WorksComputer Science
Configurational-force-driven adaptive refinement and coarsening in topology optimization
Abstract
The iterative nature of topology optimization, especially in combination with nonlinear state problems, often requires the solution of thousands of linear equation systems. Furthermore, due to the pixelated design representation, the use of a fine mesh is essential to obtain geometrically well-defined structures and to accurately compute response quantities such as the von Mises stress. Therefore, the computational cost of solving a fine-mesh topology optimization problem quickly adds up. To address this challenge, we consider a multi-level adaptive refinement and coarsening strategy based on configurational forces. Configurational forces based on the Eshelby stress predict configurational changes such as crack propagation or dislocation motion. Due to a relaxation in the calculation of (Eshelby) stresses with respect to the design variables, discrete configurational forces increase not only in highly stressed regions, but also in grey transition regions (design boundaries). For this reason they are an ideal criterion for mesh adaptivity in topology optimization, especially when avoiding stress failure is a priority. By using configurational forces for refinement, we obtain a high-resolution structure where the refined mesh is present along the design boundaries as well as in stress-critical regions. At the same time, multilevel coarsening using the same criterion drastically minimizes the computational effort.
Introduction
The rapid development of computational resources has enabled the execution of increasingly complex and numerically demanding simulations, even on personal computers. These advances in hardware have been accompanied by intensified research on computational methods, which has led to an increasing role for computer simulations in product development. This is due to the low costs, reliability, accuracy, and efficiency of such simulations. However, the fundamental challenge remains the trade-off between solution accuracy and computational efficiency, which is exemplified by the degree of discretization of the system to be solved. It is evident that the finer the discretization, the more accurate the solution. However, this increased accuracy is associated with a higher computational cost, regardless of the specific computational method employed. This is particularly evident in topology (or any structural) optimization, where linear systems need to be solved hundreds of times, sometimes in a nested fashion, e.g., for each optimization iteration multiple load steps and Newton–Rhapson iterations are required for highly nonlinear boundary value problems. Therefore, a relevant research direction in the world of computer simulation is focused on discretization techniques that aim to maximize the accuracy/cost ratio. In the context of the finite element method (FEM), many commercial programs offer advanced preprocessing tools with different meshing algorithms. However, these almost exclusively produce unstructured meshes, which introduce an additional level of complexity in topology optimization as compared to the original approach based on a structured mesh. For simulations involving iterative solving of linear systems, i.e. quasi-static nonlinear elastic problems or time discretized problems, there are several techniques that modify the mesh for each iteration.
In the context of topology optimization, mesh adaptivity is additionally focused on maximizing geometric accuracy, furthermore to the ratio of numerical accuracy and computational cost. Several works deal with mesh refinement strategies for tria-based meshes [1, 2, 3], (classical) structured meshes with hanging node constraints (with coarsening) [4, 5, 6, 7] or unstructured quad-based meshes [8]. The adaptivity criteria can be primarily divided into density-based [4, 3, 8], which prioritize the geometric accuracy of the designs, and error-based [5, 9], which prioritize the computational accuracy. The work of Muñoz et al. [7] tackles both density-based and error-based adaptivity to address both geometry representation and accurate stress computation to meet the constraints. In combination with shape optimization employing an embedding domain discretization (EDD), adaptive meshing has been applied to the variable shape in Stankiewicz et al. [10]. Moreover, in [11, 12, 13] mesh adaptivity in the EDD-based shape optimization was introduced for both the structured mesh of the computational domain and the variable shape.
In the following, we employ the concept of configurational forces for mesh adaptivity in topology optimization. Configurational (or material) forces are a well-established driving criterion for fracture propagation (Maugin, 1995; Steinmann, 2000; Miehe and Gürses, 2007; Kuhn and Müller, 2016; Ballarini and Royer-Carfagni, 2016; Moreno-Mateos and
2016; Ballarini and Royer-Carfagni, 2016; Moreno-Mateos and Steinmann, 2024). Furthermore, configurational forces have been successfully explored for r- and h-adaptation of meshes in the context of hyperelasticity {citep}[20], poromechanics {citep}(Na et al., 2019), phase field fracture {citep}[22], isogeometric analysis {citep}[23], elastoplasticity {citep}[24], fracture mechanics {citep}Qinami et al. [25] or truss optimization {citep}[26]. In this article, we show that configurational forces are also a powerful criterion for mesh adaptivity in topology optimization and offer a promising alternative to density-based and stress-based adaptivity techniques. We propose a robust adaptivity strategy that utilizes the deal.II finite element library to handle refinement and coarsening with hanging node constraints {citep}[27]`.
This article is organized as follows. In § 2 we motivate the use of configurational mechanics and re-iterate its essential concepts. We also discuss the meaning and role of configurational forces. In § 3 we introduce our methodology for mesh adaptivity using configurational forces and adapt the configurational forces to incorporate the pseudo-density variables. We address the possible edge cases and how deal.II handles mesh coarsening and refinement. In § 4, two examples, a cantilever beam and a U-beam, are studied to evaluate the introduced mesh adaptivity based on configurational forces. A thorough comparison with other known adaptivity strategies is given. Concluding remarks are given in § 5.
