See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358488790

Robust preconditioning for a mixed formulation of phase-field fracture

problems

Preprint · February 2022

DOI: 10.48550/arXiv.2202.04191

CITATIONS READS
0 285

3 authors:

Timo Heister Katrin Mang

Texas A&M University Leibniz Universität Hannover
60 PUBLICATIONS 4,275 CITATIONS 19 PUBLICATIONS 204 CITATIONS

SEE PROFILE SEE PROFILE

Thomas Wick
Leibniz Universität Hannover
272 PUBLICATIONS 6,075 CITATIONS

SEE PROFILE

All content following this page was uploaded by Thomas Wick on 18 March 2022.

The user has requested enhancement of the downloaded file.

 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD
FRACTURE PROBLEMS

TIMO HEISTER

Clemson University, O-110 Martin Hall, Clemson, SC

[email protected]
arXiv:2202.04191v1 [math.NA] 8 Feb 2022

KATRIN MANG

Leibniz Universität Hannover, Welfengarten 1, 30167 Hannover, Germany

[email protected]

THOMAS WICK

Leibniz Universität Hannover, Welfengarten 1, 30167 Hannover, Germany

[email protected]

Abstract. In this work, we consider fracture propagation in nearly incompressible and (fully) incompress-
ible materials using a phase-field formulation. We use a mixed form of the elasticity equation to overcome
volume locking effects and develop a robust, nonlinear and linear solver scheme and preconditioner for the
resulting system. The coupled variational inequality system, which is solved monolithically, consists of three
unknowns: displacements, pressure, and phase-field. Nonlinearities due to coupling, constitutive laws, and
crack irreversibility are solved using a combined Newton algorithm for the nonlinearities in the partial differ-
ential equation and employing a primal-dual active set strategy for the crack irreverrsibility constraint. The
linear system in each Newton step is solved iteratively with a flexible generalized minimal residual method
(GMRES). The key contribution of this work is the development of a problem-specific preconditioner that
leverages the saddle-point structure of the displacement and pressure variable. Four numerical examples in
pure solids and pressure-driven fractures are conducted on uniformly and locally refined meshes to investi-
gate the robustness of the solver concerning the Poisson ratio as well as the discretization and regularization
parameters.

1. Introduction
Phase-field fracture modeling [37, 43] emerged from a variational formulation introduced in [20, 10] is an
attractive model approach to simulate crack propagation in solids. To date, displacement-based formulations
have been used in the large majority of investigations [8, 43, 37, 35, 19, 51, 13, 2, 12]. However, considering
(nearly) incompressible solids, these models are subject to locking effects, i.e., the values of the displacement
field are underestimated. For this reason, mixed phase-field formulations have been recently developed
in [40, 39] in which classical ideas from (non-fractured) solids were employed by introducing a Lagrange
multiplier for the pressure variable. With help of a mixed form we get a stable problem formulation up to
the incompressible limit [5]. However only sparse direct solvers, e.g., [47], were utilized in these previous
studies for solving the arising linear equation systems.
The main purpose of the current work is to propose (for the first time) a preconditioned iterative linear
solver for solving mixed formulations of phase-field fracture problems. Therein, we deal with three unknowns,
namely displacements, pressure, and phase-field, U := (u, p, ϕ). For classical, displacement-based (u, ϕ)
formulations iterative linear and multigrid methods are known. We also note that we consider fractures
in pure solids as well as pressurized cracks (Sneddons’s test, see e.g., [49, 9, 48]). The reader should not
1
2 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

confound the pressure p introduced due to the mixed formulation with the pressure ρ, which is imposed
inside the crack region in pressurized (i.e., pressure-driven) configurations.
The first study with a clear focus on linear solvers is [19]. Therein, a nonlinear Gauss-Seidel scheme was
proposed together with a Schur complement based preconditioner for the linear systems. A parallel GMRES
(generalized minimal residual) solver with diagonal preconditioner with algebraic multigrid preconditioning
was developed in [26, 27]. Earlier versions were used in [25, 38], however without studying the parallel
performance and scalability. A GMRES solver with a matrix-free geometric multigrid preconditioner was
later suggested in [32] with a subsequent parallel version in [31]. An overall summary of these developments
can be found in the PhD thesis of Jodlbauer [30].
A Galerkin finite element discretization yields a nonlinear system of the form MU = F with a (3×3) block
matrix M ∈ Rn×n , U and F ∈ Rn . For inf-sup stability a Taylor-Hood element Qc2 /Qc1 is used for the (u, p)
system. Here, Qc2 denotes a continuous finite element space with bi-quadratic finite elements (we restrict
the discussion to quadrilateral finite elements here). We note that computational comparisons to stabilized
low-order equal-order finite elements were undertaken in [39]. However, it was found that this approach can
not be recommended for the mixed phase-field fracture formulation combined with high Poisson ratios. The
stabilizing terms contain ∇p with mesh-dependent coefficients leading to large gradients in the crack region.
The discretized system is nonlinear, for which we employ Newton’s method as a nonlinear solver. Inside,
the linear system is non-symmetric and therefore, we use a GMRES method. The key contribution is the
development of a block triangular preconditioner. Individual blocks are approximated with inner solves using
the conjugate gradient method (CG) and algebraic multigrid (AMG) from the ML package [52, 21]. The
mixed form of the elasticity equation has a saddle point structure, which allows to reuse spectral approxi-
mations for the inverse matrices from the Stokes problem [6, 7, 17]. All ingredients of the preconditioner can
be parallelized and the developed code is parallel (since extended from pfm-cracks [27] with scalability tests
undertaken in [26]). However, we decided to focus on the various challenges in robustness of the (3 × 3) block
system. Therefore, parallel computing studies with scalability tests are outside the scope of this paper.
The main challenge in developing the preconditioner is the interaction of various model, discretization,
and material parameters to obtain a robust approach. These are the spatial discretization parameter h
and the Poisson ratio ν (related to the Lamé coefficient λ) up to the incompressible limit ν = 0.5, and the
regularization parameter κ and the crack bandwidth . We note that the basis of this work was developed
in Section 6 of the PhD thesis of the second author [39] and some preliminary results were published in [24].
The outline of this paper is as follows: In Section 2, the notation and governing equations are introduced.
Next, Section 3 is the main part in which we first summarize the discretization and nonlinear solver. Then,
the iterative solver and a Schur-type preconditioner are derived. Afterward, in Section 4 four numerical
experiments are conducted to substantiate the performance of our algorithmic developments. Our work is
summarized in Section 5.

2. Notation and governing equations

Let Ω be an open and smooth two-dimensional domain and T := (0, Tend ) is a time (i.e., loading) interval
with the partition 0 =: t0 < t1 < . . . < tN := Tend . The lower-dimensional crack is approximated by a
phase-field indicator function ϕ : (Ω × T ) → [0, 1] with ϕ = 0 in the crack and ϕ = 1 in the unbroken area.
The bandwidth of the zone between broken and unbroken is named . Further, a displacement function is
defined as u : (Ω×T ) → R2 . In the following, the scalar-valued L2 -product
R
is denoted by (x, y) := Ω
x·y dΩ,
whereas the vector-valued L2 -product is described by (X, Y ) := Ω X : Y dΩ, with the Frobenius product
R

X : Y of two vectors X and Y . We define the usual Sobolev spaces V := H01 (Ω)2 , W := H 1 (Ω) and a convex
subset K := {ϕ ∈ W| 0 ≤ ϕ ≤ ϕn−1 ≤ 1 a.e. in Ω} ⊂ W and U := L2 (Ω). Further, the degradation function
is defined as g(ϕ) = (1−κ)ϕ2 +κ, where κ is a sufficiently small regularization parameter. The stress tensor is
defined as σ(u) := 2µElin (u)+λtr(Elin (u))I with a linearized strain tensor Elin (u) := 12 (∇u+∇uT ), material
dependent Lamé coefficients λ and µ, and the two-dimensional identity matrix I. The critical energy release
rate is denoted as Gc . Based on this notation, the pressurized phase-field fracture model in its classical form
can be formulated as follows [54]:
Problem 1 (Pressurized phase-field fracture).
Let a (constant) pressure ρ ∈ L∞ (Ω) and the initial value ϕ(0) := ϕ0 be given. Given the previous timestep
data ϕn−1 := ϕ(tn−1 ) ∈ K. Find u := un ∈ V and ϕ := ϕn ∈ K for loading steps n = 1, 2, . . . , N with
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 3

{u, ϕ} ∈ V × K such that

 
g(ϕ̃)σ(u) , Elin (w) + (ϕ̃2 ρ, ∇ · w) = 0 ∀w ∈ V,
(1 − κ)(ϕσ(u) : Elin (u), ψ−ϕ) + 2(ϕρ∇ · u, ψ − ϕ)
 1 
+Gc − (1 − ϕ, ψ − ϕ) + (∇ϕ, ∇(ψ − ϕ)) ≥ 0 ∀ψ ∈ K.

In the elasticity part, a linear-in-time extrapolation with ϕ̃ := ϕ̃(ϕn−1 , ϕn−2 ) is used in the phase-field
variable ϕ to obtain a convex functional [25]. Therein, for ϕn−2 at n = 1, we set ϕ−1 := ϕ0 .
Based on Problem 1 and following [40], we introduce a pressure p := λ tr(Elin (u)), which is a Lagrange
multiplier. As mentioned in the introduction, p and the crack pressure ρ (see for instance [54] and therein
itself denoted as p) should not be mixed up.

