 Research article
 Open Access
 Published:
Enhanced numerical integration scheme based on imagecompression techniques: application to fictitious domain methods
Advanced Modeling and Simulation in Engineering Sciences volume 7, Article number: 21 (2020)
 The Correction to this article has been published in Advanced Modeling and Simulation in Engineering Sciences 2020 7:34
Abstract
In the present work, we propose a new approach, the socalled compressed adaptive integration scheme (CAIS), for the computation of the stiffness and mass matrices in fictitious domain methods requiring the integration of discontinuous functions. The novel approach extends the conventional quadtreedecompositionbased adaptive integration scheme (AIS) by an additional step, in which established imagecompression techniques are exploited to decrease the number of integration subcells. The benefits of the CAIS are manifold: First, the compression of the subcells inevitably leads to significant savings in terms of computational time required by the numerical integration. Second, the compression procedure, which is executed directly after the quadtreedecomposition algorithm, can be easily included in existing codes. Third, if applied to polynomial integrands, the CAIS yields exactly the same accuracy as the conventional AIS. Finally, the fourth advantage is seen in the fact that the CAIS can readily be combined with other approaches seeking a reduction of the number of integration points such as the BooleanFCM. The efficiency of the CAIS approach is presented in the context of the FCM based on Cartesian meshes applied to problems of linear elastostatics and modal analysis, while it is also suitable for the quadrature in other fictitious domain approaches, e.g., CutFEM and cgFEM.
Introduction
In engineering and physics one often deals with continuous field problems governed by partial differential equations (PDEs). These problems are often boundary value problems (BVP), where the primary variables have to satisfy a set of PDEs, while on the boundary of the domain of interest specific values of the field variables, formulated in form of boundary conditions (BCs), are prescribed [1]. Since these problems cannot be solved analytically in most of the cases, they have to be solved numerically.
Although the traditional FEM [2, 3] has proven its power throughout the years for solving BVPs, various extensions were developed in order to obtain an even higher accuracy, a lower computational time and/or a wider field of application. Such an extension is the fictitious domain approach (FDA) [4,5,6,7], also known as unfitted finite element method [8] or immersed boundary method [9], where the BVP is solved on an unfitted computational mesh independent from the original domain while maintaining the same robustness and accuracy as the FEM [10]. Such a formulation can be highly beneficial compared to the conventional FEM when it comes to automatic mesh generation, moving boundaries and simulation of complex media involving holes, inclusions and cracks. There are several possibilities of constructing an unfitted mesh, see Fig. 1 for details. One idea is to reuse an already existing FEM mesh and inserts extra features, such as voids as depicted in Fig. 1b and exemplarily discussed in [11]. Other approaches include embedding the original domain into a larger domain, which can be discretized by distorted or regular elements depending on the shape of the embedding domain, as shown in Fig. 1c, d, respectively. There is a vast variety of different methods that are based on the fundamental idea of the FDA, differing in their assumption regarding the topology of the unfitted mesh, the type of shape functions and their typical areas of application. However, it is worth to emphasize, that almost all of the different methods can adapt most of the features of the other ones. Well known methods based on the FDA are for example the extended finite element method (XFEM) [12,13,14,15] and generalized finite element method (GFEM) [15,16,17,18]; both of which are based on the partition of unity method (PUM) [19, 20] enabling a sophisticated tool for modelling problems with singularities, kinks, jumps or other nonsmooth features. An indepth mathematical foundation of the FDA and of the implementation of boundary conditions over the unfitted meshes can be found in the literature of the cut finite element method (CutFEM) [21,22,23,24], where often triangular elements are used. While the above approaches do not require the computational mesh to be rectilinear, the fixed grid finite element method (FGFEM) [25,26,27] assumes the mesh to be strictly Cartesian and uses a less accurate, yet very fast, integration scheme based on precomputation, making the FGFEM especially suitable for topology optimization [28, 29]. Similar to the FGFEM, the Cartesian grid finite element method (cgFEM) [30] is also restricted to rectilinear elements, however, it is combined with a hierarchic datastructure for adaptive hrefinement steered either by the curvature of the boundary or by appropriate error estimators. Finally, we mention the finite cell method (FCM) [31,32,33,34,35], which is a combination of highorder shape functions known from pFEM resulting in an exponential convergence [33] and unfitted (often Cartesian) meshes enabling a fast discretization of the computational domain [10, 31]. Please note that while the FCM can also be straightforwardly used in combination with distorted elements our focus in this article is on Cartesian meshes. A detailed discussion of the FCM, which is the basis of our implementation, will be provided in “The finite cell method for linear elastostatics” section.
However, regardless of the governing equations of the problem under consideration and of the chosen FDA, all of the above methods give rise to a non boundaryconforming mesh by definition, resulting in elements cut by the boundary. Consequently, compared to the FEM, the difficulty in these methods is shifted, on the one hand, to the implementation of the boundary conditions [4,5,6, 21, 22, 36,37,38], and on the other hand, to the computation of the cellmatrices. In this contribution, our focus is on the latter issue. A detailed discussion of different approaches that are currently used for the integration of the element matrices is provided in “Review of numerical integration methods: discontinuous integrands” section.
Review of numerical integration methods: discontinuous integrands
In order to illustrate the problem regarding the computation of the element matrices in the FDAbased methods, we consider a continuous polynomial function \(\mathcal {P}(\varvec{\xi })\) over a quadrilateral element \(\Omega ^{(\mathrm {e})} \in [1,1] \times [1,1]\) (Fig. 2a), which is divided by the boundary \(\partial \Omega \) intersecting the element into the arbitraryshaped disjoint physical \(\Omega _{\mathrm {phys}}\) and fictitious \(\Omega _{\mathrm {fict}}\) subdomains. Generally in the FDA, one is interested in the integral of \(\mathcal {P}(\varvec{\xi })\)^{Footnote 1} only over \(\Omega _{\mathrm {phys}}\), however, the integration itself should be performed over the entire domain \(\Omega ^{(\mathrm {e})}\). For eliminating the contribution of the integral over \(\Omega _{\mathrm {fict}}\), the indicator function \(\alpha (\varvec{\xi })\) is introduced, which is essentially a step function, having the value 1 in \(\Omega _{\mathrm {phys}}\) and the value 0 in \(\Omega _{\mathrm {fict}}\) (Fig. 2b). Then, the desired integral value I can be computed according to Eq. (1) [39], where the integrand on the right hand side is discontinuous (Fig. 2c) due to \(\alpha \).
The numerical computation of discontinuous integrals is an interesting topic with numerous applications which is still actively researched to all methods based on unfitted meshes. In the following, we briefly discuss the main approaches that are currently used in FDAs. Further details on these approaches can be found in Refs. [33, 40, 41]. Note that the discussed methods may originate in a specific field, nonetheless, they are generally applicable.
Precomputation
If the computational mesh consists of identical elements, certain parts of the integrals can be precomputed and reused while integrating over the individual cut elements. Both of the following two approaches assume a Cartesian mesh (see Fig. 1d).
It is a widely used practice in the FGFEM [25, 26, 29], that the term \(\int _{\Omega _{\mathrm {e}}} \mathcal {P}(\varvec{\xi })~\mathrm {d}\Omega \) on the right hand side of Eq. (1) is assumed to be the same for each element regardless of the topology of discontinuity and therefore, it can be easily precomputed. Then, for the computation of the discontinuous integral, the precomputed term can be reused for each cut element while scaled accordingly by the ratio \(A_{\mathrm {phys}}/A^{(\mathrm {e})}\), where \(A_{\mathrm {phys}}\) and \(A^{(\mathrm {e})}\) are the areas of the domains \(\Omega _{\mathrm {phys}}\) and \(\Omega ^{(\mathrm {e})}\), respectively. Clearly, such a basic computation of the discontinuous integral is related to serious errors, since it does not take the shape and position of \(\Omega _{\mathrm {phys}}\) into account, but only its area. Being aware of this, Maan et al. [26] used some additional equations based on the topology of the cut elements in order to determine a sufficient mesh density leading to smaller integration errors.
Another integration scheme based on the precomputation approach was proposed by Yang et al. [42,43,44], where one is interested in solving an integral given in Eq. (2). At this point only a generic discussion of their approach is given, independent of the area of application.^{Footnote 2}
Here, \(\Omega ^{(\mathrm {e})}\) is assumed to be composed of uniformly distributed rectangular regions as exemplarily depicted in Fig. 3a. In Eq. (2), \(\beta \) is a discontinuous function which is piecewise constant over the individual subregions, cf. Fig. 3b. Assuming that the distribution of subdomains is the same for all integration domains \(\Omega ^{(\mathrm {e})}\), the integral of \(\mathcal {P}\) can be precomputed over the individual domains. Then, the integral over a given element (see Fig. 3c) is computed by summing the precomputed integrals scaled by the corresponding value of the function \(\beta \), being constant for each subdomain.
Local integration mesh
Another idea is to introduce a local integration mesh (LIM) in the cut cells which serves integration purposes only and does not increase the degrees of freedom (DOF) of the global system. Then, the integral over the entire cell \(I^{(c)}\) is obtained by summing the integrals \(I^{(\mathrm {sc})}\) over the \(n_{\mathrm {sc}}\) integration subcells^{Footnote 3} constituting the LIM, while the subcells may or may not conform to the discontinuity within the cell.
Nonboundaryconforming LIM
In terms of a nonconforming LIM, Parvizian et al. [31] proposed a basic implementation for Eq. (3) by distributing \(n \times n\) subcells of uniform size in the cell, over which a low order integration, such as the trapezoidal rule can be used (Fig. 4a). A more sophisticated approach is the adaptive integration scheme (AIS), which generates an LIM by successively creating smaller and smaller subcells in the vicinity of the discontinuity by using spacepartitioning techniques such as the quadtreedecomposition (QTD) [32], yielding an integration mesh, exclusively consisting of squareshaped subcells (Fig. 4b). The BFCM approach by Abedian et al. [45] is very similar to the QTDbased AIS when applied to complex regions, however, due to its improved strategy based on Boolean operations and areacalculations of \(\Omega _{\mathrm {phys}}\) and \(\Omega _{\mathrm {fict}}\), it requires significantly fewer IPs while maintaining the same accuracy (see “Combination with the BFCM” section).
Boundaryconforming LIM
A more accurate resolution of the discontinuity is possible using boundaryconforming subcells. Such an approach is the adaptive anisotropic integration scheme proposed by Legrain and Moës [46], generating a mesh of distorted subcells with increased density and decreased size in the vicinity of the discontinuity. A significantly smaller number of subcells is required when using a Delaunay triangulation to create an integration mesh that consists of triangular subcells with straight edges, as it was done by Nadal et al. [30]. They also note that for higher order elements also a higher order boundary representation is required for an optimal convergence rate of the simulation. Therefore, the need for subcells with curved edges enabling a higher approximation accuracy of the boundary emerges. In the works by Cheng and Fries [14] and Fries and Omerović [47], the discontinuity \(\partial \Omega \) is assumed to be given in an implicit form by the isocontour \(\phi (\varvec{x}) = 0\) of the scalarvalued levelset function and for the mapping of the subcells Lagrangian elements with enriched nodes are used [48], enabling a more accurate representation of the boundary. While the approach discussed in Ref. [14] is restricted to twodimensional problems (Fig. 4c), an extension of the basic formulation to threedimensions is proposed in Ref. [47]. Kudela et al. [49, 50] follow the twodimensional approach of Ref. [14], however, they assume a parametric formulation of a twodimensional boundary and use the blending function method [3, 51] (Fig. 4c). Finally, the smart octree approach by Kudela et al. [52] is an extension of this idea to threedimensions, which is even capable of handling sharp geometric features on the interface, such as nodes and edges. In case of a parametric boundary, it is enough to find the intersection parameters by which the curve enters and leaves a cell. Finding implicitly defined boundaries renders a more challenging task, involving the solution of a nonlinear equation system (EQS), which can be achieved by the NewtonRaphson method for example [47].
Divergence theorem
Another approach is to utilize the divergence theorem (DT) [53, 54] in order to transform an integral over a two or threedimensional domain into closed surface or contour integrals, respectively, and thus reducing the dimensionality of the integration by one. The main problem one is faced with during this approach in the FEM, is that the antiderivative of the original integrand has to be computed. While Dasgupta [55] solves this integral symbolically, Sudhakar et al. [56] apply the DT on polynomial functions preintegrated over arbitrary polyhedra without using costly symbolic computations. Then, the integral over the triangular and quadrilateral facets is computed using Gaussian quadrature rules. Finally, Duczek et al. [57] apply the DT in the framework of the FCM and assume a predefined polynomial integrand, that can be (pre)computed symbolically. Note that this is a perfectly legitimate assumption in the context of the FCM, where the mesh consists of a regular arrangement of rectilinear cells with a constant Jacobian.
Moment fitting