Why configurational mechanics
In the conventional understanding of continuum mechanics, considering finite deformations, we are concerned with finding the spatial deformation of a continuum under spatial (deformational) forces, be it boundary tractions or a body-distributed force, e.g. gravity. By assuming quasi-static loading conditions, we distinguish between the material (reference) configuration of such a system, which corresponds to the undeformed (unloaded) state of a continuum body, and the spatial configuration that is in equilibrium for the external loads. For an elastic structure this equilibrium corresponds to a minimum of the total potential energy, i.e. the sum of the internal energy and external potential energies.
The core assumption of conventional deformational continuum mechanics is that the material configuration is fixed, i.e. it does not allow any geometrical changes. In contrast, configurational continuum mechanics not only allows for such configurational changes, but in fact is primarily concerned with them. In other words, the material configuration is no longer fixed and allows energy release (in accordance with the second law of thermodynamics) in the form of configurational changes such as crack propagation or defect motion.
Analogously, if spatial (deformational) forces are the driving mechanism behind the spatial deformation of a body, we recognize material (configurational) forces that are the driving mechanism behind any configurational changes. Configurational forces provide us with the necessary information to understand what are the energetic implications resulting from configurational changes such as crack propagation or dislocation motion. This information is highly useful in determining regions where the risk of such configurational change is highest. However, the application of configurational forces is more versatile than merely the prediction of material defects. Therefore, in this article, we demonstrate the power of configurational forces as a criterion for finite element mesh adaptivity in the context of topology optimization.
Essential concepts
To provide a thorough understanding of configurational forces, basic concepts from classical deformational continuum mechanics for the spatial (deformational) problem are first reviewed. Subsequently, the analogous quantities for the material (configurational) problem are introduced, which form the core of configurational mechanics.
Deformational problem
Let (Figure 1) be material configuration of a continuum body (subscript 0 indicates the material configuration). The body is composed by the continuum points with . Then the material configuration results from the material placement of every into . In response to spatial forces, in the form of tractions and body forces , the body undergoes deformation described by the spatial deformation map resulting in the spatial configuration . The body in the spatial configuration, which is the actual deformed state of the body, is described by the spatial placement .1
Figure 1. Basic kinematics of the deformational problem in continuum mechanics.
In classical deformational continuum mechanics of solids, all spatial quantities, their gradients and variations are usually evaluated with respect to the fixed material placement .1 Thus, the spatial deformation gradient is defined as .
To obtain the balance equations for the deformational problem, we introduce the total energy density consisting of the internal potential per unit reference volume and the external potential per unit reference volume. The total potential energy is given by
The common virtual work statement is then obtained via the stationarity condition, which requires the spatial variation of the total potential energy at fixed material placement to vanish, i.e.
where the derivative yields the two-point Piola stress tensor and the derivative is the spatial body force. In localized form, using Cauchy’s theorem, , where is the outward unit normal to the reference boundary , we obtain the localized equilibrium of deformational forces
where the divergence operator is evaluated with respect to the material placement .
Configurational problem
Now, to reiterate, the essence of configurational mechanics lies in the consideration of configurational changes. Here we consider the spatial placement of the body (Figure 2) as fixed and search for the change of the material configuration, described by the material deformation map , analogously parametrized in the spatial placement . Naturally, we consider the material deformation gradient , where the gradient operator is evaluated with respect to the fixed spatial placement .
Figure 2. Basic kinematics of the configurational problem in continuum mechanics.
Next, we re-parametrize and re-express the total potential energy density per unit volume in the spatial configuration so that . The total potential energy is then given by
For the configurational problem, however, the stationarity of the total potential energy via the material variation at fixed spatial placement is not guaranteed, i.e. . This is best understood by considering an example of a crack propagation problem, where a propagating crack is an irreversible change in the material configuration itself, resulting in the release of stored energy.
Energy release and Eshelby stress
To produce spontaneous (negative) energy release , configurational tractions must oppose the material configurational changes. The material variation of Eq. (4) produces
where is a two-point (Piola-like) stress tensor with spatial reference and is the material body force with spatial reference. By Piola transformation we obtain a material stress tensor with material reference, the so-called Eshelby stress: . For a detailed explanation of the relations between material and spatial references, see Steinmann [15]; Steinmann et al. [29].
Due to the localized force balance , which can be related to in Eq. (3) (see Steinmann [28]), only the first term remains, representing the energy release due to configurational changes
where the material traction is the projection of the Eshelby stress to the outwards-pointing normal
Remark 1. An important observation that indicates the physical significance of the introduced energy release is its relation to the classical definition of the (vector-valued) -integral. This relation, as shown in Steinmann et al. [29] for a singularity inside , is given by
The Eshelby stress can alternatively be represented in energy-momentum format (Eshelby, [30]; Hill, [31]) as follows
The beauty of this formulation is that the Eshelby stress can be computed directly from a spatial motion problem in a post-processing step.
Role of configurational forces
The configurational problem, which leads to the concept of energy release (Eq. (6)) and its close relation to the well-known -integral (Remark 1), forms a basis for the so-called Configurational Force Method (alt.: Material Force Method). To compose discrete configurational forces, we consider a discrete (FE) system for the weak form of the configurational problem with material reference
where is an element shape function associated with the th node and is the assembly operator over all finite elements. Assuming a homogeneous material and no body forces, we arrive at the definition of a discrete configurational force for the th global node
Since configurational forces are the configurational equivalent of deformational forces in a deformational problem, they serve as the driving forces for the assessment and propagation of material defects, like cracks or dislocations. This is due to the fact that they are power conjugated to configurational changes of the material nodal positions . For a deeper understanding of the physical meaning of configurational forces, refer to Maugin [14]; Steinmann et al. [29].
Remark 2. In the context of fracture mechanics, where are understood as design variables (which enable configurational changes), configurational forces are equivalent to the sensitivities in linear elastic setting as described in Van Goethem and Novotny [32].
Configurational forces can be easily computed in the post-processing step of a deformational problem using the energy-momentum format of the Eshelby stress (Eq. (9)). This is the basis of the Configurational Force Method. Some application examples of the Configurational Force Method include finite element mesh adaptivity (Mueller et al. [20]; Scherer et al. [33]), fracture mechanics (Steinmann, [15]; Steinmann et al., [34]; Mueller and Maugin, [35]; Nguyen et al., [36]; Moreno-Mateos and Steinmann, [19]), or crystal defects (Steinmann, [37]).
h-Adaptive meshing using configurational forces
As shown in previous works [38, 20, 33], configurational forces appear to be an excellent criterion for mesh control within a finite element context. Firstly, they provide a physics-based information on potential configurational changes such as defect motion, being power conjugated to the material variations of a discretized structure. Secondly, as purely discrete (nodal) quantities, their values strictly depend on the finite element discretization, i.e. their magnitude is proportional to the size of elements contributing to a particular node. In general, we can say that configurational forces provide an information about potential energy release due to configurational changes within the finite element discretization.
As previously discussed, mesh adaptivity in the context of topology optimization has been addressed either for improved design resolution using density-based criteria or for computational accuracy and efficiency using error-based criteria. In the following, we exploit configurational forces for mesh adaptivity in topology optimization to simultaneously target improved design resolution and computational accuracy and efficiency. This is possible due to the nature of configurational forces, since all hypothetical defects (within homogeneous materials and without pre-existing defects) can only occur at the boundary (they indicate where the boundary is) and especially in highly stressed regions (they indicate where the computational accuracy is needed).
The strategy presented in this work addresses topology optimization using structured meshes. For problems involving unstructured meshes, in addition to h-adaptivity, r-adaptivity presents an attractive alternative, as it can leverage both the magnitude and orientation of configurational forces [38, 33].
Eshelby stress in topology optimization
The Eshelby stress Eq. (9) and consequently the configurational forces Eq. (11) have to be adapted to the pseudo-density field of the SIMP (Solid Isotropic Material with Penalization) method [39, 40]. We use a variant of the relaxation according to Cheng and Guo [41] and interpolate the Eshelby stress as follows
where denotes a regularized pseudo-density . The regularization procedure follows filtering through the Helmholtz-type PDE filter [42] with boundary energy term as proposed by Wallin et al. [43] to account for improved behavior at the boundaries of the design domain. The smoothed Heaviside projection [44] is then applied to the filtered density to provide a crisp design representation. The Eshelby stress is calculated in a standard way according to Eq. (9). The configurational forces then simply utilize the regularized Eshelby stress tensor
Refinement and coarsening criteria
In the following we define simple criteria for the configurational force-based refinement and coarsening
where is the maximum value of a configurational force in the current optimization iteration, i.e. it changes for every iteration. For each finite element , the criterion is evaluated for the nodes belonging to it. The parameters and are user-defined multiplicators to set the threshold for refinement and coarsening, respectively. The adaptivity is implemented according to the following procedure:
In Algorithm 1, a finite element cell is marked for refinement if the magnitude of the configurational force at any of its vertices satisfies the refinement criterion from Eq. (14). However, a finite element cell is marked for coarsening only if the magnitude of the configurational force at
Input: Number of elements , Number of vertices per element Output: Refined and/or coarsened mesh for to do for to do if then else if then if is true then Refine the cell else if then Coarsen the cell
Algorithm 1: Mesh adaptivity in topology optimization based on configurational forces.
all of its vertices satisfies the coarsening criterion from Eq. (14). In Figure 3 the logic of Algorithm 1 is shown for better understanding.
Figure 3. Possible cases of criteria fulfillment for refinement and coarsening based on Algorithm 1. A cell is marked for refinement if at least one of its nodes fulfills the refinement criterion from Eq. (14), see the green force in the upper left image. A cell is marked for coarsening if all of its nodes satisfy the coarsening criterion from Eq. (14), see the blue forces in the bottom-left image.
The reason why the coarsening criterion is more conservative (all nodes of a cell must satisfy the coarsening criterion instead of just one) is due to the simple fact that in a structured mesh, four adjacent cells belonging to a so-called parent cell must be coarsened at the same time, see Figure 4. Thus, coarsening a cell affects a four times larger area and may affect regions where coarsening is not desired. Therefore, it seems reasonable to approach cell coarsening with more caution.
Figure 4. Coarsening is applied to four cells belonging to the same parent (inactive) cell, which is marked with a blue frame. If one of the child cells is marked for coarsening, deal.II takes care of marking the remaining three cells for coarsening, provided they were not marked for refinement.
According to the sequence of if conditions in Algorithm 1, refinement has priority over coarsening. Figure 5 shows a case where this is relevant, i.e. where two cells belonging to the same parent cell (a potentially larger cell that would be "created" during coarsening) are marked for refinement and coarsening, respectively. This order follows the obvious reasoning that we prioritize geometric and numerical accuracy over computational efficiency.
Figure 5. An edge case where two adjacent cells are marked for refinement and coarsening. In this case, only the refinement will occur, since the coarsening would affect the cell marked for refinement and therefore will not happen.
Handling multilevel coarsening and refinement
Our structural optimization software utilizes the deal.II library [27] and its dependencies. Thus, the mesh model and compatibility aspects of the adaptive mesh are handled by deal.II internal functionalities. In general, the mesh object stores not only the active elements (actually used in the simulation), but also all its parent elements, representing the refinement history in a tree-like structure. This allows multi-level coarsening as the inactive parent elements are still available and can be reactivated, for more information see Bangerth et al. [27].
Since the mesh is perfectly structured and consists entirely of quadrilateral elements, an adaptively refined and coarsened mesh must deal with hanging nodes (Figure 6). The hanging nodes are geometrically constrained by the neighboring nodes , where is the discrete solution at a newly introduced hanging node.

Figure 6. Adaptively refined mesh with hanging nodes (green). Red nodes: the unchanged, original nodes before refinement; purple nodes: newly created nodes with normal connectivity; green nodes: the hanging nodes that have no connection to the element of the lower refinement level.
Compatibility criteria ensuring that refinement and coarsening do not lead to unwanted situations, such as adjacency between elements with a refinement level difference of two or the presence of so-called unrefined islands, are available and active by default.
Examples
To demonstrate the practical implications of configurational force-based (CNF-based) mesh adaptivity in topology optimization, two example studies are presented, a cantilever benchmark and a U-beam structure with stress singularity. To evaluate the merits of CNF-based mesh adaptivity, we compare it with density-based (DENS-based) and von Mises stress-based (VNM-based) adaptivity criteria. The following examples are modeled in nonlinear elastic setting, employing the Neo-Hookean material model with the Lamé constants: and . The finite element system applies biquadratic quadrilateral elements, resulting in 18 degrees of freedom per element.³
Comparison criteria
DENS-based mesh adaptivity is given by the following formula
where is the filtered density field in the element . Naturally, DENS-based refinement is primarily concerned with the geometric accuracy of the design by targeting the grey transition regions within the design. By choosing the filtered density , as opposed to the raw density or the projected density , we ensure the presence of grey densities () at the boundary of the design.
VNM-based mesh adaptivity criterion employs von Mises stress () and targets numerical accuracy, especially for stress-constrained problems or structures with stress concentration regions. We choose the same strategy as for the CNF-based adaptivity of Eq. (14)
Here, the multiplicators and are the same as for CNF-based adaptivity. Our numerical experiments confirmed that reasonable values are and for both mesh adaptivity strategies.
Cantilever
For CNF-based mesh adaptivity, it seems reasonable to target problems that contain stress concentrations and/or are stress constrained. This is due to the nature of configurational forces, which often serve as a criterion for configurational changes within a structure. However, to provide a better reference, we first consider the classic benchmark problem, which is the compliance minimization of a cantilever (see Figure 7), given as

Figure 7. Setup of the cantilever benchmark and the initial mesh with a refinement level 1.
where is the compliance functional, is the displacement and is the volume function. To define the compliance, we assume the absence of body forces. For a discussion on well-posedness of Eq. (17), see Bendsoe and Sigmund [45].
The following test cases are primarily concerned with the final mesh and the resulting geometric accuracy and computational efficiency. Therefore, the obtained designs are presented in the format shown in Figure 8. That is, instead of showing the standard black and white density distribution, we have adopted a blue to grey scale for densities, which is applied to the element edges instead of the element fill. This way a clean representation of the final design is shown, displaying not only the final design (density distribution) but also the final mesh in both the solid and void phase.