Problem 2 (Pressurized phase-field fracture in mixed form).

Let ρ ∈ L∞ (Ω) be given and the initial value ϕ(0) := ϕ0 be given. Given the previous time step data
ϕn−1 ∈ K. Find u := un ∈ V, p := pn ∈ U and ϕ := ϕn ∈ K for loading steps n = 1, 2, . . . , N with
U := {u, p, ϕ} ∈ V × U × K such that
 
g(ϕ̃)σ(u, p) , Elin (w) + (ϕ̃2 ρ, ∇ · w) = 0 ∀w ∈ V,
1
g(ϕ̃)(∇ · u, q) − ( p, q) = 0 ∀q ∈ U,
λ
(1 − κ)(ϕσ(u, p) : Elin (u), ψ−ϕ) + 2(ϕρ∇ · u, ψ − ϕ)
 1 
+Gc − (1 − ϕ, ψ − ϕ) + (∇ϕ, ∇(ψ − ϕ)) ≥ 0 ∀ψ ∈ K,


where the stress tensor is defined as σ(u, p) := 2µElin (u) + pI.

Remark. It is clear that by setting ρ = 0, we obtain a phase-field formulation for fracture in pure solids.
With this, we can investigate our preconditioner for both situations, namely fracture in solids and pressurized
cracks.

3. Discretization and numerical solution

For the spatial discretization of Problem 2, we employ a Galerkin finite element method in each incremental
step, where the domain Ω is partitioned into quadrilaterals [14] with the discrete spaces Vh , Uh , and the
convex set Kh ⊂ Wh . To fulfill a discrete inf-sup condition, stable Taylor-Hood elements with continuous
bi-quadratic shape functions (Qc2 ) for the displacement field u and bilinear shape functions (Qc1 ) for the
pressure variable p and the phase-field variable ϕ are used as in [40].

3.1. Nonlinear solver. The nonlinear solution algorithm is based on a combined method. First, nonlin-
earities arising from the PDE (partial differential equation) are treated with a standard line-search assisted
Newton scheme. The crack irreversibility is handled with a primal-dual active set method. The combination
of both techniques yields one single nonlinear Newton iteration; see [25] for further details.

Problem 3 (Discretized pressurized phase-field fracture in mixed form).

Define unh := uh (tn ), pnh := ph (tn ) and ϕnh := ϕh (tn ) at the loading step tn . Let ϕ̃h := ϕ̃h (ϕn−1
h , ϕn−2
h )
be the discrete linear-in-time extrapolation. Find Uhn := (unh , pnh , ϕnh ) ∈ Vh × Uh × Kh for all loading steps
n = 1, 2, .., N such that

A(Uhn )(Ψh − Φnh ) = A1 (Uhn )(wh ) + A2 (Uhn )(qh ) + A3 (Uhn )(ψh − ϕnh ) ≥ 0
4 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

with Φnh = (0, 0, ϕnh ) and for all Ψh := (wh , qh , ψh ) ∈ Vh × Uh × Kh , and where
A1 (Uhn )(wh ) = g(ϕ̃h ) (σ(unh , pnh ), Elin (wh )) + (ϕ̃2h ρ, ∇ · wh ),
1
A2 (Uhn )(qh ) = g(ϕ̃h ) (∇ · unh , qh ) − (pnh , qh ),
λ
A3 (Uhn )(ψh − ϕnh ) = (1 − κ) (ϕnh σ(unh , pnh ) : Elin (unh ), ψh − ϕnh )
+ 2(ϕnh ρ∇ · unh , ψh − ϕnh )
 
1 n n n n
+ Gc (− (1 − ϕh ), ψh − ϕh ) +  (∇ϕh , ∇(ψh − ϕh )) ,

where σ(unh , pnh ) := 2µElin (unh ) + pnh I.
In order to treat the inequality constraint in Kh , we employ a primal-dual active set method as explained
in [25] and use the function space Wh for approximating ϕ. Then, at each loading step n, we have the
following Newton iteration indexed by k. We set as initial guess Uhn,0 := Uhn−1 and iterate for k = 1, 2, 3, . . .:
∇A(Uhk )(δUhn,k , Ψ) = −A(Uhn,k )(Ψ) ∀Ψ ∈ Vh × Uh × Wh .

The directional derivative ∇A(Uhn,k )(δUhk , Ψ) in direction δUhk for Uhn,k ∈ Vh × Uh × Wh is given by
∇A(Uhn,k )(δUhk , Ψ) = g(ϕ̃nh ) σ(δukh , δpkh ), Elin (wh )

1
+ g(ϕ̃nh ) ∇ · δukh , qh − (δpkh , qh )
 λ 
+ (1 − κ) ϕh 2µ(Elin (δukh ) : Elin (un,k
n,k
h ) + E (u
lin h
n,k
) : Elin (δuk
h )), ψh

+ 2(ϕn,k k n,k k n,k

h ρ∇ · δuh , ψh ) + (1 − κ)(ϕh δph I : Elin (uh ), ψh )
 
+ (1 − κ) δϕkh σ(un,k n,k n,k k n,k
h , ph ) : Elin (uh ), ψh + 2(δϕh ρ∇ · uh , ψh )
 
1 k
δϕh , ψh + Gc  ∇δϕkh , ∇ψh .


+ Gc

3.2. Linear solution and Schur-type preconditioning. For the arising linear systems MδU = F inside
Newton’s method, a GMRES method is used, which is right-preconditioned [47] with a Schur-type precondi-
tioner P −1 . As usual, the goal when developing P −1 is to have the eigenvalues of (MP −1 ) be independent
of discretization, regularization parameters and coefficients of the problem.

3.2.1. Preconditioning the (3×3) linear system. The system matrix Mmixed of the mixed phase-field fracture
from the modified mixed problem formulation has the following block structure [24]:
 uu
M up M uϕ g(ϕ̃)Au g(ϕ̃)B T 0
  
M
(1) Mmixed =  M pu M pp M pϕ  =  g(ϕ̃)B − λ1 Mp 0  ,
ϕu ϕp ϕϕ
M M M E F L
where block Au is the mass matrix of the displacements, B and B T are symmetric off-diagonal blocks coupling
u and p, and Mp is the mass matrix of the pressure variable. The blocks E, F and L from Equation (1)
consist of the entries from the phase-field equation, where L is Laplacian-like. For the entry-wise definition
of the blocks, we refer to [39, page 169].
A typical block factorization of the system matrix yields the preconditioner (details can be found in [39,
Chapter 6])
g(ϕ̃)Au g(ϕ̃)B T 0
 

Pmixed =  0 S 0 ,
0 0 L
for Mmixed , where S is the Schur complement block defined as
1
S = − Mp − g(ϕ̃)B T · [g(ϕ̃)Au ]−1 · g(ϕ̃)B.
λ
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 5

It is not feasible to construct S −1 or even S exactly, as this would result in a dense matrix. This means that
−1
exact evaluation of Pmixed is also not a feasible option, but it helps us to design an appropriate preconditioner
−1 −1
by approximating the action of Pmixed with an operator P̂mixed defined below. Note that all eigenvalues of
−1
Mmixed Pmixed are equal to one and GMRES would converge in at most two iterations [44, 6].
−1
Without considering the last row and column of Mmixed and Pmixed (the phase-field), this is a typical
saddle-point problem with a penalty term, where block triangular preconditioners are a common choice [6],
first considered by Bramble and Pasciak in 1988 [11], and frequently used for Stokes-type problems [15] and
the Oseen equations [33], where mesh-independent convergence can be observed.
−1
To be able to efficiently apply Pmixed , we require approximations of the inverses of the Laplacian-like
matrices L, of g(ϕ̃)Au , and of the Schur complement matrix S. With spectrally equivalent approximations,
this would result in an optimal preconditioner [6] yielding an eigenvalue distribution independent of mesh
size h and other problem parameters and therefore constant GMRES iterations numbers independent of
mesh size and problem parameters. Since multigrid methods allow for mesh-independent convergence [23],
algebraic or geometric multigrid methods are the method of choice.
The approximation of the inverse of S turns out to be more challenging. It is well-known, e.g., [53], for
inf-sup stable discretizations of the linear elasticity problem, the Schur-complement is spectrally equivalent
to the mass matrix. In our case, for λ → ∞, g(ϕ̃) acts like a varying viscosity. It is common to scale the
mass matrix with the inverse of the viscosity for Stokes interface problems [45] or variable viscosity Stokes
problems, e.g., [22, 41], which yields
  −1
−1 1 g(ϕ̃)
Ŝ := − + Mp ,
λ 2µ
as an approximation of the inverse of S in our situation. Under sufficient regularity, κ > 0 and if the
coefficient can be assumed to be constant, Ŝ −1 is spectrally equivalent to S −1 [45]. For the incompressible
limit ν = 0.5, the Schur complement approximation becomes
 −1
g(ϕ̃)
Ŝ −1 = − Mp .
2µ
Remark (Differences to Stokes-type problems). Commonly, this Schur complement approach is used for
Stokes-type problems and incompressible fluid dynamics, see, e.g. [18]. Even if the elasticity part of the con-
sidered phase-field fracture problem has a similar saddle-point structure, aside from the phase-field function,
material and regularization parameters complicate the situation: λ → ∞ leads to a purely κ-dependent block
Ŝ −1 , and κ → 0 increases the condition number of the block (g(ϕ̃)Au )−1 in the crack, where ϕ = 0. While
the approximation of Ŝ −1 is spectrally equivalent with respect to the mesh size, it is not robust with respect to
large viscosity variations, or in our case minimum and maximum value of g(ϕ̃) throughout the domain. For
the Stokes interface problem with a viscosity jump with single interface, the scaled mass matrix is spectrally
equivalent independent of the magnitude of the jump [45], which is the case in our situation. This will be
visible in Section 4. We hypothesize that a better Schur complement could be a weighted BFBT preconditioner
presented in [46], but a thorough investigation is future work.

3.2.2. Preconditioning algorithm. As discussed above, the evaluation of the preconditioner

 
(g(ϕ̃)Au )−1 −A−1 T −1
u B Ŝ 0
−1
P̂mixed =  0 Ŝ −1 0 
 
−1
0 0 L

requires efficient approximations to the exact inverses of Au , Ŝ, and L. Iterative solvers like GMRES of
−1
course only require the result of a matrix-vector product with the preconditioner P̂mixed , see [47] and inside
our basis software deal.II [4], see [36].
First, we approximate L−1 by a single V -cycle of algebraic multigrid (AMG). Second, for (g(ϕ̃)Au )−1 we
use an inner Conjugate Gradient (CG) solve, which, in turn, is preconditioned by one V -cycle of algebraic
multigrid. Finally, the action of Ŝ −1 is either done using a single V -cycle of AMG or, in Figures 4 and 8,
using CG preconditioned by AMG.
6 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

−1
With this, the matrix-vector product P̂mixed ~x with ~x = (xu , xp , xϕ )T given as
 
(g(ϕ̃)Au )−1 xu − (g(ϕ̃)Au )−1 g(ϕ̃)B T Ŝ −1 xp
−1
P̂mixed ~x =  Ŝ −1 xp ,
 
−1
L xϕ
is built up step by step. In deal.II [3, 4], the preconditioners given to solver classes need a vmult() member
function [16]. Then, our final algorithm is designed as follows:
−1
Algorithm 1. Evaluation of P̂mixed ~x:
(1) Approximate Ŝ −1 via AMG and compute q := Ŝ −1 xp ;
(2) Compute r := xu − g(ϕ̃)B T q;
(3) Approximate (g(ϕ̃)Au )−1 via CG preconditioned with AMG and compute s := (g(ϕ̃)Au )−1 r;
(4) Approximate L−1 via AMG and compute t := L−1 xϕ ;
(5) Return the result (s, q, t)T .

4. Numerical tests
In this section, we consider four different numerical experiments to substantiate our algorithmic develop-
ments and to investigate the performance of the nonlinear solver, linear solver and preconditioner.

4.1. Test cases and presentation of our results. To facilitate the readability of the tables from the
next sections, we give an overview, how to read them. For the four tests, we conduct numerical studies with
different emphases: we investigate robustness in h, κ, λ, , we use different models (‘primal’ from Problem 1
versus ‘mixed’ from Problem 2) and different finite element discretizations. In the top row of each table,
we summarize the key aspect of the current numerical study: the name of the example, the observed task,
the modeling, and – if required – further test-specific settings. The white rows in the tables correspond to
results based on the primal phase-field fracture model (solved with pfm-cracks [27]) or to reference values.
The colored rows belong to computations based on the mixed model and Qc2 Qc1 Qc1 for ν = 0.2 (yellow),
ν = 0.4999 (blue) and ν = 0.5 (red). A more saturated shading denotes a finer mesh size.
The four test configurations with attributes are given in the following:
• Section 4.3: a hanging block with an initial slit for ν = 0.2, 0.4999 and 0.5, uniform mesh refinement,
mixed (Qc2 Qc1 Qc1 ) versus primal (Qc2 Qc1 ),  fixed and  = 2 h, κ = 10−2 ;
• Section 4.4: Sneddon’s test [50, 49] for ν = 0.2, 0.4999 and 0.5, uniform mesh refinement, mixed
(Qc2 Qc1 Qc1 ),  fixed and  = 2 h, κ = 10−2 , 10−8 ;
• Section 4.5: Sneddon’s test layered [5] for ν = 0.2, 0.4999 and 0.5 in the inner domain, adaptive mesh
refinement (geometric), mixed (Qc2 Qc1 Qc1 ),  = h, κ = 10−2 , 10−8 ;
• Section 4.6: single-edge notched tension test for ν = 0.3, 0.45, and 0.49, adaptive mesh refinement
(predictor-corrector scheme), mixed (Qc2 Qc1 Qc1 ),  = 4 h, κ = 10−8 .
With the help of numerical studies, we investigate the robustness of the new Schur-type preconditioner via
evaluating the required number of linear iterations for different mesh sizes, Poisson ratios, κ, and different
finite element discretizations. Besides, we discuss challenges and point out difficulties.

4.2. Implementation details. The software developed for this paper is a major extension built upon pfm-
cracks [26, 27], which is an open-source code available at https://github.com/tjhei/cracks. This project
is built on the finite element library deal.II [3], which offers scalable parallel algorithms for finite element
computations. The deal.II library in turn uses functionality from other libraries such as Trilinos [28, 29] for
linear algebra, including the Trilinos ML AMG preconditioner [52, 21]. The GMRES stopping criterion is a
relative tolerance of 10−5 . CG uses a relative tolerance of 10−6 for the inner solves with a maximum of 200
iterations. The Newton iteration stops when an absolute tolerance of 10−7 is reached. We use four CPUs
on a single machine with four Intel E7 v3 CPUs for all computations.

4.3. Hanging block with initial slit. As a first test configuration, we consider a hanging block test with
an initial geometrical slit of length 2.0 mm with an interpolated initial condition ϕ = 0 in the crack; see
Figure 1. The force acting on the hanging block is reduced to f = −8.0 · 10−7 N/mm2 . In Figure 1 on the
right, the solution of the phase-field function is given on the deformed block for ν = 0.2 on a uniform refined
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 7

4.0 mm

2.0 mm

point (0,1.99)
×
ϕ=0

2.0 mm

body force

Figure 1. Left: geometry and boundary conditions of a hanging block with a slit. Right:
deformed geometry with phase-field solution; 41 924 degrees of freedom (DoFs). In the
geometrically pre-refined slit we interpolate ϕ = 0 as an initial condition. No propagating
crack. Applied force f = −8.0 · 10−7 N/mm2 .

mesh with 41 924 degrees of freedom (DoFs). We evaluate the displacement in the y-direction in a certain
point (0, 1.99) on the lower opening crack lip.
Table 1. A hanging block with an initial slit for ν = 0.2 and ν = 0.4999, uniform mesh
refinement, mixed (Qc2 Qc1 Qc1 ) versus primal (Qc2 Qc1 ),  fixed, κ = 10−2 . Problem size in
#DoFs, average number of GMRES iterations (∅lin) per Newton/active set (AS) step, the
average number of inner CG iterations (∅CG) per linear iteration, number of Newton/AS
(#AS), and goal functional displacement in a certain point (uy (0, 1.99)). Applied force
f = −8 · 10−7 N/mm2 .

Hanging block slit: robustness in h and λ; mixed versus primal; κ = 10−2

model FE ν h  #DoFs ∅lin ∅CG #AS uy (0, 1.99)
mixed Qc c c
2 Q1 Q1 0.2 0.353 0.707 2 804 4 24 3 -0.3871
mixed Qc c c
2 Q1 Q1 0.2 0.176 0.707 10 724 4 25 3 -0.5189
mixed Qc c c
2 Q1 Q1 0.2 0.088 0.707 41 924 10 32 32 -0.4919
mixed Qc c c
2 Q1 Q1 0.2 0.044 0.707 165 764 4 36 31 -0.0825
mixed Qc c c
2 Q1 Q1 0.2 0.022 0.707 659 204 8 50 53 -0.0824
mixed Qc c c
2 Q1 Q1 0.2 0.011 0.707 2 629 124 8 79 38 -0.0815

primal [25] Qc c
2 Q1 0.2 0.353 0.707 2 507 1 - 3 -0.3368
primal [25] Qc c
2 Q1 0.2 0.176 0.707 9 619 1 - 3 -0.4479
primal [25] Qc c
2 Q1 0.2 0.088 0.707 37 667 5 - 5 -0.4434
primal [25] Qc c
2 Q1 0.2 0.044 0.707 149 059 5 - 38 -0.0818
primal [25] Qc c
2 Q1 0.2 0.022 0.707 593 027 7 - 35 -0.0820
primal [25] Qc c
2 Q1 0.2 0.011 0.707 2 365 699 8 - 35 -0.0810
mixed Qc c c
2 Q1 Q1 0.4999 0.353 0.707 2 804 10 24 3 -0.2181
mixed Qc c c
2 Q1 Q1 0.4999 0.176 0.707 10 724 9 25 3 -0.2869
mixed Qc c c
2 Q1 Q1 0.4999 0.088 0.707 41 924 6 32 29 -0.1295
mixed Qc c c
2 Q1 Q1 0.4999 0.044 0.707 165 764 7 38 36 -0.0576
mixed Qc c c
2 Q1 Q1 0.4999 0.022 0.707 659 204 10 52 38 -0.0585
mixed Qc c c
2 Q1 Q1 0.4999 0.011 0.707 2 629 124 11 80 41 -0.0578
primal [25] Qc c
2 Q1 0.4999 0.353 0.707 2 507 1 - 3 -0.2077
primal [25] Qc c
2 Q1 0.4999 0.176 0.707 9 619 1 - 3 -0.2788
primal [25] Qc c
2 Q1 0.4999 0.088 0.707 37 667 4 - 4 -0.2789
primal [25] Qc c
2 Q1 0.4999 0.044 0.707 149 059 5 - 31 -0.0587
primal [25] Qc c
2 Q1 0.4999 0.022 0.707 593 027 7 - 31 -0.0583
primal [25] Qc c
2 Q1 0.4999 0.011 0.707 2 365 699 8 - 34 -0.5076

Tables 1 and 2 show the iteration numbers of numerical tests for the hanging block with a slit for three
Poisson ratios ν and h refinement. For the incompressible limit ν = 0.5, Table 2 presents the results for 
fixed, and further in the pink rows, results for  = 2 h are listed. The nearly constant number of GMRES
iterations confirms the robustness in  for ν = 0.5, tested for the hanging block with a slit on five levels of
uniform refined meshes; see the last five rows in Table 2.
8 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

Table 2. A hanging block with an initial slit for ν = 0.5, uniform mesh refinement, Qc2 Qc1 Qc1
elements,  fixed and  = 2 h, κ = 10−2 . Problem size in #DoFs, average number of GMRES
iterations (∅lin) per Newton/active set (AS) step, the average number of inner CG iterations
(∅CG) per linear iteration, number of Newton/AS (#AS), and goal functional displacement
in a certain point (uy (0, 1.99)). Applied force f = −8 · 10−7 N/mm2 .