A third approach is to use the moment fitting (MF) method, that derives an individual quadrature rule for an arbitrarily shaped domain, in our case \(\Omega _{\mathrm {phys}}\), by solving moment fitting EQS. Generally the positions of the quadrature nodes are not known a priori and therefore one has to solve a nonlinear EQS, in order to obtain both the nodal positions \(\varvec{x}_i\) and their corresponding weights \(w_i\). In cases where there are more nodes than moments, the solution can be obtained according to the leastsquares Newton’s method, as it was done by Mousavi et al. [58]. In addition, Xiao et al. [59] combine the leastsquares Newton’s method with a node elimination scheme, which results in a decreased number of IPs. However, these approaches become computationally expensive, if applied to integrals over numerous arbitrary domains, which is exactly the case when cells are intersected by arbitrary shaped domain boundaries [60, 61]. To circumvent this problem, a set of quadrature nodes with predefined positions can be used, resulting in a linear EQS, where only the weights are unknown. While Müller et al. [60] distribute the fixed IPs in the embedding domain \(\Omega _{\mathrm {e}} \supset \Omega _{\mathrm {phys}}\), Mousavi and Sukumar [62], Sudhakar et al. [63] and Joulaian et al. [61] use IPs located strictly within the domain of interest \(\Omega _{\mathrm {phys}}\). A major difference between the last three methods is their assumption of the geometry description. In Ref. [62], \(\Omega _{\mathrm {phys}}\) is assumed to be a convex polytope, while Ref. [63] computes the predefined nodal positions in both convex and concave polytopes. The MF method was applied by Hubrich et al. [64] and by Joulaian et al. [61] for threedimensional problems with curved surfaces approximated by a loworder triangular mesh and tested in the context of FCM. It was shown, that the number of required IPs in the MF is significantly lower, than in the AIS, while maintaining the same accuracy. However, on the contrary, the derivation of individual quadrature rules for every cut cell comes with a significant computational overhead, which makes the MF best suitable for nonlinear problems, where the already derived quadrature rules can be reused multiple times [61].
Equivalent polynomials
In the context of the equivalent polynomial (EP) approach, one seeks to solve a discontinuous integral in Eq. (1) over the embedding domain \(\Omega _{\mathrm {e}}\), by replacing the discontinuous term \(\alpha \) with an equivalent polynomial \(\mathcal {E}\), which is continuous over \(\Omega _{\mathrm {e}}\). Thus, it allows a standard Gaussian quadraturebased integration over the domain of interest without the need of a domain partitioning procedure [39, 65].
There are various studies available in the literature [39, 40, 65, 66], that use the EP approach. The key difference between the existing solutions lies in the formulation of the equivalent polynomial \(\mathcal {E}\), the required topology of the discontinuity and the computation of the integrals of the basis function over the physical domain. The works by Ventura [66] and by Ventura and Benvenuti [65], formulated in the context of the XFEM, assume elements intersected by a single straight line in two or by a single plane in threedimensional settings. On the one hand, Ref. [66] divides the integration region \(\Omega _{\mathrm {e}}\) into two subregions defined by the positive and negative domains according to the Heaviside function H making the integrand discontinuous, and integrates over the subdomains analytically using the DT. On the other hand, Ventura and Benvenuti [65] circumvent the need for subdivision of \(\Omega _{\mathrm {e}}\) by replacing H with its a priori known continuous regularized counterpart \(H_{\rho }\). Abedian et al. [40] applied the notions of Ref. [66] in the framework of the FCM and furthermore suggested a spacetreedecompositionbased improvement, where lower order equivalent polynomials can be used for each subcell. However, according to Ref. [39], in this case not only that an extra spacetreedecomposition is involved, but an EQS for each subcell has to be computed as well and therefore this approach does not pay off compared to the traditional QTDbased AIS. Finally, Abedian and Düster [39] describe both \(\mathcal {P}\) and \(\mathcal {E}\) on the basis of Legendre polynomials, whose orthogonality property enables an easy computation of the coefficients of the equivalent Legendre polynomial, without the cost of solving a system of equations.
Motivation
Clearly, there is a wide range of methods capable of computing discontinuous integrals, which all have their own fields of application. Nonetheless, their success is strongly restricted to specific scenarios regarding the shape of the physical domain and the geometry description of the boundary. Obviously, the DTbased methods yield the exact solution with marginal effort, if the integrand is known and if the boundary is defined as a polygonal chain in twodimension or as a triangulated surface in threedimension. Similarly, if the discontinuity happens to be given in a parametric form, an LIM with boundary conforming subcells can be easily created using the blending function method, however, for implicit boundaries, creating boundary conforming subcells renders a much more challenging task. This is also the case for the MF and EP methods, where one way or the other, an integral over an arbitrarily shaped physical domain has to be computed, while the MF even requires a solution of an additional EQS as well, for deriving the unique quadrature rules. For most cases, if the boundary is not given in the required form, additional iterations have to be performed to capture the boundary in a way, that can be handled by the given method. Not only that all these features increase the preparationtime of the integration techniques, but they also make them less robust.
In contrast to the previously mentioned methods, the LIMbased AIS relies on a simple and straightforward algorithm (“Traditional adaptive integration scheme (AIS)” section), which neither requires any iterations nor a solution of an EQS and yet is still capable of approximating the discontinuity independent of the geometry description of the boundary. Clearly, the main advantage of this approach is that it can be robustly used for a whole variety of problems. Although the AIS leaves more detailed information about the boundary unexploited and creates a rather coarse approximation of it, its fundamental idea aligns really well to the basic concept of the FCM, being an easy and effortless meshing combined with exponential convergence rates. Unfortunately, for an accurate solution, the AIS often results in a large number of IPs and thus lengthy computational times. Therefore, in this contribution we propose the compressed adaptive integration scheme (CAIS) (“Compressed adaptive integration scheme (CAIS)” section), in order to enable a method, that has all the benefits of the AIS, however, with reduced computational costs during the numerical integration.
The finite cell method for linear elastostatics
A detailed introduction to the fundamentals of the FCM was already elaborated in many contributions, such as Refs. [31,32,33,34,35]. Therefore, we only briefly sketch the fundamentals of the FCM. For the sake of simplicity, we restrict ourselves to twodimensional linear elastostatics, although the formulation of problems with governing equations of different problems, such as dynamics [67, 68], fluid mechanics [69] or fracture mechanics [70, 71], just to name a few, is also possible.
Note that the novel integration scheme being proposed in “Compressed adaptive integration scheme (CAIS)” section is not restricted to a specific FDA but can be utilized in any method with unfitted meshes in a straightforward manner. However, throughout this paper we present our approach in the framework of the FCM, where the CAIS can reach its maximum potential due to the higher order shape functions and the Cartesian meshes which are commonly used. Such a mesh generation goes handinhand with the voxelmodels obtained for computed tomography (CT) scans, thus, homogenization of highly complex structures [72,73,74] and applications to biomechanics [32, 36, 42,43,44, 70, 75] are major topics in the context of the FCM. Note that these are not the only fields where the capabilities of the FCM have been proven and exploited: multiphysics problems, such as thermo [37] and electromechanics [76], topology optimization [77], geometrically nonlinear problems. Refs. [78,79,80,81,82] as well as physically nonlinear problems [83,84,85] have also been simulated successfully. Besides static analyses, the extension and application of the FCM to dynamic problems [67, 68, 76, 86,87,88], to fluid mechanics [69] and last but not least, to phasefield modelling [70, 71] has been also in focus of current research activities with promising results. Note that besides the simple Cartesian mesh and hierarchic Legendre shape functions [33], also alternative mesh generation techniques, such as polygonal [89] and tetrahedral meshes [11, 70] as well as alternative shape functions, such as hierarchic Bsplines [78, 85] and Lagrangian shape functions based on the Gauss–Legendre–Lobatto (GLL) nodal distribution (leading to the spectral cell method—SCM) [67, 76, 86] have been also investigated and successfully implemented in the FCM.
Linear elastostatics
Consider a physical domain \(\Omega _{\mathrm {phys}}\) with its corresponding boundary \(\partial \Omega \). The domain is subjected to body forces \(\varvec{b}\) while on the Neumann boundary\(\Gamma _{\mathrm {N}} \subset \partial \Omega \) surface tractions \(\varvec{t}\) are applied. Then, without derivation, based on the principle of virtual work [2, 31] the weak formulation of the problem can be expressed by the equality of the bilinear \(\mathcal {B}(\varvec{u},\delta \varvec{u})\) and linear \(\mathcal {F}(\delta \varvec{u})\) functionals
where \(\varvec{u}\) is the displacement field, \(\delta \varvec{u}\) the virtual displacement field, \(\varvec{C}\) the constitutive matrix in Voigt notation and \(\varvec{L}\) the standard straindisplacement operator. In this case, the bilinear functional \(\mathcal {B}(\varvec{u},\delta \varvec{u})\) corresponds to the virtual internal energy, while the linear functional \(\mathcal {F}(\varvec{u})\) expresses the virtual work of the external loads. The boundary \(\partial \Omega = \Gamma _{\mathrm {N}} \bigcup \Gamma _{\mathrm {D}}\) is divided into the disjoint sets (\(\Gamma _{\mathrm {N}} \bigcap \Gamma _{\mathrm {D}} = \emptyset \)) of the Dirichlet \(\Gamma _{\mathrm {D}}\) and Neumann boundaries \(\Gamma _{\mathrm {N}}\), on which boundary conditions (BCs) for the BVP are applied. While the Neumann BC in regard of the prescribed tractions \(\varvec{\hat{t}}(\varvec{x})\), \(\forall \varvec{x} \in \Gamma _{\mathrm {N}}\) is already included in Eq. (5), the DirichletRB regarding the prescribed displacements \(\varvec{\hat{u}}(\varvec{x})\) have to be realized as an additional equation \(\varvec{u}(\varvec{x}) = \varvec{\hat{u}}(\varvec{x})\), \(\forall \varvec{x} \in \Gamma _{\mathrm {D}}\), in order to prevent rigid body motions.
Fictitious domain approach
In the FCM, the physical domain \(\Omega _{\mathrm {phys}}\) is embedded into the larger extending domain \(\Omega _{\mathrm {e}}\) of a simple and regular shape (Fig. 5), and thus the fictitious domain is defined as: \(\Omega _{\mathrm {fict}}=\Omega _{\mathrm {e}} \setminus \Omega _{\mathrm {phys}}\). In order to distinguish between physical and fictitious domains, we use the indicator function \(\alpha (\varvec{x})\) given in Eq. (6). To avoid a possible illconditioning of the system matrices, instead of setting \(\alpha = 0\) in \(\Omega _{\mathrm {fict}}\), we assign a small value \(\alpha<< 1\). In practice, one may use \(\alpha < 10^{10}\) [77]. However, according to Fries et al. [15], a possible illconditioning can be expected only in cases where cells exhibit a large difference between the areas of \(\Omega _{\mathrm {phys}}\) and \(\Omega _{\mathrm {fict}}\) within them.
Since the fictitious domain has no physical meaning, it should have no effect on the behaviour of physical domain. Therefore, the material in the fictitious domain should be significantly more compliant than that of the physical domain. Without derivation, the original problem stated in Eq. (5) can be also formulated in the extended domain using the indicator function \(\alpha \), as long as \(\alpha (\varvec{x})=0\) for all \(\varvec{x} \in \Omega _{\mathrm {fict}}\) holds [31]:
Discretization of the weak form
The advantage of the aforementioned embedding in \(\Omega _{\mathrm {e}}\) is manifold: On the one hand, the regular shape of \(\Omega _{\mathrm {e}}\) enables a straightforward and effortless discretization via Cartesian meshes (Fig. 5), which is of great advantage over the FEM, where automatic mesh generation can be challenging to achieve and in some cases, specifically in 3D problems, even impossible [2]. On the other hand, due to the quadratic/cubic shape of the resulting elements of the FCM (Fig. 5), the geometry mapping \(\varvec{Q}_{\varvec{\xi }\rightarrow \varvec{x}}\) from the reference to global space has a constant Jacobian matrix \(\varvec{J}_{\varvec{\xi }\rightarrow \varvec{x}}\). This is greatly beneficial regarding the computation of the element matrices, since the Ansatz space remains unchanged by the mapping and an exact numerical integration is possible, when polynomial shape functions are used. On the contrary, in the FEM, the elements can be distorted in various ways, inhibiting an exact numerical integration due to the mapping procedure and thus affecting the quality of the solution [90]. For sake of completeness, we mention the polygonal [89] as well as the tetrahedral extensions of the FCM [11, 70], where the above mentioned beneficial features of the traditional FCM regarding the simplicity of the mapping and the Jacobian matrix do not apply. Consequently, the polygonal and tetrahedral FCM are only advantageous for special applications, as for example the consideration of pores in already existing FEmeshes [91]. Furthermore, if the given problem contains nonsmooth features, introduced by material interfaces, cracks or sharp corners within the elements, special FCmeshes are required, based on the localenrichment [92, 93] or multilevel hprefinement strategy [75, 94].
Unlike in the FEM, the generated elements do not conform to the boundary \(\partial \Omega \). In order to distinguish them from the traditional finite elements (FEs), they are commonly referred to as finite cells (FCs). This notation is also used throughout this paper. Nevertheless, despite of this difference, the FCM is identical to the FEM in many important aspects: The FCs are still defined in the reference space with the corresponding local coordinates \(\varvec{\xi } = [\xi , \eta ]^T\), where \(\xi ,~\eta \in [1,1]\). Additionally, the displacement field \(\varvec{u}(\varvec{x})\) of the individual cells is approximated analogously to the conventional FEM using the same families of shape functions. Thus, the global equation system
whose solution is the global displacement vector \(\varvec{U}\), is obtained by the assembly of cellspecific stiffness matrices \(\varvec{K}^{(c)}\) and load vectors \(\varvec{F}^{(c)}\). Thus, already established solvers known from many FEpackages can be used straightforwardly [3, 95]. As an example, the stiffness matrix \(\varvec{K}^{(c)}\) of a specific cell is given below
where \(\varvec{B}\) is the socalled straindisplacement matrix, containing derivatives of the shape functions. Furthermore, \(\varvec{J}_{\varvec{\xi }\rightarrow \varvec{x}}\) is the Jacobian matrix of the already mentioned geometry mapping \(\varvec{Q}_{\varvec{\xi }\rightarrow \varvec{x}}\), establishing the relation between the local \(\varvec{\xi }\) and global \(\varvec{x}\) coordinates of a given cell, according to Eq. (10). Due to the simple shape of the cell in the global space, the mapping is a linear function, characterized by the global coordinates \(x_1\) and \(y_1\) of the lower left corner as well as by the height \(h_{\mathrm {x}}\) and width \(h_{\mathrm {y}}\) of the cell. It is easy to see, that in case of such a simple mapping, \(\mathrm {det}(\varvec{J}_{\varvec{\xi }\rightarrow \varvec{x}})\) is indeed constant and depends on the width \(h_{\mathrm {x}}\) and height \(h_{\mathrm {y}}\) of the cell only.
Due to the extension to \(\Omega _{\mathrm {e}}\) and the non boundaryconforming nature of the discretization, the generated cells can be classified^{Footnote 4} as follows:

