betaFeedback

WorksComputer Science

Calibration and Validation of a Phase-Field Model of Brittle Fracture within the Damage Mechanics Challenge

Abstract

In the context of the Damage Mechanics Challenge, we adopt a phase-field model of brittle fracture to blindly predict the behavior up to failure of a notched three-point-bending specimen loaded under mixed-mode conditions. The beam is additively manufactured using a geo-architected gypsum based on the combination of bassanite and a water-based binder. The calibration of the material parameters involved in the model is based on a set of available independent experimental tests and on a two-stage procedure. In the first stage an estimate of most of the elastic parameters is obtained, whereas the remaining parameters are optimized in the second stage so as to minimize the discrepancy between the numerical predictions and a set of experimental results on notched three-point-bending beams. The good agreement between numerical predictions and experimental results in terms of load-displacement curves and crack paths demonstrates the predictive ability of the model and the reliability of the calibration procedure.

Highlights

Calibration and Validation of a Phase-Field Model of Brittle Fracture within the Damage Mechanics Challenge

Jonas Heinzmann, Pietro Carrara, Chenyi Luo, Manav Manav, Akanksha Mishra, Sindhu Nagaraja, Hamza Oudich, Francesco Vicentini, Laura De Lorenzis

  • The phase-field approach is used to model brittle fracture of additively manufactured orthotropic rock.

  • We propose a calibration procedure based on a set of independent experimental tests.

  • The behavior of an additional test is blindly predicted using the calibrated parameters.

  • Experimental and numerical load–displacement curves and crack paths are compared.

  • The fracture behavior is studied for different degrees of mode mixity.

Introduction

The damage mechanics challenge (DMC), issued during a workshop held at Purdue University in 2019 [1], aims at comparing the performances of different computational approaches and calibration strategies in predicting the behavior up to failure of a notched beam subjected to a mixed-mode three-point-bending (TPB) test [2]. The beam is made of a geo-architected layered gypsum created using an additive manufacturing process [3, 4] based on the crystallization of bassanite powder by means of a proprietary water-based binder. While the experimental data for the DMC test geometry are not provided to the participants before the submission of their predictions, a set of independent characterization test results is made available [5, 6].

This paper presents the modeling approach, the calibration strategy and the numerical results obtained by the GRIPHFiTH team from the Computational Mechanics Group of ETH Zürich. We adopt the phase-field approach to brittle fracture, originally proposed by Bourdin et al. [7] as the regularization of the variational reformulation of Griffith’s brittle fracture in [8], and later re-interpreted as a special class of gradient damage models [9]. In models of this type, beside the displacement field, an additional field variable, denoted as phase-field or damage variable dd, smoothly varies between 0 and 1 to denote the transition of the material state between pristine and fully broken conditions, respectively. The numerical implementation of the approach does not require remeshing upon crack propagation nor a priori knowledge of the crack path; crack nucleation, branching, merging and any other topological changes are automatically described with no need for ad-hoc criteria. These features, along with a strong mathematical background, make the phase-field approach very attractive to model various fracture phenomena including brittle [7, 10, 11], ductile [12, 13, 14], anisotropic [15, 16, 17, 18, 19] and fatigue fracture [20, 21, 22, 23], as well as fracture in heterogeneous materials [24, 25, 26].

In this paper, we base our predictions on the assumption of a macroscopically homogeneous orthotropic material undergoing brittle fracture and we calibrate the material parameters following a procedure subdivided in two stages. In the first stage, we obtain an estimate for most of the elastic parameters using the results of plane wave propagation velocity tests and unconfined compression tests, part of the available DMC data. This allows us to limit the number of free parameters to be optimized during the second stage to two, namely one elastic parameter and the fracture toughness. The optimization procedure of the second stage aims at minimizing the mismatch between the experimental load-deflection curves from a set of notched TPB tests, also part of the available DMC data, and their numerical prediction, as measured by a properly defined cost function. The adopted model and the calibrated parameters are first verified by comparing the numerical and the experimental results for the same notched TPB specimens used for calibration, in terms of both load-deflection curves and crack paths. Then, the behavior of the specimen with the DMC test geometry is blindly predicted; finally, the corresponding numerical results are compared to the experimental results made available by the organizers after submission of the team predictions.

This paper is structured as follows: § 3 briefly overviews the DMC tasks and the experimental data available for model calibration. § 4 illustrates the adopted model and the related numerical aspects, while § 5 is devoted to the calibration of the material parameters. § 6 presents the comparison between numerical and experimental data for the notched TPB tests used for calibration and for the DMC test geometry. Conclusions are drawn in § 7.

Damage mechanics challenge

In this section, we summarize the main tasks assigned to the DMC participants and the characterization tests used to calibrate the adopted model. For a more detailed description, the interested reader should refer to [27].

Damage mechanics challenge task

The main assigned task involves the prediction of the behavior up to complete failure of a notched beam made of an additively manufactured geomaterial subjected to a monotonic TPB test, whose geometry and loading conditions are summarized in Figure 1a. The beam is supported and loaded by means of three aluminum rods covering its complete width. In particular, the two lower rods act as supports, while the third one transmits the load to the midspan section of the beam. The test is performed in displacement control by imposing the midspan deflection at a rate of 0.03 mm/min, hence sufficiently slow to consider the test as quasi-static. The notch is directly included in the specimen geometry during the manufacturing process and has a width of 1 mm, is off-centered by 9.52 mm with respect to the midspan position, and its longitudinal axis forms an angle of about 63.4∘63.4^\circ with the longitudinal axis of the beam. Also, its height is linearly tapered and decreases from 5.08 mm to 2.54 mm across the notch width (Figure 1a). This setup aims at inducing a complex mixed-mode fracture condition including all three modes, namely opening or mode I, sliding or mode II and tearing or mode III.

Image not extracted

Figure 1. Geometry and loading conditions of the notched beams for (a) the DMC test and (b) the characterization tests. All the beams have the same width W=12.7W = 12.7 mm, height H=25.4H = 25.4 mm and length L=76.2L = 76.2 mm; the midspan is indicated with a dashed-dotted line.

The complexity of the assigned task stems not only from the mode mixity but also from the material features. The beam is obtained through an additive manufacturing process, in which layers of bassanite (i.e., calcium sulfate hemi-hydrate) powder with a thickness of about 0.1 mm are bonded using a proprietary water-based binder that, reacting with the bassanite, produces gypsum as by-product [3, 4, 28]. The binder is sprayed on the bassanite layer in bands with a width of about 0.5 mm through a nozzle, similarly as in an ink-jet printer, creating the so-called mineral bands [3]. The resulting material is clearly heterogeneous and anisotropic for different reasons: firstly, the binding of the bassanite powder leaves pores at the mineral bands scale. Secondly, the deposition of the binder induces a preferential growth of the gypsum crystals along the spraying direction and an irregular distribution of bonded material at the interfaces between adjacent mineral bands and between successive depositions of bassanite layers (also termed in short deposition layers) [3]. The presence of interfaces and of a preferential direction for the growth of the gypsum crystals induces a macroscopic material behavior typical of orthorhombic systems, namely an orthotropic behavior [3]. In the adopted coordinate system, the bassanite layers are deposited in the xx-yy plane (brown planes in Figure 1a), while the binder is sprayed along the xx-direction (brown lines in Figure 1a), which is thus the direction of the mineral bands. To facilitate the identification of the directions related to the manufacturing process we define here the material directions as

  • V-direction: binder spraying or mineral bands direction, corresponding to the xx-direction;

  • T-direction: direction within the deposition layer plane and orthogonal to the mineral bands, corresponding to the yy-direction; and

  • H-direction: direction orthogonal to the deposition layer plane, corresponding to the zz-direction,

as indicated in Figure 1a.

Experimental characterization tests

To characterize the material behavior, an extensive set of experimental data is provided in [5, 6, 28]. For the sake of clarity we summarize here only the data used in the following, which include

  • load-displacement, digital image correlation and crack surface roughness data of TPB tests on beams differing from the DMC test beam only by the notch geometry2. These beams comprise (Figure 1b) i. four specimens labeled as HC, designed to induce a crack under mode-I conditions; ii. four specimens labeled as HB, designed to induce a crack under mixed-mode (I+II\mathrm{I}+\mathrm{II}) conditions with mode-I predominance; iii. four specimens labeled as HA, designed to induce a crack under mixed-mode (I+II\mathrm{I}+\mathrm{II}) conditions with mode-II predominance; iv. four specimens labeled as H45, designed to induce a crack under mixed-mode (I+III\mathrm{I}+\mathrm{III}) conditions;

  • five load-displacement curves for each material orientation obtained from unconfined compression tests;

  • density measurements on three cubic samples;

  • longitudinal and shear wave velocities from ultrasonic measurements for material directions V, T and H.

Modeling and computational aspects

This section briefly summarizes the model and the computational strategy used to carry out the DMC task assignment.

Phase-field model of brittle fracture

In the following, we assume linearized kinematics, rate independence, quasi-static and isothermal conditions and negligible body forces. Moreover, since the dimensions of the heterogeneities generated by the manufacturing process are much smaller than the specimen dimensions, we assume a homogeneous material featuring orthotropic elasticity and fracture.