Hanging block slit: robustness in h, λ and  for ν = 0.5; mixed; κ = 10−2

model FE ν h  #DoFs ∅lin ∅CG #AS uy (0, 1.99)
mixed Qc2 Qc1 Qc1 0.5 0.353 0.707 2 804 9 23 3 -0.0578
mixed Qc2 Qc1 Qc1 0.5 0.176 0.707 10 724 9 24 3 -0.2835
mixed Qc2 Qc1 Qc1 0.5 0.088 0.707 41 924 7 32 33 -0.0955
mixed Qc2 Qc1 Qc1 0.5 0.044 0.707 165 764 6 37 38 -0.0584
mixed Qc2 Qc1 Qc1 0.5 0.022 0.707 658 436 9 53 36 -0.0583
mixed Qc2 Qc1 Qc1 0.5 0.011 0.707 2 629 124 11 80 39 -0.0579
mixed Qc2 Qc1 Qc1 0.5 0.353 0.707 2 804 9 23 3 -0.2166
mixed Qc2 Qc1 Qc1 0.5 0.176 0.353 10 724 7 25 4 -0.1033
mixed Qc2 Qc1 Qc1 0.5 0.088 0.176 41 924 6 30 14 -0.0701
mixed Qc2 Qc1 Qc1 0.5 0.044 0.088 165 764 5 36 109 -0.0572
mixed Qc2 Qc1 Qc1 0.5 0.022 0.044 658 436 7 40 805 -0.0516

Remark (High iteration numbers in the primal-dual active set method). In Table 2 in the pink rows, many
active set/Newton iterations are required for  → 0. Here, not the Poisson ratio is responsible, but the
refinement in h and . For finer meshes with small , the active set algorithm oscillates between a certain
non-equal number of active nodes from the constraint. This effect leads to high total Newton iterations, even
if the Newton algorithm converges fast; see also [25, Figure 14].

The number of CG iterations does not depend significantly on the size of κ for this test setup. Further,
the required CG iterations seem to be independent of λ but sensitive to the mesh size. Aside from the
robustness in h and λ, we confirm the robustness in κ for the hanging block test with a slit. Details on that
can be found in [39, page 105].

4.4. Sneddon’s pressure-driven cavity. As a second example, we consider a benchmark test [48], which
is motivated by the book of Sneddon [50] and Sneddon and Lowengrub [49]. We restrict ourselves to a 1d
fracture C on a 2d domain Ω = (−10, 10)2 as depicted on the left in Figure 2. In this domain, an initial
crack with length 2l0 = 2.0 and thickness h of two cells is prescribed with the help of the phase-field function
ϕ, i.e., ϕ = 0 in the crack and ϕ = 1 elsewhere. As boundary conditions, the displacements u are set to zero
on ∂Ω. We use homogeneous Neumann conditions for the phase-field variable, i.e., ∂n ϕ = 0 on ∂Ω. The
driving force is given by a constant pressure ρ = 10−3 Pa in the interior of the crack. An overview of the
parameter setting is given in Figure 2 on the right.

(−10, 10) (10, 10)

Parameter value
Ω (−10, 10)2
domain Ω
h test-dependent
l0 1.0
GC 1.0 kN/mm
crack C
E 1.0MPa
transition zone of size  µ 0.42 N/mm2
λ test-dependent
ν 0.2, 0.4999, 0.5
ρ 10−3 Pa

(−10, −10) (10, −10)

Figure 2. Left: geometry of the two-dimensional Sneddon’s test in 2d. Domain of size
(−10, 10)2 with a pressurized fracture. Right: setting of material and numerical parameters
for Sneddon’s benchmark test.
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 9

Two quantities of interest are discussed: the crack opening displacement (COD) and the total crack
volume (TCV). The analytical solution (from [49]) can be computed via
1
x2 2

pl0
CODref = 2 0 1 − 2 ,
E l0
where E 0 := E
1−ν 2 , E is the Young modulus and ν is the Poisson ratio. The TCV can be computed numerically
with
Z
TCV = u(x, y) · ∇ϕ(x, y)d(x, y).
Ω

The analytical solution (from [49]) is given by

2πpl02
.
TCVref =
E0
In Table 3, for κ = 10−8 , the average number of CG iterations increases with a decreasing mesh size. We
observe an increase in the CG iteration numbers in particular for the incompressible limit ν = 0.5 and finer
meshes, where we finally do not get convergence in the solver for smaller h. Already for ν = 0.4999 and a
problem size of less than 300 000 DoFs, the average number of CG iterations is above 100.
Remark (Difficulties considering small κ). In Table 3, compared to Table 5, we can evaluate the impact of
the setting of κ. We compute Sneddon’s test for different mesh sizes h, fixed bandwidth , for three Poisson
ratios ν = 0.2, 0.4999, and ν = 0.5, and for a small and large regularization parameter κ = 10−2 and
κ = 10−8 to evaluate its impact on the behavior of the CG solver. These solver dependencies on κ have
a natural correspondence in error estimates. For a decoupled linearized system, such estimates are shown
in [54, Section 5.5]. A numerical error analysis for this test on a good choice of κ can be found in [34].
Further, we observe an increased number of CG iterations for high Poisson ratios. The number of GMRES
and AS iterations do not differ significantly for different κ.

Table 3. Sneddon’s pressure-driven cavity in 2d. Average number of GMRES iterations

(#lin) per Newton step (#AS), the average number of CG iterations (∅CG) per linear
iteration. Based on the newly developed mixed model with Qc2 Qc1 Qc1 elements for different
problem sizes and setting of the length scale parameter  for three Poisson ratios. Quantities
of interest: CODmax and TCV and κ = 10−8 . Uniform refined meshes.

Sneddon’s test: robustness in h, λ, ; κ = 10−8 ; mixed

FE ν h  #DoFs ∅lin ∅CG #AS CODmax TCV
Qc2 Qc1 Qc1 0.2 0.707 1.414 16 484 3 26 4 0.00282 0.0240
Qc2 Qc1 Qc1 0.2 0.353 1.414 64 964 6 28 6 0.00270 0.0189
Qc2 Qc1 Qc1 0.2 0.176 1.414 257 924 9 35 4 0.00260 0.0164
Qc2 Qc1 Qc1 0.2 0.088 1.414 1 027 844 12 31 5 0.00252 0.0150
ref. [49] 0.2 0.0019200 0.00603
Qc2 Qc1 Qc1 0.4999 0.707 1.414 16 484 3 31 6 3.0383e-05 0.000257
Qc2 Qc1 Qc1 0.4999 0.353 1.414 64 964 7 46 8 3.6024e-05 0.000254
Qc2 Qc1 Qc1 0.4999 0.176 1.414 257 924 6 107 39 3.9899e-05 0.000252
Qc2 Qc1 Qc1 0.4999 0.088 1.414 1 027 844 5 57 24 4.2265e-05 0.000250
ref. [49] 0.4999 0.0015001 0.004713
Qc2 Qc1 Qc1 0.5 0.707 1.414 16 484 3 31 3 2.9937e-20 7.1504e-20
Qc2 Qc1 Qc1 0.5 0.353 1.414 64 964 6 25 2 1.3258e-19 2.3835e-19
Qc2 Qc1 Qc1 0.5 0.176 1.414 257 924 5 59 7 1.9309e-19 7.8981e-19
Qc2 Qc1 Qc1 0.5 0.088 1.414 1 027 844 - - - - -
Qc2 Qc1 Qc1 0.5 0.707 1.414 16 484 11 37 3 2.4585e-15 1.2562e-14
Qc2 Qc1 Qc1 0.5 0.353 0.707 64 964 6 32 3 2.3632e-18 1.0069e-17
Qc2 Qc1 Qc1 0.5 0.176 0.353 257 924 10 30 3 6.5953e-18 1.4749e-16
Qc2 Qc1 Qc1 0.5 0.088 0.176 1 027 844 14 38 3 1.2397e-18 2.6778e-18
ref. [49] 0.5 0.0015000 0.0047124

This observation is confirmed by the numerical results from Table 4, where a CG solver preconditioned
with AMG is used to approximate Ŝ −1 . The numerical results in Table 4 are based on the same tests as in
10 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

Table 4. Sneddon’s pressure-driven cavity in 2d. Average number of GMRES iterations

(#lin) per Newton step (#AS), the average number of CG iterations (∅CG) per linear
iteration, CG plus AMG is used for (g(ϕ̃)Au )−1 and Ŝ −1 . Based on the newly developed
mixed model with Qc2 Qc1 Qc1 elements for different problem sizes and setting of the length
scale parameter  for two Poisson ratios. Uniform refined meshes.

Sneddon’s test: robustness in h, λ, ; κ = 10−8 ; mixed; CG+AMG for Ŝ −1

FE ν h  #DoFs ∅lin ∅CG (g(ϕ̃)Au )−1 ∅CG Ŝ −1 #AS
Qc2 Qc1 Qc1 0.4999 0.707 1.414 16 484 3 26 1 3
Qc2 Qc1 Qc1 0.4999 0.353 1.414 64 964 8 56 6 8
Qc2 Qc1 Qc1 0.4999 0.176 1.414 257 924 6 106 6 38
Qc2 Qc1 Qc1 0.4999 0.088 1.414 1 027 844 6 42 6 69
Qc2 Qc1 Qc1 0.5 0.707 1.414 16 484 10 36 1 3
Qc2 Qc1 Qc1 0.5 0.353 1.414 64 964 6 26 6 8
Qc2 Qc1 Qc1 0.5 0.176 1.414 257 924 6 63 6 37
Qc2 Qc1 Qc1 0.5 0.088 1.414 1 027 844 7 41 6 101

Table 3 but for ν = 0.4999 and ν = 0.5. The number of linear iterations is moderate, and at most six CG
iterations are needed for Ŝ −1 .
As expected in Tables 3, 4 and 5, considering the quantities of interest CODmax and TCV, they get
vanishingly small for high Poisson ratios. This is what we expected for incompressible solids: a closed
domain does not change its volume; the opening of the initial crack in the interior of the domain is avoided.
For ν = 0.2, the quantities of interest are acceptable compared to the reference values. Also for ν = 0.2,
since all computations are conducted with uniformly refined meshes, moderate problem sizes, and fixed ,
we cannot expect excellent results in the quantities of interest.

Table 5. Sneddon’s pressure-driven cavity. Average number of GMRES iterations (#lin)

per Newton step (#AS), average number of CG iterations (∅CG) per linear iteration. Based
on the newly developed mixed model with Qc2 Qc1 Qc1 elements for different problem sizes and
setting of the length scale parameter  for three Poisson ratios. Quantities of interest:
CODmax and TCV and κ = 10−2 . Uniform refined meshes.

Sneddon’s test: robustness in h, λ, ; κ = 10−2 ; mixed

FE ν h  #DoFs ∅lin ∅CG #AS CODmax TCV
Qc2 Qc1 Qc1 0.2 0.707 1.414 16 484 2 16 4 0.00248 0.0224
Qc2 Qc1 Qc1 0.2 0.353 1.414 64 964 8 18 4 0.00227 0.0173
Qc2 Qc1 Qc1 0.2 0.176 1.414 257 924 9 18 15 0.00206 0.0145
Qc2 Qc1 Qc1 0.2 0.088 1.414 1 027 844 15 28 5 0.00190 0.0129
ref. [49] 0.2 0.0019200 0.0060
Qc2 Qc1 Qc1 0.4999 0.707 1.414 16 484 13 16 3 3.0833e-05 0.000269
Qc2 Qc1 Qc1 0.4999 0.353 1.414 64 964 8 18 14 3.1739e-05 0.000242
Qc2 Qc1 Qc1 0.4999 0.176 1.414 257 924 6 18 93 3.3667e-05 0.000224
Qc2 Qc1 Qc1 0.4999 0.088 1.414 1 027 844 7 26 65 3.4560e-05 0.000216
ref. [49] 0.4999 0.0015001 0.004713
Qc2 Qc1 Qc1 0.5 0.707 1.414 16 484 9 14 3 1.7339e-19 5.8895e-19
Qc2 Qc1 Qc1 0.5 0.353 1.414 64 964 9 18 14 2.3734e-19 5.6268e-18
Qc2 Qc1 Qc1 0.5 0.176 1.414 257 924 11 18 14 5.6547e-20 6.0823e-18
Qc2 Qc1 Qc1 0.5 0.088 1.414 1 027 844 5 26 39 7.7351e-19 2.2733e-17
Qc2 Qc1 Qc1 0.5 0.707 1.414 16 484 9 14 3 1.5881e-19 5.8895e-19
Qc2 Qc1 Qc1 0.5 0.353 0.707 64 964 6 17 3 1.9290e-19 1.8057e-18
Qc2 Qc1 Qc1 0.5 0.176 0.353 257 924 6 18 3 4.1847e-19 2.1156e-18
Qc2 Qc1 Qc1 0.5 0.088 0.176 1 027 844 10 26 3 2.4801e-18 9.5514e-18
ref. [49] 0.5 0.0015000 0.0047124

In Table 5, the same computations are conducted as in Table 3 and Table 4 for κ = 10−2 to discuss the
statement of Remark 4.4. The COD values are close to the reference values. Here, a large regularization
parameter κ = 10−2 stabilizes the block (g(ϕ̃)Au )−1 . Further, the linear iterations are stable, and also the
inner CG iterations are relatively constant. In the last four rows of Table 5, similar to Table 2, results of four
tests with  = 2 h are listed to check the robustness in  for ν = 0.5, which can be confirmed for Sneddon’s
benchmark test.
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 11

4.5. Sneddon’s pressure-driven cavity, layered. As a fourth test case, the pressure-driven cavity from [48]
is modified similarly to [5]. We consider a two-dimensional domain Ω = (−20, 20)2 . In contrast to the pre-
vious Sneddon test, a compressible layer of size 10 is added around the incompressible domain to allow
deforming of the solid on a finite domain. So the Poisson ratio changes over the domain for the layered
Sneddon test. We expect to get better results concerning COD and TCV on a finite domain compared to
the reference values on an infinite domain. A sketch of the geometry is given in Figure 3 on the left. The
setting of the material and numerical parameters is the same as in the previous section.

(−20, 20) (20, 20)

compressible layer

(in)compressible domain

crack C
transition zone of size 

(−20, −20) (20, −20)

Figure 3. Left: Geometry of the two-dimensional Sneddon’s test with a compressible layer
of size 10. Further, the inside of the initial crack is assumed to be compressible [5]. Right:
Close zoom-in to the geometrically refined mesh around the crack, used in Table 7.

In Figure 3 on the right, a zoom-in snapshot of the inner domain is given to show the geometric refinement
for the tests in Tables 6, 7, and 8. Aside from the adaptively refined mesh, we set  = h, depending on the
current mesh size. The total numbers of degrees of freedom (#DoFs) on Ω are listed in the numerical results
in Tables 6 to 8.
In Table 6, the results for the Sneddon test in 2d with a compressible layer around a possibly incompressible
domain are given for three Poisson ratios and adaptively refined meshes, with  = h, and κ = 10−2 . We choose
κ = 10−2 to avoid the effects of κ on the inner CG iterations. For large κ, the computed quantities of interest
CODmax and TCV do not converge to the correct physics (κ ≈ 0), however they still converge, but to values
corresponding to large κ material’s physics. In Table 6, the numbers of GMRES iterations are moderate for
ν = 0.2. For higher Poisson ratios, we observe high linear iteration numbers. The incompressibility and the
mesh adaptivity seem to significantly impact the linear solver. We observe the same effects for κ = 10−8 in
Table 7.
In Table 7, the numerical results of the same tests are given as in Table 6 for κ = 10−8 . Analogously
to Table 4, Table 8 contains the numerical results for the Sneddon test layered for high Poisson ratios and
small κ. In contrast to Table 7, we approximate Ŝ −1 with a CG solver which is preconditioned with AMG.
The results of CODmax and TCV in Tables 7 and 8 look promising for all three Poisson ratios. For ν = 0.5
the solver does not converge with sufficiently small κ and h → 0. An explanation is given in Remark 4.4
(Section 4.4). In Table 8 for high Poisson ratios, the modified approximation of Ŝ −1 changes the behavior
of the linear solver. With a relative tolerance of 10−6 for the preconditioned CG solver for (g(ϕ̃)Au )−1 and
Ŝ −1 , we observe that more GMRES iterations are required. The number of linear iterations is relatively high,
but nearly constant for ν = 0.4999 and ν = 0.5. The number of linear iterations increases for higher Poisson
ratios with adaptive refined meshes and  = h. The results of CODmax and TCV match the manufactured
reference values.
In Figure 4, the solutions of ux , uy , p, and ϕ are presented as zoom-in snapshots for ν = 0.5 with a
compressible layer, based on Table 8. Especially the pressure field (upper left snapshot) is expected to have
zero values in the interior of the crack and the maximal values in the crack tip on the left and the right of
the pre-defined initial crack. Further, in Figure 4, the mesh on the finest refinement level is given on the
bottom left. On the bottom right, the crack zone is shown, on which the computed solutions are presented
above to get an impression of the mesh size around the fracture.
12 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

Table 6. Sneddon’s pressure-driven cavity layered. Average number of GMRES iterations

(∅lin) per Newton step (#AS). Computations based on the newly developed mixed model
with Qc2 Qc1 Qc1 elements for different problem size,  = h for three Poisson ratios. Quantities
of interest: CODmax and TCV and κ = 10−2 . Geometrically refined mesh in the area around
the crack zone as depicted in Figure 3.

Sneddon layered adaptive: robustness in h, λ, ;  = h κ = 10−2 ; mixed

FE ν h #DoFs ∅lin #AS CODmax TCV
Qc2 Qc1 Qc1 0.2 0.353 257 924 18 3 0.00214077 0.0097207
Qc2 Qc1 Qc1 0.2 0.176 263 604 25 4 0.00188194 0.0069353
Qc2 Qc1 Qc1 0.2 0.088 282 484 20 3 0.00163459 0.0055415
Qc2 Qc1 Qc1 0.2 0.044 350 804 17 3 0.00136002 0.0044906
Qc2 Qc1 Qc1 0.2 0.022 610 164 19 4 0.00104379 0.0034504
Qc2 Qc1 Qc1 0.2 0.011 1 620 244 24 6 0.00071731 0.0024168
Qc2 Qc1 Qc1 0.2 0.0055 5 606 324 28 6 0.00043963 0.0015294
ref. [49] 0.2 0.00192000 0.0060318
Qc2 Qc1 Qc1 0.4999 0.353 257 924 40 2 0.00205349 0.0108334
Qc2 Qc1 Qc1 0.4999 0.176 263 604 52 2 0.00168136 0.0069892
Qc2 Qc1 Qc1 0.4999 0.088 282 484 58 3 0.00143863 0.0052347
Qc2 Qc1 Qc1 0.4999 0.044 350 804 58 3 0.00122931 0.0041947
Qc2 Qc1 Qc1 0.4999 0.022 610 164 62 4 0.00099810 0.0033284
Qc2 Qc1 Qc1 0.4999 0.011 1 620 244 150 4 0.00073786 0.0024681
Qc2 Qc1 Qc1 0.4999 0.0055 5 606 324 318 7 0.00048545 0.0016595
ref. [49] 0.4999 0.00150019 0.0047130
Qc2 Qc1 Qc1 0.5 0.353 257 924 40 2 0.00205332 0.0108334
Qc2 Qc1 Qc1 0.5 0.176 263 604 52 2 0.00168117 0.0069889
Qc2 Qc1 Qc1 0.5 0.088 282 484 59 3 0.00143847 0.0052343
Qc2 Qc1 Qc1 0.5 0.044 350 804 57 4 0.00122920 0.0041944
Qc2 Qc1 Qc1 0.5 0.022 610 164 65 4 0.00099804 0.0033282
Qc2 Qc1 Qc1 0.5 0.011 1 620 244 155 4 0.00073784 0.0024681
Qc2 Qc1 Qc1 0.5 0.0055 5 606 324 271 7 0.00048545 0.0016595
ref. [49] 0.5 0.00150000 0.0047124

Table 7. Sneddon’s pressure-driven cavity layered. Average number of GMRES iterations

(∅lin) per Newton step (#AS). Computations with Qc2 Qc1 Qc1 elements for different problem
size,  = h for three Poisson ratios. Quantities of interest: CODmax and TCV and κ = 10−8 .
Geometrically refined mesh as depicted on the right in Figure 3.

Sneddon layered adaptive: robustness in h, λ, ;  = h, κ = 10−8 ; mixed

FE(u, p, ϕ) ν h #DoFs ∅lin #AS CODmax TCV
Qc2 Qc1 Qc1 0.2 0.353 257 924 10 3 0.00242526 0.0107193
Qc2 Qc1 Qc1 0.2 0.176 263 604 20 3 0.00221789 0.0080340
Qc2 Qc1 Qc1 0.2 0.088 282 484 18 3 0.00208683 0.0069646
Qc2 Qc1 Qc1 0.2 0.044 350 804 29 6 0.00200814 0.0064862
Qc2 Qc1 Qc1 0.2 0.022 610 164 26 4 0.00196329 0.0062530
Qc2 Qc1 Qc1 0.2 0.011 1 620 244 32 3 0.00193890 0.0061344
Qc2 Qc1 Qc1 0.2 0.0055 5 606 324 40 3 0.00192609 0.0060733
ref. [49] 0.2 0.00192000 0.0060318
Qc2 Qc1 Qc1 0.4999 0.353 257 924 40 2 0.00223914 0.0116630
Qc2 Qc1 Qc1 0.4999 0.176 263 604 74 5 0.00187365 0.0077192
Qc2 Qc1 Qc1 0.4999 0.088 282 484 229 4 0.00168693 0.0060788
Qc2 Qc1 Qc1 0.4999 0.044 350 804 511 4 0.00159278 0.0053537
Qc2 Qc1 Qc1 0.4999 0.022 610 164 601 6 0.00154436 0.0050158
Qc2 Qc1 Qc1 0.4999 0.011 1 620 244 565 5 0.00151941 0.0048527
Qc2 Qc1 Qc1 0.4999 0.0055 5 606 324 641 5 0.00150668 0.00477248
ref. [49] 0.4999 0.00150019 0.0047130
Qc2 Qc1 Qc1 0.5 0.353 257 924 40 2 0.00223891 0.0116629
Qc2 Qc1 Qc1 0.5 0.176 263 604 73 5 0.00187338 0.0077187
Qc2 Qc1 Qc1 0.5 0.088 282 484 227 4 0.00168667 0.0060782
ref. [49] 0.5 0.0015000 0.0047124
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 13

Table 8. Sneddon’s pressure-driven cavity layered with Qc2 Qc1 Qc1 elements and  = h.
Average number of GMRES iterations (∅lin) per Newton step (#AS). CG plus AMG is
used for (g(ϕ̃)Au )−1 and Ŝ −1 : the average number of CG iterations for (g(ϕ̃)Au )−1 is 38
for ν = 0.4999 and 36 for ν = 0.5. The average number of CG iterations for Ŝ −1 is 8 for
ν = 0.4999 and 7 for ν = 0.5. Computations for different problem size,  = h for three
Poisson ratios. Quantities of interest: CODmax and TCV and κ = 10−8 . Geometrically
refined mesh as depicted on the right in Figure 3.

Sneddon layered adaptive: robustness in h, λ, ;  = h, κ = 10−8 ; mixed; CG for two blocks

FE(u, p, ϕ) ν h #DoFs ∅lin #AS CODmax TCV
Qc2 Qc1 Qc1 0.4999 0.353 257 924 40 2 0.00223914 0.0116630
Qc2 Qc1 Qc1 0.4999 0.176 263 604 70 4 0.00187365 0.0077192
Qc2 Qc1 Qc1 0.4999 0.088 282 484 165 4 0.00168693 0.0060788
Qc2 Qc1 Qc1 0.4999 0.044 350 804 153 4 0.00159278 0.0053537
Qc2 Qc1 Qc1 0.4999 0.022 610 164 145 5 0.00154436 0.0050158
Qc2 Qc1 Qc1 0.4999 0.011 1 620 244 139 5 0.00151941 0.0048527
Qc2 Qc1 Qc1 0.4999 0.0055 5 606 324 148 5 0.00150668 0.0047724
ref. [49] 0.4999 0.00150019 0.0047130
Qc2 Qc1 Qc1 0.5 0.353 257 924 40 2 0.00223891 0.0116629
Qc2 Qc1 Qc1 0.5 0.176 263 604 70 4 0.00187338 0.0077187
Qc2 Qc1 Qc1 0.5 0.088 282 484 200 4 0.00168667 0.0060782
Qc2 Qc1 Qc1 0.5 0.044 350 804 229 4 0.00159254 0.0060782
Qc2 Qc1 Qc1 0.5 0.022 610 164 238 7 0.00154414 0.0050151
Qc2 Qc1 Qc1 0.5 0.011 1 620 244 220 5 0.00151920 0.0048521
Qc2 Qc1 Qc1 0.5 0.0055 5 606 324 222 5 0.00150648 0.0047718
ref. [49] 0.5 0.00150000 0.0047124

4.6. Single edge notched pure tension test. As the last example, we use the single-edge notched tension
test from Miehe et al. [42] testing with three Poisson ratios. We use the predictor-corrector scheme from
Heister et al. [25] for two steps of adaptive mesh refinement on four times uniformly refined mesh with a
phase-field threshold of 0.5. The parameter setting is the same as in [42] but we use the mixed problem
formulation and discretization from Section 2 and vary the Poisson ratio; see Table 9.

Table 9. Parameter setting for three tests with different Poisson’s ratios for the single-edge
notched tension test with κ = 10−8 , and  = 4 h. The maximal number of DoFs is given in
the last column for the test cases. For all tests, four uniform (h = 0.011) and two adaptive
refinement steps are conducted with a phase-field threshold of 0.5 for predictor-corrector.

ν µ λ #DoFs
0.3 80.77 · 103 121.15 · 103 19 584
0.45 80.77 · 103 726.93 · 103 19 704
0.49 80.77 · 103 3957.73 · 103 19 498

We consider the bulk and crack energy as two further numerical quantities of interest. The bulk energy
EB can be computed via
Z
EB (u, ϕ) = (g(ϕ̃)ψ(Elin (u)) d(x, y),
Ω
where the strain energy functional is defined as

1 2
ψ(Elin (u)) := µ tr Elin (u)2 + λ tr (Elin (u)) .
2
Here, no manufactured reference values are provided and we only present values computed numerically.
Further, we compute the crack energy EC via
(ϕ − 1)2
Z  
GC
EC (u, ϕ) = + |∇ϕ|2 d(x, y).
2 Ω 
Again, no manufactured reference values are provided. At least for ν = 0.3, we can compare our results for
EB and EC with reference values from the literature, e.g., [1, 40]. In Figures 5 and 6, on the left side, the
bulk and the crack energy are plotted versus the incremental step number. On the right of Figures 5 to 6,
14 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

Figure 4. Sneddon 2d layered. Upper four snapshots: zoom-in solutions from left to right,
and top to bottom: the pressure field, the phase-field, the displacement in the x-direction
and the displacement in the y-direction. The solutions are for ν = 0.5 from Table 8 on the
finest level with Qc2 Qc1 Qc1 elements. The solutions fit the reference values from [5]. Lower
two snapshots show on the left the whole domain (−20, 20)2 with the adaptively refined
mesh in the last refinement step. On the bottom right, a zoom-in snapshot of the crack
zone is given on the domain, where the upper snapshots are taken.

the average number of linear iterations and the number of Newton/AS steps are plotted. The number of
linear iterations behaves differently for ν = 0.3 from the results for higher Poisson ratios. While for ν = 0.3,
in Figure 5 on the right, the linear iterations decrease if the crack starts propagating, in Figure 6, the linear
iterations increase up to an average of more than 70 iterations at the end of the crack simulations.
In Figure 7, snapshots of the pressure field and phase-field are given for ν = 0.49, where – to the author’s
knowledge – no reference values are available in the literature. The crack paths look similar as for ν = 0.3,
but a slight asymmetry is visible in the crack path. We decided to present the crack path during the
simulation to depict the pressure field with the maximal value in front of the crack tip while the pressure
values in the crack are zero. The computed bulk and crack energies in Figure 5 fit well to results in the
literature, e.g., [25]. The bulk energy increases until the critical energy release rate is reached, and the crack
energy increases when the crack propagates while the bulk energy releases. Also, in Figure 6, the bulk and
crack energy curves fit the observed crack pattern in Figure 7. For ν = 0.49 with snapshots in the last
column in Figure 7, no comparable results in the literature are available. The crack pattern differs from
the snapshots for smaller Poisson ratios. We observe that the crack has an orientation to the upper left
corner, and a second crack develops from the singularity in the corner, where non-homogeneous Dirichlet
boundary conditions and Neumann boundary conditions meet. In the first column in Figure 7, the pressure
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 15

16

EB , ν = 0.3 14
1.5

Bulk/Crack energy [J]

EC , ν = 0.3
12

∅lin/#AS
1 10 ∅lin, ν = 0.3
#AS, ν = 0.3
8
0.5
6

0 4

0 20 40 60 80 0 20 40 60 80
Incremental step Incremental step

Figure 5. Left: bulk (EB ) and crack energy (EC ) for the single-edge notched tension test,
AT2 functional, adaptive refined meshes. The incremental step size was 10−4 s for the first
58 steps and reduced to 10−5 s. Right: number of linear iterations on average per Newton
step (∅lin), and number of active set/Newton steps (#AS) against the incremental steps.
The crack starts propagating at incremental step 66, mesh refinement starts at step 57.
ν = 0.3.
80

2 EB , ν = 0.49 ∅lin, ν = 0.49

Bulk/Crack energy [J]

EC , ν = 0.49 60 #AS, ν = 0.49

1.5
∅lin/#AS
40
1

0.5 20

0
0
0 20 40 60 80 0 20 40 60 80
Incremental step Incremental step

Figure 6. Left: bulk (EB ) and crack energy (EC ) for the single-edge notched tension test,
AT2 functional, adaptive refined meshes. The incremental step size was 10−4 s for the first
48 steps and reduced to 10−5 s. Right: number of linear iterations on average per Newton
step (∅lin), and number of active set/Newton steps (#AS) against the incremental steps.
The crack starts propagating at incremental step 60, mesh refinement starts at step 52.
ν = 0.49.

and phase-field solution is given for ν = 0.3 after total failure. The crack propagates from the center of the
geometry to the left boundary, as we expect it. Further, one can see a pure zero pressure field after total
failure.

5. Conclusions
In this work, a preconditioner for a mixed formulation phase-field fracture model that is robust in h, , and
λ was developed and tested on four numerical examples for different Poisson ratios up to the incompressible
limit, namely ν → 0.5 yielding λ → ∞. For the first test case, a hanging block with a slit, we confirmed the
robustness and efficiency of the physics-based preconditioner, discretized with Qc2 Qc1 Qc1 finite elements. To
the best of the authors’ knowledge, in the last test case the well-known single edge notched tension test was
considered for higher Poisson ratios for the first time. For ν = 0.49 a non-symmetric crack behavior and
crack initiation from the upper left corner singularity was observed. In Sneddon’s test case and κ = 10−8 ,
an impact of κ on the condition of the κ-dependent block entries of the system matrix could be explicitly
seen.
It is well-known that from a phase-field perspective the regularization parameter  is challenging, in
particular its choice in relation to h. However, we found in this paper that from a preconditioner perspective
the first regularization parameter κ (in the bulk term of the displacement equation) causes difficulties instead.
Basically, we deal with an elliptic (Laplacian) term where diffusion ranges from κ ≈ 10−8 in the crack
region to 1, a difference of 8 orders of magnitude. We expect that a carefully designed geometric multigrid
preconditioner or a weighted BFBT preconditioner might handle this situation better. We emphasize that
16 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

Figure 7. Snapshots of the solution for the single-edge notched tension test with ν =
0.3, 0.45, and 0.49 from left to right. Adaptive mesh refinement with predictor-corrector.
Pressure field (first row) and phase-field (second row) on deformed two-dimensional domain
in incremental step 88 for ν = 0.3, step 70 for ν = 0.45, and step 75 for ν = 0.49.

having robustness in h,  and λ is a significant contribution, which has not yet been studied so far in
the published literature. A second future extension would be thermodynamically consistent constitutive
materials laws, namely incorporating stress splitting in σ(u, p).

Acknowledgements
K. Mang thanks Clemson University for the financial support for a one-month research stay.
T. Heister was partially supported by the National Science Foundation (NSF) Award DMS-2028346,
OAC-2015848, EAR-1925575, by the Computational Infrastructure in Geodynamics initiative (CIG), through
the NSF under Award EAR-0949446 and EAR-1550901 and The University of California – Davis, and by
Technical Data Analysis, Inc. through US Navy STTR Contract N68335-18-C-0011.
Clemson University is acknowledged for generous allotment of compute time on Palmetto cluster.
The second and third authors were supported by the German Research Foundation, Priority Program
1748 (DFG SPP 1748) within the subproject Structure Preserving Adaptive Enriched Galerkin Methods for
Pressure 3D Fracture Phase-Field Models (WI 4367/2-1) with the project number 392587580.

References
[1] M. Ambati, T. Gerasimov, and L. De Lorenzis. Phase-field modeling of ductile fracture. Computational Mechanics,
55(5):1017–1040, 2015.
[2] Marreddy Ambati, Tymofiy Gerasimov, and Laura De Lorenzis. A review on phase-field models of brittle fracture and a
new fast hybrid formulation. Computational Mechanics, 55(2):383–405, 2015.
[3] Daniel Arndt, Wolfgang Bangerth, Bruno Blais, Thomas C. Clevenger, Marc Fehling, Alexander V. Grayver, Timo Heister,
Luca Heltai, Martin Kronbichler, Matthias Maier, Peter Munch, Jean-Paul Pelteret, Reza Rastak, Ignacio Thomas, Bruno
Turcksin, Zhuoran Wang, and David Wells. The deal.II library, version 9.2. Journal of Numerical Mathematics, 28(3):131–
146, 2020.
[4] Daniel Arndt, Wolfgang Bangerth, Denis Davydov, Timo Heister, Luca Heltai, Martin Kronbichler, Matthias Maier,
Jean-Paul Pelteret, Bruno Turcksin, and David Wells. The deal.II finite element library: Design, features, and insights.
Computers & Mathematics with Applications, 81:407–422, 2021.
[5] Seshadri Basava, Katrin Mang, Mirjam Walloth, Thomas Wick, and Winnifried Wollner. Adaptive and pressure-robust
discretization of incompressible pressure-driven phase-field fracture. arXiv preprint arXiv:2006.16566, 2020.
[6] Michele Benzi, Gene H Golub, and Jörg Liesen. Numerical solution of saddle point problems. Acta numerica, 14:1–137,
2005.
[7] Daniele Boffi, Franco Brezzi, and Michel Fortin. Finite elements for the stokes problem. Mixed Finite Elements, Compat-
ibility Conditions, and Applications: Lectures given at the CIME Summer School held in Cetraro, Italy, June 26-July 1,
2006, page 45, 2008.
[8] B. Bourdin. Numerical implementation of the variational formulation for quasi-static brittle fracture. Interfaces and free
boundaries, 9:411–430, 2007.
 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS 17

[9] B. Bourdin, C. Chukwudozie, and K. Yoshioka. A variational approach to the numerical simulation of hydraulic fracturing.
SPE Journal, Conference Paper 159154-MS, 2012.
[10] B. Bourdin, G.A. Francfort, and J.-J. Marigo. Numerical experiments in revisited brittle fracture. Journal of the Mechanics
and Physics of Solids, 48(4):797–826, 2000.
[11] James H Bramble and Joseph E Pasciak. A preconditioning technique for indefinite systems resulting from mixed approx-
imations of elliptic problems. Mathematics of Computation, 50(181):1–17, 1988.
[12] Mats Kirkesaether Brun, Thomas Wick, Inga Berre, Jan Martin Nordbotten, and Florin Adrian Radu. An iterative
staggered scheme for phase field brittle fracture propagation with stabilizing parameters. Computer Methods in Applied
Mechanics and Engineering, 361:112752, 2020.
[13] S. Burke, Ch. Ortner, and E. Süli. An adaptive finite element approximation of a variational model of brittle fracture.
SIAM J. Numer. Anal., 48(3):980–1012, 2010.
[14] Philippe G. Ciarlet. The finite element method for elliptic problems. North-Holland, Amsterdam [u.a.], 2. pr. edition, 1987.
[15] Thomas C. Clevenger and Timo Heister. Comparison between algebraic and matrix-free geometric multigrid for a stokes
problem on adaptive meshes with variable viscosity. Numerical Linear Algebra with Applications, page e2375, 2021.
[16] The step-20 tutorial program of deal.II. [Online; accessed 15-January-2022].
[17] D. Drzisga, L. John, U. Rüde, B. Wohlmuth, and W. Zulehner. On the analysis of block smoothers for saddle point
problems. SIAM Journal on Matrix Analysis and Applications, 39(2):932–960, 2018.
[18] Howard C Elman, David J Silvester, and Andrew J Wathen. Finite elements and fast iterative solvers: with applications
in incompressible fluid dynamics. Numerical Mathematics and Scie, 2014.
[19] Patrick Farrell and Corrado Maurini. Linear and nonlinear solvers for variational phase-field models of brittle fracture.
International Journal for Numerical Methods in Engineering, 109(5):648–667, 2017.
[20] G.A. Francfort and J.-J. Marigo. Revisiting brittle fracture as an energy minimization problem. Journal of the Mechanics
and Physics of Solids, 46(8):1319–1342, 1998.
[21] Michael W Gee, Christopher M Siefert, Jonathan J Hu, Ray S Tuminaro, and Marzio G Sala. Ml 5.0 smoothed aggregation
user’s guide. Technical report, Technical Report SAND2006-2649, Sandia National Laboratories, 2006.
[22] Piotr P. Grinevich and Maxim A. Olshanskii. An iterative method for the stokes-type problem with variable viscosity.
SIAM Journal on Scientific Computing, 31(5):3959–3978, January 2009.
[23] Wolfgang Hackbusch. Multi-grid methods and applications, volume 4. Springer Science & Business Media, 2013.
[24] Timo Heister, Katrin Mang, and Thomas Wick. Schur-type preconditioning of a phase-field fracture model in mixed form.
PAMM, 21(1):e202100065, 2021.
[25] Timo Heister, Mary F Wheeler, and Thomas Wick. A primal-dual active set method and predictor-corrector mesh adap-
tivity for computing fracture propagation using a phase-field approach. Computer Methods in Applied Mechanics and
Engineering, 290:466–495, 2015.
[26] Timo Heister and Thomas Wick. Parallel solution, adaptivity, computational convergence, and open-source code of 2d and
3d pressurized phase-field fracture problems. PAMM, 18(1):e201800353, 2018.
[27] Timo Heister and Thomas Wick. pfm-cracks: A parallel-adaptive framework for phase-field fracture propagation. Software
Impacts, 6:100045, 2020.
[28] Michael A Heroux, Roscoe A Bartlett, Vicki E Howle, Robert J Hoekstra, Jonathan J Hu, Tamara G Kolda, Richard B
Lehoucq, Kevin R Long, Roger P Pawlowski, Eric T Phipps, et al. An overview of the trilinos project. ACM Transactions
on Mathematical Software (TOMS), 31(3):397–423, 2005.
[29] Michael A Heroux, Roscoe A Bartlett, Vicki E Howle, Robert J Hoekstra, Jonathan J Hu, Tamara G Kolda, Richard B
Lehoucq, Kevin R Long, Roger P Pawlowski, Eric T Phipps, et al. Trilinos web page, 2021. [Online; accessed 25-November-
2021].
[30] D. Jodlbauer. Parallel Multigrid Solvers for Nonlinear Coupled Field Problems. PhD thesis, Johannes Kepler University
Linz, 2021.
[31] D. Jodlbauer, U. Langer, and T. Wick. Parallel matrix-free higher-order finite element solvers for phase-field fracture
problems. Mathematical and Computational Applications, 25(3):40, 2020.
[32] Daniel Jodlbauer, Ulrich Langer, and Thomas Wick. Matrix-free multigrid solvers for phase-field fracture problems. Com-
puter Methods in Applied Mechanics and Engineering, 372:113431, 2020.
[33] Axel Klawonn and Gerhard Starke. Block triangular preconditioners for nonsymmetric saddle point problems: field-of-
values analysis. Numerische Mathematik, 81(4):577–594, 1999.
[34] Leon Kolditz and Katrin Mang. On the relation of gamma-convergence parameters for pressure-driven quasi-static phase-
field fracture. Examples and Counterexamples, 2:100047, 2022.
[35] Alena Kopaničáková and Rolf Krause. A recursive multilevel trust region method with application to fully monolithic
phase-field models of brittle fracture. Computer Methods in Applied Mechanics and Engineering, 360:112720, 2020.
[36] Martin Kronbichler, Timo Heister, and Wolfgang Bangerth. High accuracy mantle convection simulation through modern
numerical methods. Geophysical Journal International, 191(1):12–29, 2012.
[37] Charlotte Kuhn and Ralf Müller. A continuum phase field model for fracture. Engineering Fracture Mechanics, 77(18):3625–
3634, 2010.
[38] S. Lee, M. F. Wheeler, and T. Wick. Pressure and fluid-driven fracture propagation in porous media using an adaptive
finite element phase field model. Computer Methods in Applied Mechanics and Engineering, 305:111 – 132, 2016.
[39] K. Mang. Phase-field fracture modeling, numerical solution, and simulations for compressible and incompressible solids.
PhD thesis, Leibniz University Hannover, 2021.
[40] Katrin Mang, Thomas Wick, and Winnifried Wollner. A phase-field model for fractures in nearly incompressible solids.
Computational Mechanics, 65(1):61–78, 2020.
 18 ROBUST PRECONDITIONING FOR A MIXED FORMULATION OF PHASE-FIELD FRACTURE PROBLEMS

[41] D.A. May, J. Brown, and L. Le Pourhiet. A scalable, matrix-free multigrid preconditioner for finite element discretizations
of heterogeneous stokes flow. Computer Methods in Applied Mechanics and Engineering, 290:496–523, June 2015.
[42] Christian Miehe, Martina Hofacker, and Fabian Welschinger. A phase field model for rate-independent crack propagation:
Robust algorithmic implementation based on operator splits. Computer Methods in Applied Mechanics and Engineering,
199(45-48):2765–2778, 2010.
[43] Christian Miehe, Fabian Welschinger, and Martina Hofacker. Thermodynamically consistent phase-field models of fracture:
Variational principles and multi-field fe implementations. International Journal for Numerical Methods in Engineering,
83(10):1273–1311, 2010.
[44] Malcolm F. Murphy, Gene H. Golub, and Andrew J. Wathen. A note on preconditioning for indefinite linear systems.
SIAM J. Sci. Comput., 21:1969–1972, 2000.
[45] Maxim A. Olshanskii and Arnold Reusken. Analysis of a stokes interface problem. Numer. Math., 103(1):129–149, mar
2006.
[46] Johann Rudi, Georg Stadler, and Omar Ghattas. Weighted bfbt preconditioner for stokes flow problems with highly
heterogeneous viscosity. SIAM Journal on Scientific Computing, 39(5):S272–S297, 2017.
[47] Yousef Saad. Iterative methods for sparse linear systems. SIAM, 2003.
[48] J. Schröder, T. Wick, S. Reese, P. Wriggers, R. Müller, S. Kollmannsberger, M. Kästner, A. Schwarz, M. Igelbüscher,
N. Viebahn, H. R. Bayat, S. Wulfinghoff, K. Mang, E. Rank, T. Bog, D. d’Angella, M. Elhaddad, P. Hennig, A. Düster,
W. Garhuom, S. Hubrich, M. Walloth, W. Wollner, Ch. Kuhn, and T. Heister. A selection of benchmark problems in solid
mechanics and applied mathematics. Archives of Computational Methods in Engineering, 28(2):713–751, 2021.
[49] I. N. Sneddon and M. Lowengrub. Crack problems in the classical theory of elasticity. SIAM series in Applied Mathematics.
John Wiley and Sons, Philadelphia, 1969.
[50] Ian Naismith Sneddon. The distribution of stress in the neighbourhood of a crack in an elastic solid. Proc. Roy. Soc.
London Ser. A, 187:229–260, 1946.
[51] E. Tanné, T. Li, B. Bourdin, J.-J. Marigo, and C. Maurini. Crack nucleation in variational phase-field models of brittle
fracture. Journal of the Mechanics and Physics of Solids, 110:80–99, 2018.
[52] Ray S Tuminaro and Charles Tong. Parallel smoothed aggregation multigrid: Aggregation strategies on massively parallel
machines. In SC’00: Proceedings of the 2000 ACM/IEEE Conference on Supercomputing, pages 5–5. IEEE, 2000.
[53] Rüdiger Verfürth. Error estimates for a mixed finite element approximation of the stokes equations. RAIRO. Analyse
numérique, 18(2):175–182, 1984.
[54] Thomas Wick. Multiphysics Phase-Field Fracture: Modeling, Adaptive Discretizations, and Solvers, volume 28. De
Gruyter, 2020.

View publication stats