Available online at www.sciencedirect.

com
ScienceDirect

Comput. Methods Appl. Mech. Engrg. 368 (2020) 113129

www.elsevier.com/locate/cma

A multimesh finite element method for the Navier–Stokes equations

based on projection methods
Jørgen S. Dokkena,b ,∗, August Johanssona,c , André Massingd,e , Simon W. Funkea
a Simula Research Laboratory, Martin Linges vei 25, 1364 Fornebu, Norway
b Department of Engineering, Cambridge University, Cambridge CB2 1PZ, United Kingdom
c SINTEF Digital, Mathematics and Cybernetics, PO Box 124 Blindern, 0314 Oslo, Norway
d Department of Mathematical Sciences, Norwegian University of Science and Technology, NO-7491, Trondheim, Norway
e Department of Mathematics and Mathematical Statistics, Umeå University, SE-90187, Umeå, Sweden

Received 5 December 2019; received in revised form 4 May 2020; accepted 8 May 2020
Available online 5 June 2020

Abstract
The multimesh finite element method is a technique for solving partial differential equations on multiple non-matching
meshes by enforcing interface conditions using Nitsche’s method. Since the non-matching meshes can result in arbitrarily cut
cells, additional stabilization terms are needed to obtain a stable method. In this contribution we extend the multimesh finite
element method to the Navier–Stokes equations based on the incremental pressure-correction scheme. For each step in the
pressure-correction scheme, we derive a multimesh finite element formulation with suitable stabilization terms. The proposed
scheme is implemented for arbitrary many overlapping two dimensional domains, yielding expected spatial and temporal
convergence rates for the Taylor–Green problem, and demonstrates good agreement for the drag and lift coefficients for the
Turek–Schäfer benchmark (DFG benchmark 2D-3). Finally, we illustrate the capabilities of the proposed scheme by optimizing
the layout of obstacles in a two dimensional channel.
⃝c 2020 Elsevier B.V. All rights reserved.

Keywords: Navier–Stokes equations; Multimesh finite element method; Incremental pressure-correction scheme; Nitsche’s method; Projection
method

1. Introduction
A variety of physical processes that are relevant in science and engineering can be described by partial differential
equations (PDEs). To find numerical approximations to the solution of these PDEs, a wide range of discretization
methods relies on meshes to discretize the physical domain. To be able to obtain high quality approximations of
the physical system, high quality meshes are usually required.
Mesh generation is expensive, both computationally and in terms of human resources as it can require human
intervention. Examples of industrial applications where this is relevant is in the generation of biomedical image
data [1], and for creation of complex components used in engineering [2]. In addition, mesh generation is a particular
∗ Corresponding author at: Department of Engineering, Cambridge University, Cambridge CB2 1PZ, United Kingdom.
E-mail addresses: [email protected] (J.S. Dokken), [email protected] (A. Johansson), [email protected],
[email protected] (A. Massing), [email protected] (S.W. Funke).

https://doi.org/10.1016/j.cma.2020.113129
0045-7825/⃝ c 2020 Elsevier B.V. All rights reserved.
2 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

challenge for problems where the geometry of the domain is subject to changes during the simulation. This occurs
for instance in fluid–structure interaction problems, where mechanical structures can undergo large deformations, but
also in optimization problems where the shape is the design variable. When large domain deformations occur, even
advanced mesh moving algorithms might be pushed beyond their limit, and the only resort is a costly remeshing
step.
One potential strategy to facilitate the mesh generation for complex domains is to design discretization methods
which admit arbitrary shaped, polygonal mesh elements. This idea has led to the development of several new
discretization frameworks, a non-exhaustive list includes discontinuous Galerkin (DG) methods on polyhedral
grids [3–6], the Hybrid-Higher Order method [7,8], the Virtual Element Method [9,10], and Mimetic Finite
Difference method [11,12].
Another alternative to ease the burden of mesh generation is to decouple the geometric description of the physical
domain from the definition of the approximation spaces as much as possible. This can be done by using a union
of overlapping non-matching meshes to represent the computational domain, and this technique has been studied
in a wide variety of settings. In the setting of finite volume and finite difference methods, the idea of decoupling
the computational domain was already studied in the 1980s using domain decomposition techniques. It has later
gone under the name of Chimera [13–15] and Overset [16,17] methods. See also e.g. [18–21] for the finite element
setting. A recent overview may be found in [22].
Several fictitious domain formulations where Lagrange multipliers are used to enforce boundary and interface
conditions have been proposed in literature, see e.g. [23–25]. In [26], enriched finite element function spaces were
introduced to handle crack propagation. This led to the development of XFEM, which has been used for a large
variety of problems [27–31]. As opposed to enriching the finite element function space as in XFEM, the method
proposed in [32,33] uses Nitsche’s method [34] for enforcing boundary and interface conditions weakly. For the
interface problem, two meshes are allowed to intersect, meaning that there is a doubling of the degrees of freedom
in the so-called cut cells. These methods form the basis of CutFEM, see e.g. [35–37] and references therein. Worth
mentioning is that classical DG methods [38], as well as recent formulations of the finite cell method [39–42] and
the immersogeometric analysis technique [43,44] also make use of Nitsche’s method.
There are vast numbers of other methods for problems where the discretization of the computational domain is
based on non-matching meshes, using either finite differences, finite volumes or finite elements. For example, there
is the classical immersed boundary method [45–47], immersed interface methods [48,49] and the s-version of the
finite element method [50,51], to name a few. An overview of recent developments can be found in [52].
For the generality of a method based on overlapping meshes, the placement of the meshes should be arbitrary.
This means that arbitrarily small intersections can occur, which can fatally influence the discrete stability as well
as lead to arbitrarily large condition numbers. One approach to resolve this is suitable preconditioning [53,54]. In
the multimesh FEM, which is the method used in this paper, this is addressed by adding suitable stabilization terms
over the cut cells. As in CutFEM, Nitsche’s method is used to enforce continuity over these artificial interfaces
formed by the intersecting meshes. In [55,56] it is shown that the multimesh FEM is stable both in the sense of
coercivity and condition number for the Poisson problem for arbitrarily many number of intersecting meshes and
with arbitrary mesh sizes. The Stokes equations have been studied in [57,58].
Methods using overlapping meshes, such as the multimesh FEM, offer several potential advantages over standard,
single-mesh techniques. First, complex geometries can be broken down into individual subdomains which are
easier to mesh, and which can be re-used if a component occurs more than once in the geometry. Secondly, the
subdomains can be easily rearranged during a simulation, which can be helpful both during an initial design face, as
well for automated design optimization at a later design phase. Finally, the overlapping mesh method is beneficial
when mesh-deformation or remeshing is necessary, since the deformation or remeshing can be restricted to those
subdomains that require treatment. Overall, this results in a reduction of the computational effort, and preserves the
mesh quality longer compared to mesh deformation on the entire geometry [59].
In this paper, we explore the multimesh FEM in the setting of the Navier–Stokes equations. The Navier–
Stokes equations are non-linear, transient and the pressure and velocity have a non-trivial coupling. Because of
this complexity, a popular approach is to split the problem into several simpler equations which are consistent on
the operator level. The original splitting scheme was proposed by Chorin and Temam [60,61], using an explicit
time stepping. This scheme was later improved by Goda [62] and made popular by Van Kan [63], adding a
correction step for the velocity at each time step. This is known as the incremental pressure-correction scheme
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 3

(IPCS). Following [62], an alternative formulation called the IPCS on rotational form was proposed by Timmermans
et al. [64], avoiding numerical boundary layers. A thorough overview of error-estimates for these splitting schemes
can be found in [65].
The IPCS is composed of three equations,
• the tentative velocity step (a reaction–diffusion–convection equation),
• the pressure-correction step (a Poisson equation),
• the velocity update step (a projection).
In this paper, we present the appropriate multimesh Nitsche and stabilization terms for these three equations.
We present two schemes for multimesh FEM based on IPCS using the second order backward difference and
Crank–Nicolson temporal discretization schemes.
The paper is structured as follows. We review the classical IPCS in Section 2.1 in the setting of the second
order backward difference and Crank–Nicolson temporal discretization schemes. Then, in Section 3 we present
an equivalent formulation for a domain decomposed into M disjoint domains. In Section 4, we present the
multimesh finite element method for each of the steps in the IPCS algorithm. Following, we briefly discuss
some implementation aspects in Section 5. In Section 6, we present several numerical results for two dimensional
domains. First, in Section 6.1, the multimesh IPCS is employed to solve the 2D Taylor–Green flow problem. We
obtain expected spatial and temporal convergence rates. Second, in Section 6.2 the Turek–Schäfer benchmark (DFG
benchmark 2D-3) is presented. We compare lift and drag coefficients for the multimesh IPCS with results from a
standard FEM. Finally, in Section 6.3, we present an application example considering optimization of the position
and orientation of six obstacles in a channel flow. Concluding the paper, Section 7 summarizes and indicates future
research directions.

2. The Navier–Stokes equations and the incremental pressure-correction scheme

This section provides a brief introduction to the incremental pressure-correction scheme (IPCS) [62], which is an
operator splitting scheme for solving the Navier–Stokes equations. We will throughout this paper restrict ourselves
to the setting where the spatial domain Ω ⊂ Rd , d = 2, 3, is stationary and polygonal.
We start by considering the Navier–Stokes equations: Find the velocity field u and pressure field p such that
∂t u + u · ∇u − ν∆u + ∇ p = f in Ω × (0, T ), (2.1a)
∇·u=0 in Ω × (0, T ), (2.1b)
u=g on ∂ΩD × (0, T ), (2.1c)
(ν∇u − p Id) · n = 0 on ∂ΩN × (0, T ), (2.1d)
u = u0 on Ω × {0}. (2.1e)
Here, T > 0 is the end time, ν is the kinematic viscosity and n is the outer normal vector field on the
domain boundary ∂Ω . Vector-valued functions are denoted ∫ partition ∂Ω as ∂Ω = ∂ΩD ∪ ∂ΩN where
in bold. We
∂ΩD ∩ ∂ΩN = ∅. If ∂ΩD = ∂Ω , we also require that Ω p dx = 0 and ∂Ω g(·, t) · n ds = 0 for all t ∈ (0, T ).
∫

2.1. Incremental pressure-correction scheme

In this subsection, we present two IPCS variations based on the second order backward difference (BDF2) and
Crank–Nicolson (CN) approximations of the temporal derivative. IPCS decomposes Eq. (2.1) into three, uncoupled
equations for the velocity un = u(t n ) and pressure p n = p(t n ) for each time step n = 1, . . . , N with t n = t0 + nδt.
We assume that u0 = u0 and p 0 = p(·, 0) are given. In case p 0 is unknown, it can be found by solving
(∇ p 0 , ∇q)Ω = (ν∆u − u · ∇u + f , ∇q)Ω ∀q. (2.2)
This result is obtained by taking the ∇· of the momentum equation to obtain a Poisson problem for the initial
pressure, exploiting that ∇ · ∂t u = 0 thanks to the incompressibility constraint, and finally deriving a corresponding
weak formulation by multiplying with q and integrating over Ω . Suitable boundary conditions for Eq. (2.2) are not
obvious and are discussed for instance in [66–68]. In the case ∂ΩN = ∅, the weak formulation (2.2) is supplemented
with a homogeneous Neumann condition for pressure, which is then consistent with the imposed boundary condition
in the pressure-correction step.
4 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

2.1.1. Second order backward difference scheme (BDF2)

At time-step n + 1, we have the following algorithm.
Step 1 (Tentative velocity step). Find the tentative velocity u∗ solving
3u∗ − 4un + un−1
+ [u∗ · ∇u∗ ] AB − ν∆u∗ + ∇ p n = f n+1 in Ω , (2.3a)
2δt
u∗ = g(·, t n+1 ) on ∂ΩD , (2.3b)
ν∇u · n = p n
∗ n
on ∂ΩN , (2.3c)
∗ ∗ AB
where [u · ∇u ] is a 2nd order approximation of the non-linear convection term based on an Adams–Bashforth
extrapolation [69]. If un−1 is not known, a standard procedure is to perform an initial time-step with IPCS and
implicit Euler discretization of the time derivative and advection terms.
In this paper, we consider operator splitting schemes which use either a fully explicit or a semi-implicit
linearization of the convection term. The fully explicit linearization is
[u∗ · ∇u∗ ] AB = 2un · ∇un − un−1 · ∇un−1 , (2.4)
following [70], while for the semi-implicit linearization we use
[u∗ · ∇u∗ ] AB = (2un − un−1 ) · ∇u∗ . (2.5)
In the remaining parts of the section, either linearization can be employed. In the numerical results, we will explicitly
state which of the Adams–Bashforth approximations is used.
Step 2 (Projection step). Find un+1 and φ such that
3un+1 − 3u⋆
= −∇φ in Ω , (2.6a)
2δt
∇·u n+1
=0 in Ω , (2.6b)
(u n+1 ∗
−u )·n=0 on ∂ΩD , (2.6c)
φ=0 on ∂ΩN , (2.6d)
and set p n+1 = p n + φ. Alternatively, the projection step can be rewritten as a Poisson problem for the
pressure-correction and a subsequent update of the velocity. More precisely, we solve
Step 2a (pressure-correction). Find φ satisfying
3
−∆φ = − ∇ · u⋆ in Ω , (2.7a)
2δt
∇φ · n = 0 on ∂ΩD , (2.7b)
φ=0 on ∂ΩN , (2.7c)
and set p n+1 = p n + φ.
Step 2b (Velocity update step). Finally, we compute the velocity approximation un+1 at t n+1 by
2δt
un+1 = u⋆ − ∇φ. (2.8)
3
2.1.2. Second order scheme using Crank–Nicolson (CN)
An alternative second order accurate scheme is obtained by using Crank–Nicolson for the time discretization:
Step 1 (Tentative velocity step). Find the tentative velocity u∗ solving
u∗ − un 1
+ [u∗ · ∇u∗ ] AB − ν∆(u∗ + un ) + ∇ p n−1/2 = f n+1/2 in Ω , (2.9a)
δt 2
u = g(·, t ) on ∂ΩD ,
∗ n+1
(2.9b)
1
ν∇(u∗ + un ) · n = p n−1/2 n on ∂ΩN , (2.9c)
2
where [u∗ · ∇u∗ ] AB is an Adams–Bashforth approximation of the convection term that can be either explicit
3 1
[u∗ · ∇u∗ ] AB = un · ∇un − un−1 · ∇un−1 (2.10)
2 2
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 5

or implicit
1 1
[u∗ · ∇u∗ ] AB = (3un − un−1 ) · ∇(u∗ + un ). (2.11)
2 2
Step 2a (pressure-correction). Find φ satisfying
1
−∆φ = − ∇ · u⋆ in Ω , (2.12a)
δt
∇φ · n = 0 on ∂ΩD , (2.12b)
φ=0 on ∂ΩN , (2.12c)
n+1/2 n−1/2
and set p =p + φ.
Step 2b (Velocity update step). Finally, we obtain the velocity un+1 at t n+1 by
un+1 = u⋆ − δt∇φ. (2.13)
There are other related splitting schemes, such as the IPCS on rotational form, proposed by Timmermans
et al. [64]. The derivations in the following sections apply to this scheme in a similar way.

3. Incremental pressure-correction scheme for multiple domains

In this section we describe how the IPCS with BDF2 (Eqs. (2.3), (2.7) and (2.8)) and CN (Eqs. (2.9), (2.12) and
(2.13)) is altered by introducing a decomposition of Ω into M overlapping domains. The definitions and notation
are similar to [55,56] and are included here in brevity for completeness.
• Let Ω be composed of M polygonal predomains Ω̂i , i = 1, . . . , M, such that Ω = ∪i=1 M
Ω̂i . The predomains
may intersect each other in an arbitrary way.
• The predomains are placed in an ordering such that we say that Ω̂ j is on top of Ω̂i if j > i. Fig. 1(a) illustrates
such an ordering for three predomains. Here Ω̂1 = Ω .
• Let Ωi be the visible part of Ω̂i defined as
M
⋃
Ωi = Ω̂i \ Ω̂ j , i = 1, . . . , M − 1. (3.1)
j=i+1
M M
Thus, {Ωi }i=1 form a partition of Ω such that Ω = ∪i=1 Ωi and Ωi ∩ Ω j = ∅ if i ̸= j. Furthermore, Ω̂ M = Ω M
since this is the top domain. See Fig. 1(b) for an example.
The predomains Ω̂i intersect each other and form artificial interfaces. We say that these are artificial since they
are only a numerical construction. To define these interfaces we use the following.
• Partition the boundary ∂ Ω̂i into three non-intersecting parts as
∂ Ω̂i = ∂ Ω̂i,D ∪ ∂ Ω̂i,N ∪ ∂ Ω̂i,Γ , i = 1, . . . , M, (3.2)
where ∂ Ω̂i,{D,N} = ∂ Ω̂i ∩ ∂ Ω̂{D,N} . In the configuration shown in Fig. 1(a) we may note that ∂ Ω̂1,Γ = ∅ and
∂ Ω̂i,D = ∂ Ω̂i,N = ∅ for i = 2, 3.
• Let the interface Γi be defined by
M
⋃
Γi = ∂ Ω̂i,Γ \ Ω̂ j , i = 1, . . . , M − 1. (3.3)
j=i+1

and let
Γi j = Γi ∩ Ω j , i > j, (3.4)
be a partition of Γi . Note that it is natural to partition Γi using the domains Ω j below Γi .
We denote the normal ni to be the outer-pointing normal of Ωi .
With these definitions, we now formulate an IPCS such that each step is solved on the visible domains Ωi ,
i = 1, . . . , M, and add interface conditions on Γi j to ensure that the solution is equivalent to applying IPCS to Ω .
6 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

Fig. 1. Visualization of (a) predomains and (b) the corresponding visible part of each predomain.

First, we consider the modified scheme with the BDF2 time discretization. For the tentative velocity step (2.3), this
yields:
3ui∗ − 4uin + uin−1
+ [ui∗ · ∇ui∗ ] AB − ν∆ui∗ + ∇ pin = f n+1 in Ωi , (3.5a)
2δt
ui∗ = g(·, t n+1 ) on ∂Ωi,D , (3.5b)
ν∇ui∗ · ni = pin ni on ∂Ωi,N , (3.5c)
∗
Ju K = 0 on Γi j , (3.5d)
∗
Jν∇u · ni K = 0 on Γi j , (3.5e)

for i, j = 1, . . . , M, i > j. The velocity ui and pressure pi are functions defined on Ωi , and the jump operator is
defined on Γi j as

JvK = vi − v j , i > j. (3.6)

Similarly, for the pressure-correction step, we obtain additional interface conditions

3
−∆φi = − ∇ · ui∗ in Ωi , (3.7a)
2δt
∇φi · ni = 0 in ∂Ωi,D , (3.7b)
φi = 0 in ∂Ωi,N . (3.7c)
JφK = 0 on Γi j (3.7d)
J∇φ · ni K = 0 on Γi j . (3.7e)
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 7

As in the original IPCS, the pressure is then updated with pin+1 = pin + φi for i = 1, . . . , M. Finally, the tentative
velocity update step is
2
uin+1 = ui∗ − δt∇φi , (3.8)
3
for i = 1, . . . , M.
For the IPCS with the Crank–Nicolson time-discretization one obtains identical interface conditions at the
artificial interfaces Γi j as above. The details are hence not presented here explicitly for brevity.

4. Multimesh finite element formulations of the incremental pressure-correction schemes

In this section we explain how to find an approximate solution to the multiple domain IPCS presented in Section 3
using multimesh finite element methods. We begin with reviewing the notation for defining discrete function spaces
spanned by finite elements on multiple meshes, following [55,56].
• Let T̂i be a quasi-uniform [71] mesh of Ω̂i with mesh parameter h i = maxT ∈T̂i diam(T ), i = 1, . . . , M.
• Let
Ti = {T ∈ T̂i : T ∩ Ωi ̸= ∅}, i = 1, . . . , M, (4.1)
be the active meshes. These are of particular importance, since the finite element spaces will be constructed
on these meshes.
• Let
⋃
Ωh,i = T, i = 1, . . . , M, (4.2)
T ∈Ti

denote the active domains, i.e., the domains defined by the active meshes Ti .
Note that Ωh,i typically extends beyond the corresponding domain Ωi , as shown in Fig. 2, since it also includes
all elements that are partially visible. We will denote these parts of the elements as overlaps, and to obtain stable
scheme we will need to define stabilization terms on these overlaps.
• Let Oi denote the overlap domain defined by
Oi = Ωh,i \ Ωi , i = 1, . . . , M − 1, (4.3)
and let
Oi j = Oi ∩ Ω j = Ωh,i ∩ Ω j , i < j, (4.4)
be a partition of Oi . Note that, unlike for the interfaces Γi Eq. (3.3), it is natural to partition Oi using the
domains Ω j above Oi .
Since we have Γi j with i > j (cf. Eq. (3.4)), but Oi j with i < j, the corresponding jump operator Eq. (3.6) on Oi j
is actually
JvK = vi − v j , i < j. (4.5)
However, we will use the same notation for both jump operators, since it will be clear from the context which one
is used.
With this notation, we can now define appropriate finite element spaces. First, we associate with each active
mesh Ti the space of continuous, piecewise polynomials of order k ⩾ 1,
k
Vi,h = {v ∈ C(Ωh,i ) : v|T ∈ Pk (T ) ∀ T ∈ Ti }, i = 1, . . . , M. (4.6)
Then the corresponding multimesh finite element space is simply defined as the direct sum of the individual spaces,
M
⨁
Vhk = k
Vh,i . (4.7)
i=1
Note that this direct sum is the algebraic construction also known as the external direct sum [72]. If the polynomial
order is not important or clear from the context, we simply drop the superscript k.
8 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

Fig. 2. Illustration of the active and overlap domains for a multimesh consisting of three meshes based on the predomains in Fig. 1(a).
Note that the visible domain Ωi is a subset of the active domain Ωh,i , except for the topmost domain, where Ωh,3 = Ω3 . (a) illustrates the
active domain Ωh,1 of T̂1 . (b) illustrates the active domain Ωh,2 of T̂2 . (c) illustrates the overlap domains for all three meshes.

Now the multimesh function spaces for the velocity and pressure are based on the multimesh realization of the
standard inf–sup stable Taylor–Hood velocity and pressure spaces [71]:

V h = [Vhk ]d , Q h = Vhk−1 , (4.8)

g
for k > 1. Moreover, we adopt the notation V h and V 0h to indicate the incorporation of Dirichlet data on the
physical boundaries in the velocity test and trial function spaces. Similar notation will be used for the pressure test
and trial function spaces.
Since the meshes are not disjoint, multimesh functions are multi-valued in the regions where the meshes overlap.
For this reason, one may interpret v ∈ V h as a function in [L 2 (Ω )]d by the inclusion E : V h ↪→ [L 2 (Ω )]d by
(Ev)|Ωi = v i |Ωi , i = 1, . . . , M. In the implementation we define the embedding by selecting the top-most mesh
where there are overlaps, i.e., we let

(Ev)(x) = max v i (x). (4.9)

i, x∈Ωh,i

Corresponding definitions are made for the pressures.

From a theoretical perspective, this does not matter if the method is convergent, but there may be some influence
on accuracy. Best practice should perhaps be to select the finest mesh and do the evaluation there, since the
solution is perhaps more likely to be more correct. However, this has not been thoroughly investigated in this paper.
However, we note that for the two numerical examples involving embedded obstacles in fluid flow, the top-most
meshes, describing the obstacles, have a finer mesh resolution than the background domain. Analogously to the
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 9

jump operator (3.6), we define the average operator as

1
⟨v⟩ = (vi + v j ), (4.10)
2
where vi and v j are finite element functions represented on the active meshes Ti and T j .

4.1. Variational form for the multimesh tentative velocity step

Initially, we will describe the variational form for the tentative velocity step using the BDF2 temporal
discretization and an implicit Adams–Bashforth approximation, as described in Section 2.1.1. Then, we will point
to the differences that occur with a CN discretization (Section 2.1.2).
To derive the variational formulation of the tentative velocity step (3.5), we will use Nitsche’s method [34] com-
monly used in DG-type methods for interface problems to weakly enforce the interface conditions
(3.5d)–(3.5e) over Γi, j . We start by multiplying Eq. (3.5a) with a test function v and integrate over the visible
g
domains Ωi , i = 1, . . . , M yielding the problem: Find u∗ ∈ V h such that for all v ∈ V 0h
M ( M
3u∗ − 4un + un−1
∑ ) ∑
+ [ui · ∇ui ] − ν∆u + ∇ p , v ,v Ω .
∗ ∗ AB ∗ n
( n+1 )
= f (4.11)
i=1
2δt Ωi i=1
i

Next, we integrate the diffusion and pressure terms by parts:

M ( M
3u∗ − 4un + un−1
∑ ) ∑
+ [ui∗ · ∇ui∗ ] AB , v p ,∇ · v Ω
( n )
−
i=1
2δt Ωi i=1
i

(4.12)
M
∑ M ∑
∑ i−1 ∫ M
∑
ν∇u∗ , ∇v J p n ni − ν∇u∗ · ni · vK ds = f n+1 , v .
( ) ( ) ( )
+ Ωi
+ Ωi
i=1 i=2 j=1 Γi j i=1

Note that boundary terms over ∂Ωi,N vanish due to the Neumann condition, and the boundary terms over ∂Ωi,D
vanish since the test functions are zero at these boundaries. Using the identity JabK = JaK ⟨b⟩ + ⟨a⟩ JbK and the
interface-condition Eq. (3.5e), we split Eq. (4.12) into a linear form ltent on the right-hand side
M ( M
4un − un−1
∑ ) ∑
,v ,v Ω
( n+1 )
ltent (v) = + f
i=1
2δt Ωi i=1
i

M M ∑
i−1 (
∑ ∑ ) )
pn , ∇ · v ( p n ni , JvK)Γi j + J p n ni K, ⟨v⟩ Γ ,
( ) ⟨ ⟩ (
+ Ωi
− (4.13)
ij
i=1 i=2 j=1

and an intermediate bilinear form a on the left-hand side,

M M
∑ 3 ( ∗ ) ∑
a(u∗ , v) = u ,v Ω + [ui · ∇ui∗ ] AB , v Ω
( ∗ )

i=1
2δt i
i=1
i

M
∑ M ∑
∑ i−1
ν∇u∗ , ∇v ( ν∇u∗ · ni , JvK)Γi j .
( ) ⟨ ⟩
+ Ωi
− (4.14)
i=1 i=2 j=1

As in classical symmetric interior penalty based methods for diffusion–convection problems [73], we modify the
intermediate bilinear form by adding a symmetry and a penalty term of the form
M ∑
∑ i−1
− Ju K, ⟨ν∇v · ni ⟩ Γ + αt ν ⟨h⟩−1 Ju∗ K, JvK Γ .
{ ( ∗ ) ( ) }
(4.15)
ij ij
i=2 j=1

In a similar fashion, the convection term is brought into skew-symmetric form which results in an additional
convection related interface term, see also Remark 4.1. As a result, we obtain the final, but unstabilized bilinear
10 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

form for the multimesh formulation of the tentative velocity step,

M M M
∑ 3 ( ∗ ) ∑ ∑
atent (u∗ , v) = u ,v Ω + [ui · ∇ui∗ ] AB , v Ω + ν∇u∗ , ∇v Ω
( ∗ ) ( )

i=1
2δt i
i=1
i
i=1
i

M ∑
∑ i−1
( ν∇u∗ · ni , JvK)Γi j + Ju∗ K, ⟨ν∇v · ni ⟩ Γ
{⟨ ⟩ ( ) }
−
ij
i=2 j=1
M ∑
∑ i−1
αt ν ⟨h⟩−1 Ju∗ K, JvK Γ − ( (2un − un−1 ) · ni Ju∗ K, ⟨v⟩)Γi j .
{ ( ) ⟨ ⟩ }
+ (4.16)
ij
i=2 j=1
Note that we so far have followed closely the same approach used in standard DG methods to derive weak
formulations for diffusion–convection boundary and interface problems. But in contrast to, e.g., the classical
symmetric interior penalty method, the meshes in the multimesh mesh can overlap arbitrarily, which can lead to
a number of stability issues in the presence of so-called small cut elements T ∈ Ti where |T ∩ Ω | ≪ |T |. To
counteract small cut element related instabilities, we add a stabilization form in the overlap regions Oi j , which for
atent has the form
M−1 M
∑ ∑ 3
stent (u, v) = βp (JuK, JvK)Oi j + βt (νJ∇uK, J∇vK)Oi j .
{ }
(4.17)
i=1 j=i+1
2δt

A more detailed justification of the overlap stabilizations (4.17) and its counterparts in the following IPCS sub-steps
is given in Section 4.4, including a short discussion regarding the stability parameters αt , β p etc.
g
To this end, the final stabilized multimesh finite element method for the tentative velocity reads: Find u∗ ∈ V h
such that
atent (u∗ , v) + stent (u∗ , v) = ltent (v), ∀v ∈ V 0h . (4.18)
For the Crank–Nicolson scheme (Section 2.1.2), one obtains a similar left hand side of the problem, with different
weights on the diffusive, temporal and convection terms, as shown in Section 2.1.2. Further, due to the centered
difference of the diffusive term, one obtains the following additional terms on the right hand side l(v)
⎛ ⎞
M M ∑ i−1 ∫
1 ⎝∑ ∑
lC N (v) = (−ν∇un , ∇v)Ωi + J(ν∇un · ni ) · vK ds ⎠ . (4.19)
2 i=1 i=2 j=1 Γij

Note that the interface integrals over Γi j are non-zero, due to the multimesh discretization, where Eq. (3.5e) is
enforced weakly.

Remark 4.1. The convection related interface term (4.16) is similar to the central flux formulation in DG
methods [74] for convection problems with a divergence-free velocity field. As the velocity field in (2.5) constructed
from the Adams–Bashforth approximation is typically not divergence free, one might use the Temam device to
guarantee that the convective part of atent is skew-symmetric. A detailed description can be found for instance
in [75] for DG based projection methods. For our BDF2 based multimesh formulation of the tentative velocity step
Eq. (4.16), this amounts to adding
M M ∑
i−1
∑ 1 ∑ 1
((∇ · b)u , v)Ωi −
∗
(JbK · ni , u∗ · v h )Γi j
⟨ ⟩
(4.20)
i=1
2 i=2 j=1
2

with b = 2un − un−1 . Similar adjustments are needed if the Crank–Nicolson time-stepping is used. In our numerical
experiments, the Temam device had very little effect on the stability of the overall formulation and was therefore
not included in most of the presented experiments. See Table 3 for details.

4.2. Variational form for the multimesh pressure-correction step

The pressure-correction equation (3.7) is a Poisson equation, which has been explored extensively in the
multimesh setting, see [56] for an overview. As for the tentative velocity step, we can derive a weak formulation
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 11

for (3.7a)–(3.7e) by first using Nitsche’s method to enforce interface conditions (3.7d)–(3.7e) weakly and then add
suitable overlap stabilization to render the method robust with respect to the overlap configuration. The resulting
stabilized multimesh formulation for the pressure-correction step then reads: Find φ ∈ Q 0h such that
acorr (φ, q) + scorr (φ, q) = lcorr (q), ∀q ∈ Q 0h , (4.21)
where
M
∑ M ∑
∑ i−1
⟨∇q · ni ⟩ , JφK Γ + JqK, ⟨∇φ · ni ⟩ Γ
{( ) ( ) }
acorr (φ, q) = (∇φ, ∇q)Ωi −
ij ij
i=1 i=2 j=1
M ∑
∑ i−1
αc ⟨h⟩−1 JφK, JqK Γ ,
( )
+ (4.22a)
ij
i=2 j=1
M−1
∑ M
∑
scorr (φ, q) = βc (J∇φK, J∇qK)Oi j , (4.22b)
i=1 j=i+1
M
∑ 3 (
∇ · u∗ , q Ω .
)
lcorr (q) = − (4.22c)
i=1
2δt i

Similarly, the CN-discretization of the pressure-correction equations only have a different scaling of the right hand
side, and can be written in the same form as Eq. (4.22).

4.3. Variational form of the multimesh velocity update

Finally, the velocity update step Eq. (3.8) is solved through the following stabilized multimesh version of an
g
L 2 -projection: Find un+1 ∈ V h such that
aup (un+1 , v) + sup (un+1 , v) = lup (v), ∀v ∈ V 0h , (4.23)
where the bilinear forms aup , sup and the linear form lup are given by respectively,
M
∑
aup (un+1 , v) = (un+1 , v)Ωi (4.24a)
i=1
M−1
∑ M
∑
sup (un+1 , v) = β p (Jun+1 K, JvK)Oi j , (4.24b)
i=1 j=i+1
M ( )
∑ 2δt
lup (v) = u∗ − ∇φ, v . (4.24c)
i=1
3 Ωi

As for the pressure-correction scheme, only the right hand side of the equation changes for the CN-scheme.

4.4. On the role of the overlap stabilizations

In this section, we want to provide the reader with some motivation and theoretical background for the choices
of our overlap stabilizations. A main feature in the multimesh finite element method is that the overlapping meshes
can be placed arbitrarily. Consequently, in each mesh Ti except for the top one T M , arbitrarily small cut elements
T ∈ Ti might appear where |T ∩ Ωi | ≪ |T |. As the definitions of atent , acorr and aup involve only integration over
Ωi , their associated norms will not provide enough control over multimesh finite element functions in the presence
of small cut elements. Thus, certain stability properties such as the discrete coercivity of the bilinear forms or the
condition number of the associated system matrices will be highly dependent on the particular overlap configuration.
The role of the overlap stabilizations is to render our formulations geometrically robust with stability properties
similar to the standard, single mesh case, regardless of the particular overlap configuration. This is achieved through
overlap stabilizations as they extend norms relevant for theoretical analysis from the physical domains Ωi to their
12 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

corresponding active mesh Ti . For instance, using multiple applications of the triangle inequality, it can be easily
verified that
M
∑ M
∑ M−1
∑ ∑
ν∥∇v h ∥2L 2 (T ) ∼ ν∥∇v h ∥2L 2 (Ω ) + ν∥J∇v h K∥2Oi j , (4.25)
i i
i=1 i=1 i=1 j=i+1

for some constant which depends only linearly on the number M of overlapping meshes [55]. Having regained
control over ∇v h on the entire active meshes Ti enables us to follow, roughly speaking, the standard analysis
of DG methods. This has successfully been illustrated for multimesh formulations of the Poisson and Stokes
problem [35,55,57], where overlap stabilizations played a crucial role to prove optimal and geometrically robust
estimates for the a priori error and the condition number of system matrix. And it is precisely the Poisson multimesh
method with overlap stabilization from [55] we employed here in Section 4.2 for the multimesh formulation of the
pressure-correction step.
Similarly, in the velocity update Eq. (4.23), a stabilized multimesh L 2 -projection is employed. Here, the key
observation is that
M
∑ M
∑ M−1
∑ ∑
∥v h ∥2L 2 (T ) ∼ ∥v h ∥2L 2 (Ω ) + ∥Jv h K∥2Oi j . (4.26)
i i
i=1 i=1 i=1 j=i+1

It is not difficult to show that as a result of (4.26), the stabilized L 2 -projection has similar properties as involving
standard L 2 -projections on each single mesh Ti . A similar approach was taken in [76]. In particular, the projection
error satisfies the same a priori error estimates as the L 2 -projection, and the corresponding mass matrix is
similarly well-conditioned as a standard finite element mass matrix. Finally, as the system matrix for the multimesh
formulation of the tentative velocity step Eq. (4.23) mainly comprises scaled mass and stiffness matrices, the overlap
stabilization sup is simply a suitably scaled combination of the overlap stabilizations for the pressure-correction and
velocity update multimesh formulations.
As the result of the previous discussion, we can also shed some light on suitable choices for the stability
parameters αt , αc , β p etc. First note that the constant in the norm equivalence in Eqs. (4.25) and (4.26) depend
only linearly on the number M of overlapping meshes. Consequently, setting the overlap stabilization parameters
β p , . . . , βc ∼ 1 in our examples is a reasonable choice. Our method is rather insensitive to the choice of the overlap
stabilization parameters, but for rather larger choices in the order of 10,000, a locking of the finite element functions
in the overlap regions Oi j can be observed, as the finite element functions defined on non-matching meshes are
then “overfitted” in the overlap region through a too strong penalization of the volume jumps on Oi j .
The penalty parameters αt and αc correspond to the penalty parameter in the classical symmetric interior
penalty method [77] and Nitsche-type methods [34,78] which through the trace inverse inequalities depend on
the polynomial order, shape regularity of the mesh, and the dimension of the space Rd [79]. Consequently, αt
and αc inherit the same dependencies, in addition to a mild linear dependency on the number of simultaneously
overlapping meshes. In practice, this means that our choice of αt and αc is similar to the penalization parameters
in Nitsche-type methods on single fitted meshes, with a similar impact on the accuracy of the numerical solution.

5. Implementation and creation of holes

The multimesh finite element method is implemented in FEniCS [80,81]. The implementation and numerical
examples presented in this paper is limited to two dimensional meshes. However, multimesh has been explored in
the setting of three dimensional problems in [57,82,83]. The code to reproduce the numerical results is available
on Zenodo [84].
In the multimesh implementation in FEniCS, the active domains Ωh,i are denoted with cell-markers on T̂i , as
shown in Fig. 2. This implies that we do not alter the meshes, and they include both the active and inactive cells.
The corresponding multimesh function space is built over the whole mesh, and the inactive degrees of freedom are
treated as identity rows in the arising linear systems. A benefit of this approach is that one can change the positioning
of the top meshes, without remeshing the lowermost mesh. Only mesh intersections and new cell-markers have to
be determined.
To be able to efficiently create holes and to simplify the mesh generation, we extend the lowermost domain, such
that Ω ⊆ Ω̂1 (as opposed to Ω = Ω̂1 ). In Fig. 4 an example for such a selection of meshes is visualized, where
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 13

the top mesh describes an elliptic obstacle, and the bottom mesh describes a channel. Then, by changing the status
of the cells that are overlapped or cut by the obstacle from active to inactive, we obtain the meshes describing our
physical domain. A more detailed description of this process can be found in [59].

6. Numerical results
This section presents several validations of the multimesh IPCS from Section 4. First, a Taylor–Green flow
problem with known analytical solution is used to check spatial and temporal convergence for the proposed
multimesh schemes. Then the results for the Turek–Schäfer benchmark are presented and relevant numerical
quantities are compared to values obtained with a single mesh implementation in FEATFLOW [85]. Finally, the
Navier–Stokes multimesh method is used in an optimization setting to demonstrate the flexibility of the proposed
method with regard to larger mesh deformations. We note that all the schemes used in the numerical examples
obeys a Courant–Friedrichs–Lewy condition (CFL) [86] of the form δt ⩽ Ch.

6.1. Taylor–Green flow

This section considers the two dimensional Taylor–Green flow [87], which is an analytical solution to the
Navier–Stokes problem (2.1) given by
( )
ue = − sin(π y) cos(π x)e−2π νt , sin(π x) cos(π y)e−2π νt ,
2 2
(6.1a)
1 ( )
pe = − cos(2π x) + cos(2π y) e−4π νt ,
2
(6.1b)
4
f e = (0, 0)T . (6.1c)
We solve this problem in the domain Ω = [−1, 1]2 , T = 1, with a kinematic viscosity ν = 0.01, and with Dirichlet
boundary conditions
∫ on the entire boundary, i.e., u = ue on ∂Ω . To obtain a unique pressure solution, we further
require that Ω p dx = 0.
The domain Ω was decomposed into three predomains as shown in Fig. 1(a). The Taylor–Hood finite element
pair P2 − P1 was employed unless stated otherwise. For the spatial convergence analysis, the predomains where
meshed with increasing resolution. The cell diameters of the coarsest multimesh are 0.25, 0.177, and 0.259 for the
blue, green, and red predomains, respectively. The resulting mesh is shown in Fig. 2. For each spatial refinement
level, denoted as L x , the cell diameter was halved. Similarly, for the temporal convergence analysis, we define a
sequence of decreasing time-steps. The coarsest time-step used was δt = 0.1. For each temporal refinement level,
denoted as L t , the time-step was halved. The stabilization parameters in Eqs. (4.17), (4.22b) and (4.24c) were set
to αt = αc = 50 and β p = βt = βc = 10. As in traditional Nitsche methods, the α parameters scales with k 2 , k
being the polynomial degree of the function space [74]. For the β parameters, 10 is a common choice in literature.
The initial conditions for u0 , u−1 and p 0 were obtained by interpolating the analytical solution at the appropriate
time steps t = 0 and t = −δt into the corresponding multimesh function space.
To measure the error of the discrete solutions, we define appropriate error norms. Specifically, we consider the
space–time L 2 norm ∥ · ∥ L 2 (Ω)×L 2 (0,T ) and the H 1 -space L 2 -time norm as ∥ · ∥ H 1 (Ω)×L 2 (0,T ) . For the different norms,
0
we expect the following behavior, see [65]:
∥u − ue ∥ L 2 (Ω)×L 2 (0,T ) ≲ (h 3 + δt 2 ), (6.2a)
∥u − ue ∥ H 1 (Ω)×L 2 (0,T ) ≲ (h + δt),
2
(6.2b)
0

∥ p − pe ∥ L 2 (Ω)×L 2 (0,T ) ≲ (h + δt).

2
(6.2c)
The convergence rates are computed as followed. Denote ui, j the discrete velocity solution for the ith spatial
refinement level and the jth temporal refinement level. Then, the spatial convergence rate is computed as
( )
∥ui, j − ue ∥
eocx = log / log(2), (6.3)
∥ui+1, j − ue ∥
the temporal convergence rate is computed as
( )
∥ui, j − ue ∥
eoct = log / log(2), (6.4)
∥ui, j+1 − ue ∥
14 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

and the spatial–temporal convergence rate is computed as

( )
∥ui, j − ue ∥
eocxt = log / log(2). (6.5)
∥ui+1, j+1 − ue ∥
In Eqs. (6.3)–(6.5), ∥ · ∥ denotes either the aforementioned L 2 − L 2 norm or the H01 − L 2 norm. We use the
corresponding definitions of convergence rates for the pressure pi, j at the ith spatial refinement level and the jth
temporal refinement level.
The resulting space–time errors and convergence rates using the P2 − P1 finite element pair and the BDF2-
scheme with an implicit Adams–Bashforth approximation of the convection term, is shown in Table 1. In Table 1(a)
we observe that we obtain the expected temporal convergence rate for the first three refinements with the finest
spatial discretization. Similar observations are for the first three spatial convergence rates with the finest temporal
discretization. We observe a reduction in order of convergence in both eocx and eoct for fine discretizations, as the
temporal and spatial error is of the same order of magnitude. For the pressure, shown in Table 1(b), the convergence
rates have the same behavior. Finally, we note that the temporal convergence rate of the H01 norm of the velocity
in Table 1(c) is heavily influenced by the spatial discretization.
To eliminate spatial discretization errors, we use the same mesh configuration as above, but a higher order
function space pair, P4 − P3. Also, to observe larger temporal changes, we change the temporal discretization to
δt = 0.5, T = 6. The errors, and corresponding convergence rates are visualized in Table 2. Here we observe the
expected temporal convergence rate eoct for all temporal refinement levels for all error measures. Additionally, as
we have eliminated the spatial discretization error, we observe no spatial convergence as the temporal discretization
is of a larger magnitude.
As noted in Remark 4.1, one can ensure that the convective part of the tentative velocity equation is skew-
symmetric by adding the two additional terms Eq. (4.20) to Eq. (4.16), the so-called Temam device. To compare
the effects of the Temam device a convergence study corresponding to Table 1 was performed with this modified
tentative velocity. The results are presented in Table 3, and we observe errors and convergence rates similar to
Table 1. In the remaining numerical examples in the paper, we do not employ a Temam device.
In addition to these results, we have performed the same convergence study for the error at the end time, where
we have obtained similar convergence rates as for the space–time norms. For the Crank–Nicholson scheme with an
implicit Adams–Bashforth approximation, the same convergence study was performed, yielding similar results as
for the BDF2 scheme. We do not present tables for any of these cases in the paper.

6.2. Turek–Schäfer benchmark (flow around a cylinder)

In this section, we consider the Turek–Schäfer benchmark [88] for unsteady flow around a cylinder with Reynolds
number 100 for a fixed time interval T = [0, 8]. The problem consists of a cylinder placed in a channel, as shown
in Fig. 3. The outlet condition is chosen as the natural boundary condition Eq. (2.1d) and the inlet condition is
g(0, y, t) = (4U (t)y(H − y)/H 2 , 0),
where U (t) = 1.5 sin(π t/8). For the other boundary conditions as well as the dimensions, see Fig. 3. The kinematic
viscosity ν = 0.001 and the fluid density ρ = 1.
For this problem, we use a multimesh consisting of two meshes, one describing the channel and one describing
the obstacle, as shown in Fig. 5. The cells of the background mesh that is inside the obstacle are deactivated and not
part of the active mesh as discussed in [59]. The total active degrees of freedom in the velocity and pressure space
is 15,114. There are 1789 deactivated degrees of freedom, due the marking of the obstacle, and cells fully covered
by the top mesh. The meshes are visualized in Fig. 4. We choose the temporal discretization δt = 1/1600, similar
to [88]. We use the same multimesh stabilization parameters as for the Taylor–Green problem in Section 6.1.
For this benchmark, the representative quantities are the drag and lift coefficients over the cylinder for the full
time interval, as well as the pressure difference between (0.15, 0.2) and (0.25, 0.2). In two dimensions, the drag
and lift coefficient can be written as the following [88].
∫
2
CD (u, p, t, ∂Ω S ) = ρνn · ∇u t S (t)n y − p(t)n x ds,
( )
(6.6a)
ρ LUmean ∂Ω S
2
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 15

Table 1
Error and convergence rates for the BDF2-scheme with an implicit Adams–Bashforth approximation with P2–P1 elements. The spatial
convergence rates and corresponding errors for the finest temporal discretization level is highlighted in boldface. Similarly the temporal
convergence rates and corresponding errors for the finest spatial discretization level is highlighted in italics. The combined space–time error
and convergence rates are underlined.
(a) Errors and convergence rates for ui, j with t = (0, 1) in the L 2 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 5 eoct
−3
0 5.07 · 10−2 7.67 · 10−3 4.97 · 10−3 5.03 · 10−3 5.07 · 10−3 5.08 · 10 −
1 5.16 · 10−2 6.56 · 10−3 1.54 · 10−3 1.29 · 10−3 1.31 · 10−3 1.32 · 10−3 1.94
2 5.21 · 10−2 6.32 · 10−3 9.78 · 10−4 3.62 · 10−4 3.27 · 10−4 3.31 · 10−4 2.00
3 5.28 · 10−2 6.24 · 10−3 8.26 · 10−4 1.92 · 10−4 8.95 · 10−5 8.22 · 10−5 2.01
4 5.31 · 10−2 6.43 · 10−3 7.62 · 10−4 1.37 · 10−4 4.40 · 10−5 2.23 · 10−5 1.88
5 5.32 · 10−2 6.84 · 10−3 7.70 · 10−4 1.02 · 10−4 2.85 · 10−5 1.06 · 10−5 1.07
eocx − 2.96 3.15 2.92 1.84 1.43
eocxt − 2.95 2.75 2.35 2.13 2.05
(b) Errors and convergence rates for pi, j with t = (0, 1) in the L 2 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 5 eoct
−3
0 3.49 · 10−2 9.01 · 10−3 4.94 · 10−3 4.54 · 10−3 4.49 · 10−3 4.48 · 10 −
1 3.55 · 10−2 7.59 · 10−3 2.13 · 10−3 1.28 · 10−3 1.18 · 10−3 1.17 · 10−3 1.94
2 3.65 · 10−2 7.26 · 10−3 1.68 · 10−3 5.28 · 10−4 3.24 · 10−4 2.99 · 10−4 1.97
3 3.77 · 10−2 7.18 · 10−3 1.58 · 10−3 4.05 · 10−4 1.31 · 10−4 8.14 · 10−5 1.88
4 3.85 · 10−2 7.29 · 10−3 1.55 · 10−3 3.76 · 10−4 9.94 · 10−5 3.27 · 10−5 1.32
5 3.96 · 10−2 7.51 · 10−3 1.55 · 10−3 3.65 · 10−4 9.18 · 10−5 2.46 · 10−5 0.41
eocx − 2.40 2.28 2.09 1.99 1.90
eocxt − 2.20 2.18 2.05 2.03 2.01
(c) Errors and convergence rates for ui, j with t = (0, 1) in the H01 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 5 eoct
−2
0 1.34 · 100 3.19 · 10−1 8.27 · 10−2 6.40 · 10−2 6.51 · 10−2 6.49 · 10 −
1 1.35 · 100 3.31 · 10−1 7.10 · 10−2 1.93 · 10−2 1.59 · 10−2 1.64 · 10−2 1.98
2 1.33 · 100 3.32 · 10−1 7.37 · 10−2 1.58 · 10−2 4.56 · 10−3 3.94 · 10−3 2.06
3 1.31 · 100 3.31 · 10−1 7.41 · 10−2 1.64 · 10−2 3.64 · 10−3 1.11 · 10−3 1.83
4 1.28 · 100 3.30 · 10−1 7.40 · 10−2 1.66 · 10−2 3.79 · 10−3 8.64 · 10−4 0.36
5 1.27 · 100 3.29 · 10−1 7.38 · 10−2 1.66 · 10−2 3.83 · 10−3 9.02 · 10−4 −0.06
eocx − 1.95 2.16 2.15 2.12 2.09
eocxt − 2.02 2.17 2.17 2.11 2.07

∫
2
CL (u, p, t, ∂Ω S ) = − ρνn · ∇u t S (t)n x + p(t)n y ds,
( )
(6.6b)
ρ LUmean
2
∂Ω S

where u t S is the tangential velocity component at the interface of the obstacle ∂Ω S , defined as u t S = u · (n y , −n x ),
Umean = 1 the average inflow velocity, and L the length of the channel.
The flow and pressure at the final time t = 8 for the implicit Crank–Nicholson scheme is visualized in Fig. 5.
We compare our numerical values with those obtained from the FEATFLOW web page [85]. For this comparison,
we consider two schemes:
• The BDF2 scheme with an explicit Adams–Bashforth approximation
• The Crank–Nicholson scheme with an implicit Adams–Bashforth approximation
The computed drag and lift coefficients and the pressure difference are shown in Fig. 6a, alongside with the data
obtained from FEATFLOW [85]. The absolute error between the multimesh simulation and the FEATFLOW data is also
shown. We observe that the lift coefficient has a slight phase shift and a lower amplitude than the FEATFLOW data.
We also perform an experiment using a finer discretization to closer match the number of degrees of freedom
in [88]. The refined problem now contains a total of 32,271 degrees of freedom and the results are shown in Fig. 6b.
16 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

Table 2
Error and convergence rates for the BDF2-scheme with an implicit Adams–Bashforth approximation with P4–P3 elements. The spatial
convergence rates and corresponding errors for the finest temporal discretization level is highlighted in boldface. Similarly the temporal
convergence rates and corresponding errors for the finest spatial discretization level is highlighted in italics. The combined space–time error
and convergence rates are underlined.
(a) Errors and convergence rates for ui, j with t = (0, 6) in the L 2 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 eoct
−1
0 4.66 · 10−1 4.65 · 10−1 4.65 · 10−1 4.65 · 10−1 4.65 · 10 −
1 1.51 · 10−1 1.51 · 10−1 1.51 · 10−1 1.51 · 10−1 1.51 · 10−1 1.62
2 4.10 · 10−2 4.07 · 10−2 4.06 · 10−2 4.06 · 10−2 4.06 · 10−2 1.89
3 1.08 · 10−2 1.04 · 10−2 1.04 · 10−2 1.04 · 10−2 1.04 · 10−2 1.96
4 3.03 · 10−3 2.64 · 10−3 2.62 · 10−3 2.62 · 10−3 2.62 · 10−3 1.99
eocx − 0.20 0.01 0.00 0.00
eocxt − 1.63 1.89 1.96 1.99
(b) Errors and convergence rates for pi, j with t = (0, 6) in the L 2 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 eoct
0 2.53 · 10−1 2.53 · 10−1 2.53 · 10−1 2.53 · 10−1 2.53 · 10−1 −
1 7.88 · 10−2 7.87 · 10−2 7.87 · 10−2 7.87 · 10−2 7.87 · 10−2 1.68
2 2.12 · 10−2 2.11 · 10−2 2.11 · 10−2 2.11 · 10−2 2.11 · 10−2 1.90
3 5.49 · 10−3 5.42 · 10−3 5.42 · 10−3 5.42 · 10−3 5.42 · 10−3 1.96
4 1.50 · 10−3 1.37 · 10−3 1.37 · 10−3 1.37 · 10−3 1.37 · 10−3 1.98
eocx − 0.13 0.00 0.00 0.00
eocxt − 1.68 1.90 1.96 1.98
(c) Errors and convergence rates for u i, j with t = (0, 6) in the H01 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 eoct
0
0 2.47 · 100 2.44 · 100 2.43 · 100 2.43 · 100 2.43 · 10 −
1 7.52 · 10−1 7.49 · 10−1 7.48 · 10−1 7.48 · 10−1 7.48 · 10−1 1.70
2 2.01 · 10−1 2.00 · 10−1 2.00 · 10−1 2.00 · 10−1 2.00 · 10−1 1.90
3 5.39 · 10−2 5.11 · 10−2 5.10 · 10−2 5.10 · 10−2 5.10 · 10−2 1.97
4 1.97 · 10−2 1.29 · 10−2 1.29 · 10−2 1.29 · 10−2 1.29 · 10−2 1.98
eocx − 0.61 0.00 0.00 0.00
eocxt − 1.72 1.90 1.97 1.98

Fig. 3. Geometrical setup of the Turek–Schäfer benchmark.

Due to the explicit handling of the convection term in Eq. (2.1), we set δt = 1/2000 to ensure that the CFL
condition [86] is fulfilled. We observe in Fig. 6b that the phase shift and dampening disappears for both the
Crank–Nicolson and BDF2 multimesh scheme, and that the error decreases with one order of magnitude.

6.3. Positional optimization of six obstacles

In this section, we use the multimesh Navier–Stokes splitting scheme in an optimization setting to demonstrate
the flexibility of the proposed method with regards to larger mesh deformations. The goal of this section is to find the
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 17

Table 3
Error and convergence rates for the BDF2-scheme with an implicit Adams–Bashforth approximation and a Temam device with P2-P1
elements. The spatial convergence rates and corresponding errors for the finest temporal discretization level is highlighted in boldface.
Similarly the temporal convergence rates and corresponding errors for the finest spatial discretization level is highlighted in italics. The
combined space time error and convergence rates are underlined.
(a) Errors and convergence rates for ui, j with t = (0, 1) in the L 2 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 5 eoct
−3
0 4.78 · 10−2 7.87 · 10−3 4.97 · 10−3 5.03 · 10−3 5.07 · 10−3 5.08 · 10 −
1 4.82 · 10−2 6.47 · 10−3 1.55 · 10−3 1.29 · 10−3 1.31 · 10−3 1.32 · 10−3 1.94
2 4.96 · 10−2 6.12 · 10−3 9.76 · 10−4 3.62 · 10−4 3.27 · 10−4 3.31 · 10−4 2.00
3 5.08 · 10−2 6.00 · 10−3 8.23 · 10−4 1.92 · 10−4 8.95 · 10−5 8.22 · 10−5 2.01
4 5.12 · 10−2 6.17 · 10−3 7.61 · 10−4 1.37 · 10−4 4.41 · 10−5 2.22 · 10−5 1.89
5 5.10 · 10−2 6.58 · 10−3 7.77 · 10−4 1.03 · 10−4 2.85 · 10−5 1.06 · 10−5 1.07
eocx − 2.95 3.08 2.92 1.85 1.43
eocxt − 2.89 2.73 2.35 2.12 2.06
(b) Errors and convergence rates for pi, j with t = (0, 1) in the L 2 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 5 eoct
−3
0 3.59 · 10−2 9.23 · 10−3 4.95 · 10−3 4.54 · 10−3 4.49 · 10−3 4.48 · 10 −
1 3.67 · 10−2 7.85 · 10−3 2.15 · 10−3 1.28 · 10−3 1.18 · 10−3 1.17 · 10−3 1.94
2 3.78 · 10−2 7.53 · 10−3 1.70 · 10−3 5.29 · 10−4 3.24 · 10−4 2.99 · 10−4 1.97
3 3.88 · 10−2 7.45 · 10−3 1.60 · 10−3 4.06 · 10−4 1.31 · 10−4 8.14 · 10−5 1.88
4 3.95 · 10−2 7.57 · 10−3 1.57 · 10−3 3.77 · 10−4 9.94 · 10−5 3.27 · 10−5 1.32
5 4.03 · 10−2 7.79 · 10−3 1.57 · 10−3 3.66 · 10−4 9.19 · 10−5 2.46 · 10−5 0.41
eocx − 2.37 2.31 2.10 1.99 1.90
eocxt − 2.19 2.21 2.07 2.03 2.01
(c) Errors and convergence rates for ui, j with t = (0, 1) in the H01 − L 2 space–time norm.
L t ↓ \L x → 0 1 2 3 4 5 eoct
−2
0 1.32 · 100 3.11 · 10−1 8.26 · 10−2 6.40 · 10−2 6.51 · 10−2 6.49 · 10 −
1 1.32 · 100 3.18 · 10−1 6.98 · 10−2 1.92 · 10−2 1.59 · 10−2 1.64 · 10−2 1.98
2 1.32 · 100 3.17 · 10−1 7.22 · 10−2 1.57 · 10−2 4.56 · 10−3 3.94 · 10−3 2.06
3 1.30 · 100 3.16 · 10−1 7.26 · 10−2 1.63 · 10−2 3.63 · 10−3 1.11 · 10−3 1.83
4 1.28 · 100 3.15 · 10−1 7.24 · 10−2 1.65 · 10−2 3.78 · 10−3 8.63 · 10−4 0.36
5 1.26 · 100 3.15 · 10−1 7.23 · 10−2 1.64 · 10−2 3.82 · 10−3 9.01 · 10−4 −0.06
eocx − 2.00 2.12 2.14 2.10 2.08
eocxt − 2.05 2.14 2.15 2.11 2.07

Fig. 4. The multimesh used for the Turek–Schäfer benchmark.

optimal placement and orientation of 6 obstacles to maximize the drag over the obstacles (6.6a). One application
where such problems arise are in the layout optimization of tidal-turbines [89]. We denote the center of the ith
obstacle by ci and the orientation by θi . The boundary of the ith obstacle is denoted by ∂Ωi . This is formulated
18 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

Fig. 5. The velocity magnitude and pressure field visualized at the end time for the implicit Crank–Nicholson scheme. Note that the values
of the inactive dofs inside the obstacle are (0, 0) for the velocity and 0 for the pressure.

mathematically as: Find ci , θi , i = 1, . . . 6 such that

∫ 1 6
∑
max J (c1 , . . . , c6 , θ1 , . . . , θ6 ) = CD (u, p, t, ∂Ωi ) dt, (6.7a)
c1 ,...,c6 ,θ1 ,...,θ6 0.1 i=1

subject to Eq. (2.1),

∥ci − c j ∥l 2 > di j , i, j = 1, . . . , 6, i ̸= j, (6.7b)
(0, 0) < (l, h) ≤ ci ≤ (l + l1 , h + h 1 ) < (L , H ), (6.7c)
0 ≤ θi ≤ 2π, (6.7d)
where CD is the drag coefficient defined in Eq. (6.6a), di j denotes the minimal distance between the center of the
ith and jth obstacle, [l, l + l1 ] × [h, h + h 1 ] denote the bounded area of the optimization parameters, and L and H
the length and width of the channel, respectively.
For all boundaries but the outlet, we prescribe Dirichlet boundary conditions:
{ (
πt
)
sin 2·0.1 t ∈ [0, 0.1)
g(y, t) = for (x, y) ∈ ∂Ωin ,
1 t ∈ [0.1, 1]
M
⋃
g(y, t) = 0 for (x, y) ∈ ∂Ωw ∂Ωi .
i=1

Here ∂Ωin denotes the inlet (left side) and ∂Ωw denotes the top and bottom wall of the channel. The boundary
condition enforced on the outlet (right side) is the do-nothing condition, as specified in Eq. (2.1d).
The computational domain is described in Fig. 7, where the green area visualizes the box constraint Eq. (6.7c).
The time discretization parameter δt = 0.01, the kinematic viscosity ν = 0.001, the source term f = (0, 0) and
the multimesh stabilization parameters were the same as in the previous examples. The domain parameters in our
example were set to L = 2, H = 1.5, l = 0.3, l1 = 0.9, h = 0.4, h 1 = 0.7, di j = 0.183 for i, j = 1, . . . , 6, i ̸= j.
The obstacles are ellipses with r y = 0.05, r y = 0.025.
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 19

Fig. 6. Numerical results of the Turek–Schäfer benchmark for two different multimesh discretizations (sub-figures a and b). For each
discretization, we compare the BDF2 multimesh scheme with an explicit Adams–Bashforth discretization (MM BDF2), a Crank–Nicolson
multimesh scheme with an implicit Adams–Bashforth discretization (MM CN), and a reference solution computed with FEATFLOW [85] (Turek
Level 4). The plots visualize the drag coefficient CD , the lift coefficient CL and the pressure difference ∆ p = p(0.15, 0.2) − p(0.25 − 0.2)
for t ∈ [0, 8], as well as their absolute errors. We observe that the magnitude of the error in all quantities in Fig. 6b are reduced with one
order compared to Fig. 6a.
20 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

Fig. 7. The physical domain, a channel including 6 obstacles, described by 8 meshes. The background mesh (in black) describes the fluid
channel, where the inlet is at the left side of the channel, the outlet at the right hand side, and rigid walls on the top and bottom. The
second mesh (in green), visualizes the area the obstacles (in red) are bounded to. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

Fig. 8. The functional value for each IPOPT iteration. The sporadic decrease in the functional from one iteration to another is explained
by an increase in the barrier parameter, which are enforcing the non-collision and box constraints.

We use the multimesh Crank–Nicholson splitting scheme with an implicit Adams–Bashforth approximation for
the numerical simulation of the state constraint Eq. (2.1). To solve the optimization problem, we use IPOPT [90]. A
finite difference gradient, with ϵ = 10−3 is supplied to IPOPT. This means that IPOPT will solve the Navier–Stokes
problem to evaluate the functional, and to approximate the derivative of the functional with respect to movement
and rotation. IPOPT uses an interior point line search filter method to find a local maximum of the functional that
satisfies Eqs. (6.7b)–(6.7d). The optimization algorithm was terminated manually after 40 iterations, as no further
increase was observed. After 40 iterations, the functional value J had increased from 2.64 to 13.09. In Fig. 8 the
functional value is plotted against the IPOPT iterations. From iteration 25 to 40, there is less improvement in the
functional value than for the first 25 iterations, suggesting that we are close to a local minimum. Note that sometimes
the functional value decreases from one iteration to another when the barrier parameters change in IPOPT which
enforce Eqs. (6.7b)–(6.7d). Further, due to the finite difference approximation of the gradient, we cannot expect
to converge the discrete local minimum arbitrarily close. The initial and final configuration of the obstacles are
visualized in Fig. 9. We observe that no remeshing or mesh deformation schemes are needed to update the domain,
as they can move independently of each other.
A breakdown of the time-consumption of a forward simulation is visualized in Table 4. The forward simulation
is split into four core components: Each of the steps in the splitting scheme, and the mesh update procedure. Each
of the three steps are then further split into an assembly and solve step, while the mesh update step is split into
the movement of the meshes, and the recomputation of intersections and marking of degrees of freedom inside the
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 21

Fig. 9. The initial and final configuration of the turbines. The final configuration was reached after terminating IPOPT at 40 iterations. Then
the functional had increased from 2.64 to 13.09. The termination is due to the finite difference approximation of the gradient, which is not
discretely consistent without changing the step size in the finite difference operation.

Table 4
Timings for a forward run of the optimization problem with the implicit Adams–Bashforth approximation and Crank–Nicholson temporal
discretization. Each of the three steps of the splitting scheme is split into an assemble and a solve operation. The assembly operation
generates the linear system and applies boundary conditions. The solve operation solves the corresponding linear system. The mesh update
step consists of two operations, translating and rotating all of the meshes, and computing the intersection between the meshes and deactivating
the dofs inside the obstacles. The total time corresponds to the time a full forward simulation with 100 time steps. It is averaged over 5
runs.
Tentative velocity Pressure correction Velocity update Mesh update
Operation Assembly (s) Solve (s) Assembly (s) Solve (s) Assembly (s) Solve (s) Update (s) Intersections (s)
One call 1.71 · 10−1 2.40 · 10−1 2.80 · 10−2 1.54 · 10−2 9.11 · 10−2 1.30 · 10−1 4.63 · 10−5 1.69 · 10−2
Total 4.12 · 101 3.16 · 100 1.42 · 101 1.70 · 10−2

obstacles. We observe that the first step is the most time-consuming step, as we have to reassemble the left hand
side of the linear system for each time step with the implicit Adams–Bashforth approximation. The second-most
expensive step is the velocity update steps, since the velocity function space is higher order than the pressure space.
The mesh update barely takes any time, while the intersection computation, done once per forward run, takes as
much time as solving the pressure-correction equation at a single time step.

7. Conclusions
In this paper, we demonstrated how widely used incremental pressure-correction methods can be combined
with the recently developed multimesh finite element method to solve the Navier–Stokes equations on multiple,
overlapping meshes. Detailed numerical studies of the proposed schemes were conducted using two standard
benchmark problems. For the analytical Taylor–Green vortex problem in two dimensions, the numerical experiments
demonstrated that our multimesh formulations of the incremental pressure-correction method exhibit similar
temporal and spatial convergence rates as their single mesh counterparts. For the Turek DFG-2D benchmark
problem, the obtained drag and lift coefficients, as well as the pressure difference at the finest refinement level
were within the recommended reference interval given in [88]. Moreover, in comparison with the well established
FE-software FEATFLOW [85], we achieved the same level of accuracy for the these benchmark quantities while using
a comparable number of degree of freedoms: ∼32,000 active dofs plus ∼4100 inactive dofs in our formulation vs.
∼42,000 dofs in the single mesh solver FEATFLOW.
In comparison with traditional single mesh flow solvers, a main advantage of the proposed multimesh method is
that it allows for fast manual or automatic modifications of the computational domain, for instance by deforming,
22 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

translating, scaling, rotating or remeshing individual meshes in the multimesh hierarchy. Thanks to the overlap
stabilization, the convergence and stability properties of the resulting multimesh discretization are insensitive to the
particular overlap configuration.
Possible applications include streamlining the mesh generation component of simulation pipelines, as well as
development of new solvers for flow problems where the domain might change substantially in the course of the
simulation. For instance, efficient solvers handling large deformation in fluid–structure interaction problems might be
devised by combining splitting schemes [91,92] with a hybrid Eulerian ALE approach similar to [93,94]. Also shape
optimization for fluid problems as demonstrated in Section 6.3 and [59] can greatly benefit from the more flexible
geometry handling in the multimesh approach. Future work will include a full shape analysis of the proposed method
for the Navier–Stokes problem. In particular, we are interested in the optimize-then-discretize strategy replacing the
finite difference approximation of the shape gradient employed in this paper.
The flexible and automatic handling of complex geometries in the multimesh method is achieved by using
computational geometry routines which automatically compute the relevant intersections. Thus, to provide an
advantageous alternative to single mesh based discretizations, one must assure that the overhead caused by the
additional geometric computations does not outweigh the benefits of the multimesh approach. A breakdown of the
time-consumption presented in Section 6.3 indicates that for 2D positioning optimization problem, the additional
mesh intersection computations introduce only a small overhead compared to the overall simulation time.
Finally, for complex industrial application, a three dimensional implementation of the multimesh finite element
method would be of great interest. Benchmarks for an earlier prototype implementation presented in [95,96]
considered only two overlapping 3D meshes, but clearly demonstrated that an efficient implementation of 3D mesh
intersection algorithms is feasible. For high-performance computations involving a large number of processors, a
major challenge is the distribution of meshes which a priori are unaware of each other, but possible solutions for
similar load balancing algorithms are presented for instance in [97].

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could
have appeared to influence the work reported in this paper.

Acknowledgments
The authors would like to acknowledge Kristian Valen-Sendstad and Alban Souche at Simula Research
Laboratory for many fruitful discussions regarding splitting schemes. This work was supported by the Research
Council of Norway through a FRIPRO grant, project number 251237. André Massing gratefully acknowledges
financial support from the Swedish Research Council under Starting Grant 2017-05038.

References
[1] L. Antiga, J. Peiró, D.A. Steinman, From image data to computational domains, in: Cardiovascular Mathematics, Springer, 2009,
pp. 123–175, http://dx.doi.org/10.1007/978-88-470-1152-6_4.
[2] T.J. Hughes, J.A. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement,
Comput. Methods Appl. Mech. Engrg. 194 (39) (2005) 4135–4195, http://dx.doi.org/10.1016/j.cma.2004.10.008.
[3] P. Antonietti, M. Verani, C. Vergara, S. Zonca, Numerical solution of fluid–structure interaction problems by means of a high order
Discontinuous Galerkin method on polygonal grids, Finite Elem. Anal. Des. 159 (2019) 1–14, http://dx.doi.org/10.1016/j.finel.2019.02.
002.
[4] P. Antonietti, I. Mazzieri, High-order Discontinuous Galerkin methods for the elastodynamics equation on polygonal and polyhedral
meshes, Comput. Methods Appl. Mech. Engrg. 342 (2018) 414–437, http://dx.doi.org/10.1016/j.cma.2018.08.012.
[5] P.F. Antonietti, S. Giani, P. Houston, hp-version composite discontinuous Galerkin methods for elliptic problems on complicated
domains, SIAM J. Sci. Comput. 35 (3) (2013) A1417–A1439, http://dx.doi.org/10.1137/120877246.
[6] A. Cangiani, Z. Dong, E.H. Georgoulis, P. Houston, hp-Version Discontinuous Galerkin Methods on Polygonal and Polyhedral Meshes,
in: Springer Briefs in Mathematics, Springer, 2017, http://dx.doi.org/10.1007/978-3-319-67673-9.
[7] D.A. Di Pietro, A. Ern, A hybrid high-order locking-free method for linear elasticity on general meshes, Comput. Methods Appl.
Mech. Engrg. 283 (2015) 1–21, http://dx.doi.org/10.1016/j.cma.2014.09.009.
[8] D.A. Di Pietro, A. Ern, S. Lemaire, A Review of Hybrid High-Order Methods: Formulations, Computational Aspects, Comparison
with Other Methods, Springer International Publishing, 2016, pp. 205–236, http://dx.doi.org/10.1007/978-3-319-41640-3_7.
[9] L. Beirão da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L. Marini, A. Russo, Basic principles of virtual element methods, Math.
Models Methods Appl. Sci. 23 (01) (2013) 199–214, http://dx.doi.org/10.1142/S0218202512500492.
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 23

[10] L. Beirão da Veiga, F. Brezzi, L. Marini, A. Russo, The Hitchhiker’s guide to the virtual element method, Math. Models Methods
Appl. Sci. 24 (08) (2014) 1541–1573, http://dx.doi.org/10.1142/S021820251440003X.
[11] F. Brezzi, K. Lipnikov, V. Simoncini, A family of mimetic finite difference methods on polygonal and polyhedral meshes, Math.
Models Methods Appl. Sci. 15 (10) (2005) 1533–1551, http://dx.doi.org/10.1142/S0218202505000832.
[12] K. Lipnikov, G. Manzini, M. Shashkov, Mimetic finite difference method, J. Comput. Phys. 257 (2014) 1163–1227, http://dx.doi.org/
10.1016/j.jcp.2013.07.031.
[13] J.L. Steger, F.C. Dougherty, J.A. Benek, A Chimera grid scheme. [multiple overset body-conforming mesh system for finite difference
adaptation to complex aircraft configurations], Advances in grid generation, in: Proceedings of the Applied Mechanics, Bioengineering,
and Fluids Engineering Conference, Houston, TXw, 1983, pp. 59–69, URL https://ntrs.nasa.gov/search.jsp?R=19840028795.
[14] J. Steger, The Chimera method of flow simulation, in: Workshop on Applied CFD, Vol. 188, Univ of Tennessee Space Institute, 1991.
[15] F. Brezzi, J.-L. Lions, O. Pironneau, Analysis of a Chimera method, C. R. Acad. Sci., Paris I 332 (7) (2001) 655–660, http:
//dx.doi.org/10.1016/S0764-4442(01)01904-8.
[16] D.M. Belk, The role of overset grids in the development of the general purpose CFD code, in: Surface Modeling, Grid Generation,
and Related Issues in Computational Fluid Dynamic (CFD) Solutions, 1995, pp. 193–204, URL https://ntrs.nasa.gov/search.jsp?R=
19950022317.
[17] W. Chan, R. Gomez, S. Rogers, P. Buning, Best practices in overset grid generation, in: 32nd AIAA Fluid Dynamics Conference and
Exhibit, 2002, http://dx.doi.org/10.2514/6.2002-3191.
[18] V. Girault, B. Rivière, M. Wheeler, A discontinuous Galerkin method with nonoverlapping domain decomposition for the Stokes and
Navier–Stokes problems, Math. Comp. 74 (249) (2005) 53–84, http://dx.doi.org/10.1090/S0025-5718-04-01652-7.
[19] E. Rank, Adaptive remeshing and h-p domain decomposition, Comput. Methods Appl. Mech. Engrg. 101 (1) (1992) 299–313,
http://dx.doi.org/10.1016/0045-7825(92)90027-H.
[20] R. Becker, P. Hansbo, R. Stenberg, A finite element method for domain decomposition with non-matching grids, ESAIM Math. Model.
Numer. Anal. 37 (2) (2003) 209–225, http://dx.doi.org/10.1051/m2an:2003023.
[21] M. Balmus, A. Massing, J. Hoffman, R. Razavi, D.A. Nordsletten, A partition of unity approach to fluid mechanics and fluid–structure
interaction, Comput. Methods Appl. Mech. Engrg. 362 (2020) 112842, http://dx.doi.org/10.1016/j.cma.2020.112842.
[22] G. Houzeaux, J.C. Cajas, M. Discacciati, B. Eguzkitza, A. Gargallo-Peiró, M. Rivero, M. Vázquez, Domain decomposition methods
for domain composition purpose: Chimera, overset, gluing and sliding mesh methods, Arch. Comput. Methods Eng. 24 (4) (2017)
1033–1070, http://dx.doi.org/10.1007/s11831-016-9198-8.
[23] Q.V. Dinh, R. Glowinski, J. He, V. Kwock, T.W. Pan, J. Périaux, Lagrange multiplier approach to fictitious domain methods: application
to fluid dynamics and electro-magnetics, in: Fifth International Symposium on Domain Decomposition Methods for Partial Differential
Equations, No. 55, SIAM, 1992, pp. 151–194.
[24] R. Glowinski, T.-W. Pan, J. Periaux, A fictitious domain method for Dirichlet problem and applications, Comput. Methods Appl. Mech.
Engrg. 111 (3) (1994) 283–303, http://dx.doi.org/10.1016/0045-7825(94)90135-X.
[25] R. Glowinski, T.-W. Pan, J. Periaux, A Lagrange multiplier/fictitious domain method for the Dirichlet problem — Generalization to
some flow problems, Jpn J. Ind. Appl. Math. 12 (1) (1995) 87, http://dx.doi.org/10.1007/BF03167383.
[26] N. Moës, J. Dolbow, T. Belytschko, A finite element method for crack growth without remeshing, Internat. J. Numer. Methods Engrg.
46 (1) (1999) 131–150, http://dx.doi.org/10.1002/(sici)1097-0207(19990910)46:1<131::aid-nme726>3.3.co;2-a.
[27] A. Gerstenberger, W.A. Wall, An extended finite element method/Lagrange multiplier based approach for fluid–structure interaction,
Comput. Methods Appl. Mech. Engrg. 197 (19–20) (2008) 1699–1714, http://dx.doi.org/10.1016/j.cma.2007.07.002.
[28] U.M. Mayer, A. Popp, A. Gerstenberger, W.A. Wall, 3D fluid–structure-contact interaction based on a combined XFEM FSI and dual
mortar contact approach, Comput. Mech. 46 (1) (2010) 53–67, http://dx.doi.org/10.1007/s00466-010-0486-0.
[29] L. Cattaneo, L. Formaggia, G.F. Iori, A. Scotti, P. Zunino, Stabilized extended finite elements for the approximation of saddle point
problems with unfitted interfaces, Calcolo 52 (2) (2015) 123–152, http://dx.doi.org/10.1007/s10092-014-0109-9.
[30] K. Agathos, E. Chatzi, S.P.A. Bordas, Multiple crack detection in 3D using a stable XFEM and global optimization, Comput. Mech.
62 (4) (2018) 835–852, http://dx.doi.org/10.1007/s00466-017-1532-y.
[31] L. Formaggia, C. Vergara, S. Zonca, Unfitted extended finite elements for composite grids, Comput. Math. Appl. 76 (4) (2018) 893–904,
http://dx.doi.org/10.1016/j.camwa.2018.05.028.
[32] A. Hansbo, P. Hansbo, An unfitted finite element method, based on Nitsche’s method, for elliptic interface problems, Comput. Methods
Appl. Mech. Engrg. 191 (47–48) (2002) 5537–5552, http://dx.doi.org/10.1016/S0045-7825(02)00524-8.
[33] A. Hansbo, P. Hansbo, M.G. Larson, A finite element method on composite grids based on Nitsche’s method, ESAIM-Math. Model.
Numer. 37 (3) (2003) 495–514, http://dx.doi.org/10.1051/m2an:2003039.
[34] J. Nitsche, Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwendung von Teilräumen, die keinen
Randbedingungen unterworfen sind, Abh. Math. Semin. Univ. Hambg. 36 (1) (1971) 9–15, http://dx.doi.org/10.1007/BF02995904.
[35] A. Massing, M.G. Larson, A. Logg, M.E. Rognes, A stabilized Nitsche overlapping mesh method for the Stokes problem, Numer.
Math. 128 (1) (2014) 73–101, http://dx.doi.org/10.1007/s00211-013-0603-z.
[36] E. Burman, S. Claus, P. Hansbo, M.G. Larson, A. Massing, CutFEM: Discretizing geometry and partial differential equations, Internat.
J. Numer. Methods Engrg. 104 (7) (2015) 472–501, http://dx.doi.org/10.1002/nme.4823.
[37] C. Gürkan, A. Massing, A stabilized cut discontinuous Galerkin framework for elliptic boundary value and interface problems, Comput.
Methods Appl. Mech. Engrg. 348 (2019) 466–499, http://dx.doi.org/10.1016/j.cma.2018.12.041.
[38] D.N. Arnold, F. Brezzi, B. Cockburn, L.D. Marini, Unified analysis of discontinuous Galerkin methods for elliptic problems, SIAM J.
Numer. Anal. 39 (5) (2002) 1749–1779, http://dx.doi.org/10.1137/S0036142901384162.
[39] J. Parvizian, A. Düster, E. Rank, Finite cell method, Comput. Mech. 41 (1) (2007) 121–133, http://dx.doi.org/10.1007/s00466-007-
0173-y.
24 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129

[40] M. Ruess, D. Schillinger, Y. Bazilevs, V. Varduhn, E. Rank, Weakly enforced essential boundary conditions for nurbs-embedded
and trimmed nurbs geometries on the basis of the finite cell method, Internat. J. Numer. Methods Engrg. 95 (10) (2013) 811–846,
http://dx.doi.org/10.1002/nme.4522.
[41] D. Schillinger, M. Ruess, The finite cell method: A review in the context of higher-order structural analysis of CAD and image-based
geometric models, Arch. Comput. Methods Eng. 22 (3) (2015) 391–455, http://dx.doi.org/10.1007/s11831-014-9115-y.
[42] T. Hoang, C.V. Verhoosel, F. Auricchio, E.H. van Brummelen, A. Reali, Mixed isogeometric finite cell methods for the Stokes problem,
Comput. Methods Appl. Mech. Engrg. 316 (2017) 400–423, http://dx.doi.org/10.1016/j.cma.2016.07.027, (special Issue on Isogeometric
Analysis: Progress and Challenges).
[43] V. Varduhn, M.-C. Hsu, M. Ruess, D. Schillinger, The tetrahedral finite cell method: Higher-order immersogeometric analysis on
adaptive non-boundary-fitted meshes, Internat. J. Numer. Methods Engrg. 107 (2016) 1054–1079, http://dx.doi.org/10.1002/nme.5207.
[44] T. Hoang, C.V. Verhoosel, C.-Z. Qin, F. Auricchio, A. Reali, E.H. van Brummelen, Skeleton-stabilized immersogeometric analysis
for incompressible viscous flow problems, Comput. Methods Appl. Mech. Engrg. 344 (2019) 421–450, http://dx.doi.org/10.1016/j.cma.
2018.10.015.
[45] C.S. Peskin, The immersed boundary method, Acta Numer. 11 (2002) 479–517, http://dx.doi.org/10.1017/S0962492902000077.
[46] D. Boffi, L. Gastaldi, A finite element approach for the immersed boundary method, Comput. Struct. 81 (8–11) (2003) 491–501,
http://dx.doi.org/10.1016/S0045-7949(02)00404-2.
[47] L. Heltai, F. Costanzo, Variational implementation of immersed finite element methods, Comput. Methods Appl. Mech. Engrg. 229–232
(2012) 110–127, http://dx.doi.org/10.1016/j.cma.2012.04.001.
[48] Z. Li, The immersed interface method using a finite element formulation, Appl. Numer. Math. 27 (3) (1998) 253–267, http:
//dx.doi.org/10.1016/S0168-9274(98)00015-4.
[49] Z. Li, K. Ito, The Immersed Interface Method: Numerical Solutions of PDEs Involving Interfaces and Irregular Domains, in: Frontiers
in Applied Mathematics, Society for Industrial and Applied Mathematics, 2006.
[50] J. Fish, The s-version of the finite element method, Comput. Struct. 43 (3) (1992) 539–547, http://dx.doi.org/10.1016/0045-7949(92)
90287-A.
[51] J. Fish, S. Markolefas, R. Guttal, P. Nayak, On adaptive multilevel superposition of finite element meshes for linear elastostatics, Appl.
Numer. Math. 14 (1) (1994) 135–164, http://dx.doi.org/10.1016/0168-9274(94)90023-X.
[52] S.P.A. Bordas, E. Burman, M.G. Larson, M.A. Olshanskii, Geometrically Unfitted Finite Element Methods and Applications, in: Lecture
Notes in Computational Science and Engineering, vol. 121, Springer International Publishing, 2017, http://dx.doi.org/10.1007/978-3-
319-71431-8.
[53] F. de Prenter, C. Verhoosel, E. van Brummelen, Preconditioning immersed isogeometric finite element methods with application to
flow problems, Comput. Methods Appl. Mech. Engrg. 348 (2019) 604–631, http://dx.doi.org/10.1016/j.cma.2019.01.030.
[54] F. de Prenter, C.V. Verhoosel, E.H. van Brummelen, J.A. Evans, C. Messe, J. Benzaken, K. Maute, Multigrid solvers for immersed
finite element methods and immersed isogeometric analysis, Comput. Mech. 65 (3) (2020) 807–838, http://dx.doi.org/10.1007/s00466-
019-01796-y.
[55] A. Johansson, M.G. Larson, A. Logg, MultiMesh finite elements with flexible mesh sizes, Comput. Methods Appl. Mech. Engrg.
(2020) arXiv:1804.06455, submitted for publication.
[56] A. Johansson, B. Kehlet, M.G. Larson, A. Logg, Multimesh finite element methods: Solving PDEs on multiple intersecting meshes,
Comput. Methods Appl. Mech. Engrg. 343 (2019) 672–689, http://dx.doi.org/10.1016/j.cma.2018.09.009.
[57] A. Johansson, M.G. Larson, A. Logg, High order cut finite element methods for the Stokes problem, Adv. Model. Simul. Eng. Sci. 2
(1) (2015) 1–23, http://dx.doi.org/10.1186/s40323-015-0043-7.
[58] A. Johansson, M.G. Larson, A. Logg, A multiMesh finite element method for the Stokes problem, in: E.H. van Brummelen, A. Corsini,
S. Perotto, G. Rozza (Eds.), Numerical Methods for Flows: FEF 2017 Selected Contributions, Springer International Publishing, 2019,
http://dx.doi.org/10.1007/978-3-030-30705-9.
[59] J.S. Dokken, S.W. Funke, A. Johansson, S. Schmidt, Shape optimization using the finite element method on multiple meshes with
Nitsche coupling, SIAM J. Sci. Comput. 41 (3) (2019) A1923–A1948, http://dx.doi.org/10.1137/18M1189208.
[60] A.J. Chorin, Numerical solution of the Navier–Stokes equations, Math. Comp. 22 (104) (1968) 745–762, http://dx.doi.org/10.1090/
S0025-5718-1968-0242392-2.
[61] R. Temam, Sur l’approximation de la solution des équations de Navier–Stokes par la méthode des pas fractionnaires (I), Arch. Ration.
Mech. Anal. 32 (2) (1969) 135–153, http://dx.doi.org/10.1007/BF00247696.
[62] K. Goda, A multistep technique with implicit difference schemes for calculating two- or three-dimensional cavity flows, J. Comput.
Phys. 30 (1) (1979) 76–95, http://dx.doi.org/10.1016/0021-9991(79)90088-3.
[63] J. Van Kan, A second-order accurate pressure-correction scheme for viscous incompressible flow, SIAM J. Sci. Stat. Comput. 7 (3)
(1986) 870–891, http://dx.doi.org/10.1137/0907059.
[64] L. Timmermans, P. Minev, F. Van De Vosse, An approximate projection scheme for incompressible flow using spectral elements,
Internat. J. Numer. Methods Fluids 22 (7) (1996) 673–688, http://dx.doi.org/10.1002/(SICI)1097-0363(19960415)22:7<673::AID-
FLD373>3.0.CO;2-O.
[65] J. Guermond, P. Minev, J. Shen, An overview of projection methods for incompressible flows, Comput. Methods Appl. Mech. Engrg.
195 (44) (2006) 6011–6045, http://dx.doi.org/10.1016/j.cma.2005.10.010.
[66] A. Vreman, The projection method for the incompressible Navier–Stokes equations: The pressure near a no-slip wall, J. Comput. Phys.
263 (2014) 353–374, http://dx.doi.org/10.1016/j.jcp.2014.01.035.
[67] P.M. Gresho, R.L. Sani, On pressure boundary conditions for the incompressible Navier–Stokes equations, Internat. J. Numer. Methods
Fluids 7 (10) (1987) 1111–1145, http://dx.doi.org/10.1002/fld.1650071008.
 J.S. Dokken, A. Johansson, A. Massing et al. / Computer Methods in Applied Mechanics and Engineering 368 (2020) 113129 25

[68] R.L. Sani, J. Shen, O. Pironneau, P.M. Gresho, Pressure boundary condition for the time-dependent incompressible Navier–Stokes
equations, Internat. J. Numer. Methods Fluids 50 (6) (2006) 673–682, http://dx.doi.org/10.1002/fld.1062.
[69] A. Quarteroni, R. Sacco, F. Saleri, Numerical Mathematics, Vol. 37, Springer Science & Business Media, 2010, http://dx.doi.org/10.
1007/b98885.
[70] J.-L. Guermond, Un résultat de convergence d’ordre deux en temps pour l’approximation des équations de Navier–Stokes par une
technique de projection incrémentale, ESAIM Math. Model. Numer. Anal. 33 (1) (1999) 169–189, http://dx.doi.org/10.1051/m2an:
1999101.
[71] S.C. Brenner, L.R. Scott, The Mathematical Theory of Finite Element Methods, third ed., in: Texts in Applied Mathematics, vol. 15,
Springer, 2008, http://dx.doi.org/10.1007/978-0-387-75934-0.
[72] P. Halmos, Finite-Dimensional Vector Spaces, Springer, 1974, http://dx.doi.org/10.1007/978-1-4612-6387-6.
[73] D.A. Di Pietro, A. Ern, Incompressible Flows, Springer Berlin Heidelberg, Berlin, Heidelberg, 2012, http://dx.doi.org/10.1007/978-3-
642-22980-0_6.
[74] D.A. Di Pietro, A. Ern, Mathematical Aspects of Discontinuous Galerkin Methods, Vol. 69, Springer Science & Business Media, 2011,
http://dx.doi.org/10.1007/978-3-642-22980-0.
[75] L. Botti, D.A.D. Pietro, A pressure-correction scheme for convection-dominated incompressible flows with discontinuous velocity and
continuous pressure, J. Comput. Phys. 230 (3) (2011) 572–585, http://dx.doi.org/10.1016/j.jcp.2010.10.004.
[76] E. Burman, S. Claus, A. Massing, A stabilized cut finite element method for the three field Stokes problem, SIAM J. Sci. Comput.
37 (4) (2015) A1705–A1726, http://dx.doi.org/10.1137/140983574.
[77] D.N. Arnold, An interior penalty finite element method with discontinuous elements, SIAM J. Numer. Anal. 19 (4) (1982) 742–760,
http://dx.doi.org/10.1137/0719052.
[78] P. Hansbo, Nitsche’s method for interface problems in computational mechanics, GAMM-Mitt. 28 (2) (2005) 183–206, http:
//dx.doi.org/10.1002/gamm.201490018.
[79] T. Warburton, J.S. Hesthaven, On the constants in hp-finite element trace inverse inequalities, Comput. Methods Appl. Mech. Engrg.
192 (25) (2003) 2765–2773, http://dx.doi.org/10.1016/S0045-7825(03)00294-9.
[80] M.S. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson, J. Ring, M.E. Rognes, G.N. Wells, The FEniCS
project version 1.5, Arch. Numer. Softw. 3 (100) (2015) 9–23, http://dx.doi.org/10.11588/ans.2015.100.20553.
[81] A. Logg, G.N. Wells, DOLFIN: Automated finite element computing, ACM Trans. Math. Software 37 (2) (2010) 20, http://dx.doi.org/
10.1145/1731022.1731030.
[82] A. Massing, M.G. Larson, A. Logg, Efficient implementation of finite element methods on non-matching and overlapping meshes in
3D, SIAM J. Sci. Comput. 35 (1) (2013) C23–C47, http://dx.doi.org/10.1137/11085949X.
[83] C.V. Hansen, A. Logg, C. Lundholm, Simulation of flow and view with applications in computational design of settlement layouts,
2016, arXiv:1610.02277.
[84] J.S. Dokken, A. Johansson, A. Massing, S.W. Funke, Source code for: A multimesh finite element method for the Navier–Stokes
equations based on projection methods, 2019, http://dx.doi.org/10.5281/zenodo.3564206.
[85] S. Turek, Featflow CFD benchmarking project: DFG flow around cylinder benchmark 2D-3, fixed time interval (Re = 100), 2019,
URL http://www.featflow.de/en/benchmarks/cfdbenchmarking/flow/dfg_benchmark3_re100.html. (Accessed 25 July 2019).
[86] R. Courant, K. Friedrichs, H. Lewy, Über die partiellen Differenzengleichungen der mathematischen Physik, Math. Ann. 100 (1) (1928)
32–74, http://dx.doi.org/10.1007/BF01448839.
[87] C.E. Pearson, A Computational Method for Time Dependent Two Dimensional Incompressible Viscous Flow Problems, Tech. Rep.
SRRC-RR-64-17, Sperry Rand Research Centre, 1964.
[88] M. Schäfer, S. Turek, F. Durst, E. Krause, R. Rannacher, Benchmark Computations of Laminar Flow Around a Cylinder,
Vieweg+Teubner Verlag, Wiesbaden, 1996, pp. 547–566, http://dx.doi.org/10.1007/978-3-322-89849-4_39.
[89] S. Funke, P. Farrell, M. Piggott, Tidal turbine array optimisation using the adjoint approach, Renew. Energy 63 (2014) 658–673,
http://dx.doi.org/10.1016/j.renene.2013.09.031.
[90] A. Wächter, L.T. Biegler, On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,
Math. Program. 106 (1) (2006) 25–57, http://dx.doi.org/10.1007/s10107-004-0559-y.
[91] E. Burman, M.A. Fernández, An unfitted Nitsche method for incompressible fluid–structure interaction using overlapping meshes,
Comput. Methods Appl. Mech. Engrg. 279 (2014) 497–514, http://dx.doi.org/10.1016/j.cma.2014.07.007.
[92] M.A. Fernández, M. Landajuela, Splitting schemes for incompressible fluid/thin-walled structure interaction with unfitted meshes, C.
R. Math. 353 (7) (2015) 647–652, http://dx.doi.org/10.1016/j.crma.2015.04.003.
[93] A. Massing, M.G. Larson, A. Logg, M.E. Rognes, A Nitsche-based cut finite element method for a fluid–structure interaction problem,
Commun. Appl. Math. Comput. Sci. 10 (2015) http://dx.doi.org/10.2140/camcos.2015.10.97.
[94] B. Schott, C. Ager, W. Wall, A monolithic approach to fluid–structure interaction based on a hybrid Eulerian-ALE fluid domain
decomposition involving cut elements, Internat. J. Numer. Methods Engrg. 119 (3) (2019) 208–237, http://dx.doi.org/10.1002/nme.6047.
[95] A. Massing, M.G. Larson, A. Logg, Efficient implementation of finite element methods on non-matching and overlapping meshes in
3d, SIAM J. Sci. Comput. 35 (1) (2013) C23–C47, http://dx.doi.org/10.1137/11085949X.
[96] P. Farrell, J. Maddison, Conservative interpolation between volume meshes by local Galerkin projection, Comput. Methods Appl. Mech.
Engrg. 200 (1) (2011) 89–100, http://dx.doi.org/10.1016/j.cma.2010.07.015.
[97] R. Krause, P. Zulian, A parallel approach to the variational transfer of discrete fields between arbitrarily distributed unstructured finite
element meshes, SIAM J. Sci. Comput. 38 (3) (2016) C307–C333, http://dx.doi.org/10.1137/15M1008361.