Figure 8. Example of a final design shown on top, using a black-to-white scale to fill the cells. In the bottom an alternative design representation features the mesh additionally. Here, a wireframe representation and a blue-to-gray color scale for the densities are chosen to maintain the visibility of the void regions and to avoid possible confusion with the black-to-white scale.
In the following study (see Table 1) we explored different intervals of mesh adaptation. That is, the mesh adaptation based on each of the above criteria takes place every th iteration.
| Common settings | ; ; Max ref. level = 4 | |
| Case no. | Adaptivity interval | Initial ref. level |
| Row 1 | 1 iteration | 1 |
| Row 2 | 5 iterations | 1 |
| Row 3 | 25 iterations | 1 |
| Row 4 | 5 iterations | 4 |
Table 1. Summary of the parameters studied in the cantilever benchmark. Three different cases of adaptivity interval and one case of fully refined initial mesh are compared. The row numbering corresponds to the images in Figure 9.
For three different adaptivity intervals, i.e. every 1, 5, or 25 iterations, the initial refinement level is 1. The base mesh consists of 20x10 elements and does not allow further coarsening (it does not store parent cell information for level 0 elements). Using an initial refinement level of 1 renders a mesh of 40x20 elements and provides a good starting point for further refinement while still allowing for one level of coarser elements where needed. The maximum refinement level is limited to 4. In addition, a special case is studied where we start with a mesh at refinement level 4 with a mesh adaptation interval of 5 iterations. This case is particularly interesting because, to the best of our knowledge, such a multi-level coarsening problem has not been studied before in the context of topology optimization.
Figure 9 shows the final results for the cases from Table 1. The DENS-based and CNF-based adaptivity strategies produce consistent results for all adaptivity intervals and initial refinement levels. The elements are refined along the boundaries of the structure and coarsened outside and inside the structure, which is desired for both geometric accuracy and computational efficiency. More frequent mesh refinement results in more detailed designs. The dependence of the geometric complexity on the adaptivity interval is pronounced for DENS-based and CNF-based adaptivity. The VNM-based mesh adaptivity strategy generally does not render geometrically refined structures because the stress values are highest near the Dirichlet and Neumann boundary conditions. In addition, it is highly dependent on the initial level of refinement, as for the fully refined initial structure it only coarsens void cells. However, it is important to keep in mind that VNM adaptivity is not a reasonable criterion for cantilever-type problems where the stress response is not relevant. The efficiency of VNM-based adaptivity is studied more carefully in the sequel for the U-beam problem.
Figure 9. Final designs using the wireframe representation for the cantilever study. The columns are divided according to the mesh adaptivity technique and the rows according to the mesh adaptivity interval (Int) and the initial mesh refinement level (InL).
To address computational efficiency, the mesh evolution for selected iterations is shown in Figure 10 for the cases from the second row of Table 1, i.e. for the adaptivity interval of 5 iterations. In addition, the objective function (Figure 11, top plot) the number of cells per iteration (Figure 11, middle plot) and the accumulated computation time per iteration (Figure 11, bottom plot) are plotted for all cases from Table 1. The analysis of the mesh evolution reveals the difference between DENS-based and CNF-based adaptivity. Due to the dominance of intermediate densities in the early iterations, the DENS-based adaptivity refines the entire mesh, which costs a significant amount of computation, as can be seen in Figure 11 (bottom plot) for the first 50 iterations. In the case of CNF-based adaptivity, few cells were initially refined, which did not significantly affect the performance. In the end, the CNF-based adaptivity case required only 21% of the computational time of the DENS-based adaptivity case. The evolution and final values of the objective function show no major differences between the selected methodologies, with the only stand out case being the VNM-based adaptivity producing higher compliance values due to less refined boundaries. As a result of CNF mesh adaptivity, we obtain a uniform distribution of configurational forces
Figure 10. Mesh evolution for an adaptivity interval of 5 iterations. Due to the dominance of the volume constraint, the DENS-based criterion refines the mesh globally in the initial iterations, which consumes a significant amount of computational time.

Figure 11. Evolution of the objective value (top), the number of cells (middle) and accumulated computation time (bottom) over optimization iterations for the cantilever benchmark. The coloring is assigned according to the mesh adaptivity method.
along the design boundaries, as shown in Figure 12. This is a particularly desirable state when fracture resistance is of interest.

Figure 12. Configurational forces for the final cantilever design using CNF mesh adaptivity.
A general conclusion from the cantilever study is that CNF-based mesh adaptivity is able to provide a well-defined geometry, similar to DENS-based mesh adaptivity, with improved robustness in the early iterations and resulting significant computational time saving. In contrast, VNM mesh adaptivity is not a reasonable choice for non-stress-constrained problems, so its comparison is more useful in the following example.
U-beam
A more sophisticated example in the context of configurational mechanics is a U-beam (Figure 13), which contains stress concentration regions (inner corners) that are natural candidates for material failure.

Figure 13. Setup of a U-beam example with an initial mesh of refinement level 1. By taking advantage of symmetry, only half of the structure is considered for optimization.
In this case, the introduction of a stress constraint in the optimization setup is essential to obtain a desirable structure that mitigates the
stress buildup. Therefore, the optimization problem is defined as follows
where is a volume normalized P-norm stress formulation, where the P-norm is given by
and
is a stress measure that employs the -relaxed (, Cheng and Guo [41]) von Mises stress normalized with respect to the user-defined stress limit , which is also referred to in this example by the abbreviation "SL". This formulation ensures that for a given stress limit , the stress constraint is satisfied if .
This time, the case study focuses on the stress limit and the adaptivity interval as the main parameters, see Table 2. The initial refinement level of 1 is used for all cases.
| Common set.: ; ; Init/Max ref. level | ||
| Case no. | Stress limit | Adaptivity interval |
| Row 1 | 0.3 | 5 |
| Row 2 | 0.3 | 25 |
| Row 3 | 0.5 | 5 |
| Row 4 | 0.5 | 25 |
Table 2. Summary of the parameters studied in the U-beam example. The row numbering corresponds to the images in Figure 14.
The final designs are shown in Figure 14 and the corresponding plots for the objective function, the number of cells and the accumulated computation time per iteration are shown in Figure 15, respectively. Upon visual inspection of the final designs in Figure 14, several observations can be made. DENS-based adaptivity is capable of generating geometrically detailed and smooth designs. However, for the top three cases, there were some artifacts within the solid phase, resulting in locally high refinement. The DENS-based adaptivity case with ( - stress limit) failed to refine the mesh around the stress concentration region, resulting in an inaccurate stress field and consequently failed to enforce a smooth geometry without a sharp corner. The VNM adaptivity naturally prioritized the highly stressed corner for refinement, but it did not render geometrically

Figure 14. Final designs using the wireframe representation for the U-beam study. The columns are divided according to the mesh adaptivity technique and the rows according to the stress limit (SL) and mesh adaptivity interval (Int).

Figure 15. Evolution of the objective value (top), the number of cells (middle) and accumulated computation time (bottom) over optimization iterations for the U-beam. The coloring is assigned according to the mesh adaptivity method.
accurate boundaries along low stress regions. The stricter constraint value of results in larger fully refined regions, which affects the computation time. CNF-based adaptivity renders a consistent set of results, with full refinement only along the formed structure edges for the cases with , guaranteeing both geometric and stress accuracy. Moreover, for the cases , DENS-based and VNM-based adaptivity produces complementary designs whose refinements add up to the design with CNF-based adaptivity. By evaluating the plots in Figure 15, a similar tendency as for the cantilever benchmark is observed, i.e. a large computational cost for DENS adaptivity with the adaptivity interval . The objective value plot (Figure 15, top plot) depicts no significant differences for the final iterations between the studies cases. As expected, the less restrictive stress constraints of allowed for a slightly more improved value of the final compliance.
Due to the presence of stress concentration region, the U-beam is particularly susceptible to failure. Therefore, it is essential to analyze the configurational forces in the final designs. Figure 16 shows the final designs for the case together with the configurational forces. Their magnitudes are scaled consistently for all three examples. The dependence of the configurational forces on the element size is immediately apparent, especially for the DENS-based adaptivity. This highlights the importance of an equivalently refined mesh along the structure boundaries to allow a reasonable interpretation of the configurational forces. In the DENS-based adaptivity case, the largest magnitude of the force is naturally found around the unrefined stress concentration region. In the CNF-based adaptivity case, due to the consistently refined mesh along the stress concentration region, the distribution of forces is less variable in magnitude, but still peaks in this corner.
Figure 16. Configurational forces in the final designs for the setup SL/Int = 0.5/25, between the studied mesh adaptivity methods. For better comparison, the forces use the same scale in all cases.
Conclusions and outlook
In order to obtain geometrically accurate designs with uncompromising computational cost, a novel combined mesh refinement and coarsening strategy based on configurational forces is introduced in the context of topology optimization. Configurational forces provide information about possible energy release due to configurational changes in a discrete system. Therefore, they identify critical regions in a structure that may be subject to failure due to fracture or dislocation motion. These regions usually require the highest accuracy and, consequently, an appropriate discretization. We have shown that, for topology optimization, mesh adaptivity based on configurational forces produces a highly refined mesh along the edges of the structure, providing targeted refinement for maximum geometric accuracy, similar to density-based mesh adaptivity. In addition, unlike density-based mesh refinement, high stress corners, in the example of the U-beam, are always maximally refined, ensuring highly accurate stress calculations. On the other hand, mesh refinement based on the von Mises stress value does not guarantee geometric accuracy along the edges of the structure as it only focuses on high stress regions. Furthermore, mesh adaptivity based on configurational forces is robust in terms of computational effort, providing consistent computational times regardless of the mesh adaptivity interval. Therefore, we can conclude that mesh adaptivity based on configurational forces combines the advantages of density-based and von Mises-based mesh adaptivity and is robust to different problem setups.
Since configurational forces are a vector field, they can be used for an anisotropic type of mesh adaptivity, i.e., partial cell refinement along a selected direction, e.g., only along the x-axis. Future research will focus on anisotropic mesh adaptivity to further improve computational efficiency. In addition, mesh adaptivity based on configurational forces can be easily extended to three-dimensional problems.
Acknowledgements
The authors gratefully acknowledge financial support for this work by the European Research Council (ERC) under the Horizon Europe program (Grant-No. 101052785, project: SoftFrac).
Author contributions
Conceptualization: Gabriel Stankiewicz; Methodology: Gabriel Stankiewicz; Implementation: Gabriel Stankiewicz and Chaitanya Dev; Discussions and improvement of the methodology: all authors; Writing - original draft preparation: Gabriel Stankiewicz; Writing - review and editing: all authors; Funding acquisition: Paul Steinmann, Resources: Paul Steinmann, Supervision: Paul Steinmann.
Funding
This work is supported by the European Research Council (ERC) under the Horizon Europe program (Grant-No. 101052785, project: SoftFrac).
Data availability
The version of the code, the executable, the parameter settings and the result files are available from the corresponding author upon request.
Declarations
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
Ethical approval
This study does not involve human participants, animal subjects, or ethical concerns requiring institutional approval.
Replication of results
All the presented methodology is implemented in C++ utilizing the finite element library deal.II [27] (Arndt et al. 2021). The version of the code, the executable, the parameter settings and the result files are available from the corresponding author upon request.
References
References
- [1]Costa Jr, J.C.A., Alves, M.K.: Layout optimization with h-adaptivity of structures. International Journal for Numerical Methods in Engineering 58(1), 83–102 (2003)DOI
- [2]Stainko, R.: An adaptive multilevel approach to the minimal compliance problem in topology optimization. Communications in numerical methods in engineering 22(2), 109–118 (2006)DOI
- [3]Nana, A., Cuillère, J.-C., Francois, V.: Towards adaptive topology optimization. Advances in Engineering Software 100, 290–307 (2016)DOI
- [4]Wang, S., Sturler, E., Paulino, G.H.: Dynamic adaptive mesh refinement for topology optimization. arXiv preprint arXiv:1009.4975 (2010)arxiv.org/abs/1009.4975
- [5]Troya, M.A., Tortorelli, D.A.: Adaptive mesh refinement in stress-constrained topology optimization. Structural and Multidisciplinary Optimization 58(6), 2369–2386 (2018)
- [6]Troya, M.A., Tortorelli, D.A.: Three-dimensional adaptive mesh refinement in stress-constrained topology optimization. Structural and Multidisciplinary Optimization 62(5), 2467–2479 (2020)
- [7]Muñoz, D., Albelda, J., Ródenas, J.J., Nadal, E.: Improvement in 3d topology optimization with h-adaptive refinement using the cartesian grid finite element method. International Journal for Numerical Methods in Engineering 123(13), 3045–3072 (2022)DOI
- [8]Lambe, A.B., Czekanski, A.: Topology optimization using a continuous density field and adaptive mesh refinement. International Journal for Numerical Methods in Engineering 113(3), 357–373 (2018)DOI
- [9]Zhang, S., Gain, A.L., Norato, J.A.: Adaptive mesh refinement for topology optimization with discrete geometric components. Computer Methods in Applied Mechanics and Engineering 364, 112930 (2020)DOI
- [10]Stankiewicz, G., Dev, C., Steinmann, P.: Coupled topology and shape optimization using an embedding domain discretization method. Structural and Multidisciplinary Optimization 64(4), 2687–2707 (2021)DOI
- [11]Dev, C., Stankiewicz, G., Steinmann, P.: Sequential topology and shape optimization framework to design compliant mechanisms with boundary stress constraints. Structural and Multidisciplinary Optimization 65(6), 180 (2022)DOI
- [12]Stankiewicz, G., Dev, C., Steinmann, P.: Geometrically nonlinear design of compliant mechanisms: Topology and shape optimization with stress and curvature constraints. Computer Methods in Applied Mechanics and Engineering 397, 115161 (2022)DOI
- [13]Stankiewicz, G., Dev, C., Weichelt, M., Fey, T., Steinmann, P.: Towards advanced piezoelectric metamaterial design via combined topology and shape optimization. Structural and Multidisciplinary Optimization 67(2), 26 (2024)DOI
- [14]Maugin, G.r.A.: Material forces: concepts and applications (1995)DOI
- [15]Steinmann, P.: Application of material forces to hyperelastostatic fracture mechanics. i. continuum mechanical setting. International Journal of Solids and Structures 37(48-50), 7371–7391 (2000)DOI
- [19]Moreno-Mateos, M.A., Steinmann, P.: Configurational force method enables fracture assessment in soft materials. Journal of the Mechanics and Physics of Solids 186, 105602 (2024)
- [20]Mueller, R., Gross, D., Maugin, G.: Use of material forces in adaptive finite element methods. Computational Mechanics 33, 421–434 (2004)DOI
- [22]Welschinger, F., Hofacker, M., Miehe, C.: Configurational-force-based adaptive fe solver for a phase field model of fracture. PAMM 10(1), 689–692 (2010)DOI
- [23]Hénap, G., Szabó, L.: On the configurational-force-based r-adaptive mesh refinement in isogeometric analysis. Finite Elements in Analysis and Design 124, 1–6 (2017)DOI
- [24]Hénap, G., Szabó, L.: On numerical solution of elastic-plastic problems by using configurational force driven adaptive methods. Finite Elements in Analysis and Design 92, 50–59 (2014)DOI
- [25]Qinami, A., Bryant, E.C., Sun, W., Kaliske, M.: Circumventing mesh bias by r-and h-adaptive techniques for variational eigenfracture. International Journal of Fracture 220(2), 129–142 (2019)DOI
- [26]Askes, H., Bargmann, S., Kuhl, E., Steinmann, P.: Structural optimization by simultaneous equilibration of spatial and material forces. Communications in numerical methods in engineering 21(8), 433–442 (2005)DOI
- [27]Bangerth, W., Hartmann, R., Kanschat, G.: deal. ii—a general-purpose object-oriented finite element library. ACM Transactions on Mathematical Software (TOMS) 33(4), 24 (2007)DOI
- [28]Steinmann, P.: Spatial and Material Forces in Nonlinear Continuum Mechanics: A Dissipation-Consistent Approach. Solid Mechanics and Its Applications, vol. 272, p. 395. Springer, ??? (2022)DOI
- [29]Steinmann, P., Scherer, M., Denzer, R.: Secret and joy of configurational mechanics: From foundations in continuum mechanics to applications in computational mechanics. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik: Applied Mathematics and Mechanics 89(8), 614–630 (2009)DOI
- [30]Eshelby, J.: The elastic energy-momentum tensor. Journal of elasticity 5, 321–335 (1975)
- [31]Hill, R.: Energy-momentum tensors in elastostatics: some reflections on the general theory. Journal of the Mechanics and Physics of Solids 34(3), 305–317 (1986)DOI
- [32]Van Goethem, N., Novotny, A.: Crack nucleation sensitivity analysis. Mathematical Methods in the Applied Sciences 33(16), 1978–1994 (2010)DOI
- [33]Scherer, M., Denzer, R., Steinmann, P.: Energy-based r-adaptivity: a solution strategy and applications to fracture mechanics. International journal of fracture 147, 117–132 (2007)
- [34]Steinmann, P., Ackermann, D., Barth, F.: Application of material forces to hyperelastostatic fracture mechanics. ii. computational setting. International Journal of Solids and Structures 38(32-33), 5509–5526 (2001)DOI
- [35]Mueller, R., Maugin, G.: On material forces and finite element discretizations. Computational mechanics 29, 52–60 (2002)
- [36]Nguyen, T., Govindjee, S., Klein, P., Gao, H.: A material force method for inelastic fracture mechanics. Journal of the Mechanics and Physics of Solids 53(1), 91–121 (2005)DOI
- [37]Steinmann, P.: On spatial and material settings of hyperelastostatic crystal defects. Journal of the Mechanics and Physics of Solids 50(8), 1743–1766 (2002)DOI
- [38]Thoutireddy, P., Ortiz, M.: A variational r-adaption and shape-optimization method for finite-deformation elasticity. International Journal for Numerical Methods in Engineering 61(1), 1–21 (2004)
- [39]Bendsøe, M.P.: Optimal shape design as a material distribution problem. Structural optimization 1, 193–202 (1989)
- [40]Bendsøe, M.P., Sigmund, O.: Material interpolation schemes in topology optimization. Archive of applied mechanics 69, 635–654 (1999)
- [41]Cheng, G.D., Guo, X.: ε-relaxed approach in structural topology optimization. Structural optimization 13, 258–266 (1997)
- [42]Lazarov, B.S., Sigmund, O.: Filters in topology optimization based on helmholtz-type differential equations. International journal for numerical methods in engineering 86(6), 765–781 (2011)DOI
- [43]Wallin, M., Ivarsson, N., Amir, O., Tortorelli, D.: Consistent boundary conditions for pde filter regularization in topology optimization. Structural and Multidisciplinary Optimization 62, 1299–1311 (2020)DOI
- [44]Wang, F., Lazarov, B.S., Sigmund, O.: On projection methods, convergence and robust formulations in topology optimization. Structural and multidisciplinary optimization 43, 767–784 (2011)DOI
- [45]Bendsoe, M.P., Sigmund, O.: Topology Optimization: Theory, Methods, and Applications. Springer, ??? (2013)