1.
Physical cells are located entirely in the physical domain and their stiffness matrices \(\varvec{K}_{\mathrm {phys}}^{(c)}\) are identical to the ones of the FEs, due to the fact that \(\alpha =1\) in Eq. (9). Assuming a polynomial Ansatz space and constant material properties, \(\varvec{K}_{\mathrm {phys}}^{(c)}\) can be integrated via Gaussian quadrature to machine precision [88]. Furthermore, due to their uniform shape and size, a precomputation of \(\varvec{K}_{\mathrm {phys}}^{(c)}\) is possible, which then applies to all physical cells (see dark gray cells in Fig. 6) [25].

2.
Fictitious cells, which are located entirely in the fictitious domain, have no physical meaning and therefore, they can be excluded from the simulation (see light grey cells in Fig. 6).

3.
Cut cells,^{Footnote 5} as the name implies, are intersected by the boundary \(\partial \Omega \), and hence they contain pieces of both the physical and fictitious domains (see yellow cells in Fig. 6). The difficulty of computing accurate results for Eq. (9) in this case arises due to the discontinuous nature of \(\alpha \) within the cell (see Fig. 2), as discussed in “Review of numerical integration methods: discontinuous integrands” section.
In the FCM, significant drawbacks of the FEM, such as the typically demanding discretization procedure and integration error caused by the distorted elements in the initial mesh, can be avoided. However, due to the cut cells present in the FCM (and generally in all fictitious domain methods), the main challenge is shifted to the implementation of the BCs and to the computation of the cell matrices of cut cells [31, 33]. While an exhaustive introduction was given regarding the computation of cell matrices in “Introduction”, we only briefly mention the enforcement of the BCs in the FCM and refer the reader to the pertinent literature.
Inhomogeneous Neumann BCs can be implemented by considering the last term in Eq. (7), expressing an integral over the part of the boundary \(\Gamma _{\mathrm {N}} \subset \partial \Omega \), located within the cut cells. A detailed description of the implementation of such BCs for threedimensional triangulated surfaces can be found in [32]. On the other hand, homogeneous Neumann BCs require no special treatment, assuming that the fictitious domain has a material with zero stiffness.
Methods such as the penalty method, Lagrange multiplier method and Nitsche’s method can be used to enforce Dirichlet BCs in a weak sense by focusing on finding the stationary point of the original problem augmented by a constraint potential. Detailed explanations of these methods can be found in Refs. [4, 37, 38, 96].
Adaptive integration schemes
In “Motivation” section, after discussing the different approaches typically used in the fictitious domain concept for computing discontinuous integrands, the benefits of the AIS compared to the other methods were stated. In this section, the AIS is studied in greater detail (“Traditional adaptive integration scheme (AIS)” section), then, the CAIS approach is proposed and investigated with regard to the reduction of the computational costs compared to the AIS (“Compressed adaptive integration scheme (CAIS)” section). Finally, in “Combination with other approaches” section, the performance of the CAIS is further enhanced by combining it with other methods discussed in the literature.
Traditional adaptive integration scheme (AIS)
The computation of cellmatrices of the broken cells is challenging, since it involves discontinuous integrands, where a strong discontinuity is introduced by the indicator function \(\alpha \) (“Review of numerical integration methods: discontinuous integrands” section). At this point we recall the general form of such discontinuous integrals given in Eq. (1), while a specific example in form of the cell stiffness matrix is given in Eq. (9). Since the aim of this work is the improvement of the computational speed of the AIS, let us briefly outline its procedure. In the traditional AIS [32], the LIM (see “Local integration mesh” section) is created in twodimensional cases via a quadtreedecomposition (QTD), which is based on the recursive subdivision of the cut subcells in four equal sized quadrants. This procedure continues until a refinement level k defined by the user is reached. In the end, the QTD produces a nonuniform integration mesh, consisting of quadratic subcells, with decreased size and increased density in the vicinity of the discontinuity (Fig. 7) [78, 97]. Similar to the classification of cells, the resulting subcells as well are classified as physical (dark grey), fictitious (light grey) and cut/broken (yellow). At this point we emphasize again, that the created mesh serves integration purposes only and it does not introduce new DOFs to the global system. The equivalent threedimensional version of the QTD is the octreedecomposition (OTD), during which every cut subcell divides into eight equal sized cubeshaped octants.
Now, in terms of the stiffness matrix \(\varvec{K}^{(c)}\), the AIS being based on an LIM (“Local integration mesh” section), is computed according to Eq. (3) over the \(n_{\mathrm {sc}}\) subcells as follows
where the stiffness matrix \(\varvec{K}^{(c,sc)}\) of a particular subcell sc in a given cell c is
Note that the above integral is computed in the local coordinates \(\varvec{r} = [r,s]^T\) of the subcells, hence the integrand in Eq. (9) is extended by the term \(\det (\varvec{J}^{(sc)}_{\varvec{r}\rightarrow \varvec{\xi }})\), where \(\varvec{J}^{(sc)}_{\varvec{r}\rightarrow \varvec{\xi }}\) is the Jacobian of the geometry mapping \(\varvec{\xi } = \varvec{Q}_{\mathrm {\varvec{r}\rightarrow \varvec{\xi }}}(\varvec{r})\), defining the position of a subcell in \(\varvec{\xi }\) (see Fig. 8) [32]. Again, due to the simple shape of the subcells,^{Footnote 6} similar to Eq. (10), the mapping is linear and has a constant Jacobian. In Eq. (13), \(\xi _1\) and \(\eta _1\) refer to the lower left corner of a given subcell, while \(h_{\xi }\) and \(h_{\eta }\) are its height and width measured in the cell (Fig. 8).
In the AIS, the integration over the individual subcells is computed via Gaussian quadrature, as depicted in Fig. 8, where the values of the integrand at the IPs in the physical and fictitious domains are scaled by \(\alpha \) according to Eq. (6). The resolution of the mesh and therefore the accuracy of the integration is determined by the refinement level k. Due to \(\alpha \) being constant within the physical and fictitious subcells, applying Gaussian quadrature rules to the subcells for computing Eq. (12) yields exact results, assuming \(\varvec{B}^T \varvec{C} \varvec{B}\) is a polynomial function. However, the application of Gaussian quadrature rules to the cut cells results in significant errors as it is applied to nonpolynomial functions, in this case discontinuous functions. Since the boundary in the cut subcells is captured by the IPs, the integration error can be decreased when using sufficiently small subcells and/or a highorder Gaussian quadrature rule [77]. While using the first approach the total number of IPs increases exponentially with \(\mathcal {O}(2^k)\), the latter method yields a quadratic increase \(\mathcal {O}(n_{\mathrm {dir}}^2)\), where \(n_{\mathrm {dir}}\) refers to the number of IPs in the \(\xi \) and \(\eta \)directions for each subcell. Note that despite these challenges it has been mathematically shown that if the integration error is below the discretization error, i.e., the discretization error dominates the overall numerical error, the theoretical optimal convergence known from the FEM is also attained in the FCM [10].
The main advantage of the AIS lies in its simplicity and robustness. While the former attribute enables a straightforward implementation, the latter one results from its capability of handling any kind of geometry representation by requiring insideoutside tests^{Footnote 7} only. If two points in the same cell belonging to two different domains are found, the cell must be cut. Its weaknesses lie in its rather poor boundary approximation and in the exponentially increasing rate of \(n_{\mathrm {sc}}\) for incrementally increasing k. Thus, not only the time required by the QTD grows exponentially, but also the number of the IPs when integrating over the subcells, leading to extreme computation costs for the AIS [40, 52, 57].
Compressed adaptive integration scheme (CAIS)
At the end of “Traditional adaptive integration scheme (AIS)” section, we stated the drawbacks of the AIS. In order to reduce the computational effort needed for an accurate numerical integration, different approaches have been developed [38, 40, 45, 83, 84], which aim at the reduction of the IPs during the AIS. In this section we present a new approach, the compressed adaptive integration scheme (CAIS), which seeks the reduction of the IPs by applying imagecompression techniques on the set of \(n_{\mathrm {sc}}\) subcells \(\varvec{\mathcal {S}}=\{\mathcal {S}_i \}_{i=1}^{n_{\mathrm {sc}}}\) resulting from the QTD. An example of a cut cell as well as its QTD are depicted in Fig. 9a, b, respectively. The compression itself is realized as an additional step being executed after the QTD, resulting in a new set of nonoverlapping rectangular subcells \(\varvec{\mathcal {R}}=\{\mathcal {R}_j \}_{i=1}^{n_{\mathrm {R}}}\), where \(n_{\mathrm {R}} \le n_{\mathrm {sc}}\), as depicted in Fig. 9c. This also implies, that the QTD is still required. Finally, the integration itself is then carried out over \(\varvec{\mathcal {R}}\) via Gaussian quadrature, analogous to the standard AIS, while maintaining the same accuracy. Consequently, the compression step can be straightforwardly implemented as an intermediate step between the QTD and the actual numerical integration, without the need for major modifications in the already existing codes. Note that the CAIS is only concerned with the number and shape of the subcells, and it is independent of the family of shape functions or nodal distributions that are employed for the approximation of the solution field.
Shape of the subcells
Due to the rectangular shape of the subcells in \(\varvec{\mathcal {R}}\) (see for example Fig. 9c), the determinant of the Jacobian matrix \(\varvec{J}_{\varvec{r}\rightarrow \varvec{\xi }}\) of the geometry mapping \(\varvec{Q}_{\varvec{r}\rightarrow \varvec{\xi }}\) is still constant (see Eq. (13) for \(h_{\xi } \ne h_{\eta }\)), enabling a simple and exact integration over \(\varvec{\mathcal {R}}\) via Gaussian quadrature, if a polynomial Ansatz space is used. Note that an exact integration is also possible for subcells with curved edges, if and only if the mapping \(\varvec{Q}_{\varvec{r}\rightarrow \varvec{\xi }}\) is polynomial—otherwise an exact numerical integration is not possible. Nonetheless curved subcells require an increased number of IPs for the numerical integration, while for square and rectangularshaped subcells no additional IPs are needed. Furthermore, it is well known, that among other parameters, the aspect ratio of twodimensional FEs has an effect on the convergence properties of the solution [3]. While allowing elongated rectangular subcells, they can have extreme aspect ratios (see for example Fig. 9c), we would like to emphasize, that the created subcells serve for integration purposes only and regardless of the aspect ratios, the parent cell, which holds actual DOFs, is not affected in its shape. In conclusion, when using rectangular subcells, an exact integration via Gaussian quadrature is still possible, without the need for any additional IPs and without affecting the convergence of the solution.
Accuracy
The compression should only affect the speed of the AIS but not its accuracy and therefore, the equality regarding the integral value I over a given cell
has to be demanded. The above equation holds, if (i) a lossless compression is used and if (ii) the compression targets only the physical and fictitious subcells, over which the integrand is constant, while the cut subcells are left unchanged. This can be explained by the following reasoning. The boundary is resolved by the IPs in the cut subcells, i.e., the IPs are partly located in the physical domain, while the rest is distributed in the fictitious one. Although the Gaussian quadrature is not suitable at all for the discontinuous integrals over the cut subcells, by using small enough subcells along the boundary the problematic area can be restricted to a negligible domain and the error can be kept below a certain threshold as already described in Sec. 5.1. The smaller the cut subcells along the boundary and the denser the distributed IPs are, the smaller the error becomes. Since for an accurate integration it is crucial to have enough IPs capturing the discontinuity consequently, it is not recommended to compress the cut subcells.
Imagecompression techniques
Compressing the subcells in \(\varvec{\mathcal {S}}\) (see for example Fig. 10a) by merging adjacent subcells into larger blocks is a challenging task due to the nonuniform sizes of the subcells. Instead, the set \(\varvec{\mathcal {S}}\) will be transformed by the function T to an equivalent pixelimage \(\mathcal {I}\), where the individual pixels covered by the fictitious, physical and cut subcells are assigned by the pixelvalues 0, 1 und 2, respectively (see Fig. 10b). This and the following steps of the proposed approach are sketched in the flowchart in Fig. 11. There, also the used compression techniques are listed (RLE, IBR and MRP) that are discussed in “Compression techniques” section. Note that the QTD is not only a space partitioning method, but also an imagecompression technique [98], thus, \(\varvec{\mathcal {S}}\) can be seen as a compressed version of \(\mathcal {I}\) and applying T means no more, than reconstructing\(\mathcal {I}\) from its compressed state. Clearly, T can be performed with negligible effort. The image should not contain more information than necessary and therefore its resolution \(2^k \times 2^k\) is defined by the refinement level k of the QTD. Thus, the smallest subcell in \(\varvec{\mathcal {S}}\) corresponds to a single pixel in \(\mathcal {I}\). Once \(\mathcal {I}\) is created, other compression techniques can be applied to the pixelregions defined by the values 1 and 0. Then, the results are transformed from the pixelspace back to the localspace of the cell by the function \(T^{1}\), in order to obtain the position of the subcells \(\varvec{\mathcal {R}}\) in \(\varvec{\xi }\). Since both the QTD and the used compression methods represent lossless compression techniques, the pixelimage can be restored and recompressed without any error, which is a major requirement for Eq. (14) to hold.
It is important to emphasize, that since we are dealing with imagecompression, most likely all methods would significantly reduce the number of pixels in the image. However, our goal is not only to compress the image, but to apply a compression that outperforms the QTD in terms of the number of the resulting subcells.
Alternative solutions for deriving the pixelimage
The question naturally arises, whether the need for the costly QTD could be eliminated by either (i) directly creating the pixelimage from the topology of the given cut cell or—which would be even better—(ii) by utilizing the other compression techniques for creating the local mesh directly from the topology of the cut cell, thus additionally eliminating the steps of the reconstruction as well as of the compression.
The first option is possible when we distribute \(2^k \times 2^k\) equalsized subcells in an isotropic manner in the cell and evaluate whether they belong to the physical, fictitious or cut domains. Note that in this case k is no longer the refinement level, only an integer defining the sidelength of all the subcells by \(l=2/(2^k)\). In this case, there is a onetoone correspondence between the subcells and the pixels of the image. Matter of fact, such a local mesh was already proposed by Parvizian et al. [31], however they have directly used the mesh for the numerical integration by the trapezoidal rule. According to our tests on implicitly defined geometries, it is much more timeefficient to use the QTD rather than an isotropic distribution of subcells, as it increases the density of subcells only along the boundary but not in the entire cell. Consequently, less time is spent on determining the status of the subcells, i.e., whether they are physical, fictitious or cut. Therefore, even if the QTD requires an additional step of reconstructing the pixelimage using T, it is still a computationally more efficient solution. Constructing a bounding box around the discontinuity shrinks the domain in which the subcells (pixels) have to be checked. However, the resulting reduction in computational effort is heavily dependent on the topology of the discontinuity. The method based on the use isotropic subcells only becomes efficient when the insideoutside tests have to be ran on a small portion of the subcells, which is for example the case when the boundary \(\partial \Omega \) is defined by straight lines or by planar surfaces. Note that one may modify the QTD in such a way, that it already yields \(\mathcal {I}\) instead of \(\varvec{\mathcal {S}}\). However, in this case, the transformation T is not eliminated, it is just integrated in the QTD.
Unfortunately there is no record or evidence for the second option to be computationally cheaper or even possible. Although other compression techniques such as the RLE, the IBR and the MRP (“Compression techniques” section) generally yield a higher compression than the QTD, they are rather unsuitable for the partitioning of the cell and for iteratively capturing the boundary.
In conclusion, for the construction of the initial mesh, the QTD seems to be the best candidate combining the abilities of spacepartitioning and datacompression. While the first property is required for the creation of an initial local mesh independent from the geometry description of the boundary and from the connectivity of the physical and fictitious domains in the cell, the second feature enables the reconstruction of the pixelimage. Note that having an initial LIM, where the subcells can take any position and form within the cell, could not be transformed to a corresponding pixelimage \(\mathcal {I}\) and would make the process of compressing the subcells significantly harder.
Performance
We have to stress again that the main goal is not only to achieve a high compression of \(\mathcal {I}\), but to outperform the QTD in terms of the number IPs. Therefore, it is reasonable to evaluate the results with respect to the QTD. In the following, let us use the compression ratio of the number of the IPs
where the function \(n_{\mathrm {IP}}(\cdot )\) expresses the number of IPs for a given set of subcells, being deployed for the numerical integration. In that sense, the smaller the ratio \(\lambda \) (for \(\lambda <1\)), the better the compression, while \(\lambda > 1\) indicates an undesired scenario, where the number IPs is increased compared the QTD (see “Drawback of the RLE compression technique” section for such a scenario). In the framework of the FCM combined with the AIS, the complexity of \(\mathcal {I}\) depends (i) on the topological complexity of the physical boundary \(\partial \Omega \) within a given cell and (ii) on the refinement level k of the QTD. Clearly, if the adjacent pixels are likely to have different values (= highly complex boundary) or if the image consists of a few pixels only (= low k), \(\lambda \) is less significant. Note that in the context of the FCM, an accurate and efficient solution can only be achieved if the discretization and integration errors are balanced. This generally requires FCs that are sufficiently small such that the boundary variation within one cell is not too complex (the smooth shape functions need to be capable of capturing the solution fields) and additionally the refinement level k has to be high enough to ensure an accurate numerical integration. Therefore, the number of integration subcells is typically rather high such that compression algorithms can produce meaningful results, i.e., a significantly decreased number of integration domains. Similar to \(\lambda \), let us define the time compression ratio
where the function \(t(\cdot ) = t_{\mathrm {LIM}}(\cdot )+t_{\mathrm {I}}(\cdot )\) expresses the total time needed for the chosen algorithm, including both the time for creating the local mesh \(t_{\mathrm {LIM}}(\cdot )\) as well as the time required for the integration \(t_{\mathrm {I}}(\cdot )\) over the set of resulting subcells. Both the CAIS and AIS are based on a QTD requiring \(t_{\mathrm {QTD}}\) time, however, the CAIS involves an additional step of compression requiring \(t_{\mathrm {C}}\) time, i.e., for the CAIS \(t_{\mathrm {LIM}}=t_{\mathrm {QTD}}+t_{\mathrm {C}}\). Thus, it is obvious, that \(t_{\mathrm {LIM}}(\textit{CAIS}) \ge t_{\mathrm {LIM}}(AIS)\). On the other hand, due to the compression a reduced integration time \(t_{\mathrm {I}}(\varvec{\mathcal {R}}) \le t_{\mathrm {I}}(\varvec{\mathcal {S}})\) is obtained. It is clear, that as long as the following inequality holds \(t_{\mathrm {C}} < t_{\mathrm {I}}(\varvec{\mathcal {S}})t_{\mathrm {I}}(\varvec{\mathcal {R}})\), a desired overall time compression ratio \(\tau < 1\) can be achieved. An undesired scenario is indicated by \(\tau > 1\), i.e., the computational time invested in the compression is greater than the time saved when integrating over the compressed subcells.
Compression techniques
Runlength encoding (RLE)
The RLE (Fig. 12a) is one of the oldest and most straightforward approaches among the traditional imagecompression techniques. It is based on scanning the pixelimage and replacing constant valued pixel bands of unit width by their start and endpixels as well as by the pixelvalue being constant throughout the band. Considering imagecompression, one may use different scanning schemes, such as rowwise, columnswise or zigzag. In any case, the algorithm checks every single pixel in \(\mathcal {I}\) and thus its time complexity is proportional to \(\mathcal {O}(n^2)\), where the resolution of the image is \(n \times n\) pixels. Obviously, a low complexity of an image results in an improved compression [98, 99]. In the framework of the FCM, the row and columnwise scanning directions are of main concern, resulting in vertical or horizontal elongated rectangular shaped subcells \(\varvec{\mathcal {R}}\) after the transformation \(T^{1}\). The implemented code performs both a row and columnwise RLE compression of \(\mathcal {I}\) and chooses the one with the fewer number of resulting rectangles. Since the RLE performs pixelchecking operations only, it can easily compress images with multiply connected regions, i.e., domains that may enclose holeregions (see for example physical pixels in Fig. 10b) and/or consist of several regions isolated from each other.
Image block representation (IBR)
The IBR (Fig. 12b) is a very simple yet powerful extension of the RLE. It is based on a primary RLE, which is then followed by merging adjacent rectangles into larger blocks, if they have the same start and end positions. In case of a columnwise RLE, a horizontal merging procedure is invoked during the IBR, while considering a rowwise RLE, a vertical merging technique is applied. Whether the results of the horizontal or the rowwise RLE are passed on to the IBR, depends on which one of the both approaches yields less rectangular domains. As the RLE, the IBR is based solely on pixel checking operations and can be performed with marginal additional effort compared to the RLE, however, it yields significantly better results [98, 100]. Furthermore, similar to the RLE, the IBR can easily handle multiply connected regions. As outlined in “Comparison” section, while the condition \(n_{\mathrm {R}} \le n_{\mathrm {sc}}\) is not always fulfilled in certain cases for the RLE, the IBR is not prone to such artifacts and reliably yields a compression rate of the IPs \(\lambda \) below 1.
Minimal rectangular partition (MRP)
The methods presented in the previous two paragraphs significantly reduce the number of subcells and in most cases they perform considerably better than the QTD alone. Nevertheless, the fewest number of subcells results from the MRP, which guarantees a minimal partitioning into a set of nonoverlapping rectangles for twodimensional rectilinear polygons. Such a minimal partitioning can be achieved by translating the problem into the graphtheoretic realm and solving it on an equivalent bipartite graph, hence in some papers it is also referred to as a graphbased decomposition (GBD) [98, 101]. In the following, we stick to the term MRP, as it better articulates the focus of this method. Finding the minimum partition is typically an optimization problem and as such it comes at a cost. The MRP algorithm generally has a much higher complexity and slower performance compared to the simple RLE and IBR algorithms [98]. Since for the MRP a rectilinear input polygon is needed, the procedure starts with a boundary tracing [102,103,104] of the physical and fictitious pixel domains, realizing the transition from the pixelspace to the localspace, where then the MRP can be performed (Fig. 11). Since the MRP performs the partition in the geometric space, multiply connected regions require an even more complex algorithm. Since a detailed introduction to the MRP is beyond the scope of this work, for general information about the method we recommend Refs. [98, 101, 105], while for its application to the compression of integration subcells in the framework of the FCM we refer to Petö et al. [106]. For further reading about the minimal partitioning/covering problems, the interested reader is referred to Refs. [107,108,109,110,111]. Regarding graphtheoretic tools, we recommend Ref. [112].
Comparison
A visual comparison of \(\varvec{\mathcal {R}}\) obtained by the different methods when applied to Fig 10a is depicted in Fig. 12. A further investigation of the compression techniques regarding their performance as a function of the refinement level k is shown in Fig. 13. The domain for which a local integration mesh should be created is depicted in Fig. 13a, where \(\Omega _{\mathrm {fict}}\) is defined by the implicit function given in Eq. (17).
Figure 13b, depicts the \(\lambda (k)\)curves of the examined methods, which all converge to stationary values, expressing the best compression reachable by the different methods. The curves are in a good agreement with Fig. 12 as well, although the RLE shows somewhat different results for the two different domains, due to the phenomenon discussed later in “Drawback of the RLE compression technique” section. Despite the fact, that the MRP yields a minimal partition, it comes with the cost that it requires approximately one order of magnitude more time for the compression \(t_{\mathrm {C}}\) than the RLE and IBR, as shown in Fig. 13c. Nevertheless, \(t_{\mathrm {C}}\) of the different methods for reasonable refinement levels (\(k\ge 5\)) is negligible compared to \(t_{\mathrm {QTD}}\), which is unconditionally included in the CAIS for all cases. This can be also seen in Fig. 13d, where the dominance of \(t_{\mathrm {QTD}}\) is clear due to the convergence of the curves to the value 1. This is to say, the compression increases the time for the construction of the LIM only by a negligible amount.
The difference of the IBR and MRP methods becomes more present in case of highly complex domains. However, for reasonable discretizations in the context of the FCM the topology of discontinuity in the cut cells is less complex and typically it intersects only 2 edges of the cell. In such cases, the IBR and MRP both yield very similar or even identical results [98]. Furthermore, the difference in \(\lambda \), if present, vanishes for almost all cases, if the IBR and MRP compression techniques are combined with other methods also aiming at a reduction of the IPs. However, it is important to note, that the MRP is based on a complex algorithm that is significantly more cumbersome to implement and timeconsuming to execute than the one of the IBR. Since the superiority of the IBR over the RLE is also clearly demonstrated (see “Drawback of the RLE compression technique” section), we recommend the IBR approach for compressing the subcells \(\varvec{\mathcal {S}}\) resulting from the QTD.
Drawback of the RLE compression technique
Finally, we discuss a special case, where instead of reducing the number of subcells of the QTD, the RLE actually increases their number. This is a great example for illustrating that it is not enough to compress the pixelimage, but the resulting rectangular subcells \(\varvec{\mathcal {R}}\) should definitely have a lower cardinality^{Footnote 8} than the original set of subcells \(\varvec{\mathcal {S}}\) of the QTD. Figure 14a depicts the QTD with \(k=4\) of a cell being cut in the vicinity of its topright corner, resulting in a significant difference between the areas of the physical and fictitious domains within the cell. The corresponding pixelimage is shown in Fig. 14b. Although the rowwise RLE compression of the image significantly reduces the number of pixels, with its \(\varvec{\mathcal {R}}=20\) subcells compared to the QTD having only \(\varvec{\mathcal {S}}=13\) subcells,^{Footnote 9} it results in an unacceptable value of \(\lambda = 20/13 = 1.538 > 1\). Applying the IBR to the same problem avoids this negative artifact and yields a reliable compression with \(\lambda <1\). Taking (i) this robustness of the IBR into account, (ii) its generally better compression (Fig. 13b) as well as (iii) its low additional time costs (Fig. 13c, d), it is definitely superior to the RLE.
Combination with other approaches
Several approaches have been published in the wide body of literature aiming at the reduction of the IPs during the AIS. It is straightforwardly possible to combine the compression techniques outlined above with these already existing approaches, in order to achieve an even better reduction of the IPs.
Combination with the reduction of fictitious integration points approach
In order to decrease the number of IPs in the numerical integration Abedian et al. [83, 84] propose an approach consisting of two key features. The first feature is based on the idea of decreasing the number of integration points (DIP),^{Footnote 10}, which exploits the fact, that the integrand over the smaller subcells can be seen as a lower order function for which a lower order Gaussian quadrature rule suffices. In other words, the number of IPs per subcell is not only chosen depending on the polynomial degree of the shape functions but also on the size of the subcell compared to the original cell. The second feature of their approach, which we directly apply to the integration over the compressed subcells \(\varvec{\mathcal {R}}\), is based on the reduction of the fictitious integration points (RFIP)\(^{10}\). Thus, the integration over a cut cell is executed in two distinct steps:

1.
Integrating over the physical and cut subcells, while taking only the physical IPs in the cut subcells into account (see Fig. 15a).