Let us consider a body Ω\Omega with boundary ∂Ω\partial\Omega. Introducing the spatial coordinate x∈Ωx \in\Omega, we denote the displacement field as u(x)u(x) and the stress and strain fields respectively as σ(x)\sigma(x) and ε(x)=∇sym(u(x))\varepsilon(x) = \nabla_{\mathrm{sym}}(u(x)), where ∇sym(∙)\nabla_{\mathrm{sym}}(\bullet) is the symmetric gradient of (∙)(\bullet). Boundary tractions tˉ(x)\bar{t}(x) are prescribed on the Neumann part of the boundary ∂ΩN\partial\Omega_N and imposed displacements uˉ(x)\bar{u}(x) are applied to the Dirichlet part of the boundary ∂ΩD\partial\Omega_D, with ∂Ω=∂ΩN∪∂ΩD\partial\Omega= \partial\Omega_N \cup\partial\Omega_D, ∂ΩN∩∂ΩD=∅\partial\Omega_N \cap\partial\Omega_D = \varnothing. We also define nn as the outward unit normal to ∂Ω\partial\Omega.

To model the material behavior we adopt a phase-field model of brittle fracture, whereby damage and fracture are described through a phase-field or damage variable dd; this variable features a steep but smooth transition from d=0d = 0, corresponding to an intact material, to d=1d = 1, denoting a fully damaged material, over a support whose size is governed by a length-scale parameter ℓ\ell [29] (Figure 2). The phase-field approach thus smears a sharp crack over a small but not vanishing portion of the domain, Γℓ\Gamma_{\ell} (Figure 2a).

Image not extracted

Figure 2. (a) Schematic representation of the domain Ω\Omega with prescribed traction tˉ\bar{t} on the Neumann boundary ∂ΩN\partial\Omega_N and prescribed displacement uˉ\bar{u} on the Dirichlet boundary ∂ΩD\partial\Omega_D. The optimal profile of the phase-field variable along a coordinate perpendicular to the axis of the regularized crack Γℓ\Gamma_{\ell} is shown in (b) in comparison to the sharp crack representation which is recovered for a length-scale parameter ℓ→0\ell\to0.

The governing equations of the model are derived in a variational fashion from the local minimization of the total energy functional, which reads

E(u,d)=∫Ωg(d)ψ0(ε)⏟ψ(ε,d) dV+∫ΩGc4cw(w(d)ℓ+ℓ∣∇d∣2)⏟D(d) dV−∫∂ΩNtˉ⋅u dA.(1)\mathcal{E}(u,d) = \int_{\Omega} \underbrace{g(d)\psi_0(\varepsilon)}_{\psi(\varepsilon,d)} \,\mathrm{d}V + \int_{\Omega} \underbrace{\frac{G_{\mathrm{c}}}{4c_w}\left(\frac{w(d)}{\ell}+\ell\lvert\nabla d\rvert^2\right)}_{\mathcal{D}(d)} \,\mathrm{d}V - \int_{\partial\Omega_N}\bar{t}\cdot u\,\mathrm{d}A . \tag*{(1)}

The strain energy density ψ(ε,d)\psi(\varepsilon,d) depends on the monotonically decreasing scalar function g(d)=(1−d)2+g0g(d) = (1-d)^2 + g_0, termed degradation function, where the parameter g0=o(ℓ)g_0 = o(\ell) (here chosen as 10−510^{-5}) provides a small residual stiffness preventing numerical instabilities [7]. A comparison of different degradation functions is given in [30]. This function governs how the material transitions between intact and broken states and is directly applied to the undegraded elastic strain energy density ψ0(ε)\psi_0(\varepsilon), which reads

ψ0(ε)=12ε:C:ε.(2)\psi_0(\varepsilon) = \frac{1}{2}\varepsilon: \mathbf{C} : \varepsilon. \tag*{(2)}

As a consequence of the assumed orthotropic elastic behavior, the symmetric fourth-order elasticity tensor C\mathbf{C} contains nine independent components and, in Voigt notation, reads

[C]=[CxxxxCxxyyCxxzz000CxxyyCyyyyCyyzz000CxxzzCyyzzCzzzz000000Cyzyz000000Cxzxz000000Cxyxy].(3)[\mathbf{C}] = \begin{bmatrix} C_{xxxx} & C_{xxyy} & C_{xxzz} & 0 & 0 & 0 \\ C_{xxyy} & C_{yyyy} & C_{yyzz} & 0 & 0 & 0 \\ C_{xxzz} & C_{yyzz} & C_{zzzz} & 0 & 0 & 0 \\ 0 & 0 & 0 & C_{yzyz} & 0 & 0 \\ 0 & 0 & 0 & 0 & C_{xzxz} & 0 \\ 0 & 0 & 0 & 0 & 0 & C_{xyxy} \end{bmatrix}. \tag*{(3)}

The dissipated energy density D(d)\mathcal{D}(d) in (1) depends not only on the phase-field variable but also on its gradient ∇d\nabla d. The normalization constant cwc_w ensures that the energy dissipated per unit length of a regularized crack corresponds to the fracture toughness GcG_c. Its value depends on the choice of the monotonically increasing dissipation function w(d)w(d), which modulates the energy dissipation with the increase of homogeneous damage. Although alternatives are available [31, 32], the two most common dissipation functions used for brittle materials are w(d)=dw(d)=d (AT1 model), yielding cw=2/3c_w=2/3, and w(d)=d2w(d)=d^2 (AT2 model), leading to cw=1/2c_w=1/2. The main difference between the two is that the former leads to an initial linear elastic stage of behavior, whereas with the latter damage starts evolving already at vanishing stress [9].

The functional presented in (1) yields a symmetric evolution of the phase-field variable under tensile and compressive stress states. This implies that a crack can nucleate and propagate also under mainly compressive states, and that the faces of a fully developed crack can exhibit interpenetration upon closure. To overcome these issues, different approaches are available in the literature; they are mostly based on an additive split of the undegraded elastic strain energy into an active and an inactive part [33, 34, 35, 36, 37]. While this split is essential to obtain meaningful results in problems with significant compression [37], in cases with predominant tension it does not significantly modify results with respect to the symmetric model, and at the same time it leads to an increased computational cost. Therefore, since here we focus on notched TPB tests for which predominant tension is anticipated, we keep the symmetric model based on the total energy (1). A further simplifying assumption inherent to (1) is that the orthotropic behavior determined by the manufacturing process is limited to the elastic stiffness, whereas the damage behavior stays isotropic. Although phase-field models addressing a directional dependence of GcG_c have been presented in the literature [16, 18], our reasons for this assumption are that the degree of material anisotropy is found to be moderate (see § 5.2), and that the available data are insufficient to properly characterize an anisotropic fracture model.

The governing equations are derived by applying the principles of irreversibility, energy balance and first-order stability [9], leading to the equations summarized in Table 1. For the phase-field variable, beside the Neumann boundary conditions that naturally emerge from the variational formulation, we introduce in (9) an additional Dirichlet condition, denoted as internal condition, in order to prevent the evolution of spurious damage in the subset Ωp\Omega_{\mathrm{p}} of the domain. As further detailed later, Ωp\Omega_{\mathrm{p}} consists of the portions of the domain close to the support and loading rods.

Linear momentum balance
Equilibrium equation∇⋅σ=0∀x∈Ω\nabla\cdot\boldsymbol{\sigma} = 0 \quad\forall x \in\Omega (4)
Cauchy stressσ(ε,d)=∂ψ(ε,d)∂ε=g(d)∂ψ0(ε)∂ε\boldsymbol{\sigma}(\boldsymbol{\varepsilon},d) = \frac{\partial\psi(\boldsymbol{\varepsilon},d)}{\partial\boldsymbol{\varepsilon}} = g(d)\frac{\partial\psi_0(\boldsymbol{\varepsilon})}{\partial\boldsymbol{\varepsilon}} (5)
Degradation functiong(d)=(1−d)2+g0withg0=10−5g(d) = (1-d)^2 + g_0 \quad\text{with} \quad g_0 = 10^{-5} (6)
Boundary conditionsσ⋅n=tˉ∀x∈∂ΩN\boldsymbol{\sigma} \cdot\boldsymbol{n} = \bar{\boldsymbol{t}} \quad\forall x \in\partial\Omega_N<br>u=uˉ∀x∈∂ΩD\boldsymbol{u} = \bar{\boldsymbol{u}} \quad\forall x \in\partial\Omega_D (7)
Parameters<br>(Total: 9)C_{xxxx}, C_{yyyy}, C_{zzzz}, C_{xxyy}, C_{yyzz}, C_{zzxx}, C_{xyyx}, C_{yzzy}, C_{zxxz} \to C
Damage evolution
KKT conditionsd˙≥0\dot{d} \ge0<br>g′(d)ψ0+Gc4cw(w′(d)ℓ−2ℓ∇⋅(∇d))≥0g'(d)\psi_0 + \frac{G_c}{4c_w}\left(\frac{w'(d)}{\ell} - 2\ell\nabla\cdot(\nabla d)\right) \ge0<br>[g′(d)ψ0+Gc4cw(w′(d)ℓ−2ℓ∇⋅(∇d))]d˙=0\left[g'(d)\psi_0 + \frac{G_c}{4c_w}\left(\frac{w'(d)}{\ell} - 2\ell\nabla\cdot(\nabla d)\right)\right]\dot{d}=0 (8)
Boundary & internal conditionsn⋅∇d=0,∀x∈∂Ω∖(∂Ω∩ΩP)\boldsymbol{n}\cdot\nabla d=0,\quad\forall x\in\partial\Omega\setminus(\partial\Omega\cap\Omega_P)<br>d=0∀x∈ΩPd=0\quad\forall x\in\Omega_P (9)
Dissipation functionw(d)=dwithcw=23 (AT1),orw(d)=d2withcw=12 (AT2)w(d)=d\quad\text{with}\quad c_w=\frac{2}{3}\ \text{(AT1)},\quad\text{or}\quad w(d)=d^2\quad\text{with}\quad c_w=\frac{1}{2}\ \text{(AT2)} (10)
Parameters<br>(Total: 2)ℓ,Gc\ell, G_c

Table 1. Governing equations and boundary conditions of the problem.

For the present modeling choices, in absence of stress singularities within the domain the length-scale parameter ℓ\ell controls the material strength and can thus be regarded as a material parameter [11]. Conversely, in presence of stress singularities, results are expected to be fairly independent on ℓ\ell as long as it is sufficiently small compared to the smallest characteristic length in the domain. Thus, in such cases ℓ\ell plays the role of a numerical parameter determining the thickness of the regularized representation of a crack [11]. However, if ℓ\ell is comparable to the smallest characteristic length in the domain, the system response becomes dependent on its value, in which case ℓ\ell should be treated as a further model parameter [38].

Computational aspects

The solution of the non-linear system of coupled equations arising from Table 1 requires a spatial discretization method and an incremental iterative solution scheme.

For spatial discretization, we adopt the standard finite element (FE) method based on hexahedral 8-node FEs with four degrees of freedom (DOFs) per node, i.e. the three components of the displacement field u\boldsymbol{u} and the phase-field variable dd. The discretization of the domain must be sufficiently fine to properly resolve the length-scale parameter ℓ\ell. As better illustrated in the following sections, FE computations are used for both the calibration of the model, based on the TPB tests on specimens HC, HB, HA and H45, and the prediction of the TPB test on the DMC specimen. Since all these specimens feature a pre-existing notch, we set the length scale to ℓ=0.625\ell= 0.625 mm, which is considered sufficiently small to treat it as a numerical parameter (Figure 1) but still large enough to achieve a reasonable computational cost.

To reduce the computational cost, the portions of the domain where the boundary conditions are applied or where the development of a crack is expected are discretized with a finer mesh. As in the experimental results [5, 6, 28], in all numerical tests a single crack starts from the tip of the notch and propagates upwards until complete failure. Accordingly, we adopt the meshes illustrated in Figure 3 whose details in terms of number of elements and DOFs are summarized in Table 2. The notches are explicitly considered in the meshed geometries rather than modeled as a pre-crack using the phase-field variable, since they are introduced during manufacturing with a non-vanishing width. Within the plane orthogonal to the notch, the HC, HB and HA meshes feature an average size of h≃ℓ/5=0.125h \simeq\ell/5 = 0.125 mm and h≃1h \simeq1 mm in the fine and the coarse portions, respectively. In the thickness direction, the size of the elements is larger and equal to about 0.397 mm in the fine and 3.175 mm in the coarse mesh regions (Figure 3a-Figure 3c)3. In the H45 and DMC meshes, we adopt a uniform size of the elements in all directions equal to 0.125 mm and 1 mm in the fine and coarse portions, respectively (Figure 3d-Figure 3e). As visible in the details of the notches in Figure 3, the refinement of the mesh is non-conforming, leading to the presence of hanging nodes which are automatically dealt with by the used deal.ii library by means of affine constraints [39, 40].

HC testHB testHA testH45DMC test
# ELEMENTS125 392215 664354 852415 280820 063
# DOFs550 036931 7761 515 9201 798 7243 425 740

Table 2. Mesh details for the TPB tests.

Image not extracted

Figure 3. Meshes for the (a) HC, (b) HB, (c) HA, (d) H45 and (e) DMC tests.

The applied boundary and internal conditions in terms of displacements and phase-field variable are illustrated in Figure 4 on the representative example of the H45 specimen. The displacement boundary conditions are applied at the nominal contact line between loading or support rods and the specimen (Figure 4a). For the phase-field variable, we apply a homogeneous Dirichlet condition within half-cylindrical sub-volumes with radius rPr_{\mathrm{P}} centered at the lines where the displacement boundary conditions are applied (Figure 4b). This is done to prevent a spurious damage evolution in these portions of the domain, which are affected by highly localized reaction forces. By means of preliminary analyses conducted by trial-and-error, we choose rP=2.5r_{\mathrm{P}} = 2.5 mm for specimens HC, HB, and H45; for specimens HA and DMC, since here the notch is farther away from midspan and a higher peak force is expected, we choose the slightly larger value rP=3r_{\mathrm{P}} = 3 mm.

Image not extracted

Figure 4. (a) Displacement boundary conditions and (b) phase-field internal conditions illustrated exemplarily for the H45 specimen.

We adopt a staggered solution procedure, which involves the alternate solution through a Newton-Raphson approach of the equilibrium equation (4) and of the phase-field evolution equations (8) until the norms of the residuals fall below a given tolerance [7] (set to 10−610^{-6} for the Newton-Raphson scheme and 10−310^{-3} for the staggered loop). To enforce the irreversibility condition for the phase field (8), although different alternatives are available [41], we adopt the approach proposed in [33]. We perform all computations with an in-house code based on the C++ library deal.ii version 9.3.0 [39, 40], on the Euler High Performance Cluster of ETH Zürich using one single node with 128 CPU cores.

Model calibration

This section illustrates the procedure adopted to calibrate the model parameters summarized in Table 1. A total of eleven material constants are needed, including nine elastic parameters to determine the orthotropic stiffness tensor C\mathbf{C}, the length-scale parameter ℓ\ell and the fracture toughness GcG_{\mathrm{c}}, which we assume to be isotropic, as explained in § 4.1. However, as discussed in § 4.2, ℓ\ell is considered here as a numerical parameter and its value is set to ℓ=0.625\ell= 0.625 mm, which leaves us with ten remaining unknown parameters. Also, a suitable dissipation function w(d)w(d) must be selected. Our basis for calibration are the results of the unconfined compressive, notched TPB and ultrasonic wave propagation tests available in [5, 6, 28]. For the unconfined compressive and notched TPB tests we specifically consider the load FF vs. deflection uˉ\bar{u} curves (see boundary conditions in Figure 4a).

Preliminaries and calibration procedure outline

The load vs. displacement curves of the unconfined compressive (Figure 5) and notched TPB tests (Figure 6) display an initial linear elastic branch. For this reason, we select for the dissipation function the AT1 model, see also § 4.1. Note that in Figure 6 the curves for the HA specimens exhibit a higher peak force, confirming the choice of a larger rPr_{\mathrm{P}} for this geometry.

Image not extracted

Figure 5. Unconfined compressive test results along the V, T and H directions, where uˉ\bar{u} denotes the displacement and FF the reaction force.

Image not extracted

Figure 6. TPB test results for the HC, HB, HA and H45 specimens.

The calibration of the ten unknown parameters, i.e. the elastic parameters in C\mathbf{C} and the fracture toughness GcG_{\mathrm{c}}, is subdivided in two stages. First, we use the unconfined compressive and the dynamic ultrasonic test results to obtain a first approximate estimate of the elastic tensor, denoted as C~\widetilde{\mathbf{C}}. As better explained in the following, this estimate cannot be considered representative of the behavior of the DMC notched TPB specimen, however we consider the ratios between the elastic parameters to be reliable. We normalize C~ijkl\widetilde{C}_{ijkl} with C~VVVV\widetilde{C}_{\mathrm{VVVV}} and assume

CijklCVVVV=C~ijklC~VVVVwithi,j,k,l=V,T,H(4)\frac{C_{ijkl}}{C_{\mathrm{VVVV}}} = \frac{\widetilde{C}_{ijkl}}{\widetilde{C}_{\mathrm{VVVV}}} \quad\text{with} \quad i,j,k,l = \mathrm{V},\mathrm{T},\mathrm{H} \tag*{(4)}

which reduces the number of unknown parameters to two, namely CVVVVC_{\mathrm{VVVV}} and GcG_{\mathrm{c}}. In the second stage of the calibration, we optimize these remaining quantities by means of an iterative optimization procedure based on the comparison between numerical and experimental load-deflection curves from the HC, HB, HA and H45 notched TPB tests.

First estimate of the elastic parameters

We start by evaluating the main diagonal elements of C~\widetilde{\mathbf{C}} using the plane wave velocity measurements provided in [5, 6, 28], where ultrasonic waves are let to propagate in cubic samples along the directions V, T, H and receivers are placed so as to measure both longitudinal and shear wave velocities. For a monochromatic plane wave propagating in a homogeneous and infinite anisotropic elastic medium, the relation between wave velocity and elastic parameters is obtained combining the plane wave equation and the equation of motion, giving Christoffel’s equation [42]. Using Einstein’s summation convention, the latter reads [43]

(C~ijklwjwk−ρs2δil)u^l=0for i,j,k,l=V,T,H(5)\left(\widetilde{C}_{ijkl}w_jw_k-\rho s^2\delta_{il}\right)\hat{u}_l=0 \quad\text{for } i,j,k,l=\mathrm{V},\mathrm{T},\mathrm{H} \tag*{(5)}

where the vector w\mathbf{w} denotes the wave propagation direction and the vector u^\hat{\mathbf{u}} defines the direction of particle motion. We assume here that both coincide with the principal material directions V, T, H, meaning that they are either parallel for compressional waves or orthogonal to each other for shear waves. Further, δij\delta_{ij} is the Kronecker delta and ρ=1350 kg m−3\rho=1350\ \mathrm{kg\ m^{-3}} (with a standard deviation of 17 kg m−317\ \mathrm{kg\ m^{-3}}) is the density of the material obtained as the average of three measurements on cubic samples [28]. For notational convenience, we use the abbreviation smns_{mn} in the following to represent the wave velocity ss with propagation direction mm and particle movement direction nn. Hence, smns_{mn} represents the velocity of a longitudinal wave if m=nm=n and the velocity of a shear wave if m≠nm\ne n. The average values of the wave velocities obtained from [5, 6, 28] are summarized in Table 3 along with their standard deviations.

longitudinal wave velocities in m/sshear wave velocities in m/s
sVV=2730s_{VV}=2730 (st. dev. = 12)sVH=1476s_{VH}=1476 (st. dev. = 63)sVT=2020s_{VT}=2020 (st. dev. = 29)
sTT=2530s_{TT}=2530 (st. dev. = 83)sTH=1696s_{TH}=1696 (st. dev. = 71)sTV=1980s_{TV}=1980 (st. dev. = 40)
sHH=2332s_{HH}=2332 (st. dev. = 84)sHT=1745s_{HT}=1745 (st. dev. = 62)sHV=1799s_{HV}=1799 (st. dev. = 59)

Table 3. Longitudinal and shear wave velocities: average values and standard deviations [5, 6, 28]. The first index denotes the direction of wave propagation and the second index is the direction of particle movement.

Eq. (5) with the measurements in Table 3 gives [42]

C~VVVV=ρ(sVV)2=10061 MPa,C~TTVV=C~VVTT=ρ(sVT+sTV2)2=5400 MPa,C~TTTT=ρ(sTT)2=8641 MPa,C~THHT=C~HTTH=ρ(sTH+sHT2)2=3996 MPa,C~HHHH=ρ(sHH)2=7342 MPa,C~HVVH=C~VHHV=ρ(sVH+sHV2)2=3620 MPa.(6)\begin{aligned} \widetilde{C}_{VVVV} &= \rho(s_{VV})^2 = 10061\ \mathrm{MPa}, & \widetilde{C}_{TTVV} = \widetilde{C}_{VVTT} &= \rho\left(\frac{s_{VT}+s_{TV}}{2}\right)^2 = 5400\ \mathrm{MPa}, \\ \widetilde{C}_{TTTT} &= \rho(s_{TT})^2 = 8641\ \mathrm{MPa}, & \widetilde{C}_{T H H T} = \widetilde{C}_{H T T H} &= \rho\left(\frac{s_{TH}+s_{HT}}{2}\right)^2 = 3996\ \mathrm{MPa}, \\ \widetilde{C}_{HHHH} &= \rho(s_{HH})^2 = 7342\ \mathrm{MPa}, & \widetilde{C}_{HVVH} = \widetilde{C}_{VHHV} &= \rho\left(\frac{s_{VH}+s_{HV}}{2}\right)^2 = 3620\ \mathrm{MPa}. \tag*{(6)} \end{aligned}

Concerning the shear components in (13), the minor symmetry property C~ijkl=C~jilk\widetilde{C}_{ijkl}=\widetilde{C}_{jilk} leads to sjl=sljs_{jl}=s_{lj} but two separate measurements are available, which is why we use the mean value (sjl+slj)/2(s_{jl}+s_{lj})/2 for j≠lj\ne l. For the off-diagonal components of C~\widetilde{C}, wave velocity measurements not along the principal directions and more involved calculations would be necessary, which is why the remaining three components of C~\widetilde{C} are obtained following the standard relationships for orthotropic materials [44]

C~VVTT=ν~TV+ν~HVν~THξE~V=8439 MPa,C~TTHH=ν~HT+ν~HVν~VTξE~T=6887 MPa,with ξ=1−ν~VTν~TV−ν~THν~HT−ν~HVν~VH−2ν~VTν~THν~HV,C~HHVV=ν~HV+ν~TVν~HTξE~V=7513 MPa.(7)\begin{aligned} \widetilde{C}_{VVTT} &= \frac{\widetilde{\nu}_{TV}+\widetilde{\nu}_{HV}\widetilde{\nu}_{TH}}{\xi}\widetilde{E}_{V}=8439\ \mathrm{MPa},\\ \widetilde{C}_{TTHH} &= \frac{\widetilde{\nu}_{HT}+\widetilde{\nu}_{HV}\widetilde{\nu}_{VT}}{\xi}\widetilde{E}_{T}=6887\ \mathrm{MPa}, \qquad\text{with }\xi=1-\widetilde{\nu}_{VT}\widetilde{\nu}_{TV}-\widetilde{\nu}_{TH}\widetilde{\nu}_{HT}-\widetilde{\nu}_{HV}\widetilde{\nu}_{VH}-2\widetilde{\nu}_{VT}\widetilde{\nu}_{TH}\widetilde{\nu}_{HV},\\ \widetilde{C}_{HHVV} &= \frac{\widetilde{\nu}_{HV}+\widetilde{\nu}_{TV}\widetilde{\nu}_{HT}}{\xi}\widetilde{E}_{V}=7513\ \mathrm{MPa}. \tag*{(7)} \end{aligned}

It remains to explain how we obtain the Young’s moduli E~i\widetilde{E}_i and the Poisson’s ratios ν~ij\widetilde{\nu}_{ij} in (14), with i,j=V,T,Hi,j=\mathrm{V},\mathrm{T},\mathrm{H}. The elastic moduli are determined from unconfined compressive tests on cylindrical specimens with nominal length l=50.8 mml=50.8\ \mathrm{mm} and radius R=12.7 mmR=12.7\ \mathrm{mm}. We first measure the uniaxial stiffnesses K~i\widetilde{K}_i in the V, T and H directions as the slopes of the experimental FF-uˉ\bar{u} curves obtained by testing specimens aligned with the corresponding directions (Figure 5), by means of a linear regression within the linear elastic regime, taken as the range from 15 % to 65 % of the peak force value Fmax⁡F_{\max} (indicated with cross symbols in Figure 5). Then, we compute the elastic moduli as [5, 6, 28]

E~i=K~ilπR2with i=V,T,H.(8)\widetilde{E}_i=\frac{\widetilde{K}_i l}{\pi R^2}\quad\text{with }i=\mathrm{V},\mathrm{T},\mathrm{H}. \tag*{(8)}

The obtained values of E~i\widetilde{E}_i are summarized in Table 4.

E~V\widetilde{E}_{V}E~T\widetilde{E}_{T}E~H\widetilde{E}_{H}ν~VT(∗)\widetilde{\nu}_{VT}^{(*)}ν~TH(∗)\widetilde{\nu}_{TH}^{(*)}ν~VH(∗)\widetilde{\nu}_{VH}^{(*)}
1485 MPa1365 MPa1512 MPa0.6380.3380.425

Table 4. First estimate of Young’s moduli and Poisson’s ratios.

(*) Given νij\nu_{ij}, the value of νji\nu_{ji} is obtained using the relationships (17).

The Poisson’s ratios in (14) are obtained using the following equations

C~VVVV=1−ν~THν~HTξE~VC~TTTT=1−ν~HVν~VHξE~TC~HHHH=1−ν~VTν~TVξE~H(9)\begin{aligned} \widetilde{C}_{VVVV} &= \frac{1-\widetilde{\nu}_{TH}\widetilde{\nu}_{HT}}{\xi}\widetilde{E}_{V} \\ \widetilde{C}_{TTTT} &= \frac{1-\widetilde{\nu}_{HV}\widetilde{\nu}_{VH}}{\xi}\widetilde{E}_{T} \\ \widetilde{C}_{HHHH} &= \frac{1-\widetilde{\nu}_{VT}\widetilde{\nu}_{TV}}{\xi}\widetilde{E}_{H} \tag*{(9)} \end{aligned}

along with [44]

ν~VTE~V=ν~TVE~T,ν~VHE~V=ν~HVE~H,ν~THE~T=ν~HTE~H(10)\frac{\widetilde{\nu}_{VT}}{\widetilde{E}_{V}}=\frac{\widetilde{\nu}_{TV}}{\widetilde{E}_{T}}, \qquad \frac{\widetilde{\nu}_{VH}}{\widetilde{E}_{V}}=\frac{\widetilde{\nu}_{HV}}{\widetilde{E}_{H}}, \qquad \frac{\widetilde{\nu}_{TH}}{\widetilde{E}_{T}}=\frac{\widetilde{\nu}_{HT}}{\widetilde{E}_{H}} \tag*{(10)}

and the obtained values are summarized in Table 4.

As mentioned in § 5.1, the estimates of the elastic parameters in (13) and (14) cannot be directly adopted as final calibrated values due to different reasons. First, the estimates in (13) are based on Christoffel’s equation (12), thus they should be intended as dynamic elastic parameters, and it is known that the quasi-static elastic parameters are usually lower than their dynamic counterparts by a factor that, for gypsum, can be as low as 1/4 [45]. Also, Christoffel’s equation assumes that the tested material is homogeneous, while the investigated material is heterogeneous across different scales and includes interfacial zones (§ 3.1). Finally, the experimental tests adopted to estimate the elastic parameters in (14) are representative of compressive rather than tensile stress states. While a symmetric behavior under tension and compression may be expected for a homogeneous material, this assumption may easily break down for a heterogeneous material with interfacial zones.

Combining (13) and (14) with the assumption (11), the elastic stiffness tensor can be reduced to

[C]=CVVVV[10.83880.74670000.83880.85880.68450000.74670.68450.72970000000.39720000000.35980000000.5367].(11)[\mathbf{C}] = C_{VVVV} \begin{bmatrix} 1 & 0.8388 & 0.7467 & 0 & 0 & 0 \\ 0.8388 & 0.8588 & 0.6845 & 0 & 0 & 0 \\ 0.7467 & 0.6845 & 0.7297 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0.3972 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0.3598 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0.5367 \end{bmatrix}. \tag*{(11)}

Parameters optimization procedure

We now focus on the calibration of the two remaining parameters, CVVVVC_{VVVV} and GcG_{c}. As mentioned earlier, this is based on an optimization procedure which aims at minimizing the discrepancies between the load-deflection curves for the TPB characterization specimens obtained numerically and experimentally, denoted respectively as Fnum,i(uˉ∣CVVVV,Gc)F_{\mathrm{num},i}(\bar{u} \mid C_{VVVV},G_{c}) and Fexp,i(uˉ)F_{\mathrm{exp},i}(\bar{u}) with i=HC,HB,HA,H45i = \mathrm{HC},\mathrm{HB},\mathrm{HA},\mathrm{H45}.

For given values of CVVVVC_{VVVV} and GcG_{c}, the numerical curves are obtained by predicting the behavior until failure of the TPB characterization specimens, using the model and numerical setup illustrated in Sections 3.1-3.2 along with (18). Each experimental curve for a TPB specimen is taken as the average of the curves obtained from all repetitions of the respective test. To obtain these average curves, we first apply a moving average filter and then normalize the data points in each curve with the respective peak force and its associated displacement. We then shift the curves such that the initial instant of the test is the point at which this normalized force becomes and remains positive for the first time. These curves are then resampled using a common grid of normalized displacements by interpolating the normalized force between the experimental measurements of each individual test. Since the normalized curves are now sampled with a common, equidistant grid of normalized displacements, the average normalized force across the different tests can be computed for each of the sampling points. Finally, this average of the normalized and resampled curves is scaled up by the arithmetic mean of the individual normalization factors for the respective peak force and its associated displacement to obtain the average experimental curve (Figure 7a).

Image not extracted

Figure 7. (a) Illustrative scheme of experimental and average load-deflection curves FF-uˉ\bar{u} (on the example of the HC specimen). (b) Cost function for the optimization procedure. (c) Fitting of a second-order polynomial approximation of the cost function using a sampling grid of nine points and definition of the optimum.

To setup the optimization problem, we use as cost function the integral over the loading history of the mismatch between experimental and numerical force-deflection curves (Figure 7b), i.e.

ΔAi(CVVVV,Gc)=∫hist.∣Fnum,i(uˉ∣CVVVV,Gc)−Fexp,i(uˉ)∣ duˉ,for i=HC,HB,HA,H45.(12)\Delta A_i(C_{\mathrm{VVVV}},G_c)=\int_{\mathrm{hist.}} \left|F_{\mathrm{num},i}(\bar{u}\mid C_{\mathrm{VVVV}},G_c)-F_{\mathrm{exp},i}(\bar{u})\right|\,\mathrm{d}\bar{u},\quad\text{for }i=\mathrm{HC},\mathrm{HB},\mathrm{HA},\mathrm{H45}. \tag*{(12)}

Clearly, the cost function ΔAi(CVVVV,Gc)\Delta A_i(C_{\mathrm{VVVV}},G_c) cannot be obtained analytically, but it can be sampled at given values of CVVVVC_{\mathrm{VVVV}} and GcG_c. Thus, we assume that (19) can be represented by a second-order polynomial of the form

ΔAi(CVVVV,Gc)=pi,0+pi,1CVVVV+pi,2Gc+pi,3CVVVV2+pi,4CVVVVGc+pi,5Gc2,\Delta A_i(C_{\mathrm{VVVV}},G_c)=p_{i,0}+p_{i,1}C_{\mathrm{VVVV}}+p_{i,2}G_c+p_{i,3}C_{\mathrm{VVVV}}^2+p_{i,4}C_{\mathrm{VVVV}}G_c+p_{i,5}G_c^2,
for i=HC,HB,HA,H45,(13)\text{for }i=\mathrm{HC},\mathrm{HB},\mathrm{HA},\mathrm{H45}, \tag*{(13)}

and we obtain the coefficients pi,kp_{i,k} by means of a least-square fitting, minimizing the mismatch between a grid of sampled values and its representation (20) as illustrated in Figure 7c. From (20), the optimal values (CVVVV∗,Gc∗)i(C_{\mathrm{VVVV}}^*,G_c^*)_i can be obtained by solving analytically the minimization problem

(CVVVV∗,Gc∗)i=arg⁡min⁡CVVVV,Gc ΔAi(CVVVV,Gc),for i=HC,HB,HA,H45.(14)(C_{\mathrm{VVVV}}^*,G_c^*)_i=\underset{C_{\mathrm{VVVV}},G_c}{\arg\min}\ \Delta A_i(C_{\mathrm{VVVV}},G_c),\quad\text{for }i=\mathrm{HC},\mathrm{HB},\mathrm{HA},\mathrm{H45}. \tag*{(14)}

To ensure the selection of the best set of parameters, the optimization procedure is performed iteratively by progressively restricting the range of parameters (CVVVV,Gc)(C_{\mathrm{VVVV}},G_c) used to sample the cost function (19). The procedure is initialized selecting a rather broad range for (CVVVV,Gc)(C_{\mathrm{VVVV}},G_c); then, at each iteration we refine the grid of sampling points in the neighborhood of the optimum set of parameters found at the previous iteration. The process is terminated after three iterations of parameter grid refinement, whereby the first and second iterations use a grid of 3×33 \times3, and the final iteration uses a grid of 5×55 \times5 sampled points in the (CVVVV,Gc)(C_{\mathrm{VVVV}},G_c) space.

Since the proposed procedure is performed separately for all TPB characterization tests, we obtain four sets of parameters summarized in Tab. 5. Their average gives the final calibrated values CVVVV∗=2057C_{\mathrm{VVVV}}^*=2057 MPa and Gc∗=9.23⋅10−2\mathcal{G}_c^* = 9.23 \cdot10^{-2} MPa mm.

HC testHB testHA testH45 test
CVVVV∗\mathsf{C}^{*}_{VVVV} in MPa2118195820402112
Gc∗\mathcal{G}_c^* in MPa mm8.30⋅10−28.30 \cdot10^{-2}8.62⋅10−28.62 \cdot10^{-2}8.55⋅10−28.55 \cdot10^{-2}11.45⋅10−211.45 \cdot10^{-2}

Table 5. Optimized parameters for the different TPB characterization tests.

Calibrated material parameters

Following the procedure described in Sections. 4.2 and 4.3, we obtain

[C]=[2057172515360001725176714080001536140815010000008170000007400000001104]MPa(15)[\mathbf{C}] = \begin{bmatrix} 2057 & 1725 & 1536 & 0 & 0 & 0 \\ 1725 & 1767 & 1408 & 0 & 0 & 0 \\ 1536 & 1408 & 1501 & 0 & 0 & 0 \\ 0 & 0 & 0 & 817 & 0 & 0 \\ 0 & 0 & 0 & 0 & 740 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1104 \end{bmatrix} \mathrm{MPa} \tag*{(15)}

for the orthotropic elastic stiffness tensor and

Gc=9.23⋅10−2 MPa mm(16)\mathcal{G}_c = 9.23 \cdot10^{-2}\ \mathrm{MPa\ mm} \tag*{(16)}

for the fracture toughness. Note that the obtained values of the elastic parameters are about 1/5 of those obtained for C~\widetilde{\mathbf{C}}, which we explain with the model simplifications discussed in § 4.1 as well as the fact that the estimates are based on a combination of dynamic and quasi-static experiments. A visualization of the anisotropy of the elastic properties is given in Figure 8, based on the procedure and implementation presented in [46]. In particular, Figure 8a shows the Young’s modulus one would obtain when performing a uni-axial tensile test in the direction d\boldsymbol{d} (described in a spherical coordinate system), clearly indicating the three symmetry planes and that the strongest directions of the material are at an angle of roughly 45∘45^{\circ} to the manufacturing planes. To obtain the Poisson’s ratio for such uniaxial test, a normal vector giving the direction of lateral contraction is required, which is defined vertical to d\boldsymbol{d} and at an angle of φ\varphi. The averaged as well as the maximum value for the Poisson’s ratio for this angle φ\varphi around the test direction d\boldsymbol{d} are plotted in Figure 8b and Figure 8c, respectively, again indicating the three symmetry planes for orthotropic material.

Image not extracted

Figure 8. Visualization of (a) the Young’s modulus EE for different test directions d\boldsymbol{d} as well as (b) the arithmetic mean and (c) the maximum of the Poisson’s ratio ν\nu of different angles φ\varphi around it.

Comparison between numerical and experimental results

This section illustrates the comparison between experimental results and numerical predictions obtained using the calibrated parameters. We discuss first the TPB characterization tests and then the blind prediction of the DMC test following the task summarized in § 3.1.

Comparison with the TPB characterization test results

We start by comparing the experimental and numerical load-deflection curves for the HC, HB, HA and H45 specimens (Figure 9). Considering the experimental scatter and that the numerical curves are all obtained with a unique set of material parameters, we consider the agreement very satisfactory. In particular, the peak load and the initial elastic branch are well captured in all cases. After the peak, the numerical curves display an abrupt load drop due to the assumption of a perfectly brittle material. The experimental post-peak behavior is slightly smoother, probably due to cohesive-frictional phenomena neglected in the model. However, the overall agreement confirms that a brittle fracture model predicts the real behavior of the material reasonably well.

Image not extracted

Figure 9. Comparison between numerical and experimental load-deflection curves for the HC, HB, HA and H45 tests.

In the late stage of the tests, the numerical predictions present an almost horizontal plateau while the experimental curves show a slow but steady decrease of the load (Figure 9). This is mostly due to the homogeneous Dirichlet conditions imposed on the phase-field variable, which prevent the crack from reaching the upper surface of the specimen. This leads to a residual load bearing capacity that, however, amounts to less than 10 % of the peak force.

The numerical crack patterns obtained at the last load step are compared in Figure 10, where the phase-field values d≥0.25d \ge0.25 are shown. Here the geometry and loading conditions induce different types of fracture including pure mode I (Figure 10a), mixed mode I+II (Figure 10b-Figure 10c) and mixed mode I+III (Figure 10d). As expected, the crack starts at the notch tip and propagates upwards until reaching either the region with homogeneous Dirichlet conditions for the phase field (Figure 10a, Figure 10b and Figure 10d) or the upper edge of the specimen (Figure 10c). The initial crack evolution happens unstably at the peak load, with an abrupt propagation of the crack front along a large part of the beam height; very limited further propagation takes place after the peak. This agrees with the digital image correlation measurements of the experiments [5], where an abrupt propagation of the crack front can be observed.

Figure 10. Crack path at the last load step obtained numerically for the (a) HC, (b) HB, (c) HA and (d) H45 specimens in terms of the phase-field variable dd. (All points with d≤0.25d \le0.25 are hidden for the sake of clarity).

As anticipated in § 4.2, in the HC, HB and HA specimens the gradient of the phase-field variable is negligible in the thickness direction, therefore a coarser mesh can be adopted in that direction. This is not the case for the H45 specimen, where a twisting of the crack front takes place as it propagates from the notch to the upper edge of the specimen.

Figure 11 presents a comparison between the experimental and numerical crack paths within the three horizontal sections AA-AA, BB-BB and CC-CC illustrated in Figure 10. Here, the set of points describing for each test the numerical crack surface snums_{\mathrm{num}} and the experimental average surface sexps_{\mathrm{exp}} along with the range over which they vary are obtained, as detailed in Appendix § 8. To obtain a quantitative evaluation of the agreement between numerical and experimental results, the shortest distance Δs=min⁡∥snum−sexp∥\Delta s = \min\lVert s_{\mathrm{num}} - s_{\mathrm{exp}} \rVert is computed as deviation measure (see Appendix A). The agreement between experimental and numerical results is very good, with Δs<2\Delta s < 2 mm almost everywhere (Figure 11). The only exception is the AA-AA section of the HA test, where a maximum deviation of 4.68 mm is observed. This is due to the homogeneous Dirichlet condition at the contact line between loading rod and specimen, which forces the crack to propagate away from it. A complete comparison of the numerical and experimental crack surfaces is given in Figure 12, where we can appreciate that the model is able to predict with reasonable accuracy a wide range of fracture propagation modes under different degrees of mode mixity. This is particularly evident from the H45 specimen (Figure 12d), for which the proposed model and calibration procedure deliver an accurate load-deflection curve (Figure 9) as well as a good prediction of the complex crack path.

Image not extracted

Figure 11. Comparison between numerical and experimental crack paths within the three horizontal sections AA-AA, BB-BB and CC-CC illustrated in Figure 10 for the HC, HB, HA and H45 tests.

Comparison between numerical and experimental crack surfaces for the HC, HB, HA and H45 specimens

Figure 12. Comparison between numerical and experimental crack surfaces for the (a) HC, (b) HB, (c) HA and (d) H45 specimens. (Gray and colored surfaces represent experimental and numerical results, respectively).

Augmented-reality (AR) renders of the final phase field (Figure 10) along with a comparison of the processed experimental and numerical crack surfaces (Figure 12) can be accessed at https://ar.compmech.ethz.ch or using the QR-code in Appendix B.

Blind prediction of the DMC test

After the calibration of the model (Section 4) and its verification (Section 5.1), we finally present the results of the blind numerical prediction of the DMC test along with the experimental data available in [2]. The comparison between numerical and experimental load-deflection curves in Figure 13 demonstrates that the proposed approach is able to blindly predict the behavior of the DMC specimen with a very good accuracy, considering the experimental scatter. The predicted peak force is very close to the average one obtained experimentally (with a deviation of −2.5-2.5 %). For the displacement corresponding to the average peak force, the error is slightly higher but still reasonably low (+13.24+13.24 %). The displacement value at which the force drops almost to zero is very well captured (Figure 13).

Figure 13. Comparison between numerical and experimental load-deflection curves for the DMC test.

We now present the numerically obtained crack path, displayed in Figure 14 as the set of points with phase-field values d≥0.25d \ge0.25. The model predicts a quite complex crack path, with a front that twists and bends in direction of the mids­ection as the crack propagates toward the upper surface. This complicated 3D path results from the mixed-mode fracture conditions with all three modes involved (I+II+III). An AR 3D render of the phase-field crack surface can also be accessed at https://ar.compmech.ethz.ch or using the QR code in Appendix § 9.

Figure 14. Crack path at the last load step obtained numerically for the DMC specimen: (a) front, (b) back, (c) top and (d) bottom views. (All points with d≤0.25d \le0.25 are hidden for the sake of clarity).

For the three horizontal sections AA-AA, BB-BB and CC-CC (Figure 14b), Figure 15 and Figure 16 present a comparison between numerical and experimental crack paths, obtained similarly to what described in § 6.1. Also in this case an excellent agreement is achieved, with a deviation Δs\Delta s always below 2 mm. We remark that this result is particularly important since it is obtained with material parameters calibrated on a completely independent set of tests.

Figure 15. Comparison between numerical and experimental crack paths within the three horizontal sections A-A, B-B and C-C illustrated in Figure 14b.

Comparison between numerical and experimental crack surfaces for the DMC specimen

Figure 16. Comparison between numerical and experimental crack surfaces for the DMC specimen. (Gray and colored surfaces represent experimental and numerical results, respectively).

Summary and conclusions

The main objective of this work is the blind prediction of the behavior up to failure of a notched specimen, made of an artificial gypsum through additive manufacturing and tested in three-point bending. The notch geometry induces mixed-mode fracture conditions involving all three modes (I+II+III). This specific task is defined in the context of the Damage Mechanics Challenge issued by Purdue University [5, 6, 28].

To this end, a phase-field model for fracture in a brittle orthotropic material is calibrated using a set of provided experimental tests. The calibration procedure involves two stages: first, we estimate most of the elastic properties of the orthotropic material using plane wave velocity and unconfined compressive test results; then, we perform an optimization process to identify the remaining elastic parameter and the fracture toughness of the material. The optimization process is based on the minimization of a cost function encoding the mismatch between an independent set of experimental results on notched beams under three-point bending and the related numerical predictions. To verify the calibrated model, we compare numerical and experimental results for the calibration tests in terms of load-deflection curves and crack paths. Finally, we compare the blind prediction of the challenge test results to the experimental results provided by the challenge issuer. Based on this work, the following conclusions can be drawn:

  • the adopted calibration procedure is an effective way of deriving elastic and fracture parameters of the material from the diverse set of available test results, while keeping the computational cost of the inverse problem to a manageable level;

  • considering the scatter of the experimental results and the many simplifying assumptions inherent to the macroscopic model at hand, the numerical predictions are in very good qualitative and quantitative agreement with the experimental results for both calibration and challenge tests, in terms both of load-displacement curves and of crack paths;

  • the adopted phase-field model is able to quantitatively predict the propagation of cracks under different local stress conditions, including various degrees of mode mixity which may also result in complex and fully three-dimensional crack paths.

Figure A.1. (a) Crack volume (black) defined by the phase-field iso-surface with d=0.95d = 0.95 (all elements with d<0.1d < 0.1 are hidden for visibility). (b) Definition of the numerical crack surface from averaging the minimum (green) and maximum (red) crack surface coordinates with d>0.95d > 0.95. (c) Post-processing of the crack surface asperity measurements (colored) and average experimental crack surface (black), data from [5].

Crack surface post-processing

For the submission of the results in the DMC, the obtained numerical results need to be post-processed. In particular, the definition of the crack surfaces needs specific attention and a proper procedure.

In this work, a material point is regarded as broken for a phase-field value above a threshold set to dth=0.95d_{\mathrm{th}} = 0.95. This yields a crack volume enclosed into the 3D iso-surface with d=dthd = d_{\mathrm{th}} (Fig. A.1a). This 3D iso-surface is obtained by applying the ‘Contour’ filter to the numerical results in ParaView [47, 48], which automatically creates a triangulation of the crack surface. From this 3D surface, the numerical crack tip is extracted as the subset of the points enclosed in the crack volume having maximum height coordinate along the thickness direction (Fig. A.1a).

Regarding the crack surface, the requirement for the challenge submission is to present a sampling of the crack surface based on a uniform discretization with step size of 0.1 mm. To this end, the phase-field results d(x)d(x) are first resampled with the required step size using the ParaView ‘Resample With Dataset’ filter. This resampled phase-field point cloud is then analyzed using an ad hoc script to extract the two outer limit crack surfaces within which the predicted crack surface lies (indicated by the red and green grid in Fig. A.1b). These are used for the DMC submission, where the two sides of the crack surface are requested. Concerning the comparison of the crack paths a unique surface is needed, for which here the average of the two limit surfaces is used, indicated by the black surface in Fig. A.1b.

Regarding the experimental data, to obtain a unique experimental crack surface out of the multiple tests performed, the experimentally obtained crack surfaces are averaged as well. To this end, the crack surface asperity measurements of the broken TPB specimens provided in [5] are first cleaned from any offsets or outliers (either manually or automatically with the ‘Filloutliers’ function of Matlab [49]). The corrected crack surface asperity measurements are then properly placed within the xx- and zz-direction of the geometry based on the coordinates of a significant point of an edge of the notch. As illustrated exemplarily in Fig. A.1c for the H45 specimen, a representative experimental crack surface (black surface) is then identified with the average of the corrected and aligned crack surfaces of the four specimens (colored grids).

Augmented reality renders

The AR 3D renders are created following the procedure presented in [50], and can be viewed on any recent mobile phone, tablet or internet browser. Along with the comparisons of the crack surfaces, also AR renders of the final phase-field state (Figure 10, Figure 14) as well as the anisotropy visualizations (Figure 8) are shown.

The renders are accessible at https://ar.compmech.ethz.ch or using the QR-code in Fig. B.1. Further, all scripts used for an automatic model and website generation can be found at https://gitlab.ethz.ch/compmech/ar.

QR code for accessing the AR renders of the models

Figure B.1. QR code for accessing the AR renders of the models.

Team background information

Our team is comprised of doctoral students and postdoctoral researchers in the Computational Mechanics Group led by Prof. De Lorenzis at ETH Zürich. The research of the group revolves around the development of solid mechanics models, along with novel techniques for their numerical solution and experimental validation. At the moment, one of the group’s main research topics is phase-field modeling of fracture and fatigue. Previous collaborations of the team members on the phase-field approach to fracture drove the decision to explore the capabilities of the model applied to complicated 3D cracks in an anisotropic material within the Damage Mechanics Challenge.

Declaration of Interests

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

We acknowledge funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 861061 – NEWFRAC Project, and from the Swiss National Science Foundation through Grant No. 200021-219407 ‘Phase-field modeling of fracture and fatigue: from rigorous theory to fast predictive simulations’.

References

  1. [1]A. Bobet, J. Morris , H. Yoon, L. Pyrak-Nolte, Damage mechanics challenge: Simulation of failure load and crack geometry, MR23C-0110, AGU Fall Meeting, San Francisco, CA, USA, 9–13 December 2019, 2019. URL: https://www.osti.gov/servlets/purl/1643452.
  2. [2]L. Jian, L. Pyrak-Nolte, A. Bobet, H. Yoon, J. Morris, Challenge data set for the damage mechanics challenge, 2023. doi:10.4231/ZX5A-QJ30.
  3. [3]L. Jiang, H. Yoon, A. Bobet, L. Pyrak-Nolte, Mineral fabric as a hidden variable in fracture formation in layered media, Scientific Reports 10 (2020). doi:10.1038/s41598-020-58793-y.DOI
  4. [4]L. Jiang, H. Yoon, A. Bobet, L. Pyrak-Nolte, Supplementary information: Mineral fabric as a hidden variable in fracture formation in layered media, 2020. URL: https://static-content.springer.com/esm/art%3A10.1038%2Fs41598-020-58793-y/MediaObjects/41598_2020_58793_MOESM1_ESM.pdf.DOI
  5. [5]L. Jiang, L. Pyrak-Nolte, H. Yoon, A. Bobet, J. Morris, Digital image, x-ray ct, xrd and ultrasonic data sets for damage mechanics challenge on brittle-ductile material, 2021. doi:10.4231/2E8M-W085.
  6. [6]L. Jiang, L. Pyrak-Nolte, H. Yoon, A. Bobet, J. Morris, Calibration data set for damage mechanics challenge on brittle-ductile material, 2021. doi:10.4231/QF39-QQ24.DOI
  7. [7]B. Bourdin, G. Francfort, J.-J. Marigo, Numerical experiments in revisited brittle fracture, Journal of the Mechanics and Physics of Solids 48 (2000) 797–826. doi:10.1016/S0022-5096(99)00028-9.DOI
  8. [8]G. Francfort, J.-J. Marigo, Revisiting brittle fracture as an energy minimization problem, Journal of the Mechanics and Physics of Solids 46 (1998) 1319–1342. doi:10.1016/S0022-5096(98)00034-9.DOI
  9. [9]K. Pham, H. Amor, J.-J. Marigo, C. Maurini, Gradient damage models and their use to approximate brittle fracture, International Journal of Damage Mechanics 20 (2011) 618–652. doi:10.1177/1056789510386852.DOI
  10. [10]K. Pham, J.-J. Marigo, From the onset of damage to rupture: Construction of responses with damage localization for a general class of gradient damage models, Continuum Mechanics and Thermodynamics 25 (2013) 147–171. doi:10.1007/s00161-011-0228-3.
  11. [11]E. Tanné, T. Li, B. Bourdin, J.-J. Marigo, C. Maurini, Crack nucleation in variational phase-field models of brittle fracture, Journal of the Mechanics and Physics of Solids 110 (2018) 80–99. doi:10.1016/j.jmps.2017.09.006.DOI
  12. [12]M. Ambati, T. Gerasimov, L. De Lorenzis, Phase-field modeling of ductile fracture, Computational Mechanics 55 (2015) 1017–1040. doi:10.1007/s00466-015-1151-4.DOI
  13. [13]C. Miehe, F. Aldakheel, A. Raina, Phase field modeling of ductile fracture at finite strains: A variational gradient-extended plasticity-damage theory, International Journal of Plasticity 84 (2016). doi:10.1016/j.ijplas.2016.04.011.DOI
  14. [14]R. Alessi, J.-J. Marigo, S. Vidoli, Gradient damage models coupled with plasticity: Variational formulation and main properties, Mechanics of Materials 80 (2015) 351–367. doi:10.1016/j.mechmat.2013.12.005.DOI
  15. [15]J. D. Clayton, J. Knap, A geometrically nonlocal phase field theory of brittle fracture, International Journal for Fracture 189 (2014) 139–189. doi:10.1007/s10704-014-9965-1.DOI
  16. [16]S. Teichtmeister, D. Kienle, F. Aldakheel, M.-A. Keip, Phase field modeling of fracture in anisotropic brittle solids, International Journal of Non-Linear Mechanics 97 (2017) 1–21. doi:10.1016/j.ijnonlinmec.2017.06.018.
  17. [17]N. van Dijk, J. Espadas-Escalante, P. Isaksson, Strain energy density decompositions in phase-field fracture theories for orthotropy and anisotropy, International Journal of Solids and Structures 196–197 (2020) 140–153. doi:https://doi.org/10.1016/j.ijsolstr.2020.04.022.
  18. [18]S. Nagaraja, U. Römer, H. G. Matthies, L. De Lorenzis, Deterministic and stochastic phase-field modeling of anisotropic brittle fracture, Computer Methods in Applied Mechanics and Engineering 408 (2023) 115960. doi:10.1016/j.cma.2023.115960.DOI
  19. [19]B. Li, C. Maurini, Crack kinking in a variational phase-field model of brittle fracture with strongly anisotropic surface energy, Journal of the Mechanics and Physics of Solids 125 (2019) 502–522. doi:https://doi.org/10.1016/j.jmps.2019.01.010.
  20. [20]P. Carrara, M. Ambati, R. Alessi, L. De Lorenzis, A framework to model the fatigue behavior of brittle materials based on a variational phase-field approach, Computer Methods in Applied Mechanics and Engineering 361 (2020). doi:10.1016/j.cma.2019.106928.DOI
  21. [21]M. Seiler, T. Linse, P. Hantschke, M. Kästner, An efficient phase-field model for fatigue fracture in ductile materials, Engineering Fracture Mechanics 224 (2020). doi:10.1016/j.engfracmech.2019.106807.DOI
  22. [22]A. Mesgarnejad, A. Imanian, A. Karma, Phase-field models for fatigue crack growth, Theoretical and Applied Fracture Mechanics 103 (2019). doi:10.1016/j.tafmec.2019.102282.DOI
  23. [23]J. Boldrini, E. Barros de Moraes, L. Chiareli, F. Fumes, M. Bitencourt, An non-isothermal thermodynamically consistent phase field framework for structural damage and fatigue, Computer Methods in Applied Mechanics and Engineering 312 (2016) 395–427. doi:10.1016/j.cma.2016.08.030.DOI
  24. [24]A. C. Cansen-Dörr, J. Brummund, M. Kästner, Phase-field modeling of fracture in heterogeneous materials: Jump conditions, convergence and crack propagation, Archive of Applied Mechanics 91 (2021) 579–596. doi:10.1007/s00419-020-01759-3.
  25. [25]F. Vicentini, P. Carrara, L. De Lorenzis, Phase-field modeling of brittle fracture in heterogeneous bars, European Journal of Mechanics - A/Solids 97 (2023). doi:10.1016/j.euromechsol.2022.104826.DOI
  26. [26]M. Hossain, C.-J. Hsueh, B. Bourdin, K. Bhattacharya, Effective toughness of heterogeneous media, Journal of the Mechanics and Physics of Solids 71 (2014) 15–32. doi:https://doi.org/10.1016/j.jmps.2014.06.002.
  27. [27]Purdue University, Damage Mechanics Challenge, Damage mechanics challenge: Questions & answers, 2022. URL: https://www.physics.purdue.edu/damage/questions-answers.html.
  28. [28]L. Pyrak-Nolte, H. Yoon, A. Bobet, J. Morris, L. Jiang, Damage mechanics challenge information packet (version jan 15, 2022), 2022. URL: https://purr.purdue.edu/groups/damagemechanicschallenge.
  29. [29]M. Ambati, T. Gerasimov, L. De Lorenzis, A review on phase-field models of brittle fracture and a new fast hybrid formulation, Computational Mechanics 55 (2015) 383–405. doi:10.1007/s00466-014-1109-y.DOI
  30. [30]C. Kuhn, A. Schlüter, R. Müller, On degradation functions in phase field fracture models, Computational Materials Science 108 (2015) 374–384. doi:10.1016/j.commatsci.2015.05.034.DOI
  31. [31]E. Lorentz, S. Cuvilliez, K. Kazymyrenko, Convergence of a gradient damage model toward a cohesive zone model, Comptes Rendus Mécanique 339 (2011) 20–26. doi:https://doi.org/10.1016/j.crme.2010.10.010.
  32. [32]R. Alessi, J.-J. Marigo, S. Vidoli, Gradient damage models coupled with plasticity and nucleation of cohesive cracks, Archive for Rational Mechanics and Analysis 214 (2014) 575–615. doi:10.1007/s00205-014-0763-8.DOI
  33. [33]C. Miehe, M. Hofacker, F. 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 (2010) 2765–2778. doi:10.1016/j.cma.2010.04.011.DOI
  34. [34]H. Amor, J.-J. Marigo, C. Maurini, Regularized formulation of the variational brittle fracture with unilateral contact: Numerical experiments, Journal of the Mechanics and Physics of Solids 57 (2009) 1209–1229. doi:10.1016/j.jmps.2009.04.011.DOI
  35. [35]F. Freddi, G. Royer-Carfagni, Regularized variational theories of fracture: A unified approach, Journal of the Mechanics and Physics of Solids 58 (2010) 1154–1174. doi:10.1016/j.jmps.2010.02.010.DOI
  36. [36]L. De Lorenzis, C. Maurini, Nucleation under multi-axial loading in variational phase-field models of brittle fracture, International Journal of Fracture 237 (2022) 61–81. doi:10.1007/s10704-021-00555-6.
  37. [37]F. Vicentini, C. Zolesi, P. Carrara, C. Maurini, L. De Lorenzis, On the energy decomposition in variational phase-field models for brittle fracture under multi-axial stress states, International Journal of Fracture (2024). doi:10.1007/s10704-024-00763-w.DOI
  38. [38]T. K. Mandal, V. Nguyen, J.-Y. Wu, Length scale and mesh bias sensitivity of phase-field models for brittle and cohesive fracture, Engineering Fracture Mechanics 217 (2019). doi:10.1016/j.engfracmech.2019.106532.DOI
  39. [39]D. Arndt, W. Bangerth, B. Blais, M. Fehling, R. Gaspar, H. Heister, L. Heltai, U. Köcher, M. Kronbichler, M. Maier, P. Munch, J.-P. Pelteret, S. Proell, K. Simon, B. Turcksin, D. Wells, J. Zhang, The deal.II library, version 9.3, Journal of Numerical Mathematics 29 (2021) 171–186. doi:10.1515/jnma-2021-0081.
  40. [40]D. Arndt, W. Bangerth, D. Davydov, T. Heister, L. Heltai, M. Kronbichler, M. Maier, J.-P. Pelteret, B. Turcksin, D. Wells, The deal.II finite element library: Design, features, and insights, Computers & Mathematics with Applications 81 (2021) 407–422. doi:10.1016/j.camwa.2020.02.022.DOI
  41. [41]T. Gerasimov, L. De Lorenzis, On penalization in variational phase-field models of brittle fracture, Computer Methods in Applied Mechanics and Engineering 354 (2019) 990–1026. doi:10.1016/j.cma.2019.05.038.
  42. [42]G. A. Baum, C. C. Haecker, E. H. Fleischman, Measurement of the orthotropic elastic constants of paper, The role of fundamental research in paper-making (1981) 453–478. doi:10.15376/frc.1981.1.453, trans. of the VIth Fund. Res. Symp. Cambridge, 1981.DOI
  43. [43]F. I. Fedorov, Theory of Elastic Waves in Crystals, Springer New York, 1968. doi:10.1007/978-1-4757-1275-9.DOI
  44. [44]H. Altenbach, J. Altenbach, W. Kissing, Mechanics of Composite Structural Elements, 2nd ed., Springer Singapore, 2018. doi:10.1007/978-981-10-8935-0.DOI
  45. [45]A. Vidales-Barriguete, J. Santa-Cruz-Astorqui, C. Piña-Ramírez, M. Kosior-Kazberuk, K. Kalinowska-Wichrowska, E. Atanes-Sánchez, Study of the mechanical and physical behavior of gypsum boards with plastic cable waste aggregates and their application to construction panels, Materials 14 (2021). doi:10.3390/ma14092255.
  46. [46]J. Nordmann, M. Aßmus, H. Altenbach, Visualising elastic anisotropy: theoretical background and computational implementation, Continuum Mechanics and Thermodynamics 30 (2018) 689–708. doi:10.1007/s00161-018-0635-9.DOI
  47. [47]J. Ahrens, B. Geveci, C. Law, Paraview: An end-user tool for large data visualization, Visualization Handbook (2005) 717–731. doi:10.1016/B978-012387582-2/50038-1.DOI
  48. [48]U. Ayachit, The ParaView Guide: A Parallel Visualization Application, Kitware, Inc., 2015.
  49. [49]The MathWorks Inc., Matlab version: 9.14.0 (r2023a), 2023. URL: https://www.mathworks.com.
  50. [50]M. Mathur, J. M. Brozovich, M. K. Rausch, A brief note on building augmented reality models for scientific visualization, Finite Elements in Analysis and Design 213 (2023). doi:10.1016/j.finel.2022.103851.DOI

Paper details

vlnc:W_5rr8vr7z
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents