ··· z|K| τr Ar Cm F −1 Mr,Γ + Σk zk τ̃rk Ak,n
r

with size Nr (|K| + 1). More compactly,

A11 A1ϕ
 
r r
 A22
r A2ϕ
r   cc
Acϕ


.. .. 
Ar
Anr = . = r
,
 
. 
Aϕc
Aϕϕ
 |K||K| |K|ϕ
 r r n
 Ar Ar 
ϕ|K|
Aϕ1
r Aϕ2
r ··· Ar Aϕϕ
r n
Nr |K|×Nr |K| Nr ×Nr
with Acc
r ∈ R and ∈ R Aϕϕ
rthe concentrations and potentials
blocks, respectively. The last block-column of Anr is time-dependent and must be
updated in each time step tn .
The block diagonal structure of Acc
r highlights that the different concentrations
are coupled only via the potential. We define the following coupling matrices:

−Cr1,n Mrq,Γ
 
0 ··· 0
 0 ··· 0 2,n
−Cr Mrq,Γ 
. . ..
 
n

Brq =  .
 . .
. .

,

 0 · · · 0 −C |K|,n M 
 r rq,Γ 

0 · · · 0 −Cm F −1 Mrq,Γ
with size Nr (|K| + 1) × Nq (|K| + 1), or, more compactly,
cϕ
 
n 0 Brq
Brq = ϕϕ ,
0 Brq n
cϕ
with Brq ∈ RNr |K|×Nq , Brq
ϕϕ
∈ RNr ×Nq . The global matrix at time tn , coupling
variables in Ωi and Ωe , thus reads
 
Acci Acϕ
i 0 cϕ
Bie
 ϕc
Aϕϕ ϕϕ 
 n n

A Bie Ai 0 Bie
An = ni n =
i
 .

cϕ
Bei Ae  0 Bei Ae cc cϕ
Ae 
ϕϕ
0 Bei Aϕc
e Aϕϕ
e n

Let us remark that for multiple cells, i.e. Ncell > 1, Ani can be further decomposed in
Ncell diagonal blocks coupled with Ane via the coupling matrices Brq n
.
Finally, we define the solution vectors:
|K|,n
(4.1) ck,n
r = [[k]nr,1 , . . . , [k]nr,Nr ] ∈ RNr , cnr = [c1,n 2,n
r , cr , . . . , cr ] ∈ RNr |K| ,
1 With a minor abuse of notation, we use k both to identify an ionic species, i.e. k ∈ K and as
an index for elements K, i.e. k = 1, ..., |K|.
10 P. BENEDUSI ET AL

(4.2) ϕnr = [ϕnr,1 , ..., ϕnr,Nr ] ∈ RNr ,

and recast the discrete variational problem (3.5)–(3.7) as the following linear system
of size N = (Ni + Ne )(|K| + 1), for n = 1, . . . , Nt :
 
Acϕ cϕ  n   c,n 
Acc
i i 0 B ie ci fi
 ϕc
Ai Aϕϕ
i 0 B ϕϕ 
ie
ϕn  f ϕ,n 
 ni  =  i  , ⇐⇒ An−1 un = fn ,
(4.3)

cϕ  ce   f c,n 
Bei Acc Acϕ
 
 0 e e  e
0 ϕϕ
Bei Aϕc Aϕϕ ϕne feϕ,n
e e n−1
h iT
|K|,n
with frc,n = fr1,n , fr2,n , ..., fr ∈ RNr |K| and
Z Z Nr
frk,n = ([k]n−1
r (x) + frk,n (x))φr,j (x) dx − grk,n−1 (x)φr,j (x) ds ∈ RNr ,
Ωr Γ j=1
 Z Nr
frϕ,n = − hn−1 (x)φr,j (x) ds ∈ RNr .
Γ j=1

Let us remark that the choice of An−1 , instead of An , makes problem (4.3) linear.
Remark 4.1. The linear operator An in (4.3) admits a one-dimensional nullspace
corresponding to the constant potential (see Section 2.3). There are several approaches
to eliminate this nullspace, such as enforcing a zero-mean integral constraint via a
Lagrange multiplier [22]. However, this approach leads to a dense row and column
in An , severely affecting parallel matrix assembly and iterative solver performance.
Instead, we here, for the sake of simplicity and efficiency, introduce a point-wise
Dirichlet condition on ϕe : given an arbitrary xe ∈ Ωe , we impose ϕe (xe , t) = 0 for
all t ∈ (0, T ]. A third approach would be to prescribe the discrete nullspace in the
iterative solution strategy.
4.2. Solution and preconditioning strategy. To solve the non-symmetric
linear system (4.3), we employ a preconditioned GMRES method using the block
diagonal preconditioner
(4.4)
 11
Ai

 cc  .. 
.

Ai 0 0 0  
ϕϕ ϕϕ
 
 0 Ai 0 0   Ai  ∈ RN ×N ,

Pn =  
cc =
11
0 0 Ae 0 
 Ae 

ϕϕ
0 0 0 Ae n   .. 
. 
Aϕϕ
e n

composed of 2(|K| + 1) blocks, decoupling all concentrations and potentials. In case

of several disjoint intracellular domains, the intracellular blocks can be further subdi-
vided into block-diagonal matrices, with Ncell blocks each. By definition, Pn is sym-
metric and positive definite and corresponds to a simplified problem where Ωi and
Ωe are decoupled and the continuity equation contains only diffusive contributions
and no electric drifts, with each concentration evolving independently. In relation to
Remark 4.1, the constant potential null space has to be eliminated also for Pn .
In the simpler EMI setting, the elimination of off-diagonal blocks can be mo-
tivated by a spectral argument [11], demonstrating that the membrane terms are
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 11

Geometry Domains Dynamics Stimulus N (max)

Na
A Idealized ICS, ECS HH gstim ̸= 0 8.7 · 106
B Imaging Astrocyte, ECS Kir–Na/K feNa , feK ̸= 0 5.3 · 106
Na
C Imaging Dendrite, astrocytes, ECS Kir–Na/K gstim ̸= 0 2.0 · 106
Na
D Imaging Neurons, astrocyte, ECS HH, Kir–Na/K gstim ̸= 0 1.0 · 108

Table 1: Overview of model scenarios, labeled Models A–D. The idealized Model A
includes both two-dimensional and three-dimensional (d = 2, 3) setups. Models B–
D employ different image-based 3D geometries, including one or more astrocytes or
partial astrocytic processes, neurons and neuronal compartments (dendrites, axons,
spines) at different scales in addition to the contiguous extracellular space (ECS).
The ICS in Model A and the neurons in Model D are modelled using Hodgkin-Huxley
(HH) membrane dynamics (Section 2.4.1). The astrocyte membranes in Models B–D
and the dendrite membrane in Model C are modelled using the passive Kir–Na/K
model (Section 2.4.2). Each scenario including the stimulus are described in further
detail in the respective model sections.

zero-distributed ; i.e., represent a negligible contribution to the eigenvalues distribu-

tion for a mesh resolution increasing uniformly in space. As further motivation for
this choice, numerical experiments suggest that diffusion dominates the drift induced
by gradients in the electric potentials, at least for k ∈ {Na+ , K+ } at the scale and
tissue types under consideration [51]. While the extension of the spectral theory to
the KNP-EMI case has still to be developed, the numerical experiments presented in
the subsequent sections show that Pn is an effective and robust preconditioner. Let
us note that, for physiologically relevant parameters, the condition number of An can
exceed 1010 , resulting in stagnation of GMRES if no preconditioning is adopted.
In practice, the action of Pn−1 (which can be efficiently computed block-wise with
a preconditioned CG) can be approximated by a single algebraic multigrid (AMG)
iteration applied monolithically [10]. Moreover, we may use P0 as a preconditioner
for all time steps n = 1, . . . , Nt . This reduces assembly times, with little or no impact
on GMRES convergence as compared to using the time-dependent preconditioner Pn .
5. Idealized and image-based brain tissue benchmark scenarios. In this
section, we present four model scenarios for ionic electrodiffusion in brain tissue,
including neuronal (somatic, dendritic, axonal), glial (astrocytic) and extracellular
spaces at the µm scale (Table 1). These scenarios, along with sample simulation
results, are presented in some detail, with the idea that the models define and can
be reused for future benchmarks. A list of physical parameters and initial conditions
common for all models is provided in Table 2. We remark that this setting can be
further refined considering heterogeneous initial conditions and physical parameters,
depending on the cell type, and heterogeneous membrane properties within a single
cell [48]. All scenarios consider the ionic species K = {Na+ , K+ , Cl− }, while frk = 0
in (2.1) for all r and k unless otherwise indicated. We denote spatial coordinates by
x = (x, y, z). We also refer to the associated open software repository [12].
5.1. Model A. We consider an idealized geometry for a single cell represented by
a d-dimensional cuboid Ω = [0, 1]d µm, with Ωi = [0.25, 0.75]d µm and d = 2, 3 (Fig-
12 P. BENEDUSI ET AL

Parameter Symbol Value Unit Ref.

gas constant R 8.314 J/(K mol)
temperature T 300 K
Faraday constant F 9.648 · 104 C/mol
membrane capacitance CM 0.02 F
Na+ diffusion coefficient DrNa 1.33 · 10−9 m2 /s [26]
K+ diffusion coefficient DrK 1.96 · 10−9 m2 /s [26]
Cl− diffusion coefficient DrCl 2.03 · 10−9 m2 /s [26]
Na+ leak conductivity Na
gleak 1 S/m2
K+ leak conductivity K
gleak 4 S/m2
Cl− leak conductivity Cl
gleak 0 S/m2
K+ HH max conductivity ḡ K 360 S/m2 [27]
Na+ HH max conductivity ḡ Na 1200 S/m2 [27]
stimulus factor ḡstim 40 S/m2
stimulus time constant a 0.002 s
initial ICS Na+ concentration [Na+ ]0i 12 mM [44]
initial ECS Na+ concentration [Na+ ]0e 100 mM [44]
initial ICS K+ concentration [K+ ]0i 125 mM [44]
initial ECS K+ concentration [K+ ]0e 4 mM [44]
initial ICS Cl− concentration [Cl− ]0i 137 mM [44]
initial ECS Cl− concentration [Cl− ]0e 104 mM [44]
initial membrane potential ϕ0M −67.74 mV
resting membrane potential ϕrest −65 mV
initial Na+ activation m0 0.0379 [27]
initial Na+ inactivation h0 0.688 [27]
initial K+ activation n0 0.276 [27]
maximum pump rate ρpump 1.115 · 10−6 mol/m2 s [25]
ICS Na+ threshold for Na/K-pump PNa 10 mol/m3 [25]
ECS K+ threshold for Na/K-pump PK 1.5 mol/m3 [25]

Table 2: Physical parameters and initial values, based on [22].

ures 1a–1b). We consider the active Hodgkin-Huxley membrane model (Section 2.4.1)
and, as stimulus, impose a time-periodic sodium current on the entire membrane Γ:

(5.1) Na
gstim = ḡstim e−(t mod τ )/a ,

with τ = 10 ms the time interval between subsequent stimuli (Figure 1c).

We discretize Ω with a uniform grid with Nx intervals per side and subdivide
each cube into 2 triangles in 2D (Figure 1a) or 6 tetrahedra

in 3D (Figure 1b).
This tessellation results in N = (pNx + 1)d + o(pd Nxd ) (|K| + 1) total degrees of
freedom, with p the Lagrange finite element order. The lower order term o(pd Nxd ) =
(3 − d/2)(pNx )d−1 + 2(d − 2) appears since the degrees of freedom on Γ are repeated
for Ωi and Ωe . Figures 1d–1f show the evolution of concentrations and potential over
time for Nt = 300 time steps, Nx = 64, and p = 1.
5.2. Model B. We examine the electrodiffusive response of an astrocyte (a heav-
ily branching-type of glial cell) in response to local changes in extracellular potassium
and sodium concentrations resulting e.g. from surrounding neuronal activity. The
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 13

Ti
Γ
∂Ω
Te
(a) 2D geometry (b) 3D geometry (c) Stimulus

(d) Sodium (e) Potassium (f) Membrane potential

Fig. 1: Model A. (a) Sample 2D geometry (Nx = 4). (b) Sample 3D geometry
Na
(Nx = 20). (c) Membrane current stimulus (acting on Γ) gstim versus time t. (d)
Evolution of ion concentrations over time. Extracellular and intracellular quantities
are sampled respectively at xe = (0.15, 0.15) µm ∈ Ωe and xi = (0.5, 0.5) µm ∈
Ωi . (f) Evolution of membrane potential, which is homogeneous in Γ, for different
ionic models: the default Hodgkin-Huxley (HH); the Kir-Na/K model; a leak model
obtained by setting f Kir = jpump = 0 in the Kir-Na/K model.

intracellular astrocytic domain Ωi is described by a 3D geometry obtained via Ul-

traliser [1, 2], with a diameter of approximately 90 µm, centered at the origin. The
extracellular domain Ωe is constructed as a 2 µm thick layer enclosing Ωi (Figure 2a).
The astrocytic membrane dynamics are governed by the passive Kir–Na/K membrane
model (Section 2.4.2) over the entire interface Γ. As stimulus, we consider a bolus
injection of potassium and removal of sodium in a section of the extracellular space
over a time period of 2 ms:

feNa (x, t) = −1, feK (x, t) = 1, x ∈ Ωe s.t. x < 0, 0 < t ≤ 2 ms,

and frk = 0 otherwise, cf. Figure 2b. No membrane stimulus is applied (gstim
Na
= 0).
The geometry is described by a tetrahedral mesh with Ni = 475 206 and Ne =
860 025 vertices, and a total problem size of N ≈ 5.3 · 106 for p = 1. Figures 2c–2g
show the evolution of concentrations over time for ∆t = 0.1 ms and Nt = 2000.
14 P. BENEDUSI ET AL

2.0e+00

1.8
10 μm
1.6

label
1.4

1.2

(a) Astrocyte geometry (b)1.0e+00

Stimulus

(c) [K+ ]i at t = 30 s and x = 0 plane (d) [K+ ]i at t = 200 ms

(e) Sodium (f) Potassium (g) Chloride

Fig. 2: Model B. (a) Astrocyte geometry with corresponding tetrahedral mesh. (b)
Ionic source terms over time, imposed for x < 0. (c),(d) Intracellular potassium [K]i
(mM) at two different time steps. Potassium is injected in the extracellular space for
x < 0. At t = 2 ms the extracellular potassium [K]e is increased by approximately
10 mM w.r.t. the initial condition [K]0e = 4 mM. As the system evolves, potassium
is electro-diffusing in both Ωi and Ωe through Γ. (e)–(g) Time evolution of ionic
concentrations [k]r (x⋆ , t) (mM) with x⋆ ∈ Γ labeled by ⋆ in panel (c).
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 15

5.3. Model C. This model focuses on the electrodiffusive interplay between

neuronal structures, the extracellular space and surrounding glial structures at the
scale of dendritic spines. We consider a 3D geometry (Figure 3a) representing a
neuronal dendrite segment with multiple dendritic spines, surrounded by extracellular
space and glial cells, all immersed in a 8 × 8 × 5 µm cuboid [33, 9]. We remark that
even this highly non-trivial geometry represents a substantial simplification of the
dense tissue structure, and in particular that a large number of glial structures are
ignored. We apply the passive Kir–Na/K dynamics model (Section 2.4.2) at both the
astrocyte–ECS and dendrite–ECS interfaces, though with fKir = jpump = 0 for the
latter, thus in essence only considering leak channel dynamics on the dendrite–ECS
interfaces. To trigger a system response, we apply sodium membrane currents at the
dendritic spine heads Γhead ⊂ Γ (marked in green in Figure 3a), mimicking synaptic
signalling:
(
Na 2ḡstim if x ∈ Γhead ,
(5.2) gstim (x) =
0 if x ∈
/ Γhead ,

We consider a tetrahedral mesh with Ni = 193 129 and Ne = 306 478 vertices, and a
total problem size of N ≈ 2.0 · 106 for p = 1. In Figures 3b–3f, we show the resulting
concentrations and potential for Nt = 400 time steps and ∆t = 0.1 ms.
5.4. Model D. We examine electrodiffusive neuronal–ECS–astrocyte interplay
at a larger spatial scale. From a 100 × 100 × 80 µm portion of a rat somatosensory
cortex [2, 16], we select three neurons and four astrocytes and embed these in a cuboid
extracellular space (Figure 4a). For the neuronal membranes, we apply the active
Hodgkin-Huxley model (Section 2.4.1) with a periodic, localized stimulus current

(5.3) Na
gstim = 10 · ḡstim e−(t mod τ )/a for x < 10 µm,
Na
and gstim = 0 otherwise. The period τ = 10 ms. For the astrocytic membranes, we
Na
use the passive Kir–Na/K-model as previously and gstim = 0 (Section 2.4.2).
We consider tetrahedral meshes with up to Ni = 9 390 510 and Ne = 15 541 672
and problem size N ≈ 1.0 · 108 . The numerical solution is shown in Figures 4b-4f.
6. Numerical experiments. In this section, we examine the robustness and
efficiency of the proposed monolithic numerical strategy in terms of iteration count
and parallel performance when applied to Models A–C. All numerical experiments are
performed on the Norwegian supercomputer Saga2 , using up to 256 Intel Xeon-Gold
6138/6230R 2.0-2.1 GHz CPU cores with 40 cores per node.
6.1. Implementation and numerical verification. Our KNP-EMI solver
implementation [12] is based on the well-established FEniCS finite element soft-
ware [5, 37], using the multiphenics library [8] to handle variational problems coupled
across multiple subdomains. The linear algebra backend is PETSc [7]. The imple-
mentation is flexible in terms of ionic models of arbitrary complexity, and has been
verified using the method of manufactured solutions for the same benchmark prob-
lems as introduced in [22, Section 3.1], obtaining the expected convergence rates.
The image-based volumetric meshes (in Models B–D) are generated from segmented
microscopy data via unions of lower-dimensional surfaces using the computational
geometry library fTetWild [28].
2 https://documentation.sigma2.no/hpc machines/saga.html
16 P. BENEDUSI ET AL

Extracellular space (ECS)

Dendritic spines

Dendrite

Glial cells (astrocytic)

1 μm

(a) Geometry

(b) [K+ ]i in Γ at t = 20 ms (c) [K+ ]i in Γ at t = 40 ms

(d) Sodium (e) Potassium (f) Membrane potential

Fig. 3: Model C. (a) Dendritic segment (red), spines heads Γhead (green), and glial
cells (blue). A portion of the meshed extracellular domain is shown as long as a
particular of few meshed dendritic spines. (b),(c) [K]i (mM) for two different time
steps, given a membrane stimulus located in the spines heads. (d)–(f) Concentrations
[k]r (x⋆ , t) (mM) for k ∈ {K+ , Na+ } and potential ϕM (x⋆ , t) (mV) with x⋆ ∈ Γhead .
The location of x⋆ ∈ Γhead is labeled in panels (b) and (c) with ⋆.
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 17

10 μm

(a) Geometry and intracellular potassium in Γ at t = 50 ms

(b) Membrane potential at t = 0.5 ms (c) Membrane potential at t = 1 ms

(d) Sodium (e) Potassium (f) Membrane potential

Fig. 4: Model D. (a) Cortex geometry with neurons and astrocytes, with a portion
of meshed ECS. Intracellular potassium at t = 50 ms is shown. (b),(c) Spreading
electric potential on neurons from an initial localized stimulus for two different time
steps. A portion of the plane x = 10 µm is shown, as boundary of the stimulus region.
A threshold potential of 40 mV is highlighted by the color scale transition. (d),(e)
Intracellular concentrations for Nt = 1000. The solution is sampled in the membrane
of a dendrite, a soma, and an astrocyte. Legend is shared between the two images.
(f) Evolution of neuronal membrane potential.
18 P. BENEDUSI ET AL

6.2. Solver configurations. When using a GMRES method to solve the KNP-
EMI system, we restart the algorithm after 30 iterations. For each time step n, we use
the solution at n − 1 as an initial guess, substantially reducing the number of GMRES
iterations. As stopping criterion for the iterative solvers, we consider a tolerance of
10−6 for the preconditioned relative residual. To evaluate P0−1 , we use a single V-cycle
of hypre BoomerAMG [23]. We refer to the associated software repository [12] for
the exhaustive list of AMG parameters3 . In terms of computational time to solve the
system, the GMRES method is preferable compared to BiCGStab for all the reported
experiments. For comparative tests, we also use MUMPS as a direct solver.
To solve the gating variables ODEs (cf. (2.17)), we use Node = 25 intermediate
time steps of the Rush-Larsen method for each time interval, set to ∆t = 0.05 ms
(unless noted otherwise). In terms of runtime, solving the ODEs is negligible compared
to the linear system solve.

6.3. Iterative solver robustness. We begin by studying the robustness of the

proposed preconditioning strategy with respect to the spatial and temporal discretiza-
tion. We consider a single time step for Model A, with d = 2, as a flexible benchmark
example, and vary the time step ∆t, the finite element (polynomial) degree p ∈ {1, 2}
and the mesh resolution Nx ∈ {16, 32, 64, 128, 256, 512}. The resulting linear system
sizes vary from N = 4.6·103 for (Nx , p) = (16, 1) to N = 4.2·106 for (Nx , p) = (512, 2).
For ∆t ∈ {1, 10, 100} (ms), GMRES preconditioned by P0 (LU(P0 )) converges in 3
iterations for all cases. For ∆t ∈ {0.01, 0.1} (ms), each case converges in 4 iterations.
In contrast, the unpreconditioned GMRES method never converges in less than 1000
iterations.
Next, we investigate how to approximate P0−1 efficiently. For larger problems, it
is not convenient to invert P0 to full precision. Instead, a suitable multilevel approxi-
mation of P0−1 can reduce the time-to-solution while retaining the rapid convergence
of the GMRES method. We here compare a set of such approximation strategies to
black box approaches and the exact preconditioners for increasing problem sizes N , cf.
Table 3. We immediately observe that the direct solver and ILU(0) or AMG precondi-
tioners are not viable for larger N . We therefore consider two approaches where P0−1
is evaluated exactly via LU or CG (LU(P0 ) and CG(P0 ), respectively). Here, the CG
solver is applied block-wise in a field-split fashion with 2(|K| + 1) fields, and precondi-
tioned with AMG. In both cases, the convergence behavior is dramatically improved:
the number of iterations no longer increase with the problem size and remains low
(≈ 4). However, we observe only marginal gains in terms of runtime compared to
ILU(0). Finally, we consider strategies based on the multilevel approximation of P0 ,
obtained with a single AMG V-cycle (AMG(P0 )). Besides a monolithic approach,
we consider either block-wise or field-split application of AMG (AMGFS (P0 )). We
observe that both of these latter strategies maintain the robustness in terms of a
constant and low iteration count (≈ 4) for all problem sizes. Moreover, we observe a
significant improvement in run times for larger problem sizes. The best performance
is observed for AMG applied block-wise, with a 70% reduction in runtime compared
to the exact preconditioner (LU(P0 )).
We end this section by examining how the iteration counts and runtimes depend
on time evolution, focusing on the two solution strategies (LU(P0 ) and AMG(P0 )),
now distributed over 32 cores (Figure 5). For both strategies, the number of iterations
remains low (1–6 iterations) over the whole time span simulated. Moreover, the run

3 For 3D runs, we set the threshold for strong coupling to 0.5, instead of the default value (0.25).
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 19

N 17 412 67 588 266 244 1 056 772

(Nx ) (64) (128) (256) (512)
Assembly 2.5 8.2 31.3 123.3
Direct (MUMPS) 2.1 10.0 49.5 259.1
AMG 2.3 [22.5] 24.8 [61.1] 252.5 [157] 3697 [585]
ILU(0) 0.4 [15.6] 2.4 [29.5] 17.1 [63.2] 176.7 [169]
CG(P0 ) 3.0 [4.0] 10.5 [4.0] 39.8 [4.0] 174.4 [4.0]
LU(P0 ) 0.5 [4.0] 3.5 [4.0] 18.1 [4.0] 122.4 [4.0]
AMG(P0 ) 0.8 [4.3] 3.3 [4.1] 13.8 [4.0] 55.2 [4.0]
AMGFS (P0 ) 0.6 [4.0] 2.2 [4.0] 9.0 [4.0] 37.0 [4.0]

Table 3: runtimes (in serial) and when applicable, in square brackets, average number
of preconditioned GMRES iterations to convergence over n = 1, . . . , Nt , for precon-
ditioners (Model A, p = 1, d = 2, Nt = 10 time steps). In the last two rows different
inexact preconditioners are used.

time per time step (after initialization) remains bounded throughout the time course,
with AMG(P0 ) being approximately 50% faster than LU(P0 ) on average.

Fig. 5: Number of iterations (left) and runtimes (right) over Nt = 300 time steps for
Model A with d = 2 and (Nx , p) = (512, 1) (N = 1 056 772), using 32 MPI processes.

6.4. Parallel scalability: strong scaling. We now more carefully investigate

the parallel performance of the more promising preconditioned GMRES strategies
(Figure 6). We first compare their strong scaling (increasing the number of cores
while keeping the problem size fixed) for Models A, B, and C. For the smaller problem
(Model A with d = 2), we compare with the direct solver for reference. Throughout,
we also examine the scalability of the finite element assembly. Initial tests (using
Model A in 2D) indicate that all of the iterative strategies tested scale reasonably
well for a few (4–8) cores (Figure 6a). The field-split AMGFS (P0 ) is the most efficient
for up to 32 cores, but stagnates for larger core counts since each diagonal block is
solved in parallel. The monolithic AMG strategy scales better and outperforms all
other strategies for the Model A, d = 2 test case for 64 cores or more. We therefore
focus our attention on the monolithic AMG in the subsequent experiments. For the
20 P. BENEDUSI ET AL

(a) Model A (d = 2), N = 1.0 · 106 (b) Model A (d = 3), N = 4.2 · 106

(c) Model B, N = 5.3 · 106 (d) Model C, N = 2.0 · 106

Fig. 6: Parallel performance: strong scaling for Models A, B, and C. Plots show
runtimes for iterative solves, finite element assembly and total runtime versus number
of MPI processes (cores). For all models, we use p = 1 and Nt = 20 time steps.
Average AMG(P0 )-GMRES iterations over n = 1, . . . , Nt are indicated for each data
point for Models A (d = 3), B, and C.

larger problems (Models A with d = 3, B and C), the scaling of AMG(P0 ), the
assembly algorithm, and the total runtime are close to ideal for at least up to 256
cores (Figures 6b–6d). We also note that the average number of iterations remain in
the same low range (4–5) varying the number of cores.
We also conduct a focused larger-scale experiment using Model D with up to
108 degrees of freedom to study the effect of bulk- versus interface-dominated mesh
refinement on algorithmic and parallel performance (Table 4). As a baseline, we
consider Model D with the default configuration (Table 4, first column), as described
in Section 5, for 128 and 256 cores. For the case p = 2 (Table 4, second column) with
a 6.8× increase in the number of degrees of freedom and reduced sparsity, assembly
time increases by a factor of 6.3×, while the linear solver time increases by a factor
of 9.3×. On the other hand, after uniform mesh refinement (third column), which
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 21

N 6.7 · 106 4.5 · 107 4.5 · 107 1.0 · 108

Ni + Ne 1.7 · 106 1.7 · 106 1.3 · 107 2.5 · 107
Finite element degree (p) 1 2 1 1
Initialization (128 cores) 20.2 60.5 335.1 991.1
Initialization (256 cores) 19.0 31.5 143.1 317.1
Assembly (128 cores) 82.5 511.5 468.4 1064.1
Assembly (256 cores) 44.0 278.5 263.2 547.2
AMG(P0 ) (128 cores) 20.8 [4.0] 211.6 [4.0] 132.4 [4.6] 441.8 [8.0]
AMG(P0 ) (256 cores) 12.0 [3.9] 111.5 [3.9] 64.9 [3.9] 242.9 [7.8]

Table 4: Scalability and performance for Model D. Computational runtimes and, in

square brackets, average number of iterations to convergence over n = 1, . . . , Nt = 10
time steps. Total number of degrees of freedom N , intracellular and extracellular
degrees of freedom Ni + Ne , finite element (polynomial) degree p, with rows corre-
sponding to 128 and 256 MPI processes as labeled. The initialization time corresponds
to the finite element discretization setup in the first time step, thus becoming negli-
gible for large Nt . The columns give (from left to right) the results corresponding to
Model D with: (i) p = 1, (ii) p = 2, (iii) after uniform mesh refinement, and (iv) after
uniform and then interface-based mesh refinement.

yields the same increase in N and same sparsity pattern, the assembly time increases
by a factor of 5.7× (128 cores) and 6.0× (256 cores), while the GMRES runtime
increases by 6.3× and 5.4×, demonstrating near-optimal or super-optimal scalability
with respect to N . For both cases, we again observe that the average number of
GMRES iterations stays near constant, remaining at 4–5.
Finally, we study the effect of refining the mesh near the interface Γ only. This
scenario challenges the spectral analysis [10] in which the robustness of the precondi-
tioning strategy eliminating membrane blocks is established under the assumption of
uniform refinement. Indeed, the average number of iterations now increases, from 4–5
to 8, with a corresponding increase in computational cost (comparing the third and
fourth columns of Table 4). For all four columns, both the finite element assembly
and the AMG(P0 )-GMRES solver scale near-optimally when increasing from 128 to
256 cores, with an average parallel efficiency of 93% and 94% respectively.

6.5. Parallel scalability: weak scaling. Finally, we investigate the weak scal-
ing (increasing the number of cores and the problem size simultaneously) of the various
solution strategies, returning again to Model A (Figure 7). For small-sized problems
(Figure 7, left), the field-split AMGFS (P0 )-GMRES is the fastest in serial, but, again,
the monolithic AMG(P0 ) outperforms all the other approaches in terms of scaling. For
the moderate-sized problems (Figure 7b, right), near optimal scalability is observed
for the assembly, with essentially a flat curve from 64 to 256 cores. As expected, the
number of AMG(P0 )-GMRES iterations stays near constant (≈ 4), while its runtime
increases by an average factor of 1.2× as the number of cores and degrees of freedom
doubles. The increase in runtime is connected to communication overhead in matrix-
vector products and AMG setup. We refer to [18] for examples of fine tuning AMG
parameters to further improve its weak scalability.
22 P. BENEDUSI ET AL

(a) Model A (d = 2) (b) Model A (d = 3)

Fig. 7: Parallel performance: weak scaling for Model A (p = 1) with (a) d = 2,

Nt = 10, 4.5 · 103 degrees of freedom per core and (b) d = 3 and Nt = 20, and 4.0 · 104
degrees of freedom per core. Average AMG(P0 )-GMRES iterations over n = 1, . . . , Nt
are displayed as well as the total runtime.

7. Discussion, conclusions and outlook. The availability of dense brain tis-

sue reconstructions at an extreme level of detail poses a unique challenge for math-
ematical and computational modelling. Here, we presented a scalable finite element
discretization and solution algorithm to solve the KNP-EMI equations, describing
electrodiffusion, at the level of complexity required for physiologically relevant simula-
tions. In particular, we employ a continuous finite elements discretization of arbitrary
order, with possible discontinuities across interfaces, where the KNP-EMI equations
are coupled with one or more ODEs, modeling complex membrane dynamics. We
describe the algebraic structure of the arising monolithic linear system and present
a tailored multilevel solution strategy, since black-box approaches are unfeasible for
large problems.
We show robustness of the our proposed solution strategy with respect to dis-
cretization parameters, given a near uniform spatial refinement. Strong and weak
scalability of the algorithm are near-optimal, as demonstrated here when running
simulations using from 1 to 256 MPI processes for problem sizes ranging from 4.6 · 103
to 108 . Our motivation for considering this range of parallelism is four-fold: (i) these
problem sizes represent the first relevant stage for resolving the brain tissue geometry
at the tens of micrometers scale; (ii) any viable solution strategy must successfully
address this range; (iii) this is a range of computing resources readily available, in
particular 256 represents the default core limit for the Saga computing system; and
(iv) above these problem sizes, other aspects of the simulation pipeline such as e.g.,
mesh generation, labeling, and refinement appear as non-negligible potential bottle-
necks. However, our numerical results indicate that the solution algorithm can con-
tinue to scale for larger problem sizes and a larger core count. We also remark that
while we have presented physiologically relevant scenarios, more complexity could,
and should, be considered. For instance, more attention to the connectedness or
non-connectedness of the extracellular space and intracellular compartments could
yield even higher fidelity representations. Similarly, while we here consider different
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 23

non-trivial membrane mechanisms for the different cellular types, a higher degree of
specificity would be more biologically meaningful, in particular for the neuronal com-
partments (axons, soma, dendrites). Our simulation code is openly available [12], and
importantly, we envision that the model scenarios presented can be used as stepping
stones for future benchmarks and research in this emerging field.
Acknowledgments. Pietro Benedusi, Ada J. Ellingsrud and Marie E. Rognes
acknowledge support from the Research Council of Norway via FRIPRO grant #324239
(EMIx) and from the national infrastructure for computational science in Norway,
Sigma2, via grant #NN8049K. We thank Dr. Tom Bartol, Computational Neurobiol-
ogy Laboratory, Salk Institute for Biological Studies, CA, USA, for providing imaging
data and surfaces underlying Model C, as well as inspiring discussions on the topic. We
also thank Prof. Corrado Calı̀ and Prof. Abdellah Marwan, for sharing surface repre-
sentations underlying Model B and D and for insightful comments. Special thanks are
also extended to Jørgen Dokken, Miroslav Kuchta, Francesco Ballarin, Lars Magnus
Valnes, Marius Causemann, Rami Masri, Maria Hernandez Mesa, Pasqua D’Ambra,
Fabio Durastante, and Salvatore Filippone for valuable comments and pointers.
Data availability. This manuscript is accompanied by open software and data
sets available at https://doi.org/10.5281/zenodo.10728462 [12].
Conflict of interest. The authors declare that they have no conflict of interest.

REFERENCES

[1] M. Abdellah, J. J. G. Cantero, N. R. Guerrero, A. Foni, J. S. Coggan, C. Calı̀,

M. Agus, E. Zisis, D. Keller, M. Hadwiger, et al., Ultraliser: a framework for creating
multiscale, high-fidelity and geometrically realistic 3D models for in silico neuroscience,
Briefings in Bioinformatics, 24 (2023), p. bbac491.
[2] M. Abdellah, J. J. G. Cantero, N. R. Guerrero, A. Foni, J. S. Coggan, C. Calı̀,
M. Agus, E. Zisis, D. Keller, M. Hadwiger, P. J. Magistretti, H. Markram, and
F. Schürmann, Ultraliser: a framework for creating multiscale, high-fidelity and geomet-
rically realistic 3D models for in silico neuroscience, Sept. 2022, https://doi.org/10.5281/
zenodo.7105941, https://doi.org/10.5281/zenodo.7105941.
[3] A. Agudelo-Toro and A. Neef, Computationally efficient simulation of electrical activity
at cell membranes interacting with self-generated and externally imposed electric fields,
Journal of neural engineering, 10 (2013), p. 026019.
[4] P. Aitken and G. Somjen, The sources of extracellular potassium accumulation in the CA1
region of hippocampal slices, Brain research, 369 (1986), pp. 163–167.
[5] M. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson,
J. Ring, M. E. Rognes, and G. N. Wells, The FEniCS project version 1.5, Archive of
Numerical Software, 3 (2015).
[6] M. Armbruster, S. Naskar, J. P. Garcia, M. Sommer, E. Kim, Y. Adam, P. G. Haydon,
E. S. Boyden, A. E. Cohen, and C. G. Dulla, Neuronal activity drives pathway-specific
depolarization of peripheral astrocyte processes, Nature neuroscience, 25 (2022), pp. 607–
616.
[7] S. Balay, S. Abhyankar, M. F. Adams, S. Benson, J. Brown, P. Brune, K. Buschelman,
E. M. Constantinescu, L. Dalcin, A. Dener, V. Eijkhout, J. Faibussowitsch, W. D.
Gropp, V. Hapla, T. Isaac, P. Jolivet, D. Karpeev, D. Kaushik, M. G. Knepley,
F. Kong, S. Kruger, D. A. May, L. C. McInnes, R. T. Mills, L. Mitchell, T. Munson,
J. E. Roman, K. Rupp, P. Sanan, J. Sarich, B. F. Smith, S. Zampini, H. Zhang,
H. Zhang, and J. Zhang, PETSc Web page. https://petsc.org/, 2023, https://petsc.org/.
[8] F. Ballarin, multiphenics, 2024, https://multiphenics.github.io/.
[9] T. M. Bartol, D. X. Keller, J. P. Kinney, C. L. Bajaj, K. M. Harris, T. J. Sejnowski,
and M. B. Kennedy, Computational reconstitution of spine calcium transients from indi-
vidual proteins, Frontiers in synaptic neuroscience, 7 (2015), p. 17.
[10] P. Benedusi, P. Ferrari, C. Garoni, R. Krause, and S. Serra-Capizzano, Fast paral-
lel solver for the space-time IgA-DG discretization of the diffusion equation, Journal of
24 P. BENEDUSI ET AL

Scientific Computing, 89 (2021), p. 20.

[11] P. Benedusi, P. Ferrari, M. Rognes, and S. Serra-Capizzano, Modeling excitable cells with
the EMI equations: spectral analysis and iterative solution strategy, Journal of Scientific
Computing, 98 (2024), p. 58.
[12] P. Benedusi and M. E. Rognes, Scalable approximation and solvers for ionic electrodiffusion
in cellular geometries, 2024, https://doi.org/https://doi.org/10.5281/zenodo.10728462.
[13] N. Berre, M. E. Rognes, and A. Massing, Cut finite element discretizations of cell-by-cell
EMI electrophysiology models, arXiv preprint arXiv:2306.03001, (2023).
[14] W. F. Boron and E. L. Boulpaep, Medical physiology, Elsevier Health Sciences, 2016.
[15] A. P. Buccino, M. Kuchta, J. Schreiner, and K.-A. Mardal, Improving neural simulations
with the EMI model, in Modeling Excitable Tissue, Springer, 2021, pp. 87–98.
[16] C. Calı̀, M. Agus, K. Kare, D. J. Boges, H. Lehväslaiho, M. Hadwiger, and P. J.
Magistretti, 3D cellular reconstruction of cortical glia and parenchymal morphometric
analysis from serial block-face electron microscopy of juvenile rat, Progress in neurobiology,
183 (2019), p. 101696.
[17] M. Consortium, J. A. Bae, M. Baptiste, C. A. Bishop, A. L. Bodor, D. Brittain,
J. Buchanan, D. J. Bumbarger, M. A. Castro, B. Celii, et al., Functional con-
nectomics spanning multiple areas of mouse visual cortex, BioRxiv, (2021), pp. 2021–07.
[18] P. D’Ambra, F. Durastante, and S. Filippone, AMG preconditioners for linear solvers
towards extreme scale, SIAM Journal on Scientific Computing, 43 (2021), pp. S679–S703.
[19] E. J. Dickinson, J. G. Limon-Petersen, and R. G. Compton, The electroneutrality approx-
imation in electrochemistry, Journal of Solid State Electrochemistry, 15 (2011), pp. 1335–
1345.
[20] A. G. Dietz, P. Weikop, N. Hauglund, M. Andersen, N. C. Petersen, L. Rose, H. Hirase,
and M. Nedergaard, Local extracellular K+ in cortex regulates norepinephrine levels,
network state, and behavioral output, Proceedings of the National Academy of Sciences,
120 (2023), p. e2305071120.
[21] A. J. Ellingsrud, N. Boullé, P. E. Farrell, and M. E. Rognes, Accurate numerical simu-
lation of electrodiffusion and water movement in brain tissue, Mathematical Medicine and
Biology: A Journal of the IMA, 38 (2021), pp. 516–551.
[22] A. J. Ellingsrud, A. Solbrå, G. T. Einevoll, G. Halnes, and M. E. Rognes, Finite element
simulation of ionic electrodiffusion in cellular geometries, Frontiers in Neuroinformatics,
14 (2020), p. 11.
[23] R. D. Falgout and U. M. Yang, hypre: A library of high performance preconditioners, in
Computational Science—ICCS 2002: International Conference Amsterdam, The Nether-
lands, April 21–24, 2002 Proceedings, Part III, Springer, 2002, pp. 632–641.
[24] S. Farina, S. Claus, J. S. Hale, A. Skupin, and S. P. Bordas, A cut finite element method
for spatially resolved energy metabolism models in complex neuro-cell morphologies with
minimal remeshing, Advanced Modeling and Simulation in Engineering Sciences, 8 (2021),
pp. 1–32.
[25] G. Halnes, I. Østby, K. H. Pettersen, S. W. Omholt, and G. T. Einevoll, Electrodiffu-
sive model for astrocytic and neuronal ion concentration dynamics, PLoS computational
biology, 9 (2013), p. e1003386.
[26] B. Hille et al., Ion channels of excitable membranes, vol. 507, Sinauer Sunderland, MA,
2001.
[27] A. L. Hodgkin and A. F. Huxley, A quantitative description of membrane current and its
application to conduction and excitation in nerve, The Journal of physiology, 117 (1952),
p. 500.
[28] Y. Hu, T. Schneider, B. Wang, D. Zorin, and D. Panozzo, Fast tetrahedral meshing in
the wild, ACM Trans. Graph., 39 (2020), https://doi.org/10.1145/3386569.3392385, https:
//doi.org/10.1145/3386569.3392385.
[29] N. M. M. Huynh, F. Chegini, L. F. Pavarino, M. Weiser, and S. Scacchi, Convergence
analysis of BDDC preconditioners for composite dg discretizations of the cardiac cell-by-
cell model, SIAM Journal on Scientific Computing, 45 (2023), pp. A2836–A2857.
[30] K. H. Jæger, A. G. Edwards, W. R. Giles, and A. Tveito, Arrhythmogenic influence of
mutations in a myocyte-based computational model of the pulmonary vein sleeve, Scientific
Reports, 12 (2022), p. 7040.
[31] K. H. Jæger, K. G. Hustad, X. Cai, and A. Tveito, Efficient numerical solution of the EMI
model representing the extracellular space (e), cell membrane (m) and intracellular space
(i) of a collection of cardiac cells, Frontiers in Physics, 8 (2021), p. 579461.
[32] M. Kespe, S. Cernak, M. Gleiß, S. Hammerich, and H. Nirschl, Three-dimensional sim-
ulation of transport processes within blended electrodes on the particle scale, International
 A SCALABLE SOLVER FOR IONIC ELECTRODIFFUSION 25

Journal of Energy Research, 43 (2019), pp. 6762–6778.

[33] J. P. Kinney, J. Spacek, T. M. Bartol, C. L. Bajaj, K. M. Harris, and T. J. Sejnowski,
Extracellular sheets and tunnels modulate glutamate diffusion in hippocampal neuropil,
Journal of Comparative Neurology, 521 (2013), pp. 448–464.
[34] M. Kuchta, K.-A. Mardal, and M. E. Rognes, Solving the emi equations using finite element
methods, Modeling Excitable Tissue: The EMI Framework, (2021), pp. 56–69.
[35] T. Lagache, K. Jayant, and R. Yuste, Electrodiffusion models of synaptic potentials in
dendritic spines, Journal of computational neuroscience, 47 (2019), pp. 77–89.
[36] C. T. Lee, J. G. Laughlin, N. Angliviel de La Beaumelle, R. E. Amaro, J. A. McCammon,
R. Ramamoorthi, M. Holst, and P. Rangamani, 3D mesh processing using GAMer 2 to
enable reaction-diffusion simulations in realistic cellular geometries, PLoS computational
biology, 16 (2020), p. e1007756.
[37] A. Logg, K.-A. Mardal, and G. Wells, Automated solution of differential equations by the
finite element method: The FEniCS book, vol. 84, Springer Science & Business Media,
2012.
[38] Y. Mori, A multidomain model for ionic electrodiffusion and osmosis with an application to
cortical spreading depression, Physica D: Nonlinear Phenomena, 308 (2015), pp. 94–108.
[39] Y. Mori and C. Peskin, A numerical method for cellular electrophysiology based on the
electrodiffusion equations with internal boundary conditions at membranes, Communica-
tions in Applied Mathematics and Computational Science, 4 (2009), pp. 85–134, http:
//dx.doi.org/10.2140/camcos.2009.4.85.
[40] A. Motta, M. Berning, K. M. Boergens, B. Staffler, M. Beining, S. Loomba, P. Hennig,
H. Wissler, and M. Helmstaedter, Dense connectomic reconstruction in layer 4 of the
somatosensory cortex, Science, 366 (2019).
[41] C. Nicholson, G. Ten Bruggencate, H. Stockle, and R. Steinberg, Calcium and potas-
sium changes in extracellular microenvironment of cat cerebellar cortex, Journal of neuro-
physiology, 41 (1978), pp. 1026–1039.
[42] S. Niederer, Regulation of ion gradients across myocardial ischemic border zones: a biophys-
ical modelling analysis, PloS one, 8 (2013), p. e60323.
[43] M. S. Nordentoft, A. Papoutsi, N. Takahashi, M. S. Heltberg, M. H. Jensen, and R. N.
Rasmussen, Local changes in potassium ions modulate dendritic integration, bioRxiv,
(2023), pp. 2023–05.
[44] J. Pods, J. Schönke, and P. Bastian, Electrodiffusion models of neurons and extracellular
space using the Poisson-Nernst-Planck equations—numerical simulation of the intra-and
extracellular potential for an axon model, Biophysical journal, 105 (2013), pp. 242–254.
[45] R. Rasmussen, J. O’Donnell, F. Ding, and M. Nedergaard, Interstitial ions: A key regu-
lator of state-dependent neural activity?, Progress in Neurobiology, (2020), p. 101802.
[46] G. Rosilho de Souza, R. Krause, S. Pezzuto, et al., Boundary integral formulation of
the cell-by-cell model of cardiac electrophysiology, Engineering Analysis with Boundary
Elements, 158 (2024), pp. 239–251.
[47] T. Roy, J. Andrej, and V. A. Beck, A scalable DG solver for the electroneutral Nernst-Planck
equations, Journal of Computational Physics, 475 (2023), p. 111859.
[48] M. J. Sætra, G. T. Einevoll, and G. Halnes, An electrodiffusive neuron-extracellular-glia
model for exploring the genesis of slow potentials in the brain, PLoS Computational Biol-
ogy, 17 (2021), p. e1008143.
[49] M. J. Sætra, A. J. Ellingsrud, and M. E. Rognes, Neural activity induces strongly cou-
pled electro-chemo-mechanical interactions and fluid flow in astrocyte networks and ex-
tracellular space–a computational study, PLOS Computational Biology, (2023), https:
//doi.org/10.1371/journal.pcbi.1010996.
[50] S. Shirinpour, N. Hananeia, J. Rosado, H. Tran, C. Galanis, A. Vlachos, P. Jedlicka,
G. Queisser, and A. Opitz, Multi-scale modeling toolbox for single neuron and subcellular
activity under transcranial magnetic stimulation, Brain stimulation, 14 (2021), pp. 1470–
1482.
[51] A. Solbrå, A. W. Bergersen, J. van den Brink, A. Malthe-Sørenssen, G. T. Einevoll,
and G. Halnes, A Kirchhoff-Nernst-Planck framework for modeling large scale extracel-
lular electrodiffusion surrounding morphologically detailed neurons, PLoS computational
biology, 14 (2018), p. e1006510.
[52] D. Sterratt, B. Graham, A. Gillies, G. Einevoll, and D. Willshaw, Principles of com-
putational modelling in neuroscience, Cambridge University Press, 2023, https://doi.org/
10.1017/CBO9780511975899.
[53] J. Sundnes, B. F. Nielsen, K. A. Mardal, X. Cai, G. T. Lines, and A. Tveito, On the com-
putational complexity of the bidomain and the monodomain models of electrophysiology,
26 P. BENEDUSI ET AL

Annals of biomedical engineering, 34 (2006), pp. 1088–1097.

[54] A. Tveito, K. H. Jæger, M. Kuchta, K.-A. Mardal, and M. E. Rognes, A cell-based
framework for numerical modeling of electrical conduction in cardiac tissue, Frontiers in
Physics, 5 (2017), p. 48.
[55] A. Tveito, K.-A. Mardal, and M. E. Rognes, Modeling excitable tissue: the EMI framework,
Springer, 2021.
[56] V. Untiet, F. R. Beinlich, P. Kusk, N. Kang, A. Ladrón-de Guevara, W. Song,
C. Kjaerby, M. Andersen, N. Hauglund, Z. Bojarowska, et al., Astrocytic chlo-
ride is brain state dependent and modulates inhibitory neurotransmission in mice, Nature
Communications, 14 (2023), p. 1871.
[57] C. Voßen, J. P. Eberhard, and G. Wittum, Modeling and simulation for three-dimensional
signal propagation in passive dendrites, Computing and Visualization in Science, 10 (2007),
pp. 107–121.
[58] K. Xylouris, G. Queisser, and G. Wittum, A three-dimensional mathematical model of ac-
tive signal processing in axons, Computing and visualization in science, 13 (2010), pp. 409–
418.
[59] K. Xylouris and G. Wittum, A three-dimensional mathematical model for the signal propa-
gation on a neuron’s membrane, Frontiers in computational neuroscience, 9 (2015), p. 94.
[60] H. Yang, Z. Qian, J. Wang, J. Feng, C. Tang, L. Wang, Y. Guo, Z. Liu, Y. Yang,
K. Zhang, et al., Carbon nanotube array-based flexible multifunctional electrodes to record
electrophysiology and ions on the cerebral cortex in real time, Advanced Functional Mate-
rials, 32 (2022), p. 2204794.
[61] L. Yao and Y. Mori, A numerical method for osmotic water flow and solute diffusion with
deformable membrane boundaries in two spatial dimension, Journal of Computational
Physics, 350 (2017), pp. 728–746.
[62] E. Zisis, D. Keller, L. Kanari, A. Arnaudon, M. Gevaert, T. Delemontex, B. Coste,
A. Foni, A. Marwan, C. Calı̀, et al., Architecture of the neuro-glia-vascular system,
bioRxiv 2021.01.19.427241, (2021).