Arxiv:2108.02336V1 [Cs.CE] 5 Aug 2021 Approximation to Capture the True Material Response with High Accuracy Efﬁciently

AFRACTURE MULTISCALE MODELFOR PERIDYNAMIC ENRICHMENTWITHINTHE PARTITION OF UNITY METHOD: PART I

APREPRINT

Matthias Birner Institute for Numerical Simulation, University of Bonn, Bonn, Germany Fraunhofer SCAI, Sankt Augustin, Germany

Patrick Diehl Center for Computation & Technology, Louisiana State University, Baton Rouge, LA

Robert Lipton Department of Mathematics, Louisiana State University, Baton Rouge, LA Center for Computation & Technology, Louisiana State University, Baton Rouge, LA

Marc Alexander Schweitzer Institute for Numerical Simulation, University of Bonn, Bonn, Germany Fraunhofer SCAI, Sankt Augustin, Germany

August 6, 2021

ABSTRACT

Partition of unity methods (PUM) are of domain decomposition type and provide the opportunity for multiscale and multiphysics numerical modeling. Different physical models can exist within a PUM scheme for handling problems with zones of linear elasticity and zones where fractures occur. Here, the peridynamic (PD) model is used in regions of fracture and smooth PUM is used in the surrounding linear elastic media. The method is a so-called global-local enrichment strategy, see [18, 1]. The elastic fields of the undamaged media provide appropriate boundary data for the localized PD simulations. The first steps for a combined PD/PUM simulator are presented. In part I of this series, we show that the local PD approximation can be utilized to enrich the global PUM arXiv:2108.02336v1 [cs.CE] 5 Aug 2021 approximation to capture the true material response with high accuracy efficiently. Test problems are provided demonstrating the validity and potential of this numerical approach. A PREPRINT -AUGUST 6, 2021

1 Introduction fracture growth) to validate the fundamental assumptions of our approach. As this paper provides a proof of concept, the reference cases considered are not too complex Peridynamics (PD) is a non-local generalization of clas- in geometry and loading. sical continuum mechanics [58] and naturally allows for the direct modelling of fracture phenomena and damage The paper is structured as follows: Section 2 introduces within its original formulation [13]. It can give good the preliminaries of the PUM and PD theory. Section 3 agreement with experimental data [14] in complex defor- sketches the approach to use peridynamics to construct mation scenarios. However, the computational expense appropriate multiscale enrichments on a local sub-domain incurred by exclusive use of peridynamics for large prob- for the global PUM. In Section 4 numerical results are pre- lems where damage and fracture only appear across small sented, which validate isolated steps of the overall multi- regions and most of the body behaves linearly elastic, can scale enrichment scheme. Specifically, we validate that be needlessly high. Therefor adaptive approaches have PD and PUM simulate a similar enough physical problem been proposed for such problems, e.g. force-based blend- despite different model formulations, in order to be able ing [52, 53, 56], the Arlequin approach [22], the variable to combine them. The issues here primarily are the differ- horizon method [57, 41], or discrete coupling [37, 55, 61, ent treatment of boundary conditions and the underlying 38, 30, 20, 62]. These methods directly couple PD and fi- material models. Moreover, we show examples of infor- nite element (FE) models to overcome the computational mation transfer in both directions: using the PD computed expense. These so-called concurrent coupling approaches crack path in a PUM simulation and using a PUM dis- introduce a non-physical interface (or overlapping hand- placement as boundary conditions on a PD problem. The shake zone) where the two computational models, PD and results clearly show the validity and potential of the pre- FE, must agree on the computed deformation behavior in sented approach. Finally, Section 5 concludes the paper a delicate way, and it is not obvious how computational with some remarks on future research steps. artifacts at this interface can be avoided in general. For example in [42] an out-of-balance force in the coupling region due to a lack of balance between the local and non- 2 Preliminaries local tractions is reported. For more general details we refer to the review paper [6]. 2.1 Classical continuum mechanics An alternative to the concurrent coupling approaches are Before we introduce the two models, we will look briefly so-called hierarchical coupling techniques where scale into the basics of classical continuum mechanics (CCM). separation is assumed, and a damage model together with Figure 1 shows a continuum in the reference configura- 3 an elastic model are utilized successively, exchanging in- tion Ω0 ⊂ R where a material point is identified with its 3 formation in a global and accumulated fashion, see e.g. position x ∈ R . The reference configuration Ω0 refers to [19] and references therein. For instance, in multiscale the shape of the continuum at rest with no internal forces. 2 FEM like FE methods two FE calculations are carried out The deformation φ : [0,T] × R3 → R3 of a material point in a nested manner, one at the macroscale and the other at x ∈ Ω0 to the so-called current configuration Ω(t) at time the microscale (in the vicinity of every quadrature point t ∈ [0,T] is given as of the macroscale model). Obviously, such approaches (t,x) := id(x) + u(t,x) = x(t,x) (1) yield a tremendous computational cost and are typically φ not trivial to implement which prohibits their widespread where u : [0,T] × R3 → R3 refers to the displacement use. A number of multiscale PD concepts have been proposed [36, 54, 60, 3, 45, 60, 44, 32], however, many open u(t,x) := x(t,x) − x. (2) questions remain. The stretch s : [0,T] × R3 × R3 → R3 between two ma- 0 We will pursue yet another approach to overcome the terial points x and x in the reference configuration Ω0 is computational expense of PD based on the partition of defined by unity (PU) approach [47]. The overall process we envi- s(t,x,x0) := φ(t,x0) − φ(t,x) = x(t,x) − x(t,x0). (3) sion is to construct a combined PD/PU simulator which automatically determines the region where PD should be employed so that the resulting local PD approximation 2.2 Partition of unity method can be utilized to construct a respective multiscale enrichment function for the global partition of unity method Partition of unity methods are a class of methods to (PUM) approximation to incorporate the true material re- numerically solve partial differential equations (PDEs). sponse including fracture growth with high accuracy effi- They were introduced in [39] to overcome limitations in ciently. In this paper, we present the first steps towards the the choice of basis functions of classical finite element realization of this process. Namely, we study the compati- methods (FEM). The key concept of a PUM is the use of bility of the different material models and the methods re- a compactly supported PU, that covers the computational quired for combining them. To accomplish this we seek to domain Ω. To each PU function ϕi a local approxima- outline the transfer of information between them. To this tion space Vi is attached, which yields a global, finite- PU end, we study several reference cases (with and without dimensional space V = ∑i ϕiVi, that is then used in a

2 A PREPRINT -AUGUST 6, 2021

3 of dimension ni, where i are spaces of polynomials of Ω0 φ : Ω0 → R Ω(t) P P degree d i . The spaces Ei denote so-called enrichment spaces of dimension dEi . The latter are arbitrary functions locally incorporated into the simulation, which we can x choose with respect to the problem at hand. We can ei- x(t,x) ther obtain the Ei from a-priori analytic knowledge about the structure of the solution to a problem. This is e.g. done in 2D fracture mechanics problems, where we have Figure 1: The continuum in the reference configura- 3 an analytic expansion of the solution around a crack tip tion Ω0 and after the deformation φ : Ω0 → R with available. Or the enrichments themselves are results of det(grad φ) > 0 in the current configuration Ω(t) at time other simulations. In the PUM the global approximation t. space then reads as PU V := ∑ϕiVi = ∑ϕiPi + ϕiEi. (8) Galerkin approach. An advantage over classical FEMs is i i the ability to incorporate arbitrary basis functions. Here Observe, that we do not assume the discrete functions the intent is to use only a few problem specific basis func- dPi dEi ! tions, thereby requiring less degrees of freedom (DOF) PU s s t+dPi u = ∑ϕi ∑ ui ψi + ∑ ui η, (9) to attain the desired global accuracy. Well-known in- i s=1 t=1 stances of FE based PUMs are the GFEM by Duarte and i.e. the employed basis functions, to be interpolatory. That Babuskaˇ [17] and the XFEM by Moes,¨ Dolbow and Be- j lytschko [40]. The particular meshfree PUM employed is, the coefficients ui do not necessarily correspond to in this paper was introduced in [47] and is referred to function values at specific points. We therefore use the as the PUM in the following. All computations in this direct splitting of the local spaces Vi presented in [48] study were carried out using the PUMA software frame- to enforce Dirichlet boundary conditions. By the above- work [49] developed at Fraunhofer SCAI. In the follow- mentioned flat-top property of the PU, it is possible to ing we present only a very short review of the PUM and construct a basis transformation that guarantees numerical refer the reader to [47] for details. stability and linear independence of the basis. The latter is of special importance when using enrichment functions Given a computational domain Ω, we assume to have a that are computed on-the-fly. Other PUM approaches partition of unity {ϕi} with ϕi ≥ 0 and such as the finite element based GFEM or XFEM cannot construct such a transformation. ∑ϕi x = 1 ∀x ∈ Ω (4) i that covers the domain. We call the support of a PU func- Equation of motion tion ϕi a patch ωi := supp(ϕi). In the PUMA software Let us first introduce our general model problem, the framework, patches are constructed as follows. First, we equation of (linear elastic) motion. On the global do- compute a cubic bounding-box C of the domain Ω. This main Ω we consider bounding box corresponds to discretization level zero. For level l, we sub-divide the bounding-box l-times and obtain ρ(x)u¨(t,x) = ∇ · σ + b in [0,T] × Ω0, (10) the cells where b are the volume forces acting on the body, d e.g. gravity, σ is the Cauchy stress tensor and ρ the mass k k width(C) Ci = c − h,c + h , with h = (5) density. The latter is computed from the linear strain ten- ∏ i i 2l+1 k=1 sor ε via Hooke’s law where the ci are the mid-points of the cells and h the dis- σ = C : ε = 2µε(u) + λtr(ε(u)) , (11) cretization parameter usually reported for FEM meshes. I Note that this sub-division can be done uniformly on all with C denoting Hooke’s tensor, µ,λ the Lame param- cells, or locally in an adaptive fashion. We then obtain the eters, tr(·) the trace and I the identity. The linear strain tensor is computed from the displacement field u by patches ωi by scaling the cells via d 1 T k k ε u (t,x) = ∇u(t,x) + ∇u(t,x) . (12) ωi := ∏ ci − αh,ci + αh , 2 > α > 1. (6) 2 k=1 To obtain a unique solution of (10) we impose boundary On these patches ωi we construct a Shepard PU with the conditions on ∂Ω = ΓN ∪˙ ΓD help of B-spline weight functions. Note that by construc- u(0,x) = u (x) in Ω (13) tion this PU has the flat-top property, i.e. on every patch 0 D ωi we can find an open ball on which it evaluates to one. u(t,x = u¯(t,x) on (0,T] × Γ (14) Details can be found in e.g. [47, 48]. σ(t,x) · n(x) = t¯(t,x) on (0,T] × ΓN, (15) To construct a higher order basis, each PU function ϕi is where n is the outward unit normal to ΓN and u¯ and t¯ multiplied with a local approximation space are the prescribed displacement and traction conditions, s t Vi := Pi ⊕ Ei = span ψi ,ηi , (7) respectively.

3 A PREPRINT -AUGUST 6, 2021

Discretization in space where ρ(x) is the material’s density at the material point x, f : [0,T] × 3 × 3 → 3 is the pair-wise force den- The weak formulation of (10) is given by: For fixed t ∈ R R R sity, and b : [0,T]× 3 → 3 is the external force density. [0,T] find u ∈ VPU(Ω) such that R R For more details about PD we refer to [24]. The material Z Z Z Z model is included in the pair-wise force function f and ρuv¨ dx = − σu : εvdx + tv¯ ds + bvdx for this paper the bond-based softening model [33, 34] is Ω Ω ΓN Ω chosen as the constitutive law. Figure 2 shows the deriva- (16) 0 PU D tion g (r) of the potential function g(r) of this constitutive for all test functions v ∈ V that vanish on Γ , see [48] model. In the double well model, the force acting between for details. With• ˜ denoting the coefficient vector of • 0 PU material points x and x is initially elastic and then soft- with respect to V , we can write (16) as follows: ens and decays to zero as the bond stretch between points ρNa˜ = −Ku˜ + t¯˜ + b˜, (17) increases. The critical bond stretch Sc > 0 for which the force begins to soften is given by with N the mass matrix weighted by the mass density, rc a = u¨ the acceleration and K the stiffness matrix corre- Sc = p . (23) sponding to linear elasticity. |x0 − x|

Discretization in time The bond stretch S : [0,T] × R3 × R3 → R between two 0 points x and x in Ω0 is defined as We discretize (17) in time using discrete time steps tn = 0 tn−1 + δtn, where δtn denotes the current time step size. 0 u(x ,t) − u(x,t) s(t,x,x ) = ◦ e 0 , (24) First we obtain the current acceleration from the current |x0 − x| x −x displacement at tn via x0−x where ex0−x = |x0−x| is a unit vector and “◦” is the dot −1 ˜ ˜ añ = (ρN) −Ku˜ n + t¯n + bn . (18) product. The non-local force is defined in terms of a double well potential. The potential is a function of the bond Using central differences and half time steps, we compute 0 1 stretch and is defined for all x and x in Ω0 by the (n + 2 )-velocity coefficient v˜ by W ε s(t,x0,x) = vñ+ 1 = δtn+ 1 añ + vñ− 1 (19) 2 2 2 ε 0 1 p 0 0 J (|x − x|) d+1 0 g( |x − x|s(t,x ,x)) (25) and similarly the next displacement by δ ωd|x − x| ε 0 u˜ n+1 = δtn+1vñ+ 1 + u˜ n. (20) where W (S(t,x ,x)) is the pairwise force potential per 2 unit length between two points x and x0. It is described in Quasi-static problem terms of its potential function g : R → R, given by 2 In the numerical examples, we also solve the quasi-static g(r) = h(r ) (26) version of (16). There, the assumption is that the acceler- where h is concave. Here ωd is the volume of the unit ball d ation u¨ = 0 is zero. The corresponding weak form thus is: in dimension d, and δ ωd is the volume of the horizon PU δ 0 find u ∈ V (Ω) such that Bδ (x). The influence function J (|x − x|) is a measure of the influence that the point x0 has on x. Note that we Z Z Z σ u : ε v dx = tv¯ ds + bvdx (21) chose a constant influence function Jδ (|x0 −x|) = 1 for all Ω ΓN Ω simulations in this paper. Only points inside the horizon can influence x so Jδ (|x0 − x|) is nonzero for |x0 − x| < δ for all test functions v ∈ VPU that vanish on ΓD. and zero otherwise. One common example of double well potential g out of the family of potentials is: 2.3 Peridynamic theory g(r) = C(1 − exp[−βr]) (27) Peridynamics (PD) [58], a non-local generalization of where C, are material dependent parameters. Thus, the classical continuum mechanics (CCM) which allows for β pair-wise force f : [0,T] × 3 × 3 → 3 is given by discontinuous displacement fields and provides an attrac- R R R tive framework for simulating cracks and fractures. PD 0 δ f(t,x ,x) = 2∂sW (S(t,x;x))ex0−x, (28) was employed in many validations against experimental data [14]. The governing equation for bond-based peridy- where namics reads as δ 0 ∂sW (s(t,x x)) = ρ(x)u¨(x)(x,t) = δ 0 1 J (|x − x|) p 0 0 Z d+1 0 ∂sg( |x − x|S(t,x ,x)). f(t,x0 − x,u(x0,t) − u(x,t))dx0 + b(x,t) (22) δ ωd |x − x| (29) Bδ (x)

4 A PREPRINT -AUGUST 6, 2021

Ω g0(r) 0

Bδ (xi) r c c δ −r+ −r r r+

Figure 2: Plot of the derivation g0(r) of the potential function g(r) used in the cohesive force in Equation 29. The force goes smoothly to zero at ±r+. Figure 3: The material points at discrete positions xi in the domain Ω0 in the reference configuration at time t = 0. For each discrete material point xi the neighborhood Elastic Properties Bδ (xi) := { j| |x j − xi| < δ} is computed. As an example The PD model parameters C and β in Equation (27) can be the neighborhood for the discrete node xi is shown. The estimated via energy equivalence using classical material material point xi exchanges force with all other discrete properties. The first PD material constant is obtained by nodes within its neighborhood. Adapted from [8]. G C := π c (30) 4 volumes are non-overlapping Vi ∩ Vj = /0, for i 6= j and the sum of the surrounding volumes are approximately using the classical energy release rate G as the reference n c the volume of the reference configuration ∑ Vi(t) ≈ material property. The second PD material constant β is i=1 |Ω0|,∀t ∈ [0,T]. Thus, the space-discrete governing equa- obtained by tion of motion reads as 4Eν (x )u¨(x ,t) = β := (31) ρ i i C(1ν )(1 − 2ν) ∑ f(t,x j − xi,u(x j,t) − u(xi,t))Vj + b(xi,t). using the Poisson ratio ν and the Young’s modulus E from j∈Bδ (xi) the classical theory. Note that in two dimensions, a bond- (33) based PD model has a fixed Poisson ratio of 1/3 for plain Combining the discretization in time and space thus yields strain due to the pair-wise force f [31]. Therefore, only Gc k+1 k k−1 and E are needed to calibrate the softening model. For a u(t ,xi) =2u(t ,xi) − u(t ,xi) detailed discussion for the energy equivalence with linear 2 ! ∆ts k elasticity we refer to [7]. The PUM’s material model was + b(t ,xi) + ∑ f(t,xi,x) . ρ(xi) calibrated using the Poisson ratio and the Young’s mod- j∈Bδ (xi) ulus. The values used in the simulations are provided in (34) Section 4. Boundary conditions One challenge for non-local Discretization in time models is the equivalent application of local boundary The discretization in time of Equation (22) is done via a conditions. Many publications related to the applica- central difference scheme tion of local boundary conditions in non-local models are available [43], however, since we apply local traction con- u(tk+1,x) = 2u(tk,x) − u(tk−1,x) ditions in our numerical examples, we focus on the con-   version of local traction conditions within peridynamics. ∆t2 Z The external force b(x,t) is usually applied in a region of + s b(tk,x) + f(t,x0,x). ρ(x)   horizon size δ times the length of the boundary and not Bδ (x) on a single line of nodes close to the boundary to mimic (32) a traction condition. Thus, the horizon can scale to zero so that in its limit the body forces applied to the area be- Discretization in space comes a boundary traction [35]. The peridynamic forces on the crack faces also converge to zero normal traction The governing PD equation of motion (22) is discretized on the crack faces as the horizon goes to zero, see [35]. in space using a collocation approach, the so-called EMU nodal discretization [58]. Figure 3 sketches the reference Convergence One has a-priori convergence of peridy- configuration Ω0 with the discrete PD nodes X = {xi|i = namic models under space and time grid refinement this 1,...n}. For all discrete PD nodes, a surrounding vol- is shown for both finite difference and finite element dis- ume V = {Vi ∈ R|i = 1,...n} is associated such that the cretizetions, see [25], [26], and [27].

5 A PREPRINT -AUGUST 6, 2021

Damage and crack identification Algorithm 1 1: crack pos = [] Since we do not break bonds as in many other PD models 2: // Loop over every 250th time steps and soften bonds instead, see Figure 2, we use the critical 3: for all d ∈ D , , ,...,n do c p0.5 1 250 500 value r = /β where the material model transit from 4: // Loop over the damage value at each node linear behavior to softening as an indication of damage. 5: max = 0, index = −1 To define a damage field D, we define for each discrete 6: for all di ∈ D do PD node x 7: // Check if there is damage max(r) 8: // and find the max damage d(x) = (35) rc 9: if di > 1 and di > max then 10: max = di where max(r) is the maximal value over all bonds within 11: index = i the neighborhood Bδ (x). In the case that d(x) < 1 there 12: end if is no damage and once, d(x) ≥ 1 we assume damage at 13: end for the node x. Now, the damage field for all nodes is used 14: Check if we have damage and add the position of to determine the crack path, which is not directly encoded xindex in the model. Algorithm 1 sketches the extraction of the 15: if max > 0 then crack path out of series of damage fields D1,...,n for n time 16: crack pos.append(xindex) steps. First, we iterate over a sub set of damage fields 17: end if d ∈ D1,250,500,...,n, see Line 3, where we take only every 18: end for 250th time step to extract the position of the crack tip. 19: linear interpolation(crack pos) Note that this value was chosen for the inclined crack example in Section 4.4. However, for faster crack growth, this value needs to be adjusted. Second, we iterate over all the damage values di at each PD node xi, see Line 6. On the other hand, generalizations of the FEM like PUM In Line 7 the maximal damage value is searched where we can solve large linear elastic crack evolution problems assume the current position of the crack path. In Line 14 rather efficiently when a crack path evolution is available, we collect the crack positions of each time step and con- since PUM employs only a few basis functions to resolve struct a linear interpolation to generate the crack path over the known crack path. In these methods, however, the ini- all time steps, see Line 19. For the inclined crack simu- tiation of cracks and their growth over time requires ad- lation in Section 4.4 the algorithm was used twice. Once ditional modelling and is typically the most challenging on the left side and once on the right side only to extract task. the left and right crack branches separately, see Figure 15. Thus, the benefits and weaknesses of PD and PUM are Note that the limitation of this algorithm is that the crack completely complementary to each other and their merger path aligns with the PD nodes, and we do not take into promises to allow for the efficient and reliable treatment account that the crack path is typically between two PD of arbitrary crack growth by combining the best of both nodes. To clarify the Algorithm 1 was run once as post- worlds. The aim of this work is thus to construct such a processing once the PD simulation was done. For the cou- combined PD/PUM simulator which automatically deter- pled approach, the algorithm would run until the current mines the region where PD should be employed so that time step, wait for the next time step to be processed, and the local PD approximation can be utilized to enrich the so on. There is no need to run the algorithm from the global PUM approximation to capture the true material beginning for each new crack position. response with high accuracy efficiently. To this end, we More sophisticated methods to extract a crack surface need to identify dynamically within a global PUM simu- were proposed by the authors in [10, 5] which are ap- lation those regions in which cracks will initiate or grow proximations of the crack path between PD nodes, how- in the next time step, e.g. by some strain or stress based ever, these numerical models introduce an additional un- damage model. Then, we extract the respective subdo- certainty and will be investigated in the future. mains together with respective boundary conditions on the subdomain boundaries obtained from the global PUM approximation. These subdomain problems are then passed 3 Peridynamics as Multiscale Enrichments to the PD simulator to approximate the local material re- in the Partition of Unity Method sponse, including crack initiation and growth up to the next global PUM time step. Then, we need to postpro- Solving fracture mechanics with the PD model allows for cess these local PD approximations to define appropriate naturally developing and growing cracks with the down- enrichment functions for the global PUM simulation that side of high computational costs independent of the actual encode the computed crack path and the respective de- deformation, damage and crack evolution, i.e. in regions formation in the vicinity of the crack. With the help of where the material response is essentially linearly elastic these enrichment functions, we can then define a PUM the PD model does not provide any approximation benefit approximation space that is appropriate for the resolution while being much more expensive to simulate. of the global material response in the next global time

6 A PREPRINT -AUGUST 6, 2021

Algorithm 2 region we set up a local problem, using a buffer zone, 1: n = 0 such that PD has access to a volumetric boundary strip 2: // Solve initial problem with PUM to prescribe boundary conditions. Initially we solve the 3: uPUM,0 = Solution of (17) for t0 = 0 global problem with the PUM, i.e. (17) as described in 4: C0 = initial crack patch Section 2.2, disregarding the crack. Thereby, we have a 5: Identify damaged area around crack → local domain displacement field around the local problem that captures 6: Set up PD problem on local domain the overall linear elastic response of the body. The so- 7: // Divide time interval T into N synchronization time lution is evaluated to set boundary conditions on the local steps problem, where we then solve (34) using PD. From the lo- 8: // Loop over synchronization (i.e. PUM) time steps cal PD solution we extract the crack path as described in 9: while n < N do Algorithm 1. In the PUM this crack path is for now mod- 10: Extract PD problem boundary data from uPUM,n eled using a Heaviside function as an enrichment along 11: Solve PD problem (34) until time tn+1 for uPD,n+1 the crack and the Westergaard enrichment functions at the tips. With the updated crack path, the global problem is 12: // Extract crack path as in Algorithm 1 solved by the PUM again to arrive at the coupled solution. 13: Cn+1 = crack path from uPD,n+1 In dynamic simulations, this has to be done in every syn- 14: Clear enrichments in VPU chronization time step. Here, we assume that the crack 15: Enrich VPU with crack enrichments corresponding does not grow outside the local problem domain. Cn+1 Using the full PD solution, and not only the crack path, 16: // Solve PUM problem with updated enrichments as an enrichment has already been studied in [51]. The 17: uPUM,n+1 = Solution of (17) for tn+1 main difficulty here, is that the PD solution is piece-wise 18: Update local domain constant, thus has zero derivative. Further questions that 19: n = n + 1 will arise in coupling the two methods is how to handle 20: end while time step sizes between them and the initial identification of the area of interest around the crack. These issues however need to be addressed in future work. step and thus can advance the global PUM simulation in time. Thus, we envision a so-called global-local enrichment strategy, see [18, 1] and Figure 4. 4 Numerical results As a first step towards this research goal we actually em- ploy PD and PUM both globally on the complete domain For all simulations the following material properties are to identify under which conditions the deformations com- considered: material density ρ =1200kgm−3, Young’s puted by PD and PUM agree well locally, i.e. we iden- modulus E =3.25GPa, bulk modulus K =2.16GPa, a tify the complement of the local regions on which the PD Poisson ratio ν = 1/3 and a critical energy release rate model should ultimately be employed only. This valida- G =500Jm−2. All peridynamic simulations were done tion study ensures that away from damaged regions the with the author’s NLMech code [12], all PUM simula- global PUM approximation can safely replace the PD ap- tions with PUMA [49]. For the comparison, we define the proximation and thus can be used to prescribe boundary maximal magnitude of displacement as data for the local PD simulations in the future. Then, we study how information from the PD approximation can Umax = max (|u(t,xi)|) (36) be passed back into the global PUM approximation, i.e. i=1,...,n how to construct enrichments for PUM from PD, compare to compare the simulation results of both methods. For [51]. Here, we again start with a rather simple strategy the PUM results, the locations {x } refer to the grid points to validate the overall approach. To this end, we extract i of the mesh the solution is visualized on. All PUM sim- the crack path predicted by PD, see Algorithm 1, and pass ulations were run with a discretization on #ϕ = 4096 only this geometric information to PUM. On this crack ge- i patches, as further refinement did not change the results in ometry we then define the typical crack enrichments [50] any meaningful digit. We report the derived discretization based on the Heaviside and Westergaard functions and fi- parameter h for each experiment as calculated by (5). nally compare the global PD evolution with the respective PUM The local approximation spaces (7) are given by first order global PUM solution to validate our approach. polynomials and the given crack enrichments, if a crack is The coupling then works as follows. We use global-local present in the problem. There, we enrich patches inter- enrichments (GL) as introduced in [18, 1] as a coupling secting the crack, but not the crack tips with a Heaviside framework, see Figure 4 for an overview. Assume the function and patches in an area around the tips with the problem, which we refer to as the global problem, is Westergaard functions. For the non-local model we ap- given on a simple domain with a single crack [18, 1]. By ply the local traction condition highlighted in blue in Fig- e.g. some strain or stress based damage model we iden- ure 5 using the external force density b(xi,t) in a layer of tify a region of interest around the crack. This is not per- horizon-size δ along the local traction condition (which formed in this paper, but left for future work. On that corresponds to the red rectangle). The external force den-

7 A PREPRINT -AUGUST 6, 2021

solve local problem identify boundary local region conditions solve

global problem PUM PD

crack extract enrichments crack path

Figure 4: Enriching PUM with PD using a global-local approach. We are interested in the global problem, which is solved using the PUM twice. First it is solved using the initial or old crack conﬁguration to obtain boundary conditions for the local problem. Once PD provides a new crack path through the local problem, the updated global problem is solved again.

Local Non-local validate that both methods and codes simulate the same problem. However, after the bonds start to soften, i.e. when the deformation is large enough to impose a bond stretch larger than rc. The PD model simulates a different material behavior, while the PUM model stays in δ the linear elastic behavior. To investigate the difference in the displacement fields, we apply force F = 9 × 105N to stretch the bonds to ≈ 7.3% to the critical stretch rc. Table 1 lists all simulation parameters. Here, the maxi- F b(t,x) mal displacement magnitude Umax obtained by the PUM method is ±1.235 × 10−4m for the quasi-static simulation Figure 5: Left: traction condition in the local model on the −4 lower boundary in blue. Right: Application of the non- and ±1.253 × 10 m for the dynamic simulation, respec- local traction conditions in the layer of horizon size δ. If tively. The displacement field for the quasi-static case is the horizon can scale to zero so that in its limit the body shown in Figure 8a and the displacement field for the dy- forces applied to the layer of horizon size is equivalent to namic case in Figure 7. Note that we will use quasi-static the boundary traction on the right-hand side [35]. PUM simulations for the remaining numerical examples. The maximal displacement magnitude Umax obtained by the explicit PD simulations was 1.20335 × 10−4m, see Figure 8c, and the obtained damage is around 7.3%, see sity at a node x within the layer of horizon-size δ yields i Figure 8e. Note that we add the external force density, F see (37), such that the force increases linearly over time b(x ,t) = (37) i Vm such that we have the final force F at the final time T. where F is the force in Newton applied at the local trac- Next, the force is increased by four times, which results tion condition, V is the sum of all volumes Vi associated in a bond damage of 29.8%. The maximal displacement −4 to nodes xi in the layer of horizon-size δ, and m is the magnitude Umax obtained by PUM is 4.941 × 10 m and amount of nodes within the horizon δ. More theoreti- obtained by PD is 4.890 × 10−4m, respectively. The PUM cal details are shown in [35]. We utilize these non-local displacement is shown in Figure 8g and the PD displace- boundary conditions in all remaining examples. ment and PD damage in Figure 8i and Figure 8k, respectively. 4.1 Two-dimensional bar (pure linear elasticity) Next, the force is increased by eight times, which results in a bond damage of 62.7%. The maximal displacement In this first example, we use a two-dimensional bar, see −4 magnitude Umax obtained by PUM is 9.883 × 10 m and Figure 2, to investigate the effect of softening bonds in obtained by PD is 1.020 × 10−3m, respectively. The PUM the PD model. The PD potential in Figure 2 stays in the displacement is shown in Figure 8b and the PD displace- linear regime until the stretch r exceeds its critical value ment and PD damage in Figure 8d and Figure 8f, respec- c r which means that the material behaves linearly elas- tively. tic. In that case, the same material behavior as in the PUM model is simulated. Therefore, this example can

8 A PREPRINT -AUGUST 6, 2021

Next, the force is increased by twelve times, which results the corresponding damage in Figure 11b. Here, the max- in a bond damage of 106.1%. The maximal displacement imal displacement magnitude is 1.745 × 10−7m and the −3 magnitude Umax obtained by PUM is 1.675 × 10 m and damage is around 6.9%. Note that for a similar damage obtained by PD is 1.482 × 10−3m, respectively. The PUM percentage in the two-dimensional bar example without displacement is shown in Figure 8h and the PD displace- any pre-crack both methods predicted a much closer max- ment and PD damage in Figure 8j and Figure 8l, respec- imum displacement magnitude, see first row of Table 2. tively. In [16, 59] various limits are shown and in the case that the horizon δ → 0 and the nodal spacing h → 0 the Table 2 lists all results for the different bond damage val- PD non-local models converges to the continuum PDE model. ues. One clearly sees that if the bond damage is below Furthermore, the nodal spacing h has to shrink faster one, the displacement agrees, since the PD model is in the PD than the horizon δ to obtain convergence [11]. There- linear regime and both models simulate linear elasticity. fore, we decrease the horizon by 1/2 and the nodal spac- However, if the bond damage is greater than one, the dis- ing h by 1/4. The magnitude of the displacement field placement is different, since the PD model starts to soften PD is shown in Figure 11c and the corresponding damage in and the PUM model still simulates linear elasticity. Thus, Figure 11d. Here, the maximal displacement magnitude we have shown that for the linear behavior both models re- is 9.661 × 10−8m and the damage is around 7.7%. In that sults in a similar maximal displacement magnitude U . max case, the order of magnitude is the same and the maximal Note that for the dynamic PUM and the explicit PD sim- displacement magnitude differs by 0.993 × 10−8. Table 4 ulations, one could do a more detailed convergence study. gives an overview of all results. Especially for PD the suitable choice of the horizon and One can see that PD convergences slowly to the contin- the grid size is important for convergence. Another as- uum PDE model and a small horizon and nodal spacing pect is that the traction condition only converges to the is needed. In addition, a very small horizon is needed for local traction condition while the horizon scales to zero, the convergence of the traction condition. Therefore, it see [35]. For the sake of this work, we use this discretiza- is beneficial to use PD solely where crack and fractures tion in time and space to show the proof of concept of the arise and have the PUM deal with the bulk of the undam- proposed multiscale method only. aged material and provide appropriate boundary data for the local PD simulations. 4.2 Stationary Mode I crack

Figure 9 shows a square plate (0.1m × 0.1m) with an ini- 4.3 Validation of coupling PUM and PD using tial crack of length l = 0.02m. For the PDE-based PUM displacement model a traction boundary condition of ±1 × 103N is applied to the bottom of the plate on the left-hand side and In the previous examples, the focus was on the lower right-hand side of the initial crack. A Dirichlet boundary part of Figure 4 and the extracted crack path from PD condition is applied to the top of the plate and the dis- was used as the crack enrichments for the PUM. In this placement there is fixed. However, in the non-local PD example, the focus in on the upper part and displacement model boundary conditions are applied differently. For values from the global problem are used within the local the Dirichlet boundary condition all nodes within a layer PD region. Therefore, we revisit the stationary Mode I of horizon size at the top of the square plate are fixed example in Section 4.2. Figure 9 sketches the geometry in displacement. The force is applied using the external for the global problem. For the local problem a square force density in (37). Note that we add the external force area (0.02 × 0.02) including half the length of the initial such that the force increases linearly over time, such that crack is chosen, compare Figure 12. Within the square we have the final force F at the final time T. plate we define a layer of horizon size δ highlighted in green. To test local and global coupling a simplified First, we compare the displacement field of the quasi- version of Algorithm 2 is used. In Line 5 the local static and dynamic PUM simulation with an initial domain around the crack which becomes the local PD crack. Figure 10a shows the displacement magnitude domain is prescribed. Note that later a criterion, e.g. for the quasi-static PUM and Figure 10b shows the high stress concentration, will be used to determine this displacement magnitude for the dynamic PUM, respec- region. In Line 6 the local domain is discretized using tively. In both cases the maximal displacement magni- the PD code and a VTK file with the discrete PD nodes tude is 8.688 × 10−8m for the quasi-static simulation and is written to hard disk. In Line 10 the PUM code reads 8.704 × 10−8m for the dynamic simulation. Again, as the VTK file, adds displacement values for all nodes for the previous example the both values are sufficiently within the horizon layer, and write the file back to hard close. disk. In Line 11 the PD code reads this file, uses the Second, the displacement field for two different horizons PUM displacement values as boundary conditions and and nodal spacing are compared. For the first PD sim- all other initial displacement values are set to zero, and ulation the same δ and the same hPD as for the two- runs until time tn+1. Note that we linearly increase the dimensional bar in Section 4.1 was used. The magni- displacement at each node within the horizon layer such tude of the displacement field is shown in Figure 11a and that the prescribed value by PUM is reached at time tn+1.

9 A PREPRINT -AUGUST 6, 2021

y −F F 0.1 x

1 Figure 6: Sketch of the two-dimensional bar which is used to study the inﬂuence of bond softening on the displacement. Initially a force of F = ±9 × 105N is applied on the left-hand side and right-hand side of the bar. The force increases linearly to 2F, 4F, and 8F which results in a bond damages of 7.3% up to 106.1%.

Table 1: Simulation parameters for the discretization in time and space for the two-dimensional bar problem.

5 Force F =9 × 10 N Time steps tn=50000 −8 Node spacing hPD =0.0005m Time step size ts =2 × 10 s and hPUM =0.0078125m Horizon δ = 4hPD =0.002m Final time T=0.001s

mesh refinement with PD [2, 15, 3, 60] or coupling FE with PD [62] can be used to construct a suitable PD discretization in the local problem’s domain employing the prescribed displacement, especially around the crack or discontinuity This will be investigated in the second part of the paper where the focus is on transferring information from PUM to PD. Figure 7: Obtained displacement field by the dynamic PUM simulation for the load F=9 × 105N. For comparison the displacement field obtained by the quasi-static 4.4 Inclined crack PUM simulation is shown in Figure 8a. The previous examples either had no crack or the applied force was too small to start crack growth. With this example, we further stress the agreement of PUM and For this example, the algorithm terminates here, since PD simulations by introducing dynamic crack growth. the goal was to show the displacement obtained in the Figure 14 shows a square plate of length (0.1m × 0.1m) local region using PD matches the one in the same area with an initial inclined crack of length 0.02m inclined of the global region obtained by a pure PUM simulation. by Θ = 72.5°. A force F is applied in normal direction Recall that the open challenges here are to find a criterion at the top and bottom of the square. In the dynamic PD to identify the local region, find the synchronization simulations, the force is zero initially and scales linear time step size, and time step size for the local problem. to the maximal force F at the final time T. Table 5 lists However, these will be investigated in a forthcoming the simulation parameters for the discretization in time second part of the paper. and space. In PD the boundary condition is applied as described in (37). As in the previous example, we run two PD simulations with a coarse and a fine nodal Figure 13a shows the global displacement magnitude spacing. Figure 15 shows the extracted crack path for within the local area at the synchronization time step both nodal spacings. The crack path was extracted in a 0.001s. Figure 13b shows the displacement magnitude of post-processing step using the damage field D and the the local problem for δ = 0.002 and hcoarse = δcoarse/4 at coarse PD current positions of the PD nodes, see Algorithm 1. Note the final time T = 0.001s with 50000 explicit time steps. that PD has the notion of damage included in the model, Figure 13d shows the damage field. however, the crack tip or crack surface is not encoded Figure 13c shows the displacement magnitude for the lo- in the model and needs to be extracted [13]. More fine δ cal problem for δfine = 0.001 and hPD = fine/8 at the fi- sophisticated methods to extract the crack surface were nal time T = 0.001s with 50000 explicit time steps. Fig- proposed by the authors [10, 5] which are approximations ure 13e shows the damage field around the initial crack. of the crack path between PD nodes, however, these numerical models introduce an additional uncertainty and Looking at the results from a bird’s eye view, one could will be investigated in the future. say that the displacement field matches the shape and has the same order of magnitude. Looking at the damage field, one observes the kink in the crack path, which is not ob- Figure 17a shows the displacement magnitude obtained served in the global PD simulation in the previous section. by the PUM method using the crack path from the Looking closer at the PD discretization around the crack coarse nodal spacing. Figure 17c shows the displacement explains this behavior. Here, experience from adaptive magnitude obtained by PD and Figure 17e the corre-

10 A PREPRINT -AUGUST 6, 2021

Force = 9 × 105N Force = 8×9 × 105N

(a) PUM (Displacement) (b) PUM (Displacement)

(e) PD (Damage) (f) PD (Damage)

Force = 4×9 × 105N Force = 12×9 × 105N

(g) PUM (Displacement) (h) PUM (Displacement)

(i) PD (Displacement) (j) PD (Displacement)

(k) PD (Damage) (l) PD (Damage) Figure 8: Magnitude of the displacement for the two-dimensional bar, see Figure 6. The traction condition’s load various from 9 × 105N up to 12×9 × 105N to study the effect of softened bonds. As long as the PD model stays in the linear regime, the displacement agrees. Once the softening starts the results differ since the PUM still simulates linear elasticity.

11 A PREPRINT -AUGUST 6, 2021

Table 2: The maximal displacement magnitude Umax obtained by the quasi-static PUM simulation and by the explicit PD simulation for the two-dimensional bar, see Figure 6. The traction condition’s load is increased up to twelve times to showcase the inﬂuence of softened bond to the displacement ﬁeld. As long as the PD model stays in the linear regime, both methods result in a similar displacement and once the softening starts the results diverges as expected.

Load [N] Umax [m] Damage [%] PUM PD 9 × 105 1.235 × 10−4 1.203 × 10−4 7.3 4×9 × 105 4.941 × 10−4 4.890 × 10−4 29.8 8×9 × 105 9.883 × 10−4 1.020 × 10−3 62.7 12×9 × 105 1.482 × 10−3 1.675 × 10−3 106.1

Table 3: Simulation parameters for the discretization in time and space for the stationary crack problem.

3 Force F =1 × 10 N Time steps tn=50000 −8 Node spacing hPD =0.0005m Time step size ts =2 × 10 s and hPUM =0.00078125m Horizon δ = 4hPD =0.002m Final time T=0.001s

u = 0 Previously, the magnitude of the displacement was in a good agreement for the example where the PD model stayed in the linear regime, see Section 4.1, and for stationary Mode I crack, see Section 4.2. To get a 0.1 better understanding of the difference in the displacement field, we compute the point-wise absolute error εi := |UPD(xi) − UPUM(xi)|, see Figure 16. Note that y the relative error was avoided, since the displacement could be zero, which results in a division by zero. One −F 0.02 F x clearly identifies that the absolute error εi is large on the left-hand side and right-hand side at the crack tip and 0.1 there is the band going from the initial crack to the top Figure 9: Sketch of the square plate (0.1m × 0.1m) with and bottom of the plate. an initial crack of length 0.02m. For the local PUM model a constant traction boundary condition of ±1 × 103N is The obvious issues would be that the PD model left the applied to the left-hand side and right-hand side of the linear regime and the softening started. However, from initial crack. A Dirichlet boundary condition is applied the damage fields (Figure 17e and Figure 17f) we can ob- on the top of the plate and the displacement is fixed in serve that in both cases most of the domain is in the linear both directions. regime of the PD material model, as the damage is way below 20% except for the crack. We currently assert the following reasons might explain the differences in the displacement fields once crack growth is involved in the PD sponding damage. Figure 17b shows the displacement model: magnitude obtained by the PUM method using the crack path from the fine nodal spacing. Figure 17d shows the • Static vs explicit time integration: displacement magnitude obtained by PD and Figure 17f As most particle-based methods, e.g. molecular the corresponding damage. dynamics or smoothed particle hydrodynamics, explicit time integration requires very small- In the coarse case, there is only a small gap between time steps to get the system in the equilibrium. the two solutions, as PD computes a maximal dis- Meaning that in the static simulation the ap- placement magnitude of 1.140 × 10−4m and the PUM plied traction boundary condition is ”globally 1.421 × 10−4m. In the fine case, with a greater force available” in the stiffness matrix, however, in applied, the gap widens as PD computes a maximal the explicit time integration the applied body displacement magnitude of 1.166 × 10−4m whereas force travels from the boundary to the center with 3.136 × 10−4m the PDE model of the PUM reacts of the plate where the initial crack is. This linearly to the increased load. might affect the differences in the stripe from the top and bottom to the center of the plate.

12 A PREPRINT -AUGUST 6, 2021

(a) Quasi static (b) Dynamic Figure 10: Obtained displacement magnitude by PUM for the (a) quasi-static case and (b) for the dynamic case. Same as for the previous example 4.1 without any crack the displacement is close.

Table 4: The maximal obtained displacement magnitude for the stationary crack obtained by a quasi-static and dynamic PUM simulation on the left two columns. On the right two columns an explicit PD simulation on a coarse nodal spacing and ﬁner nodal spacing.

Umax [m] PUM PD quasi-static dynamic Coarse Fine 8.688 × 10−8 8.704 × 10−8 1.745 × 10−7 9.661 × 10−8

Here, we checked the sound wave speed to get namics and the effect of the boundary we pro- some estimation how far the force traveled in vide an example of a dynamic crack centered in the simulation time. The sound speed for the a strip of finite width. The velocity of a centered dynamic problem is roughly 526 meters per crack growing in a strip of finite width starting second and the time step is 2x10−8 second. For from a fixed length and zero initial velocity is ini- the simulations done here the distance of the tially accelerating until waves reflected from the crack to the boundary is around 0.05 meter and strip boundary interact with the crack tip causing the time it takes for information to travel from the acceleration to drop nearly to zero forcing the the crack tip to the boundary and back is on the crack to move with nearly steady velocity. This order of 10−4 second or 0.5x104 time steps. To phenomenon is seen in experiment [4] and simu- reduce this effect the time step size could be lated numerically [28]. reduced which would make the PD simulation even more expensive. On the other hand, the • Sharp crack vs crack zone: local region within the global PUM region Using the extracted crack path within the PUM would be much smaller but still expensive to the yields in a sharp crack, however in the PD model small time step size. the crack is instead a narrow zone of failed bonds of finite width equal to twice the peridynamic horizon δ. In addition, at the crack tip there is For the differences at the crack tips, the same ar- a process zone on the order of the horizon size gument holds. In this implementation of PD the δ where the bonds are in the process of soften- wave travels from the crack tip to the boundary ing to failure. Because of this PUM exhibits a and in the quasi-static PUM this information is strain singularity where PD does not. This dif- there immediately. To illustrate the effect of dy- ference between the strain fields surrounding the

13 A PREPRINT -AUGUST 6, 2021

coarse δcoarse δcoarse = 0.002 and hPD = /4

(a) Displacement (b) Damage fine δ δfine = 0.001 and hPD = fine/8

(c) Displacement (d) Damage Figure 11: The displacement magnitude and the damage for the stationary crack example 4.2 as computed by PD on two different nodal spacings.

14 A PREPRINT -AUGUST 6, 2021

u = 0 such that information passes automatically between the PD and PUM codes. Specifically, this requires automatic extraction of the crack path on the PD side and passing 0.02 the boundary values from the PUM to the local PD prob- 0.1 lem. Once the scheme is running, there are several more topics to investigate: Measuring the impact of global and local timestep choices. Measuring the impact of differ- 0.02 y ent kinds of boundary conditions, with which we transfer information from the global PUM simulation to the local −F F x 0.01 PD model. 0.1 Supplementary materials Figure 12: Sketch of the local problem within the global problem. The small square plate (0.02 × 0.02) is the area The PeriHPX code [12] utilizing the C++ standard library of the local problem including half of the length of the for concurrency and parallelism (HPX) [29] is available initial crack. For all PD nodes within the layer of horizon on GitHub1. NLMech has the following dependencies: size δ the displacement is prescribed by the displacement HPX 1.5.9, Blaze [23] 3.5, Blaze Iterative, YAML-CPP obtained by the global PUM simulation. 0.6.3, hwloc 2.2.0, boost 1.73.0, jemalloc 5.1.0, gcc 9.3.0, gmsh [21] 4.7.1, PCL [46] 1.11.1, FLANN 1.9.1, and VTK 9.0.1. The simulation input files are available on cohesive zone in PD and the strain field around Zenodo [9] or GitHub2. The PUMA3 software frame- sharp crack in PUM contributes to the difference work [49] used for all PUM simulations is developed at in displacement around the crack seen for these Fraunhofer SCAI. two models. Acknowledgements 5 Conclusion PD thanks the Center of Computation & Technology at In fracture mechanics problems, peridynamics allows for Louisiana State University for supporting this work. This naturally developing and growing cracks with the down- material is partially based upon work supported by the U. side of high computational costs. Partition of unity meth- S. Army Research Laboratory and the U. S. Army Re- ods however require only a few degrees of freedom to search Office under Contract/Grant Number W911NF-19- model a known crack path. The primary focus of this 1-0245. work has been to introduce a constructive multiscale enrichment approach that allows to combine the benefits of References both methods and provide a proof of concept implementation of the proposed approach. In numerical experi- [1] Matthias Birner and Marc Alexander Schweitzer. ments, we have tested all steps necessary for the proposed Global-local enrichments in puma. In International scheme in isolation. The presented results clearly show Workshop on Meshfree Methods for Partial Differ- the validity and potential of the presented approach. It ential Equations, pages 167–183. Springer, 2017. is also clear however, that in order to provide a proof of concept, not too complex geometries were considered. [2] Florin Bobaru and Youn Doh Ha. Adaptive refinement and multiscale modeling in 2d peridynamics. Future work International Journal for Multiscale Computational Engineering, 9(6), 2011. In the next part of the series, we want to further investi- [3] Florin Bobaru, Mijia Yang, Leonardo Frota Alves, gate the effect of applying a displacement or a force from Stewart A Silling, Ebrahim Askari, and Jifeng Xu. the local problem as boundary conditions on the local PD Convergence, adaptive refinement, and scaling in 1d problem, see Section 4.3 for some preliminary results. peridynamics. International Journal for Numerical Another aspect is to construct a suitable PD discretization Methods in Engineering, 77(6):852–877, 2009. in the local problem’s domain employing the prescribed displacement, especially around the crack or discontinu- [4] E Bouchbinder, T Goldman, and J Fineberg. The ity, to avoid the observed kink in the crack path. dynamics of rapid fracture: instabilities, nonlineari-

A further topic would be the automatic identiﬁcation of 1https://github.com/perihpx/ the domain where to switch to the non-local model in 2https://github.com/diehlpk/paperPUMPD a running PUM simulation, e.g. by some stress or strain 3https://www.scai.fraunhofer. based damage model. Ultimately, we want to combine all de/en/business-research-areas/ parts introduced in the series into an automated algorithm, meshfree-multiscale-methods/products/puma.html.

15 A PREPRINT -AUGUST 6, 2021

(a) Left: displacement magnitude (Umax) of the global solution. Right: Umax in the local region obtained by the PUM.

coarse δcoarse fine δ δcoarse = 0.002 and hPD = /4 δfine = 0.001 and hPD = fine/8

(b) Displacement magnitude (c) Displacement magnitude

(d) Damage (e) Damage Figure 13: Displacement magnitude and damage on the local region as obtained by PD on two different nodal spacings, coarse on the left and ﬁne on the right.

16 A PREPRINT -AUGUST 6, 2021

Table 5: Simulation parameters for the discretization in time and space for the inclined crack problem.

6 Force Fcoarse =4.25 × 10 N Final time T=0.001s 6 Force Ffine =9 × 10 N Time steps tn=50000 −8 Node spacings: Time step size ts =2 × 10 s hPUM =0.00078125m Horizon: coarse coarse hPD =0.0005m δcoarse = 4hPD fine fine hPD =0.000125m δfine = 8hPD

0.1

8 · 10−2

Θ = 72.5° · −2 0.1 6 10 0.02 4 · 10−2

y 2 · 10−2 x

0 0 2 · 10−2 4 · 10−2 6 · 10−2 8 · 10−2 0.1

coarse δcoarse 0.1 (a) δcoarse = 0.002 and hPD = /4 Figure 14: Sketch of the square plate (0.1m × 0.1m) with an initial inclined crack of length 0.02m. For the lo- 0.1 cal PUM model a constant traction boundary condition 3 of ±1 × 10 N was applied to the top and bottom of the −2 square plate. 8 · 10

6 · 10−2 Table 6: The maximal displacement magnitude Umax obtained by the quasi-static PUM simulation and by the explicit PD simulation for the inclined crack problem, see −2 Figure 14. 4 · 10 hPD Umax [m] Load [N] PD PUM 2 · 10−2 Coarse 1.140 × 10−4 1.421 × 10−4 4.25 × 106 . × −4 . × −4 × 6 Fine 1 166 10 3 136 10 9 10 0 0 2 · 10−2 4 · 10−2 6 · 10−2 8 · 10−2 0.1 fine δ (b) δfine = 0.001 and h = fine/8 ties and length scales. Rep Prog Phys, 77(4):046501, PD 2014. Figure 15: Extracted crack tip positions obtained by the PD simulation. The black line is the initial inclined crack. [5] Michael Bussler, Patrick Diehl, Dirk Pfluger,¨ Steffen The red and blue lines are the crack branches and each dot Frey, Filip Sadlo, Thomas Ertl, and Marc Alexander is the crack position extracted between the explicit time Schweitzer. Visualization of fracture progression in step 47500 to 50000 with a resolution of 500 explicit time peridynamics. Computers & Graphics, 67:45–57, steps in between. The crack tip positions was extracted 2017. using Algorithm 1. [6] Marta D’Elia, Xingjie Li, Pablo Seleson, Xiaochuan Tian, and Yue Yu. A review of local-to-nonlocal

17 A PREPRINT -AUGUST 6, 2021

ics and phase-field models for engineering fracture mechanics, Mar 2021. [14] Patrick Diehl, Serge Prudhomme, and Martin Levesque.´ A review of benchmark experiments for the validation of peridynamics models. Journal of Peridynamics and Nonlocal Modeling, 1(1):14–35, 2019. [15] Daniele Dipasquale, Mirco Zaccariotto, and Ugo Galvanetto. Crack propagation with adaptive grid refinement in 2d peridynamics. International Jour- nal of Fracture, 190(1-2):1–22, 2014. [16] Qiang Du and Xiaochuan Tian. Robust discretization of nonlocal models related to peridynamics. In Meshfree methods for partial differential equations VII, pages 97–113. Springer, 2015. [17] C. A. Duarte, I. Babuska,ˇ and J. T. Oden. Gener- alized finite element methods for three-dimensional structural mechanics problems. Computers & Struc- tures, 77(2):215–232, 2000. Figure 16: Point-wise absolute error εi := |UPD(xi) − [18] C Armando Duarte, Dae-Jin Kim, and Ivo Babuska.ˇ UPUM(xi)| between the PUM and the PD solution at the A global-local approach for the construction of final time T. Note that the largest error is at the crack and enrichment functions for the generalized fem and dominating. Thus, we us a different color scale excluding its application to three-dimensional cracks. In values close to the crack. Advances in meshfree techniques, pages 1–26. Springer, 2007. coupling methods in nonlocal diffusion and nonlocal [19] Jacob Fish. Practical Multiscaling. Wiley, 2013. mechanics. arXiv preprint arXiv:1912.06668, 2019. [20] Ugo Galvanetto, Teo Mudric, Arman Shojaei, and [7] P Diehl, R Lipton, and M Schweitzer. Numerical Mirco Zaccariotto. An effective way to couple fem verification of a bond-based softening peridynamic meshes and peridynamics grids for the solution of model for small displacements: Deducing material static equilibrium problems. Mechanics Research parameters from classical linear theory. Technical Communications, 76:41 – 47, 2016. report, Institute for Numerical Simulation, 2016. [21] Christophe Geuzaine, Jean-François Remacle, et al. [8] Patrick Diehl. Emu-nodal discretization, May 2020. Gmsh: a three-dimensional finite element mesh gen- erator with built-in pre-and post-processing facili- [9] Patrick Diehl. A Fracture Multiscale Model for Peri- ties. International journal for numerical methods dynamic enrichment within the Partition of Unity in engineering, 79(11):1309–1331, 2009. Method: Part I, July 2021. [22] Fei Han and Gilles Lubineau. Coupling of nonlo- [10] Patrick Diehl, Michael Bußler, Dirk Pfluger,¨ Steffen cal and local continuum models by the arlequin ap- Frey, Thomas Ertl, Filip Sadlo, and Marc Alexander proach. International Journal for Numerical Meth- Schweitzer. Extraction of fragments and waves af- ods in Engineering, 89(6):671–685, 2012. ter impact damage in particle-based simulations. In Meshfree Methods for Partial Differential Equations [23] Klaus Iglberger, Georg Hager, Jan Treibig, and Ul- VIII, pages 17–34. Springer, 2017. rich Rude.¨ High performance smart expression tem- plate math libraries. In 2012 International Confer- [11] Patrick Diehl, Fabian Franzelin, Dirk Pfluger,¨ and ence on High Performance Computing & Simulation Georg C Ganzenmuller.¨ Bond-based peridynamics: (HPCS), pages 367–373. IEEE, 2012. a quantitative study of mode i crack opening. Inter- national Journal of Fracture, 201(2):157–170, 2016. [24] Ali Javili, Rico Morasata, Erkan Oterkus, and Selda Oterkus. Peridynamics review. Mathematics and [12] Patrick Diehl, Prashant K. Jha, Hartmut Kaiser, Mechanics of Solids, 24(11):3714–3739, 2019. Robert Lipton, and Martin Levesque.´ An asyn- chronous and task-based implementation of peridy- [25] Prashant K Jha and Robert Lipton. Numerical anal- namics utilizing hpx—the c++ standard library for ysis of nonlocal fracture models in holder space. parallelism and concurrency. SN Applied Sciences, SIAM Journal on Numerical Analysis, 56(2):906– 2(12):2144, Dec 2020. 941, 2018. [13] Patrick Diehl, Robert Lipton, Thomas Wick, and [26] Prashant Kumar Jha and Robert Lipton. Numeri- Mayank Tyagi. A comparative review of peridynam- cal convergence of finite difference approximations

18 A PREPRINT -AUGUST 6, 2021

(a) PUM (b) PUM

coarse δcoarse fine δ δcoarse = 0.002 and hPD = /4 δfine = 0.001 and hPD = fine/8

(e) Damage (f) Damage Figure 17: Magnitude of the displacement in the last time step of the inclined crack for a coarse nodal spacing (left) and a ﬁne nodal spacing (right).

19 A PREPRINT -AUGUST 6, 2021

for state based peridynamic fracture models. Com- [39] J. M. Melenk and I. Babuska.ˇ The partition of unity puter Methods in Applied Mechanics and Engineer- finite element method: basic theory and applica- ing, 351:184–225, 2019. tions. Computer methods in applied mechanics and [27] Prashant Kumar Jha and Robert Lipton. Finite ele- engineering, 139(1-4):289–314, 1996. ment convergence for state-based peridynamic frac- [40] N. Moes,¨ J. Dolbow, and T. Belytschko. A finite el- ture models. Communications on Applied Mathe- ement method for crack growth without remeshing. matics and Computation, 2:93–128, 2020. International journal for numerical methods in engineering, 46(1):131–150, 1999. [28] Prashant Kumar Jha and Robert Lipton. Kinetic rela- tions and local energy balance for lefm from a non- [41] Jaber Nikpayam and Mohammad Ali local peridynamic model. International Journal of Kouchakzadeh. A variable horizon method for Fracture, 226:81–95, 2020. coupling meshfree peridynamics to FEM. Computer Methods in Applied Mechanics and Engineering, [29] Hartmut Kaiser, Patrick Diehl, Adrian S. Lemoine, 355:308–322, 2019. Bryce Adelstein Lelbach, Parsa Amini, Agust´ın Berge, John Biddiscombe, Steven R. Brandt, Nikunj [42] Greta Ongaro, Pablo Seleson, Ugo Galvanetto, Tao Gupta, Thomas Heller, Kevin Huck, Zahra Khatami, Ni, and Mirco Zaccariotto. Overall equilibrium in Alireza Kheirkhahan, Auriane Reverdell, Shahrzad the coupling of peridynamics and classical contin- Shirzad, Mikael Simberg, Bibek Wagle, Weile Wei, uum mechanics. Computer Methods in Applied Me- and Tianyi Zhang. Hpx - the c++ standard library chanics and Engineering, 381:113515, 2021. for parallelism and concurrency. Journal of Open [43] Serge Prudhomme and Patrick Diehl. On the treat- Source Software, 5(53):2352, 2020. ment of boundary conditions for bond-based peridy- [30] Bahattin Kilic and Erdogan Madenci. Coupling of namic models. Computer Methods in Applied Me- peridynamic theory and the finite element method. chanics and Engineering, 372:113391, 2020. Journal of Mechanics of Materials and Structures, [44] Rezwanur Rahman, John T. Foster, and A. Haque. A 5(5):707–733, 2010. multiscale modeling scheme based on peridynamic [31] IA Kunin. Theory of elastic media with microstruc- theory. International Journal for Multiscale Com- ture. nonlocal theory of elasticity. Izdat.“Nauka,”” putational Engineering, 12(3):223–248, 2014. Moscow, 1975. [45] Rezwanur Rahman, John T Foster, and Anwarul Haque. A multiscale modeling scheme based on [32] Christopher J. Lammi and Min Zhou. Multi- peridynamic theory. International Journal for Mul- scale peridynamic modeling of dynamic frac- tiscale Computational Engineering, 12(3), 2014. ture in concrete. AIP Conference Proceedings, 1793(1):100009, 2017. [46] Radu Bogdan Rusu and Steve Cousins. 3D is here: Point Cloud Library (PCL). In IEEE International [33] Robert Lipton. Dynamic brittle fracture as a small Conference on Robotics and Automation (ICRA), horizon limit of peridynamics. Journal of Elasticity, pages 1–4, Shanghai, China, May 9-13 2011. 117(1):21–50, 2014. [47] M. A. Schweitzer. A Parallel Multilevel Partition of [34] Robert Lipton. Cohesive dynamics and brittle frac- Unity Method for Elliptic Partial Differential Equa- ture. Journal of Elasticity, 124(2):143–191, 2016. tions, volume 29 of Lecture Notes in Computational [35] Robert Lipton and Prashant K. Jha. Nonlocal elas- Science and Engineering. Springer, 2003. todynamics and fracture. Nonlinear Differential [48] M. A. Schweitzer. An Algebraic Treatment of Equations and Applications, 28(Article number:23), Essential Boundary Conditions in the Particle– 2021. Partition of Unity Method. SIAM Journal on Sci- [36] David John Littlewood, Stephen D Bond, and Tim- entific Computing, 31(2):1581–1602, 2009. othy Costa. Peridynamic multiscale finite ele- [49] M. A. Schweitzer and A. Ziegenhagel. Rapid ment methods. Technical report, Sandia National enriched simulation application development with Lab.(SNL-NM), Albuquerque, NM (United States), puma. In Scientific Computing and Algorithms in 2016. Industrial Simulations, pages 207–226. Springer, [37] Wenyang Liu and Jung-Wuk Hong. A coupling ap- 2017. proach of discretized peridynamics with finite ele- [50] M.A. Schweitzer. An Adaptive hp-Version of the ment method. Computer Methods in Applied Me- Multilevel Particle–Partition of Unity Method. Com- chanics and Engineering, 245:163–175, 2012. put. Methods Appl. Mech. Engrg., 198:1260–1272, [38] Erdogan Madenci, Atila Barut, Mehmet Dorduncu, 2009. and Nam D Phan. Coupling of peridynamics [51] Marc Alexander Schweitzer and Sa Wu. A mov- with finite elements without an overlap zone. In ing least squares approach to the construction of dis- 2018 AIAA/ASCE/AHS/ASC Structures, Structural continuous enrichment functions. In Singular Phe- Dynamics, and Materials Conference, page 1462, nomena and Scaling in Mathematical Models, pages 2018. 347–360. Springer, 2014.

20 A PREPRINT -AUGUST 6, 2021

[52] Pablo Seleson, Samir Beneddine, and Serge Prud- homme. A force-based coupling scheme for peridynamics and classical elasticity. Computational Ma- terials Science, 66:34–49, 2013. [53] Pablo Seleson, Youn Doh Ha, and Samir Bened- dine. Concurrent coupling of bond-based peridynamics and the navier equation of classical elasticity by blending. International Journal for Multiscale Computational Engineering, 13(2), 2015. [54] Pablo D Seleson. Peridynamic multiscale models for the mechanics of materials: constitutive rela- tions, upscaling from atomistic systems, and interface problems. PhD thesis, The Florida State Uni- versity, 2010. [55] A Shojaei, T Mudric, M Zaccariotto, and U Galvanetto. A coupled meshless ﬁnite point/peridynamic method for 2d dynamic fracture analysis. International Journal of Mechanical Sciences, 119:419–431, 2016. [56] SA Silling. Local-nonlocal coupling in Emu/PDMS. Sandia Report (SAND2020-11382), 2020. [57] Stewart Silling, David Littlewood, and Pablo Sele- son. Variable horizon in a peridynamic medium. Journal of Mechanics of Materials and Structures, 10(5):591–612, 2015. [58] Stewart A Silling and Ebrahim Askari. A meshfree method based on the peridynamic model of solid mechanics. Computers & structures, 83(17- 18):1526–1535, 2005. [59] Xiaochuan Tian and Qiang Du. Asymptotically compatible schemes and applications to robust discretization of nonlocal models. SIAM Journal on Numerical Analysis, 52(4):1641–1665, 2014. [60] Feifei Xu, Max Gunzburger, John Burkardt, and Qiang Du. A multiscale implementation based on adaptive mesh reﬁnement for the nonlocal peridynamics model in one dimension. Multiscale Model- ing & Simulation, 14(1):398–429, 2016. [61] Mirco Zaccariotto, Teo Mudric, Davide Tomasi, Ar- man Shojaei, and Ugo Galvanetto. Coupling of fem meshes with peridynamic grids. Computer Methods in Applied Mechanics and Engineering, 330:471– 497, 2018. [62] Mirco Zaccariotto, Davide Tomasi, and Ugo Gal- vanetto. An enhanced coupling of pd grids to fe meshes. Mechanics Research Communications, 84:125–135, 2017.