2.
The contribution of the fictitious domain needs to be considered, in order to avoid adverse effects resulting from illconditioning of the system matrices. To this end, a set of IPs is distributed according to the conventional Gaussian quadrature rule in the original cell \(\Omega ^{(c)}\) and only the points located in the fictitious domain are considered during the numerical integration (see Fig. 15b).
As a result of this algorithm, fictitious subcells can be discarded and hence, only the physical subcells are taken into account during the compression, slightly reducing the time \(t_{\mathrm {C}}\). In the remainder of this work, if the RFIP extension is used, it is indicated by an asterix \(\square ^{*}\). Although not discussed in this contribution, the RFIP can be easily employed in conjunction with the RLE and MRP techniques as well [106].
Combination with the BFCM
The BFCM approach by Abedian et al. [45] extends the original FCM by Boolean operations for an improved integration efficiency. Although the BFCM yields exact results for simple regions for which quadrature formulae exist, the practical use of it lies in its combination with a spacetreedecomposition, in order to handle cut cells with arbitrary topology. In this case, the BFCM yields exactly the same results as the QTDbased FCM, however, with significantly fewer IPs. In the following, we focus on this latter approach of the BFCM. The construction of the local mesh is governed by additional features^{Footnote 11} in such a way, that the overall IPcount is reduced. For achieving this, generally overlapping subcells are created, whose Boolean meta information by the labels \(\alpha _{\mathrm {B}} =1\) and \(\alpha _{\mathrm {B}} =1\) are utilized during numerical integration. For more in depth introduction please see the original work [45]. Throughout the current contribution we refer to such a way of constructing the LIM as BQTD, while the term BFCM is used for the entire simulation, including both the BQTD and the corresponding numerical integration. The set of subcells resulting from the BQTD is denoted by \(\varvec{\mathcal {B}}\). Due to the fact that the BFCM yields the same results as the FCM but with a reduced number of IPs, the BFCM can be directly compared in terms of \(\lambda \) and \(\tau \) to the CAIS, although its not based on any compression technique.
It turns out, that the BFCM performs equally well compared to the CAIS, however, its performance can be further improved when combined with the compression techniques. Before outlining the compression in the BFCM, let us briefly recall the basic ideas of the BQTD based on a simple example depicted in Fig. 16, where \(k_{\mathrm {max}}=5\) is used. On the levels \(k < k_{\mathrm {max}}\) we use the green and red colors to depict the positive and negative subcells, corresponding to the labels \(\alpha _{\mathrm {B}} = 1\) and \(\alpha _{\mathrm {B}} = 1\), respectively. Let us denote the sets of these subcells by \(\varvec{\mathcal {B}}_{\mathrm {full}}^+\) and \(\varvec{\mathcal {B}}_{\mathrm {full}}^\). Note that these types of subcells are all cut by the boundary, however, during the numerical integration, they will be treated as intact ones free of any discontinuity. The cut subcells, for which the discontinuity will be taken into account during the numerical integration, are located at the final level \(k = k_{\mathrm {max}}\) of the BQTD. Here, the blue color represents an integration with \(\alpha = 1\) in the physical domain, while the orange color means an integration with \(\alpha = 1\) in the fictitious domain. In both cases, the IPs in the other domain are ignored. For these sets we use the labels \(\varvec{\mathcal {B}}_{\mathrm {cut}}^+\) and \(\varvec{\mathcal {B}}_{\mathrm {cut}}^\), respectively. Note that the set of cut subcells \(\varvec{\mathcal {B}}_{\mathrm {cut}}^+\) \(\cup \) \(\varvec{\mathcal {B}}_{\mathrm {cut}}^\) of the BQTD is exactly the same as the one of the QTD, however, in case of the BQTD, the domain of interest of the integration as well as the \(\alpha \)values of the IPs vary from subcell to subcell. Generally, it is clear that
and that none of these sets have any subcells in common. For the colorcoding in the above equation, please see Fig. 16. Finally, note that the BFCM is constructed in such way, that for an integration over a given cell that does not lead to an illconditioned stiffness matrix, the contribution of the fictitious domain in the cell still has to be taken into account. This is realized in the BFCM by utilizing the RFIP approach, discussed in “Combination with the reduction of fictitious integration points approach” section.
Compression steps
Now, a compression can be performed on \(\varvec{\mathcal {B}}_{\mathrm {full}}^+\) and \(\varvec{\mathcal {B}}_{\mathrm {full}}^\), since they both represent subdomains with continuous integrands.^{Footnote 12} On the other hand, just like before, we do not compress the cut subcells, in order to preserve the accuracy of the numerical integration. Compressing the subcells on every refinement level is not reasonable or sometimes not even possible, since they often do not form coherent regions (see for example Figs. 16 and 18). Instead, we execute the followings steps, also depicted Fig. 17:

1.
Transformation of the individual layers \(k_0, k_1,...,k_{\mathrm {max}}1\) by T to the pixelimages \(\mathcal {I}_0, \mathcal {I}_1,...,\mathcal {I}_{k_{\mathrm {max}}1}\). While the subcells of \(\varvec{\mathcal {B}}_{\mathrm {full}}^+\) mark pixel regions with the value 1, the subcells of \(\varvec{\mathcal {B}}_{\mathrm {full}}^\) result in regions with the value \(1\). It is important, that each image has the same resolution, which should be defined by the size of the smallest subcell participating in the compression, just as before. Consequently, the images should have a resolution of \(2^{k_{\mathrm {max}}1} \times 2^{k_{\mathrm {max}}1}\) pixels.

2.
All images are superimposed and the overlaying pixelsvalues (\(1,0,1\)) are added resulting in the image \(\mathcal {I} = \sum _{i=0}^{k_{\mathrm {max}}1} \mathcal {I}_i\). Since (i) no two subcells of the same kind can be directly located above each other and (ii) a negative subcell can be created only if there is a positive subcell already located directly beneath it, the number of positive subcells minus the number of negative subcells over an arbitrary point is always either 1 or 0. Consequently, the image \(\mathcal {I}\) consists of pixels with the values 1 and 0 only.

3.
The discussed algorithms can be applied to \(\mathcal {I}\), while compressing only the pixel region(s) indicated by the value 1. Since the BFCM also utilizes the RFIP approach, an additional integration region is introduced in the form of the entire cell, in which we integrate over the fictitious domain with \(\alpha<< 1\) in order to avoid the possible illconditioning of the system matrices.
An example for compressing the subcells of the BQTD with \(k_{\mathrm {max}} = 6\) is shown in Fig. 18, where for the physical domain a superellipseshaped region is chosen. The corresponding implicit function defining this domain is given by Eq. (19) [41].
Due to the compression, for which in the case of the current example the IBR was used, the originally 60 subcells on the levels \(k < k_{\mathrm {max}}\) have been significantly reduced to 26 and the overall subcellcount has decreased from 273 to 239. As a visualization of the flowchart in Fig. 17, the images created in case of the current example on the refinement levels as well as their sum, are depicted in Fig. 19. Note that for the current case there are no subcells created at the levels \(k=0\) and 1. Therefore, only the layers \(k=25\) are participating in the compression, resulting in the images \(\mathcal {I}_2\), \(\mathcal {I}_3\), \(\mathcal {I}_4\), and \(\mathcal {I}_5\), while \(\mathcal {I}_0\) and \(\mathcal {I}_1\) are empty images. One can see, that the final image \(\mathcal {I}\) (Fig. 19e) on which the compression algorithms can be applied, indeed contains pixels with values 1 (green) and 0 (grey) exclusively.
Comparison
Sticking to the example presented in Fig. 18, we plot the \(\lambda \)values for different refinement levels while comparing the QTDbased IBR\(^{*}\), the BQTD^{Footnote 13} and the BQTD + IBR\(^{13}\) approaches in Fig. 20a. According to the results, it is clear, that all methods yield a significant compression by reducing the number of the IPs resulting from the plain QTD to 15–30\(\%\). The BQTD actually outperforms the QTD + IBR\(^{*}\) approach, however, its efficiency can be further improved, when combined with the IBR.
Although the QTD + IBR\(^{*}\) (Fig. 20b) and BQTD + IBR (Fig. 20c) yield visually similar results, they are not the same: The BQTD + IBR results in a smaller number of subcells and while in Fig. 20b the physical subcells (gray) do not overlap with the cut subcells (yellow), in case of Fig. 20c, the positive subcells (dark green) are covered by the negative cut subcells (red). Furthermore, while in case of the QTD + IBR\(^{*}\) the domain of interest is the same for all cut subcells during the numerical integration, for the cut subcells in Fig. 20c, the domain of interest and the sign of the integration changes, depending on which is more effective in a computational sense, resulting in \(\varvec{\mathcal {B}}_{\mathrm {cut}}^+\) and \(\varvec{\mathcal {B}}_{\mathrm {cut}}^\).
Although the BQTD + IBR approach generally yields fewer IPs, it has two disadvantages when compared to the QTD: (i) performing a QTD is computationally less expensive than the BQTD, since the latter one involves the checking of extra conditions when a parent subcell produces four children and (ii) the QTD has a somewhat simpler algorithm, being easier to implement and is already widely used in the field of computational mechanics. Therefore, the best option would be to perform a traditional QTD and transform \(\varvec{\mathcal {S}}\) by an extra step to \(\varvec{\mathcal {B}}\), i.e., Fig. 20b–c. However, according to our tests, such a direct transition is not possible.
Numerical examples
In this section, we further investigate the approaches IBR\(^{*}\), BFCM and BFCM + IBR, that proved to be the most efficient ones, while fully including them into the FCM framework for the computation of problems in linear elasticity. The simulations are performed using our inhouse SCMcode written in MATLAB using Lagrangian shape functions based on the Gauss–Lobatto–Legendre (GLL) nodal distribution [67, 76, 86]. In our investigations, we will vary the cell size, the refinement level k as well as the polynomial degree \(p = p_{\xi } = p_{\eta }\). For the integration over the (sub)cells \((p+1) \times (p+1)\) IPs are used based on a Gauss–Legendre quadrature rule. If the extended domain is squareshaped, we use \(n_{\mathrm {C}}\) to define the number of cells both in x and ydirections, and thus the total number of IPs grows with \(\mathcal {O}(n_{\mathrm {c}}^2 p^2 2^k)\).
Perforated plate
As a first example, we apply the chosen methods to a statically loaded perforated plate depicted in Fig. 21, already investigated in the context of the FCM as a benchmark problem by numerous authors [31, 33, 39,40,41, 45, 77]. The material with a Young’s modulus \(E=206.9\) GPa and Poisson’s ratio \(\nu = 0.29\) is assumed to be linearelastic and isotropic. Furthermore, the plate is assumed to be in a plane stress state.
Quality of the results
First, we evaluate the quality and reliability of the results obtained when using the three methods investigated in this section. In Fig. 22, the errors in the energy norm
for various input parameters are presented, which should be minimized by the FCM solution on the energy space\(E(\Omega _{\mathrm {e}})\) over the domain \(\Omega _{\mathrm {e}}\) [3, 33]. In Eq. (20), \(\varvec{u}\) is the displacement field obtained by the FCM solution and \(\varvec{u}_{\mathrm {ref}}\) is the reference solution, obtained by pFEM using blending functions [113] for an exact geometry mapping, resulting in a strain energy of \(1/2 \cdot \mathcal {B}(\varvec{u}_{\mathrm {ref}},\varvec{u}_{\mathrm {ref}}) = 0.7021812127\) [31]. Besides investigating the global quality of the results based on \(e_{\mathrm {E}(\Omega _{\text {e}})}\), we also evaluate the solution based on pointwise values of the stressfields \(\sigma _{\mathrm {vM}}\) and \(\sigma _{\mathrm {yy}}\) along the diagonal \(\overline{AB}\) in Fig. 21, where \(\sigma _{\mathrm {vM}}\) is the von Mises stress and \(\sigma _{\mathrm {yy}}\) the stress in the ydirection.
The following observations regarding the results are made: First, the depicted global (Fig. 22) and local (Fig. 23) results are the same for all three methods. This indicates that the methods being used indeed only differ in their computational complexities and costs. Thus, we come to the conclusion, that no compromise has to be made between the accuracy and computational effort, when choosing one of the introduced compression techniques. Second, as a verification, the obtained results have been compared to the literature (see convergencecurves for \(k=3\) and \(k=7\) in Refs. [33, 41] and stresscurves depicted in Refs. [31, 45]) and to a scientific code developed at the Technical University of Munich (FCMLab [114]). In both cases, identical results have been obtained.
Time consumption
Although the investigated methods yield identical solutions, their time requirements need further investigation. The time spent on the integration \(t_{\mathrm {I}}\) is in direct correlation with \(\lambda \), since it is influenced by the number of deployed IPs, i.e., for a given number of subcells and IPs, the reduction of \(t_{\mathrm {I}}\) depends on the performance of the chosen compression algorithm. In Fig. 24, the \(\lambda \)values regarding the investigated methods are depicted for different settings in the case of the current example. Similar to what has been depicted in Fig. 20a, it is evident, that the different compression techniques are capable of reaching different stationary \(\lambda \)values, outlining the highest performance reachable by the compression techniques. The results depicted in Fig. 24 have been computed for \(p = 10\) and therefore \((p+1)\times (p+1)=121\) IPs are deployed per subcell. Please note that using a lower value of p does not affect the overall trend of the shown curves significantly. If, however, a low polynomial degree (\(p =1\), 2) is employed in conjunction with the RFIP approach, minor deviations might occur due to the fact that only a small number of IPs is distributed within the cut subcell.
Note that performing a compression increases the time spent on setting up the local integration mesh \(t_{\mathrm {LIM}}\), leading to a decrease of the positive effect of the reduction in compression time \(t_{\mathrm {C}}\). Therefore, for investigating the overall effect of the compression on the integration, we measure the time spent on the cut cells and compute \(\tau \) according to “Performance” section. The corresponding results are depicted in Fig. 25, where in general, following two domains regarding the compression can be observed:

1.
\(\tau > 1\): For lower polynomial degrees undesired results are obtained, where the application of the different techniques increases the computation time. This is due to the fact, that when only a few IPs are distributed, their reduction is less significant and therefore, the corresponding reduction in \(t_{\mathrm {I}}\) does not outweigh the increase of \(t_{\mathrm {LIM}}\) caused by the compression procedure.

2.
\(\tau < 1\): Increasing p results in higher number of distributed IPs per subcell, where the effect of the compression becomes more evident, resulting in a significant reduction in the integration time \(t_{\mathrm {I}}\), i.e., the time saved during the integration outweighs the time invested in the compression \(t_{\mathrm {C}}\). The compression ratio \(\tau \) improves exponentially with p, until the curves reach a stationary value which is specific to the chosen compression algorithm. Note that the refinement level k has a similar effect on \(\tau \) as p, since it also directly influences \(\lambda \), as it was demonstrated in Fig. 24. Therefore, regardless of p, lower levels of k result in higher stationary values of \(\tau \).
The transition point \(\tau = 1\), below which actual timesavings occur, is located for all three methods between \(p=3\) and \(p=4\) (Fig. 25). Therefore, for problems where polynomial degrees \(p \ge 4\) are used, a successful reduction in the computational time can be safely expected. Although low refinement levels k may weaken the effect of the reduction, for obtaining optimal convergence in \(e_{{\mathrm {E}}(\Omega _e)}\), it is reasonable to avoid poor refinement levels in the first place. Note that even if different discretizations are used, both the transition point and the general trend of the \(\tau \) curves remains unchanged, as demonstrated in Fig. 26 for the case of the BFCM + IBR. For the perforated plate, the most significant reduction was achieved by the BFCM combined with the IBR (\(\tau = 0.169\)), requiring only \(69 \%\) the computational time needed for the BFCM (\(\tau = 0.245\)).
Violinshaped region
As a second example, we investigate the reduction of the computational effort for the computation of the mode shapes of the violinshaped domain depicted in Fig 27a. Since the given problem is a dynamic one, the global EQS presented in Eq. (8) is modified to
where \(\varvec{M}\) is the global mass matrix and \(\ddot{\varvec{U}}\) is the second derivative of the displacement vector with respect to the time, corresponding to the nodal accelerations. The mass matrices of the individual cells \(\varvec{M}^{(c)}\) are computed as
where \(\rho \) is the density of the physical material. For obtaining the solution, the general eigenvalue problem of Eq. (21) has to be solved, while free boundary conditions are assumed. For comparison, we use the results published by Gravenkamp and Duczek [115]. According to their simulation, we assume a linear isotropic material with the following properties: \(E=12.048\) GPa, \(\nu = 0.4\) and \(\rho = 470\) kg/m\(^3\) as well as the plane stress state. Furthermore, the embedding domain is discretized with \(192\times 256 = 49,152\) cells, and the polynomial degree of the shape function is set to \(p=4\). For capturing the boundary, a refinement level of \(k=5\) is used. Using our FCsimulation, with and without compression, identical results (see Fig. 27b, c) were obtained as the ones given in Ref. [115].
The \(\tau \)values obtained for this example are depicted in Fig. 28. Considering the chosen set up of the numerical model, the computational time that is dominantly spent on the integration of the cut cells was successfully reduced by approximately 30\(\%\) compared to the QTDbased AIS, while maintaining the same accuracy. It is already a significant reduction, even though the applied compression methods are more efficient for higher polynomial degrees, as discussed in the previous section (see Fig. 25). This is due to the already mentioned fact, that for lower polynomial degrees fewer IPs are distributed in each cell and therefore, the effect of the compression is less pronounced. Note that this phenomenon does not only affect the CAIS, but also the BFCM.
For getting a clearer insight regarding the global effect of the compression, let us now evaluate \(\tau \) for the entire processing step composed of the computation and assembly of the cell matrices as well as the solution of the eigenvalue problem. Since the free vibrations of the violin are investigated, no Dirichlet BCs are applied to the system. The corresponding results are depicted in Fig. 29, where also the effect of the precomputation of the stiffness and mass matrices of the physical cells, made possible by using a Cartesian mesh, is included. It is clear, that the precomputation of the physical cell matrices and the reduction of the IPs in the cut cells affect exclusively the times \(t_{\mathrm {phys}}\) and \(t_{\mathrm {cut}}\), respectively, the time requirement of the rest of simulation remains unchanged. Although there are significantly more physical cells (22,863) than cut ones (1245) during the simulation, \(t_{\mathrm {phys}}\) is only 8.9\(\%\) of \(t_{\mathrm {cut}}\). This result can be explained by two facts: (i) the cut cells contain a total of 123,975 subcells in the uncompressed state, each of which having the same computational cost during the numerical integration as a single physical cell and (ii) \(t_{\mathrm {cut}}\) includes both \(t_{\mathrm {I}}\), and \(t_{\mathrm {LIM}}\). As a result, the precomputation improved the overall runtime only by 4\(\%\). Furthermore, although \(t_{\mathrm {cut}}\) is significantly larger than \(t_{\mathrm {phys}}\), for the chosen combination of p and k, it represents only \(53\%\) of the total processing time. Consequently, the approximately 30\(\%\) reduction in \(t_{\mathrm {cut}}\) (see Fig. 28) results without precomputation roughly in 15\(\%\) and with precomputation in approximately 20\(\%\) reduction of the processingtime, which is still a significant number. Especially, if the straightforward implementation of the IBR\(^{*}\) and the lossless compression which results in the same accuracy is taken into account. Comparing the methods to each other, the QTD + IBR\(^{*}\) has performed better than the BFCM and did just as well as the BFCM + IBR.
Disc brake
Finally, we investigate the performance of the compression techniques for the modal analysis of a disc brake depicted in Fig. 30a. The geometry is freely in the internet [116]. The original model is a threedimensional thinwalled plate, from which the top surface is extracted for our twodimensional simulation. For the simulation, a material with a Young’s modulus of \(E=200\) GPa, a Poisson’s ratio of \(\nu = 0.29\) and density of \(\rho = 7800\) kg/m\(^3\) is used. A reference solution of the problem was obtained by the commercial FEsoftware Abaqus using an unstructured mesh consisting of 278,610 quadratic quadrilateral elements,^{Footnote 14} resulting in a total of 1,007,147 DOFs. For the FC simulation, the extended domain was discretized by \(80 \times 80\) = 6400 cells, from which 344 were purely physical and 2694 are cut (see Fig. 30b). To ensure highly accurate results and thus to compensate for the rather coarse mesh, a polynomial degree \(p=9\) and a refinement level \(k=8\) was used during the simulation, resulting in a total of 503,574 DOFs. Both the computed eigenfrequencies and the mode shapes were in a good agreement with the reference solution. A visual comparison of the 6th and 15th mode shapes are depicted in Fig. 31.
The compression ratio \(\tau \) for the different approaches is depicted in Fig. 32, where \(\tau \) is evaluated both for the computation of the cut cell matrices (value written within the bars) and for the total processing (value in parenthesis outside of the bars). Unlike in case of the previous example, due to the high value of p in combination with a sufficient refinement level \((k\ge 5)\), a significant compression was possible, requiring only 20–30\(\%\) of the original time. While in case of the previous example the QTD + IBR\(^{*}\) was only slightly better than the BFCM, for the current example, the QTD + IBR\(^{*}\) (\(\tau = 0.209\)) was 30\(\%\) faster than the BFCM (\(\tau = 0.298\)). Although the BFCM + IBR has lead to the same time saving, its implementation is definitely more cumbersome than the one of the QTD + IBR\(^{*}\). Furthermore, due to the significantly larger number of cut cells than physical ones (again, unlike in the case of the previous example), the computation of the cut cells is entirely dominating the processing time. Consequently, reducing \(t_{\mathrm {cut}}\) has a major influence on the processing time, so much so, that \(\tau \) for the cut cells and \(\tau \) for the entire processing step are basically equal.
Conclusion
In this contribution, the compressed adaptive integration scheme (CAIS) was introduced and investigated, which extends the traditional AIS by compressing the subcells resulting from the quadtreedecomposition (QTD). The additional compression step results in significantly decreased computational costs as the number of integration subcells is notably reduced. Furthermore, if only polynomial shape functions are considered in conjunction with Cartesian meshes, the accuracy is identical to the conventional AIS.
In “Compression techniques”, “Comparison”, “Drawback of the RLE compression technique” section, different compression techniques were introduced and tested. The imageblock representation (IBR) approach turned out to be the most suitable candidate due to its efficiency, simplicity and robustness. It was shown in “Combination with other approaches” section, that the CAIS can be coupled with the reduction of fictitious integration points (RFIP) and Boolean finite cell method (BFCM) approaches as well, leading to even fewer IPs and consequently, decreased computational times. Then, in “Numerical examples” section, the CAIS was successfully embedded in the FCM framework and applied to twodimensional problems of elastostatics and modal analysis. By means of several numerical examples, the IBR*, BFCM, and BFCM + IBR approaches have been investigated in detail. All of these methods yielded accurate results with significantly less computational effort. Since the accuracy of these methods is basically identical, the focus was placed on the comparison of their computational time which is influenced by two factors:

1.
Time investment for setting up an appropriate local integration mesh, which depends on the speed of the partitioning and compression algorithms.

2.
Time saved during integration over the resulting subcells, which depends on the total number of integration points in the broken cells, determined by the number of subcells and distributed integration points within them.
In “Time consumption” section, it was shown, that the polynomial order of the shape functions heavily influences whether the compression algorithm renders a meaningful reduction of computational time. A low polynomial degree (in our code \(p<4\)) may increase the computational time. However, for higher order shape functions and reasonable refinement levels, which are typically used in the FCM, a significant reduction can be achieved. In less ideal cases (“Violinshaped region” section), the time spent on the cut subcells can be compressed down to 70\(\%\) and the time of the entire simulation to \(80\%\), while in ideal cases that are more suitable for the FCM (“Disc brake” section), a significant reduction was achieved requiring only 20\(\%\) of the computational time for the entire simulation.
Since the investigated methods all yield the same accuracy, no compromise between the quality of the solution and the computational time has to be made. Therefore, one may choose the simplest and most efficient approach. Although the BFCM methodology, which can be further enhanced by the combination with the IBR compression technique, saves a significant amount of time, it requires major modifications in the already existing codes regarding the construction of the local integration mesh (LIM) as well as the compression and integration over the Boolean subcells. On the other hand, the IBR\(^{*}\), which in 2 of the 3 numerical examples turned out to be just as efficient as the BFCM + IBR, requires a QTD only, which is already widely used and implemented in the context of the FCM and other fictitious domain approaches. Also the integration over the compressed subcells can be carried out without modification in the preexisting code. Therefore, we recommend the CAIS with the IBR\(^{*}\) approach, both for the FCM and for other numerical methods as well, where a QTD is performed for integration purposes.
Outlook
In this work, we have investigated the imagecompression techniques for twodimensional problems. Consequently, future work could include the extension of the introduced compression techniques to threedimensional problems, where instead of a pixel image we deal with a sequence of images resulting in a voxelated data set which can be derived from an initial OTD. This should be fairly straightforward for the RLE and IBR methods, since they only require an additional scanning and merging directions. The RLE would scan the runlengths along the x, y or zdirection, increasing its time complexity to \(\mathcal {O}(n^3)\). Then, the IBR can be carried out in two steps, assuming the primary RLE was carried out in idirection: it first merges the line segments in the ijplane, then further unites the merged blocks, if possible, by comparing the adjacent jkplanes, each of which being in a twodimensional IBRcompressed state. Although the threedimensional compression requires more time \(t_{\mathrm {C}}\) than a twodimensional one, the number of IPs also grows at a cubicrate and therefore, a significant portion of the time \(t_{\mathrm {I}}\) can be saved when a meaningful compression of the subcells is achieved. The extension of the MRP to threedimension is arguable: Not only that in this case the intersection graph (if it can be constructed in the first place) is not bipartite any more and thus König’s theorem does not apply, it is also questionable, whether it is worth the effort, based on the good performance and easy implementation of the IBR we have seen in twodimensions. Finally, the BFCM is expected to perform similar compared to the CAIS in theedimensions as well. However, it is yet to be investigated, how these different approaches perform against each other in that case.
Future work might also include the investigation of the CAIS in conjunction with other fictitious domain approaches, where the integrand in Eq. (1) is not a polynomial function. This is the case, e.g., in the polyFCM [89], where due to the rational shape functions and nonaffine geometry mapping of the polygonal elements, the integrand in Eq. (1) is a rational function. Applying Gaussian quadrature rules to rational functions inevitably leads to integration errors, which further increase when due to the compression of the subcells a reduced number of integrations points is used for the numerical integration. Further research might reveal whether the typical convergence rates can be maintained in the polyFCM when combined with the CAIS for reducing the computational time. Another interesting area of application for the CAIS is seen in unfitted triangular and tetrahedral meshes, e.g., CutFEM [24] and TetFCM [11, 69, 70]. In this case, as long as a subparametric geometry mapping preserving the straight edges and faces of the elements is used in conjunction with a polynomial approximation of the field variables [70] the integrand in Eq. (1) remains polynomial. Hence, the CAIS can be applied without loss of accuracy compared to the conventional AIS.
The shown advantages of the CAIS become even more significant, when applied to nonlinear problems or prefinement strategies. In these cases, following the reasoning put forward in Ref. [61] for the moment fitting approach, the compression has to be performed only once for each cut cell, while the compressed subcells \(\varvec{\mathcal {R}}\) can be reused in every iteration loop, making the computational overhead \(t_{\mathrm {C}}\) vanish for even lower polynomial degrees (cf. Fig. 25).
Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Change history
08 July 2020
An amendment to this paper has been published and can be accessed via the original article.
09 July 2020
An amendment to this paper has been published and can be accessed via the original article.
Notes
 1.
Note that the integrand is not necessarily a polynomial function. This is only true if the employed shape functions are polynomials and if the geometry mapping is affine. This is, e.g., the case for the FCM in conjunction with Cartesian meshes.
 2.
In the original contribution [42,43,44], the proposed integration scheme is discussed in the context of FCM applied to 3D biomedical problems. Consequently, the computational domain is described by voxelated data, often obtained by medicalimaging such as computed tomography. Assuming constant linear elastic and isotropic material properties for the individual voxels, the stiffness and mass matrices can be precomputed for the \(n_x \times n_y \times n_z\) voxels within each element \(\Omega ^{(\mathrm {e})}\) using the two Lamé constants \(\lambda \) and \(\mu \).
 3.
In the remainder, we refer to the integration subcells simply as subcells.
 4.
 5.
Often also referred to as broken cells.
 6.
Although not all subcells are of the same size, they are all quadratic.
 7.
The term insideoutside test refers to a procedure that is capable of determining whether a given point is located in the interior (inside) or exterior (outside) of the physical domain.
 8.
In mathematics, the word cardinality refers to the number of elements in a given set.
 9.
In the remainder of this work, we use \(\cdot \) to express the cardinality of a given input set.
 10.
 11.
See the mentioned three features on page 7 (883) in [45]. The features 1 and 2 represent conditions (C1 and C2) that has to be evaluated for every parent subcell, while the third feature is the already introduced RFIP method. Note that the notations C1, C2 and RFIP do not appear in the original text, it is only our interpretation for referring to those features.
 12.
Recall, in these subcells the discontinuity is not taken into account.
 13.
Note that since the BQTD already incorporates the RFIP approach, indicating this with the asterix \(\square ^{*}\) and writing BQTD\(^{*}\) and BQTD + IBR\(^{*}\) is therefore redundant.
 14.
In Abaqus these elements are referred to as CPS8 elements indicating that conventional continuum elements under plane stress conditions with 8 nodes are used. The shape functions of an 8node quadrilateral finite element are based on the Serendipity family.
Abbreviations
 AIS:

Adaptive integration scheme
 BC:

Boundary condition
 BFCM:

Boolean finite cell method
 CAIS:

Compressed adaptive integration scheme
 DIP:

Decreasing integration points
 DOF:

Degrees of freedom
 DT:

Divergence theorem
 EP:

Equivalent polynomial
 EQS:

Equation system
 FC:

Finite cell
 FCM:

Finite cell method
 FE :

Finite element
 FEM:

Finite element method
 IBR:

Image block representation
 IBR\(^{*}\) :

IBR combined with the RFIP
 IP:

Integration point
 LIM:

Local integration mesh
 MF:

Moment fitting
 MRP:

Minimal rectangular partition
 OTD:

Octreedecomposition
 QTD:

Quadtreedecomposition
 RFIP:

Reduction of fictitious integration points
 RLE:

Runlength encoding
References
 1.
Hutton DV. Fundamentals of finite element analysis. New York: McGrawHill Science/Engineering/Math; 2003.
 2.
Zienkiewicz OC, Taylor RL, Zhu JZ. The finite element method: its basis and fundamentals. Oxford: ButterworthHeinemann; 2005.
 3.
Szabó B, Babuška I. Introduction to finite element analysis. Hoboken: Wiley; 2011.
 4.
Del Pino S, Pironneau O. A fictitious domain based general PDE solver. Numerical methods for scientific computing variational problems and applications. 2003.
 5.
Ramière I, Angot P, Belliard M. A fictitious domain approach with spread interface for elliptic problems with general boundary conditions. Comput Methods Appl Mech Eng. 2007;196(4–6):766–81.
 6.
Parussini L. Fictitious domain approach via lagrange multipliers with least squares spectral element method. J Sci Comput. 2008;37(3):316–35.
 7.
Parussini L, Pediroda V. Fictitious domain approach with hpfinite element approximation for incompressible fluid flow. J Comput Phys. 2009;228(10):3891–910.
 8.
Hansbo A, Hansbo P. An unfitted finite element method, based on Nitsche’s method, for elliptic interface problems. Comput Methods Appl Mech Eng. 2002;191(47–48):5537–52.
 9.
Roma AM, Peskin CS, Berger MJ. An adaptive version of the immersed boundary method. J Comput Phys. 1999;153(2):509–34.
 10.
Dauge M, Düster A, Rank E. Theoretical and numerical investigation of the finite cell method. J Sci Comput. 2015;65(3):1039–64.
 11.
Duczek S, Duvigneau F, Gabbert U. The finite cell method for tetrahedral meshes. Finite Elem Anal Des. 2016;121:18–32.
 12.
Sukumar N, Moës N, Moran B, Belytschko T. Extended finite element method for threedimensional crack modelling. Int J Numer Methods Eng. 2000;48(11):1549–1570, 8.
 13.
Sukumar N, Chopp DL, Moës N, Belytschko T. Modeling holes and inclusions by level sets in the extended finiteelement method. Comput Methods Appl Mech Eng. 2001;190(46–47):6183–200.
 14.
Cheng KW, Fries TP. Higherorder XFEM for curved strong and weak discontinuities. Int J Numer Methods Eng. 2009;82:564–90.
 15.
Fries TP, Belytschko T. The extended/generalized finite element method: an overview of the method and its applications. Int J Numer Methods Eng. 2010;84:253–304.
 16.
Strouboulis T, Copps K, Babuška I. The generalized finite element method. Comput Methods Appl Mech Eng. 2001;190(32–33):4081–193.
 17.
Strouboulis T, Zhang L, Babuška I. Assessment of the cost and accuracy of the generalized FEM. Int J Numer Methods Eng. 2006;69(2):250–83.
 18.
Babuška I, Banerjee U. Stable generalized finite element method (SGFEM). Comput Methods Appl Mech Eng. 2012;201–204:91–111.
 19.
Melenk JM, Babuška I. The partition of unity finite element method: basic theory and applications. Comput Methods Appl Mech Eng. 1996;139(1–4):289–314.
 20.
Babuška I, Melenk JM. The partition of unity method. Int J Numer Methods Eng. 1997;40(4):727–58.
 21.
Burman E, Hansbo P. Fictitious domain finite element methods using cut elements: I. A stabilized lagrange multiplier method. Comput Methods Appl Mech Eng. 2010;199(41–44):2680–6.
 22.
Burman E, Hansbo P. Fictitious domain finite element methods using cut elements: II. A stabilized nitsche method. Appl Numer Math. 2012;62(4):328–41.
 23.
Burman E, Hansbo P. Fictitious domain methods using cut elements: III. A stabilized nitsche method for stokes’ problem. ESAIM: Math Model Numer Anal. 2014;48(3):859–74.
 24.
Burman E, Claus S, Hansbo P, Larson MG, Massing A. CutFEM: Discretizing geometry and partial differential equations. Int J Numer Methods Eng. 2014;104(7):472–501.
 25.
GarcíaRuíz MJ, Steven GP. Fixed grid finite elements in elasticity problems. Eng Comput. 1999;16(2):145–64.
 26.
Maan FS, Querin OM, Barton DC. Extension of the fixed grid finite element method to eigenvalue problems. Adv Eng Softw. 2007;38(8–9):607–17.
 27.
Daneshmand F, KazemzadehParsi MJ. Static and dynamic analysis of 2d and 3d elastic solids using the modified FGFEM. Finite Elem Anal Des. 2009;45(11):755–65.
 28.
Kim H, García MJ, Querin OM, Steven GP, Xie YM. Introduction of fixed grid in evolutionary structural optimisation. Eng Comput. 2000;17(4):427–39.
 29.
Woon SY, Querin OM, Steven GP. On improving the GA stepwise shape optimization method through the application of the fixed grid FEA paradigm. Struct Multidiscip Optim. 2003;25(4):270–8.
 30.
Nadal E, Ródenas JJ, Albelda J, Tur M, Tarancón JE, Fuenmayor FJ. Efficient finite element methodology based on cartesian grids: application to structural shape optimization. Abstract Appl Anal. 2013;1–19:2013.
 31.
Parvizian J, Düster A, Rank E. Finite cell method: \(h\) and \(p\)extension for embedded domain problems in solid mechanics. Comput Mech. 2007;41(1):121–33.
 32.
Düster A, Parvizian J, Yang Z, Rank E. The finite cell method for threedimensional problems of solid mechanics. Comput Methods Appl Mech Eng. 2008;197(45–48):3768–82.
 33.
Düster A, Rank E, Szabó B. The \(p\)version of the finite element and finite cell methods. Encyclopedia of Computational Mechanics. 2017. p. 1–35.
 34.
Schillinger D, Cai Q, Mundani RP, Rank E. A review of the finite cell method for nonlinear structural analysis of complex CAD and imagebased geometric models., Lecture notes in computational science and engineeringBerlin: Springer; 2013. p. 1–23.
 35.
Schillinger D, Ruess M. The finite cell method: a review in the context of higherorder structural analysis of CAD and imagebased geometric models. Arch Comput Methods Eng. 2014;22(3):391–455.
 36.
Ruess M, Tal D, Trabelsi N, Yosibash Z, Rank E. The finite cell method for bone simulations: verification and validation. Biomech Model Mechanobiol. 2011;11(3–4):425–37.
 37.
Zander N, Kollmannsberger S, Ruess M, Yosibash Z, Rank E. The finite cell method for linear thermoelasticity. Comput Math Appl. 2012;64(11):3527–41.
 38.
Ruess M, Schillinger D, Bazilevs Y, Varduhn V, Rank E. Weakly enforced essential boundary conditions for NURBSembedded and trimmed NURBS geometries on the basis of the finite cell method. Int J Numer Methods Eng. 2013;95(10):811–46.
 39.
Abedian A, Düster A. Equivalent legendre polynomials: numerical integration of discontinuous functions in the finite element methods. Comput Methods Appl Mech Eng. 2019;343:690–720.
 40.
Abedian A, Parvizian J, Düster A, Khademyzadeh H, Rank E. Performance of different integration schemes in facing discontinuities in the finite cell method. Int J Comput Methods. 2013;10(03):1350002.
 41.
Joulaian M. The hierarchical finite cell method for problems in structural mechanics. Ph.D. thesis, Hamburg Technical University. 2017.
 42.
Yang Z. The finite cell method for geometrybased structural simulation. Ph.D. thesis, Technical University of Munich. 2011.
 43.
Yang Z, Kollmannsberger S, Düster A, Ruess M, Garcia EG, Burgkart R, Rank E. Nonstandard bone simulation: interactive numerical analysis by computational steering. Comput Vis Sci. 2011;14(5):207–16.
 44.
Yang Z, Ruess M, Kollmannsberger S, Düster A, Rank E. An efficient integration technique for the voxelbased finite cell method. Int J Numer Methods Eng. 2012;91(5):457–71.
 45.
Abedian A, Düster A. An extension of the finite cell method using boolean operations. Comput Mech. 2017;59(5):877–86.
 46.
Legrain G, Moës N. Adaptive anisotropic integration scheme for highorder fictitious domain methods: application to thin structures. Int J Numer Methods Eng. 2018;114(8):882–904.
 47.
Fries TP, Omerović S. Higherorder accurate integration of implicit geometries. Int J Numer Methods Eng. 2015;106(5):323–71.
 48.
Hughes TJR. The finite element method: linear static and dynamic finite element analysis. New York: Dover Pubn Inc.; 2000.
 49.
Kudela L. Highly accurate subcell integration in the context of the finite cell method. Master’s thesis, Technical University of Munich. 2013.
 50.
Kudela L, Zander N, Bog T, Kollmannsberger S, Rank E. Efficient and accurate numerical quadrature for immersed boundary methods. Adv Model Simul Eng Sci. 2015;2(1):10.
 51.
Királyfalvi Gy, Szabó B. Quasiregional mapping for the \(p\)version of the finite element method. Finite Elem Anal Des. 1997;27(1):85–97.
 52.
Kudela L, Zander N, Kollmannsberger S, Rank E. Smart octrees: accurately integrating discontinuous functions in 3D. Comput Methods Appl Mech Eng. 2016;306:406–26.
 53.
Guichard D, et al. Single and multivariable calculus: early transcendentals. In: Guichard D. Creative commons attribution—Noncommercial—Share Alike 3.0. 2018.
 54.
Riley KF, Hobson MP, Bence SJ. Mathematical methods for physics and engineering: a comprehensive guide. Cambridge: Cambridge University Press; 2006.
 55.
Dasgupta G. Integration within polygonal finite elements. J Aerosp Eng. 2003;16(1):9–18.
 56.
Sudhakar Y, Moitinho de Almeida JP, Wall WA. An accurate, robust, and easytoimplement method for integration over arbitrary polyhedra: application to embedded interface methods. J Comput Phys. 2014;273:393–415.
 57.
Duczek S, Gabbert U. Efficient integration method for fictitious domain approaches. Comput Mech. 2015;56(4):725–38.
 58.
Mousavi SE, Xiao H, Sukumar N. Generalized Gaussian quadrature rules on arbitrary polygons. Int J Numer Methods Eng. 2009;82:99–113.
 59.
Xiao H, Gimbutas Z. A numerical algorithm for the construction of efficient quadrature rules in two and higher dimensions. Comput Math Appl. 2010;59(2):663–76.
 60.
Müller B, Kummer F, Oberlack M. Highly accurate surface and volume integration on implicit domains by means of momentfitting. Int J Numer Methods Eng. 2013;96(8):512–28.
 61.
Joulaian M, Hubrich S, Düster A. Numerical integration of discontinuities on arbitrary domains based on moment fitting. Comput Mech. 2016;57(6):979–99.
 62.
Mousavi SE, Sukumar N. Numerical integration of polynomials and discontinuous functions on irregular convex polygons and polyhedrons. Comput Mech. 2010;47(5):535–54.
 63.
Sudhakar Y, Wall WA. Quadrature schemes for arbitrary convex/concave volumes and integration of weak form in enriched partition of unity methods. Comput Methods Appl Mech Eng. 2013;258:39–54.
 64.
Hubrich S, Joulaian M, Düster A. Numerical integration in the finite cell method based on momentfitting. In: Proceedings of 3rd ECCOMAS young investigators conference, 6th GACM Colloquium, Aachen, Germany. 2015. p. 1–4.
 65.
Ventura G, Benvenuti E. Equivalent polynomials for quadrature in heaviside function enriched elements. Int J Numer Methods Eng. 2014;102(3–4):688–710.
 66.
Ventura G. On the elimination of quadrature subcells for discontinuous functions in the eXtended finiteelement method. Int J Numer Methods Eng. 2006;66(5):761–95.
 67.
Duczek S. Higher order finite elements and the fictitious domain concept for wave propagation analysis. VDI FortschrittBerichte Reihe 20 Nr. 458. 2014. https://opendata.unihalle.de/handle/1981185920/11873.
 68.
Mossaiby F, Joulaian M, Düster A. The spectral cell method for wave propagation in heterogeneous materials simulated on multiple GPUs and CPUs. Comput Mech. 2018;63(5):805–19.
 69.
Xu F, Schillinger D, Kamensky D, Varduhn V, Wang C, Hsu MC. The tetrahedral finite cell method for fluids: immersogeometric analysis of turbulent flow around complex geometries. Comput Fluids. 2016;141:135–54.
 70.
Varduhn V, Hsu MC, Ruess M, Schillinger D. The tetrahedral finite cell method: Higherorder immersogeometric analysis on adaptive nonboundaryfitted meshes. Int J Numer Methods Eng. 2016;107(12):1054–79.
 71.
Nagaraja S, Elhaddad M, Ambati M, Kollmannsberger S, De Lorenzis L, Rank E. Phasefield modeling of brittle fracture with multilevel \(hp\)FEM and the finite cell method. Comput Mech. 2018;63(6):1283–300.
 72.
Sehlhorst HG. Numerical homogenization strategies for cellular materials with applications in structural mechanics. Ph.D. thesis. Technical University HamburgHarburg. 2011.
 73.
Düster A, Sehlhorst HG, Rank E. Numerical homogenization of heterogeneous and cellular materials utilizing the finite cell method. Comput Mech. 2012;50(4):413–31.
 74.
Heinze S, Joulaian M, Düster A. Numerical homogenization of hybrid metal foams using the finite cell method. Comput Math Appl. 2015;70(7):1501–17.
 75.
Elhaddad M, Zander N, Bog T, Kudela L, Kollmannsberger S, Kirschke J, Baum T, Ruess M, Rank E. Multilevel \(hp\)finite cell method for embedded interface problems with application in biomechanics. Int J Numer Methods Biomed Eng. 2018;34(4):e2951.
 76.
Duczek S, Liefold S, Gabbert U. The finite and spectral cell methods for smart structure applications: transient analysis. Acta Mech. 2014;226(3):845–69.
 77.
Parvizian J, Düster A, Rank E. Topology optimization using the finite cell method. Optim Eng. 2011;13(1):57–78.
 78.
Schillinger D, Kollmannsberger S, Mundani RP, Rank E. The finite cell method for geometrically nonlinear problems of solid mechanics. In: IOP conference series: materials science and engineering, vol. 10. 2010. p. 012170.
 79.
Schillinger D, Dedè L, Scott MA, Evans JA, Borden MJ, Rank E, Hughes TJR. An isogeometric designthroughanalysis methodology based on adaptive hierarchical refinement of NURBS, immersed boundary methods, and tspline CAD surfaces. Comput Methods Appl Mech Eng. 2012;249–252:116–50.
 80.
Konyukhov A, Lorenz C, Schweizerhof K. Various contact approaches for the finite cell method. Comput Mech. 2015;56:331–51.
 81.
Bog T, Zander N, Kollmannsberger S, Rank E. Weak imposition of frictionless contact constraints on automatically recovered highorder, embedded interfaces using the finite cell finiteelementmethod. Comput Mech. 2017;61:385–407.
 82.
Bog T. Frictionless contact simulation using the finite cell method. Ph.D thesis, Technical University of Munich. 2017.
 83.
Abedian A, Parvizian J, Düster A, Rank E. The finite cell method for the \({J}_2\) flow theory of plasticity. Finite Elem Anal Des. 2013;69:37–47.
 84.
Abedian A, Parvizian J, Düster A, Rank E. Finite cell method compared to \(h\)version finite element method for elastoplastic problems. Appl Math Mech. 2014;35(10):1239–48.
 85.
Kollmannsberger S, D’Angella D, Rank E, Garhuom W, Hubrich S, Düster A, Stolfo P. Di, Schröder A. Spline and hpbasis functions of higher differentiability in the finite cell method. GAMMMitteilungen. 2019.
 86.
Duczek S, Joulaian M, Düster A, Gabbert U. Numerical analysis of Lamb waves using the finite and spectral cell methods. Int J Numer Methods Eng. 2014;99(1):26–53.
 87.
Elhaddad M, Zander N, Kollmannsberger S, Shadavakhsh A, Nübel V, Rank E. Finite cell method: highorder structural dynamics for complex geometries. Int J Struct Stab Dyn. 2015;15(07):1540018.
 88.
Duczek S, Gabbert U. The finite cell method: a higher order fictitious domain approach for wave propagation analysis in heterogeneous structures. In: Lammering R, Gabbert U, Sinapius M, Schuster T, Wierach P, editors. Lambwave based structural health monitoring in polymer composites. Cham: Springer International Publishing; 2018. p. 217–39.
 89.
Duczek S, Gabbert U. The finite cell method for polygonal meshes: polyFCM. Comput Mech. 2016;58(4):587–618.
 90.
Lee NS, Bathe KJ. Effects of element distortions on the performance of isoparametric elements. Int J Numer Methods Eng. 1993;36(20):3553–76.
 91.
Duczek S, Berger H, Gabbert U. The finite pore method: a new approach to evaluate gas pores in cast parts by combining computed tomography and the finite cell method. Int J Cast Metals Res. 2015;28(4):221–8.
 92.
Joulaian M, Düster A. Local enrichment of the finite cell method for problems with material interfaces. Comput Mech. 2013;52(4):741–62.
 93.
Joulaian M, Zander N, Bog T, Kollmannsberger S, Rank E, Düster A. A highorder enrichment strategy for the finite cell method. PAMM. 2015;15(1):207–8.
 94.
Zander N, Bog T, Kollmannsberger S, Schillinger D, Rank E. Multilevel \(hp\)adaptivity: highorder mesh adaptivity without the difficulties of constraining hanging nodes. Comput Mech. 2015;55(3):499–517.
 95.
Düster A, Bröker H, Rank E. The \(p\)version of the finite element method for threedimensional curved thin walled structures. Int J Numer Methods Eng. 2001;52(7):673–703.
 96.
Schillinger D, Ruess M, Zander N, Bazilevs Y, Düster A, Rank E. Small and large deformation analysis with the \(p\) and Bspline versions of the finite cell method. Comput Mech. 2012;50(4):445–78.
 97.
de Berg M, Cheong O, van Kreveld M, Overmars M. Computational geometry. Berlin: Springer; 2008.
 98.
Suk T, Höschl C, Flusser J. Rectangular decomposition of binary images. In: BlancTalon J, Distante C, Philips W, Popescu D, Scheunders P, editors. Advanced concepts for intelligent vision systems. Berlin: Springer; 2012. p. 213–24.
 99.
Salomon D, Motta G. Handbook of data compression. London: Springer; 2010.
 100.
Spiliotis IM, Mertzios BG. Realtime computation of twodimensional moments on binary images using image block representation. IEEE Trans Image Process. 1998;7(11):1609–15.
 101.
Eppstein D. Graphtheoretic solutions to computational geometry problems. In: Graphtheoretic concepts in computer science. Springer Berlin Heidelberg. 2010. p. 1–16.
 102.
Seo J, Chae S, Shim J, Kim D, Cheong C, Han TD. Fast contourtracing algorithm based on a pixelfollowing method for image sensors. Sensors. 2016;16(3):353.
 103.
Pavlidis T. Algorithms for graphics and image processing. Berlin: Springer; 1982.
 104.
Miyatake T, Matsushima H, Ejiri M. Contour representation of binary images using runtype direction codes. Mach Vis Appl. 1997;9(4):193–200.
 105.
Ferrari L, Sankar PV, Sklansky J. Minimal rectangular partitions of digitized blobs. Comput Vis Graph Image Process. 1984;28(1):58–71.
 106.
Pető M. Improving the eciency of the numerical integration in the finite cell method. Master’s thesis, Otto von Guericke University Magdeburg. 2019.
 107.
Wu SY, Sahni S. Covering rectilinear polygons by rectangles. IEEE Trans ComputAided Des Integr Circuits Syst. 1990;9:377–388, 05.
 108.
Wu SY, Sahni S. Fast algorithms to partition simple rectilinear polygons. VLSI Des. 1994;1(3):193–215.
 109.
Levcopoulos C, Gudmundsson J. Approximation algorithms for covering polygons with squares and similar problems. In: Randomization and approximation techniques in computer science. Springer Berlin Heidelberg. 1997. pp. 27–41.
 110.
Franzblau DS, Kleitman DJ. An algorithm for covering polygons with rectangles. Inf Control. 1984;63(3):164–89.
 111.
Gunther O. Minimum kpartitioning of rectilinear polygons. J Symb Comput. 1990;9(4):457–83.
 112.
Bondy JA, Murty USR. Graph theory. London: Springer; 2008.
 113.
Szabó B, Düster A, Rank E. The \(p\)version of the finite element method. Chapter 5. Encyclopedia of computational mechanics. 2004.
 114.
Zander N, Bog T, Elhaddad M, Espinoza R, Hu H, Joly A, Wu C, Zerbe P, Düster A, Kollmannsberger S, Parvizian J, Ruess M, Schillinger D, Rank E. FCMLab: a finite cell research toolbox for MATLAB. Adv Eng Softw. 2014;74:49–63.
 115.
Gravenkamp H, Duczek S. Automatic imagebased analyses using a coupled quadtreeSBFEM/SCM approach. Comput Mech. 2017;60(4):559–84.
 116.
Online source of the disc brake model. https://grabcad.com/library/brakestuffpunchdisc1. Accessed 16 Oct 2019.
Acknowledgements
The authors thank the German Research Foundation (DFG) for its financial support under Grant DU 1904/21. The authors also acknowledge the many fruitful discussions with Prof. Juhre (Otto von Guericke University Magdeburg) and Prof. Düster (Hamburg University of Technology) regarding the interpretation of the numerical results.
Funding
The first author is financed by the German Research Foundation (DFG) under Grant DU 1904/21.
Author information
Affiliations
Contributions
The inhouse Matlab code which is based on the Spectral Cell Method was developed by SE and extended to imagecompression based numerical integration, as proposed in this article, by MP. The application of this approach to the presented numerical examples was jointly investigated by MP and FD. The article is mainly written by MP, while for the theoretical background as well as for the writing, discussions and comments provided by SE and FD were essential for the quality of the work. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The original version of this article was revised: The errors in equation and in text have been corrected.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Petö, M., Duvigneau, F. & Eisenträger, S. Enhanced numerical integration scheme based on imagecompression techniques: application to fictitious domain methods. Adv. Model. and Simul. in Eng. Sci. 7, 21 (2020). https://doi.org/10.1186/s40323020001572
Received:
Accepted:
Published: