betaFeedback

WorksComputer Science

Design and Optimization of Functionally-graded Triangular Lattices for Multiple Loading Conditions

Abstract

Aligning lattices based on local stress distribution is crucial for achieving exceptional structural stiffness. However, this aspect has primarily been investigated under a single load condition, where stress in 2D can be described by two orthogonal principal stress directions. In this paper, we introduce a novel approach for designing and optimizing triangular lattice structures to accommodate multiple loading conditions, which means multiple stress fields. Our method comprises two main steps: homogenization-based topology optimization and geometry-based de-homogenization. To ensure the geometric regularity of triangular lattices, we propose a simplified version of the general rank-33 laminate and parameterize the design domain using equilateral triangles with unique thickness per edge. During optimization, the thicknesses and orientation of each equilateral triangle are adjusted based on the homogenized properties of triangular lattices. Our numerical findings demonstrate that this proposed simplification results in only a slight decrease in stiffness, while achieving triangular lattice structures with a compelling geometric regularity. In geometry-based de-homogenization, we adopt a field-aligned triangulation approach to generate a globally consistent triangle mesh, with each triangle oriented according to the optimized orientation field. Our approach for handling multiple loading conditions, akin to de-homogenization techniques for single loading conditions, yields highly detailed, optimized, spatially varying lattice structures. The method is computationally efficient, as simulations and optimizations are conducted at a low-resolution discretization of the design domain. Furthermore, since our approach is geometry-based, obtained structures are encoded into a compact geometric format that facilitates downstream operations such as editing and fabrication.

Introduction

Topology optimization and lattice infill are primary strategies for designing lightweight structures for additive manufacturing. The integration of these two strategies has been a topic of intensive research in the past decade [20]. Topology optimization transforms structural design into an optimization problem of determining where to place material, while lattice infill can be considered as a form of metamaterial. Topology optimization with lattice infill results in multi-scale structures that are lightweight and robust [5, 34]. For an overview of topology optimization methods for designing multi-scale structures, please refer to the recent review article by Wu et al. [36].

[email protected] (J. Wang); [email protected] (R. Westermann); [email protected] (X. Gao); [email protected] (J. Wu)

ORCID(s): 0000-0002-4607-844X (J. Wang); 0000-0002-3394-0731 (R. Westermann); 0000-0003-0829-7075 (X. Gao); 0000-0003-4237-1806 (J. Wu)

For designing load-bearing structures, tailoring the layout of lattice infill based on the stress distribution is key to achieving exceptional structural performance. Lattice infill can be categorized as isotropic or anisotropic. By aligning the anisotropy of the lattice with the principal stress direction at each location, the structural rigidity can be significantly increased [22]. This principle has been well explored and several approaches have been developed [18, 6]. Among them, an effective and efficient approach is de-homogenization [21, 9, 37]. It involves two major steps. The first step, homogenization-based topology optimization, specifies the optimal lattice infill configuration at each location in the design domain [3]. The second step, de-homogenization, is crucial in this approach and it transforms the optimized specifications into a globally consistent structure.

De-homogenization approaches can be classified, depending on the representation of optimized designs, into image-based approaches and geometry-based approaches. Image-based approaches represent the structural layout using binary density fields, while geometry-based approaches represent it using a compact geometric format. For example, a truss-like structure can be represented by a mesh where each edge in the mesh represents a truss. Geometry-based approaches offer direct control of geometric features, which is beneficial in, e.g., addressing manufacturing constraints [14]. The geometric format also offers a compact parameter space for subsequent editing operations on the optimized structures [37, 30]. From a computational perspective, image-based approaches center on computing a fine-grid scalar field, whose gradients are aligned with the orientation of optimized lattice infill from homogenization-based topology optimization, e.g., the projection-based methods [21]. Geometric approaches, in contrast, involve directly computing a set of field-aligned geometric primitives to represent the optimized lattice infill, e.g., the conforming lattice structure [37].

De-homogenization for a single loading condition is already a prosperous realm. The projection-based approach proposed by Pantz and Trabelsi [21] was simplified by Groen and Sigmund [9] and further developed by Allaire et al. [1]. It has since been extended to design shell-infill structures [11], and to work in 3D [10, 8]. Different approaches for the projection have been proposed, from related research fields such as visualization and computer graphics, using stream surfaces [25], convolutional neural networks [7], and phasor noises [33]. Geometric approaches include those based on field-aligned hex-dominant meshing [37] and streamline tracing [30]. The singularity issue in 2D has been investigated respectively in the image- and geometry-based routines by Stutz et al. [24] and Wang et al. [30].

The development of de-homogenization approaches considers almost exclusively a single loading condition, i.e., constant external loads. In many engineering applications, however, multiple loading conditions are omnipresent. For instance, in airplane design, the major loads acting on the aircraft wing vary at different stages during the flight, including lift and drag, gravitational forces (particularly when the fuel tanks are fully loaded), propulsion forces from the forward engines, and opposing thrust from thrust reversers. All these loads are exerted on the airplane during the flight, thus requiring to incorporate them into the structural design and optimization process.

De-homogenization under multiple loading conditions, however, is challenging. Firstly, the optimized metamaterial properties can be realized by different lattice configurations, i.e., the optimal lattice configurations are not uniquely defined. Furthermore, the optimized lattice configuration exhibits severe discontinuities. This further complicates the de-homogenization of a consistent structural layout. A first solution to these challenges was recently proposed by Jensen et al. [16]. They developed an image-based method, extending the projection approach in [9] from single to multiple loading conditions. The rank-3 laminate was used as the lattice infill for homogenization. It describes a composite material composed of three differently oriented material layers, where each material layer is described by an orientation variable and a width variable. I.e., a rank-3 laminate involves three independent width variables and three orientation variables. The structural designs by this approach achieve comparable mechanical performance (i.e., stiffness) with the ones obtained from conventional density-based topology optimization, while the latter incurs a significantly larger computational burden. The image-based projection method inherently requires the integrability of the optimized specifications, making it necessary to ensure the smoothness of the optimized orientation field. This, however, becomes difficult due to the non-uniqueness of the optimal solutions of the homogenization model. In addition, as the three orientation fields are de-homogenized independently, the sub-structures aligning with different orientation fields can come very close to each other, leaving small void regions in the final designs. While this feature does not necessarily impact mechanical performance, it may not be the preferred choice in certain design or manufacturing scenarios due to the increased geometric complexity it introduces. See the close-up view of Figure 1b for an illustration of irregular intersections of three orientation fields.

Image not extracted

Figure 1. Method overview. (a) The optimization problem: Two loading conditions F1F_1 (orange) and F2F_2 (green) are applied, and the bottom is fixed. (b) The optimized density layout (background) and orientations (streamlines) generated by homogenization-based topology optimization. (c) The triangular parametrization with edges aligned to the tangents of the optimized rank-3 layers. The entire design domain and the optimized density layout are partitioned into sub-domains by this field-aligned mesh. (d) The triangular lattice design obtained via de-homogenization.

In this paper, we present a novel method to generate functionally graded triangular lattice structures that are resilient to multiple loading conditions. Our method follows the established de-homogenization pipeline, consisting two steps: (1) performing homogenization-based topology optimization to obtain a set of optimized specifications; (2) converting the optimized specifications into a graded triangular lattice structure. Figure 1 shows an overview of the proposed method.

Our technical contributions are two-fold. Firstly, we propose a simple and effective infill model for multiple loading conditions — triangular lattices. It can be viewed as a simplified version of the general rank-3 laminate. In triangular lattices, the width of each edge is allowed to change independently, creating a large range of attainable metamaterial properties. More importantly, the triangles are rotatable, allowing local alignment of their anisotropy with preferred directions. Secondly, we develop a geometric approach to translate the optimized specifications into a globally conforming triangular lattice structure, whose edges align with the three optimized orientation fields and whose edge thicknesses convey the mechanical anisotropy. Our de-homogenization approach is based on field-aligned triangulation [15]. This method does not require the integrability of the optimized specifications, as it is based on a local smoothing operator, avoiding global parameterization. In addition, compared to the randomly intersected substructures, our de-homogenization using equilateral triangles restricts the geometric complexity by carefully placing the intersections, resulting in triangular lattices with equal spacing, as shown in the close-up views of Figure 1c and Figure 1d.

The remainder of this paper is organized as follows. In § 2 we first present the triangular lattices by simplifying the general rank-3 laminate, and we introduce the corresponding equation of homogenization-based topology optimization. The construction of de-homogenization is presented in § 3. In § 4, we perform numerical experiments and use them to discuss and compare the proposed method. Finally, § 5 concludes the paper with a summary and an outlook.

Topology Optimization with Triangular Lattices

Triangular Lattices

Our triangular infill model is a simplified version of the general rank-3 laminate. Stiffness-optimal structures can be described by a set of spatially varying rank-NN laminates with local periodicity [3]. Rank-NN describes a composite material composed of NN differently oriented layers. For a single loading condition, the optimal laminate can be described as two orthogonally oriented layers, i.e., rank-2 laminate, which was simplified as a square cell with a rectangular hole in homogenization-based topology optimization [3]. However, under multiple loading conditions, there are more than two major stress directions in the structure, one thus needs to resort to the anisotropic rank-3 to achieve optimal stiffness. The optimal stiffness under multiple loading conditions may also be achieved by laminates with a higher rank, yet this introduces unnecessary computational complexity and numerical instability during optimization [16].

The rank-3 laminate is constructed by piling up three layers of materials that orient differently, assuming perfect bound condition among different layers (Figure 2b). On each layer, two phases of material are considered, i.e., stiff material (+) and compliant material (−). The stiff material indicates an isotropic solid material deposition, while the compliant material varies among the different layers. A detailed description of the composition of a rank-3 laminate can be found in existing works [17, 12, 16]. It is briefly explained here for completeness.

Schematic composition of a rank-3 laminate and its geometric interpretation

Figure 2. Schematic composition of a rank-3 laminate and its geometric interpretation. (a) The design domain and boundary conditions for optimization, where each simulation cell is depicted by a set of specific specifications of a rank-3 laminate. (b) The multi-scale rank-3 laminate in a selected simulation cell. The material is composed of three differently oriented layers, shown in (c), (d), and (e), respectively. α\alpha is the relative width of each layer. n\mathbf{n} and t\mathbf{t} refer to the normal and tangent direction of a layer, respectively. (f) The proposed single-scale geometric interpretation of the rank-3 laminate as an equilateral triangle, with a unique thickness per edge.

Let a macro-cell have unit size, with the base layer being layer 3, i.e., the rank-3 laminate is constructed by starting from layer 3 (Figure 2c). The solid material deposition (ρ3\rho_3) on layer 3 is then described by the width variable α3\alpha_3 and the compliant material 1−α31-\alpha_3. Here the compliant material is rank-2. Sequentially on layer 2 with width variable α2\alpha_2 for the solid material (Figure 2d), the remaining region to fill is 1−α31-\alpha_3. The actual consumption of solid material (ρ2\rho_2) on layer 2 is (1−α3)α2(1-\alpha_3)\alpha_2. Correspondingly, the compliant material on layer 2 is rank-1, and its volume fraction is (1−α3)(1−α2)(1-\alpha_3)(1-\alpha_2). On layer 1 with width variable α1\alpha_1 (Figure 2e), the solid material consumption (ρ1\rho_1) is (1−α3)(1−α2)α1(1-\alpha_3)(1-\alpha_2)\alpha_1, and the corresponding compliant material is ‘void’, with a ratio of (1−α3)(1−α2)(1−α1)(1-\alpha_3)(1-\alpha_2)(1-\alpha_1). The ‘void’ material is assigned a small Young’s modulus E−E^{-} to avoid numerical instabilities, yet which is negligible compared to the Young’s modulus E+E^{+} of the solid material (i.e., E−=10−6E+E^{-}=10^{-6}E^{+}). The Poisson’s ratio ν0\nu_0 of void material is the same as that of the solid material. The orientation of layer nn is described by its normal nn\mathbf{n}_n and tangent tn\mathbf{t}_n, as depicted in Figure 2. Through this process, the volume fraction of a rank-3 cell with unit size can be described by

ρ=1−∏n=13(1−αn).(1)\rho= 1-\prod_{n=1}^{3}(1-\alpha_n). \tag*{(1)}

Alternatively, the volume fraction can be described by

ρ=ρ1+ρ2+ρ3.(2)\rho=\rho_1+\rho_2+\rho_3. \tag*{(2)}

Here, ρ1=(1−α3)(1−α2)α1\rho_1=(1-\alpha_3)(1-\alpha_2)\alpha_1, ρ2=(1−α3)α2\rho_2=(1-\alpha_3)\alpha_2, and ρ3=α3\rho_3=\alpha_3.

An interpretation of a rank-3 laminate as a triangular is illustrated in Figure 2f. A general rank-3 laminate can be described by six parameters, i.e., three width variables (α1,α2,α3\alpha_1,\alpha_2,\alpha_3) and three orientation variables (θ1,θ2,θ3\theta_1,\theta_2,\theta_3). However, allowing the orientation variables to change independently of each other may result in triangles with bad aspect ratios (i.e., sliver triangles). To improve the geometric regularity of the de-homogenized triangular lattice, we propose to reduce the design space by allowing only equilateral triangles, i.e., by letting θ2=θ3+π3\theta_2 = \theta_3 + \frac{\pi}{3} and θ1=θ3+2π3\theta_1 = \theta_3 + \frac{2\pi}{3}. Consequently, the independent design variables become α1\alpha_1, α2\alpha_2, α3\alpha_3 and θ3\theta_3. The restriction of orientation variables has a consequence on the mechanical performance of optimized structures. Our experiments demonstrate that the decrease in stiffness under this restriction is mild — less than 5% in various tests. This restriction allows for significant improvement in the geometric regularity, and is also beneficial for numerical convergence, as it reduces the discontinuities in the optimized laminate configurations.

Homogenization-based Topology Optimization

In topology optimization, the objective function quantifies the design target with respect to the corresponding design variables. The optimization process minimizes the objective function by iteratively adjusting the design variables. For designing stiff structures, the objective function usually considers the strain energy, also known as compliance. For multiple loading cases the contributions from different loads need to be taken into consideration concurrently. Thus, a weighted summation of the strain energy subject to each loading condition is used as a compliance metric (CC):

C=∑q=1MwqFqTUq.(3)C = \sum_{q=1}^{M} w_q \mathbf{F}_q^{T}\mathbf{U}_q. \tag*{(3)}

Here, MM is the number of loading conditions, Fq\mathbf{F}_q is the nodal force vector of the qq-th loading condition and Uq\mathbf{U}_q is the corresponding displacement vector obtained by solving the static equilibrium equation using finite element analysis. wqw_q is the weighting factor of the qq-th loading condition, with wq=1Mw_q = \frac{1}{M}, q=1:Mq = 1 : M in our implementation.

Let OO be the objective function, which is a function concerning the compliance metric CC. By taking triangular lattices as design parameters, the homogenization-based topology optimization is formulated as

min⁡α1,α2,α3,θ3O(α1,α2,α3,θ3),(4)\underset{\alpha_1,\alpha_2,\alpha_3,\theta_3}{\min}\quad O(\alpha_1,\alpha_2,\alpha_3,\theta_3), \tag*{(4)}
subject toK(α1,α2,α3,θ3)Uq=Fq,q=1:M,(5)\text{subject to}\quad\mathbf{K}(\alpha_1,\alpha_2,\alpha_3,\theta_3)\mathbf{U}_q = \mathbf{F}_q,\quad q = 1 : M, \tag*{(5)}
1Ne∑eρe−VF≤0,(6)\frac{1}{N_e}\sum_e \rho_e - V_F \le0, \tag*{(6)}
Lmin⁡≤α1,α2,α3≤Lmax⁡,(7)L_{\min} \le\alpha_1,\alpha_2,\alpha_3 \le L_{\max}, \tag*{(7)}
−4π≤θ3≤4π.(8)-4\pi\le\theta_3 \le4\pi. \tag*{(8)}

Here, K\mathbf{K} is the global stiffness matrix assembled from the element stiffness matrices Ke\mathbf{K}_e. Ke=∫ΩeBTSB dx\mathbf{K}_e = \int_{\Omega_e} \mathbf{B}^{T}\mathbf{S}\mathbf{B}\,\mathrm{d}x, with B\mathbf{B} being the element strain matrix and S\mathbf{S} the elasticity tensor of the rank-3 laminate. The elasticity tensor of the rank-nn laminate can be determined analytically [13, 17, 12]. In Appendix § 10 we summarize the derivation of S\mathbf{S} for the rank-3 laminate. NeN_e is the number of simulation cells, ρe\rho_e corresponds to the relative density value of each cell (Equation (1)). VFV_F is the given threshold of volume fraction that controls the material budget. Lmin⁡L_{\min} and Lmax⁡L_{\max} describe the lower and upper bounds of the width values during optimization, respectively, with 0<Lmin⁡<Lmax⁡<10 < L_{\min} < L_{\max} < 1. The lower bound Lmin⁡L_{\min} is introduced to avoid the appearance of large regions that are fully empty in the design domain, and the upper bound Lmax⁡L_{\max} is used to avoid fully solid regions. We set the range of θ3\theta_3 according to [17].

A hybrid scheme is utilized to solve the optimization problem. The three width variables α1\alpha_1, α2\alpha_2, α3\alpha_3 are updated by a standard optimality criteria [2], and the orientation design variable θ\theta is updated by the Method of Moving Asymptotes (MMA) [26]. We found that a moving step size of 0.01 for the width variables and π180\frac{\pi}{180} for the orientation variable resulted in a good balance between convergence speed and quality of the optimized structure. A standard convolution filter [4, 28] is applied to the width and orientation design variables for improved numerical robustness.

Orientation regularization

As mentioned in subsection 2.1, when using an anisotropic rank-3 laminate the optimal solution is not unique. I.e., while the design variables of each cell can locally converge to the optimum, they might not form a continuous design specification globally. Since this prevents the generation of a consistent structure in de-homogenization, regularizations are needed to counteract or at least mitigate this non-uniqueness issue.

We adapt the orientation regularization method in [16] to alleviate the potential discontinuity of the optimized orientation θ3\theta_3. This is achieved by including a penalty term in the objective function. Specifically, we use a per-edge penalty function, i.e., for two adjacent elements eie_i and eje_j with shared edge eije_{ij}

Peijθ=12−12cos⁡(h0⋅(θ3(ei)−θ3(ej))).(9)P_{e_{ij}}^{\theta} = \frac{1}{2} - \frac{1}{2}\cos\left(h_0 \cdot\left(\theta_3(e_i) - \theta_3(e_j)\right)\right). \tag*{(9)}

For each cell eie_i, its adjacent cells are defined as the cells sharing an element edge with eie_i. PeijθP_{e_{ij}}^{\theta} with values between 0 and 1 quantifies the orientation deviation between two adjacent cells. Considering only the orientation variable, a rotation of an equilateral triangle by π3\frac{\pi}{3} is invariant. Thus we set h0=6h_0 = 6. The orientation deviation between adjacent cells becomes zero if θ3(ei)−θ3(ej)=kπ3\theta_3(e_i)-\theta_3(e_j)=k\frac{\pi}{3} (kk is an integer). Accordingly, the largest deviation occurs when θ3(ei)−θ3(ej)=kπ3+π6\theta_3(e_i)-\theta_3(e_j)=k\frac{\pi}{3}+\frac{\pi}{6}, leading to Peijθ=1P_{e_{ij}}^{\theta}=1.

Let NeiN_e^i be the number of cells adjacent to the considered cell eie_i, and PeijθP_{e_{ij}}^{\theta} the orientation deviation of a cell eie_i to its jj–th adjacent cell. Then, a global measure of the orientation deviation across the design domain is given by

P(θ)=∑i=1Ne∑j=1NeiPeijθ.(10)P^{(\theta)} = \sum_{i=1}^{N_e} \sum_{j=1}^{N_e^i} P_{e_{ij}}^{\theta}. \tag*{(10)}

In the end, the objective function for optimization is given by a weighted summation of the compliance and the orientation regularization

O(α1,α2,α3,θ3)=WCC∗+(1−W)PP∗.(11)O(\alpha_1,\alpha_2,\alpha_3,\theta_3) = W\frac{C}{C_*} + (1-W)\frac{P}{P_*}. \tag*{(11)}

The weighting factor WW takes on values in (0,1](0, 1]. C∗C^{*} and P∗P^{*}, respectively, refer to the compliance and the orientation regularization corresponding to the initial values of the design variables. Orientation regularization can be disabled by setting W=1W = 1. In all of our experiments, we have set W=0.5W = 0.5.

Initialization

The outcome of the non-convex optimization depends on the initial values of the width and orientation variables Φ\boldsymbol{\Phi}. Different initializations may lead to different density layouts and also affect the convergence behavior of the optimization process. In previous works, these dependencies have been demonstrated for single [31, 32] and multiple [16] loading conditions. The initialization is particularly important since the optimal solution of a rank-3 laminate is not unique, i.e., different values of the width and orientation variables may lead to the same elasticity tensor [27].

We take inspiration from the work by Jensen et al. [16], where the initial principal stress directions are used to construct the initialization. We adapt it for our triangular infill model. This adaptation considers that only the orientation of layer 3 is considered independently in our framework, i.e., only a single direction field is involved. Specifically, we first compute the principal stress fields of the fully solid domain subject to the MM loading conditions respectively, resulting in 2×M2 \times M principal stress values and the corresponding principal stress directions for each simulation cell. For each simulation cell, then, we choose the principal stress direction θp\boldsymbol{\theta}_p vector whose associated absolute principal stress value is largest as the initialization of the orientation of layer 3.

It’s worth noting that the orientation variable in rank-3 is represented by the normals of each layer, while the ‘dominant’ principal stress direction is supposed to indicate the layering orientation, i.e., the layer tangent. Thus, θp+π2\boldsymbol{\theta}_p + \frac{\pi}{2} is used as the initial value of θ3\boldsymbol{\theta}_3 in the optimization. For the initialization of the width variable, we simply let αn=VF3\alpha_n = \frac{V_F}{3}, n=1:3n = 1 : 3. Figure 3a shows the principal stress-based initialization Φstress\boldsymbol{\Phi}_{stress} via a set of streamlines generated from the direction field θp+π2\boldsymbol{\theta}_p + \frac{\pi}{2}. The optimization histories of the compliance CC and the orientation regularization PθP_{\theta} are shown in Figure 3b. The optimized design is shown in Figure 1. For comparison, Figure 3b also shows the optimization history of the compliance when a uniform initialization Φuniform\boldsymbol{\Phi}_{uniform} is used, where θ3=π2\theta_3 = \frac{\pi}{2}.

Streamlines generated from the initial orientation field and optimization histories

Figure 3. (a) Streamlines generated from the initial orientation field. (b) Optimization history of compliance and orientation regularization over the entire optimization process.

Here, to release more design space, orientation regularization is omitted. It is seen that the optimization process converges to a smaller compliance value when using Φstress\Phi_{stress} instead of Φuniform\Phi_{uniform}.

Figure 3a shows that the proposed initialization does not necessarily yield a globally continuous direction field. This, however, does not pose a problem in our approach. On the one hand, due to the use of orientation regularization, the direction field will be smoothed to the maximum extent during optimization. On the other hand, our proposed geometry-based de-homogenization is less sensitive to the potential local discontinuity.

De-homogenization using Field-aligned Triangulation

The proposed de-homogenization converts the spatially varying optimized specifications (α1\alpha_1, α2\alpha_2, α3\alpha_3, θ3\theta_3) into an explicitly represented structural design. This consists of three steps. Firstly, we create an equilateral triangle mesh with its edges following the layer orientations θ3\theta_3. Secondly, we assign a thickness to each edge of the mesh, determined based on the optimized layer widths (α1\alpha_1, α2\alpha_2, α3\alpha_3), creating a spatially graded triangular lattice structure. Last, we create a global lattice structure by correcting the inconsistency of edge thickness at the intersections, i.e., the nodes of the triangle mesh.

Field-aligned Triangulation

In the first step, instead of employing streamlines to obtain the mesh [29], which is hard to capture the common intersections for any three streamlines, we use a field-aligned triangulation algorithm [15] to generate an orientation-aligned equilateral triangle mesh. Here, we give an overview of this triangulation method and fit it into the current context. For the detailed mechanism of this method, refer to the original paper [15]. The complete pipeline of this triangulation method is composed of three subsequent stages:

  • Orientation field optimization. A direction field is computed. It consists of a set of directions to which the edges of the output mesh should be aligned. The smoothness of the orientation field is achieved by a Gauss-Seidel process.

  • Position field optimization. A local parameterization (vertex positions) is computed so that the edges between vertices are aligned with the optimized orientation field.

  • Mesh extraction. The computed orientation and position fields are turned into a graph structure from which a triangular mesh is constructed.

In the original field-aligned triangulation, the direction field is a smooth 6-RoSy field2 that aligns with the features on the boundary domain, obtained by optimizing from a set of randomly distributed vectors. In the context of de-homogenization, the direction field represents the optimized layer orientations. Thus, the original method is adjusted to take the optimized layer orientations as input, and then perform position field optimization and mesh extraction sequentially. This is illustrated in Figure 4. Given a grid mesh discretization of the domain, guided by our generated orientation field that is visualized by streamlines, we apply the triangulation method to convert the orientation field into a triangle mesh with its edges following the orientation field. In the following, we elaborate on the conversion of this

Domain discretization, position optimization, and output mesh

Figure 4. (a) The domain discretization (background) and streamlines in the three different orientation fields of an optimized 6-RoSy field. (b) The randomly distributed positions (and respective directions) used for position optimization. The distribution is generated by considering a pre-defined edge length of the output mesh and the resolution of the input grid. (c) The position field after optimization and the representative positions (green dots) used to construct the output mesh. (d) The output mesh whose edges align with the optimized orientation field. Orientations are indicated by arrows, and edge colors distinguish corresponding layer orientations. To reduce visual clutter, only three of the six pairwise bidirectional 6-RoSy field directions are illustrated.

Anisotropic triangle mesh

The de-homogenization is conducted in a per-triangle manner. As shown in Figure 5a, each triangular element covers a certain region in the domain. The target deposition ratio vk∗v_k^* of the kk-th triangular element is measured by ρkMk\frac{\rho_k}{M_k}, where MkM_k is the number of cells located in the region covered by the triangular element, and ρk\rho_k is the sum of the density values over all these cells. The available material in each triangular region should be re-distributed so that (i) a binary material layout is generated, (ii) a continuous transition at the element boundaries is obtained, and (iii) the specific depositions (layer widths optimized with respect to the object’s compliance) on each oriented layer of the cells is reflected in the binary material layout. I.e., the underlying anisotropy of the optimized cells should be preserved in the de-homogenized lattice elements. In cases where the triangular mesh has a similar or higher resolution than the rank-3 simulation cells, the specifications of the cells are resampled using bilinear interpolation so that each triangular element covers at least a certain number of cells, (e.g., ≥10\ge10).

Diagrammatic view of per-lattice de-homogenization

Figure 5. Diagrammatic view of per-lattice de-homogenization. (a) The field-aligned triangular element with the covered cells. θ~ij\tilde{\theta}_i^j refer to the tangents of layers i∈1,2,3i \in1,2,3 in element jj. Three arrow glyphs on each cell indicate the tangents. Corresponding width variables αij\alpha_i^j are encoded into the arrow lengths. (b) The representative values of widths and orientations of the cells in the triangular element. (c) Relation of the representative values of orientations of three layers to the corresponding edges of the triangular element by detecting their orientation deviations. (d) The de-homogenized triangular lattice element whose unique edge thicknesses reflect the representative values of widths. The solid regions correspond to the deposition ratio indicated in (a).

De-homogenization, in particular, requires that the de-homogenized triangular element maintains the deposition ratio after de-homogenization, i.e.,

ρkMk=A−Avoid(t)A(12)\frac{\rho_k}{M_k} = \frac{A-A_{\mathrm{void}}(t)}{A} \tag*{(12)}

Here, AA is the area of the triangular element, and AvoidA_{void} is the area of the void region left in the triangular element after de-homogenization (Figure 5d). tt is the thickness parameter that needs to be determined to satisfy the equation. The void area AvoidA_{void} is up to the edge thicknesses of the triangular element. To convey the anisotropy via the edge thickness, tt is scaled for each edge by the scaling factors α1a\alpha_1^a, α2a\alpha_2^a and α3a\alpha_3^a (Figure 5b), which are determined by the representative values of the actual material deposition on each layer.

For each cell jj in a triangular element, the material depositions of each layer are evaluated according to Equation (2), i.e.,

α^1j=(1−α3j)(1−α2j)α1jα^2j=(1−α3j)α2jα^3j=α3j.(13)\begin{aligned} \hat{\alpha}_1^j &= (1-\alpha_3^j)(1-\alpha_2^j)\alpha_1^j \\ \hat{\alpha}_2^j &= (1-\alpha_3^j)\alpha_2^j \\ \hat{\alpha}_3^j &= \alpha_3^j. \tag*{(13)} \end{aligned}

α1j\alpha_1^j, α2j\alpha_2^j, and α3j\alpha_3^j are the corresponding width variables of layer 1, 2, and 3 of the jj-th cell, respectively. From Equation (13), the representative values (α1a\alpha_1^a, α2a\alpha_2^a and α3a\alpha_3^a) of the actual material deposition along each layering orientation in a triangular element are given by α1a=∑j=1Mkα^1j\alpha_1^a=\sum_{j=1}^{M_k}\hat{\alpha}_1^j, α2a=∑j=1Mkα^2j\alpha_2^a=\sum_{j=1}^{M_k}\hat{\alpha}_2^j, and α3a=∑j=1Mkα^3j\alpha_3^a=\sum_{j=1}^{M_k}\hat{\alpha}_3^j, respectively. In practice, α1a\alpha_1^a, α2a\alpha_2^a and α3a\alpha_3^a are normalized by max⁡(α1a,α2a,α3a)\max(\alpha_1^a,\alpha_2^a,\alpha_3^a).

Relating each edge of the triangular element to the corresponding layer orientation is required to convey the anisotropy of the underlying cells. However, since the orientations vary across the cells, in general, representative orientations first need to be computed. Here, we make use of a weighted average strategy, i.e., the actual material deposition on each layer is used to weight the corresponding orientation. With θ~ij\tilde{\theta}_i^j referring to the tangent of layer ii of the jj-th cell within the triangular element, the weighted average is computed as

θ~ia=∑j=1Mkθ~ijα^ij∑j=1Mkα^ij,i=1:3.(14)\tilde{\theta}_i^a = \frac{\sum_{j=1}^{M_k}\tilde{\theta}_i^j\hat{\alpha}_i^j}{\sum_{j=1}^{M_k}\hat{\alpha}_i^j}, \qquad i = 1:3. \tag*{(14)}

We let the mesh edges correspond to θ~1a\tilde{\theta}_1^a, θ~2a\tilde{\theta}_2^a or θ~3a\tilde{\theta}_3^a, depending on which edge they have the least directional deviation, see Figure 5c.

Lattice structure

When considering triangular elements the per-lattice de-homogenization does not necessarily produce a well-posed structural design. This is because gaps are left at the mesh vertices when the thicknesses of mesh edges sharing a vertex vary significantly, see Figure 6a. To avoid this issue, we introduce additional small triangular patches to fill these gaps.

The structural designs after de-homogenization

Figure 6. The structural designs after de-homogenization. (a) Original result by only directly performing the per-lattice de-homogenization. (b) Improved result by the introduced triangular patches.

In particular, we first determine those target vertices that need to be fixed. This requires to inspect — and repair upon validation — three vertices in each lattice. At the vertex p1p_1 of a considered lattice, a gap-filling triangle S^p^1p^2p^3\hat{S}_{\hat{p}_1\hat{p}_2\hat{p}_3} is considered, see Figure 6b. p2p_2 and p3p_3 are the vertices of the corresponding internal empty triangles of the two adjacent lattices that share the same edges with the considered lattice. Then, it is checked whether the corresponding vertex of the internal empty triangle of the considered lattice p^1\hat{p}_1 is within the triangle Sp1p2p3S_{p_1p_2p_3}. If p^1\hat{p}_1 is outside of Sp1p2p3S_{p_1p_2p_3}, then no gap is to be filled. Otherwise, a solid triangle S^p^1p^2p^3\hat{S}_{\hat{p}_1\hat{p}_2\hat{p}_3} is placed, where p^2\hat{p}_2 and p^2\hat{p}_2 are the intersection points of p2p3p_2p_3 and the corresponding edges of the internal empty triangle of the considered lattice. In this way, it is ensured that there is no material overlap in the final structural design. The improved structural design is shown in Figure 6b. The additional material consumption caused by gap-filling triangles is compensated by scaling down the deposition ratio in Equation (12).

Results

In the following, we evaluate the proposed method with a variety of experiments. Firstly, a parameter study is performed to showcase the behavior of our method with respect to different settings. This includes the different strategies for setting the initialization, with/without the orientation regularization, and the different resolutions of the triangular parametrization. Secondly, we compare our results to those of alternative approaches, including the uniform triangular lattice structure, the porous infill produced by density-based topology optimization under a local volume constraint [34], and the alternative de-homogenization approach [16]. We then consider two practical examples to further verify the usability of our method. Finally, the adaptivity to situations with concurrently varying loading and constraint conditions is demonstrated via a specific test case. To differentiate our approach from alternative approaches, our generated structural design is presented by a density layout in black together with the corresponding triangular mesh in green. In this way, it is highlighted that our approach results in both a structural design and an explicit geometric format.

In all examples, the design domains are discretized by Cartesian grids with unit size. Young’s modulus and Poisson’s ratio are set to 1.0 and 0.3, respectively. Except for the field-aligned triangular meshing part, which is performed in the open source software Instant-meshes3 present by Jakob et al. [15], all other parts of the approach, including homogenization-based topology optimization and de-homogenization, is implemented in MatLab. We let the optimization process terminate after 300 iterations. The lower and upper bounds of the relative width of each cell are set to Lmin=0.1L_{min} = 0.1 and Lmax=0.5L_{max} = 0.5 if there is no special explanation. Correspondingly, the relative density value varies from 0.271 to 0.875 according to Equation (1). All experiments have been carried out on a desktop PC with an Intel Xeon CPU at 3.60GHz.

Performance statistics for all examples can be found in Table 1. In this table, we also provide the resulting compliance C∗C^{*} of homogenization-based topology optimization without restricting the layer orientations, i.e., all three orientation variables (θ1:3\theta_{1:3}) can be adjusted arbitrarily during optimization. All other settings are kept the same as the ones directing to C0C_{0}. By comparing C0C_{0} to the corresponding C∗C^{*}, we see that our additional orientation constraint does increase the compliance, whereas, it’s within 5% (measured by C0−C∗C∗\frac{C_{0}-C^{*}}{C^{*}}) in all of the experiments using the proposed Φstress\Phi_{stress} and orientation regularization.

ExamplesHomogenization-based Opti.De-homogenization
C∗C^{*}C0C_{0}V0V_{0}t0t_{0} (s)CCVVtt (s)ξ\xi
FemurFigure 1h73.3476.860.5036684.370.50259.78%
FemurFigure 7c75.2683.250.5035692.170.502710.93%
FemurFigure 7f71.2573.300.5035484.190.503014.86%
FemurFigure 7g73.3476.860.5036685.010.502910.60%
WheelFigure 9b7.788.030.502808.300.50273.29%
BeamFigure 10b2.612.660.302122.780.31218.00%
Bracket 1Figure 11c19.4619.920.5032620.450.50282.67%
Bracket 2Figure 11f26.6727.450.5064129.500.50377.47%
TriangleFigure 12b11.0511.070.5053912.120.503310.15%

Table 1. Performance statistics for homogenization-based topology optimization and field-aligned triangulation-guided de-homogenization. C0C_{0} and V0V_{0}, respectively, are the compliance and volume fraction of the optimal layout resulting from homogenization-based topology optimization. Accordingly, CC and VV correspond to the de-homogenized result. The design deviation of the de-homogenized result from the optimal layout is evaluated by ξ=C⋅V−C0⋅V0C0⋅V0\xi= \frac{C \cdot V-C_{0} \cdot V_{0}}{C_{0} \cdot V_{0}}. t0t_{0} and tt are the time consumptions of homogenization-based topology optimization and de-homogenization, respectively. Provided the same volume fraction, C∗C^{*} is the resulting compliance of homogenization-based topology optimization without restricting the layer orientations.

In the design of methodology, our method performs homogenization-based topology optimization on a simulation mesh (Rsim\mathcal{R}_{sim}) with low resolution to obtain the optimized specification, and the de-homogenized structural designs are described in a compact geometric format, i.e., a triangular mesh with each edge being assigned a unique thickness. Herein, to facilitate evaluating and comparing the obtained structural designs, we also introduce a high-resolution validation mesh (Rfine\mathcal{R}_{fine}). This is because the obtained structural design can be composed of very fine geometric details, which can only be captured by a sufficiently high resolution in compliance evaluation using finite element analysis. To alleviate the potential inconsistency of boundary conditions among different resolutions, we borrow the idea of mesh hierarchy in the geometric multi-grid solver that is frequently employed in high-resolution topology optimization [35]. Specifically, the design domain is discretized into a fine Cartesian mesh as Rfine\mathfrak{R}_{fine}, from which, a relatively coarse resolution grid (Rsim\mathfrak{R}_{sim}) and corresponding boundary conditions are constructed for homogenization-based topology optimization and the subsequent de-homogenization. This is achieved by the restriction operation in the geometric multi-grid solver. When evaluating the compliance of the obtained design, the triangular lattice structure is discretized according to Rfine\mathfrak{R}_{fine}, by identifying the cells of Rfine\mathfrak{R}_{fine} that are located on the obtained triangular lattice edges.

To ensure consistent boundary conditions and discretization when comparing C0C_{0} and CC, C0C_{0} is re-evaluated after optimization by projecting the optimized specifications onto Rfine\mathfrak{R}_{fine} through bilinear interpolation. For density-based topology optimization, Rfine\mathfrak{R}_{fine} is directly used in the simulation to provide sufficient design space. Thus, it is ensured that the resolution and boundary conditions are consistent when comparing our method to other alternatives. Table 2 provides information regarding the resolution and number of elements of Rfine\mathfrak{R}_{fine} and Rsim\mathfrak{R}_{sim}, as well as the number of triangular lattices involved in de-homogenization. The number of triangular lattices is determined as follows: First, a reference length for the triangular lattice edges is specified. This reference considers the preferred fineness of the structural design and a possible size limitation stemming from a downstream manufacturing task. Since the triangular lattices are from a set of approximately equilateral triangular elements, a reference area of the triangular lattices can then be derived from the reference length. Finally, the number of triangular lattices is computed by dividing the total area of the design domain by the reference area.

ExamplesRfine\mathcal{R}_{fine} #ResolutionRfine\mathcal{R}_{fine} #ElementsRsim\mathcal{R}_{sim} #ResolutionRsim\mathcal{R}_{sim} #Elements#Lattice
FemurFigure 1a920×1200920 \times1200632,468116×150116 \times15010,118684 (Figure 1f)
WheelFigure 9a1600×16001600 \times16001,930,188100×100100 \times1007,7401,637 (Figure 9b)
BeamFigure 10a1600×8001600 \times8001,280,000100×50100 \times505,000699 (Figure 10b)
Bracket 1Figure 11a2296×13282296 \times13282,181,496144×84144 \times848,7492,101 (Figure 11c)
Bracket 2Figure 11d2332×10002332 \times10001,144,564292×126292 \times12618,2711,167 (Figure 11f)
TriangleFigure 12a1360×11841360 \times1184790,132170×148170 \times14812,6521,065 (Figure 12b)

Table 2. Statistics for the validation meshes (Rfine\mathcal{R}_{fine}) used to evaluate the final structural designs, the simulation meshes (Rsim\mathcal{R}_{sim}) used for homogenization-based topology optimization, and the triangular lattice elements (#Lattice) used for de-homogenization.

Design verification. With the ‘Femur’ model shown in Figure 1 as an example, de-homogenization is performed using several test settings to verify our design decisions. Specifically, the results when using a uniform initialization Φuniform\Phi_{uniform} and of original homogenization-based topology optimization, i.e., no orientation regularization is used, are given Figure 7a-Figure 7c. Figure 7d-Figure 7f shows the results the principal stress-guided initialization Φstress\Phi_{stress} is used without orientation regularization. In Figure 7g-Figure 7i, the settings in homogenization-based topology optimization remain the same as in Figure 1, i.e., the initialization Φstress\Phi_{stress} and orientation regularization is used, but a higher resolution of the triangular parametrization is chosen. Figure 7(c), (Figure 7f), and (Figure 7i) use the consistent resolution for creating the field-aligned triangular mesh.

Image not extracted

Figure 7. The homogenization-based results (optimized orientations and density layout) and the corresponding de-homogenization results of three different settings are demonstrated in each box. In the left box (a-c), the homogenization-based topology optimization is conducted with the uniform initialization (Φuniform\Phi_{uniform}). In the middle one (d-f), Φuniform\Phi_{uniform} is replaced with the proposed initialization Φstress\Phi_{stress}, whereas other settings are kept the same. In the right box, both the Φstress\Phi_{stress} and orientation regularization are considered in the homogenization-based topology optimization. Among (c), (f), and (i), the resolution of the used triangular meshes is consistent.

The test results in Figure 7 and Figure 1, and the statistics in Table 1 demonstrate that the design deviation between homogenization-based topology optimization and de-homogenization results do not exceed 15%, and most are around at 10%, regardless of whether the optimized orientation field is continuous or not. This demonstrates that the proposed de-homogenization has relaxed requirements on the optimized specifications.

In terms of design quality measured by compliance and geometric regularity, however, different design decisions give rise to different results. Compared to homogenization-based topology optimization using Φuniform\Phi_{uniform} (Figure 7a, Figure 7b), we see that Φstress\Phi_{stress} does improve the convergence behavior (Figure 7d, Figure 7e). The compliance is reduced from 83.25 to 73.30 (Figure 7b, Figure 7e). However, Φstress\Phi_{stress} incurs more discontinuities in the optimized orientation field. Though the de-homogenized lattice structure corresponding to Φuniform\Phi_{uniform} (Figure 7c) is less stiff than the one obtained with Φstress\Phi_{stress} (Figure 7f), the former design gives better geometric regularity. The difference in the setting for Figure 7d, Figure 7e and Figure 7g, Figure 7h is that the orientation regularization is introduced in the latter. As a result, Figure 7g behaves more continuously than Figure 7d, yet Figure 7e is slightly stiffer than Figure 7h. This is because the orientation regularization acts like an additional constraint in the optimization process. The same phenomenon is kept in the de-homogenized results. Interestingly, the corresponding de-homogenization result Figure 7f is only slightly stiffer than Figure 7i compared to their difference in the homogenization-based results. This is regardless of the introduced geometric regularity. This behaviour is interesting due to the severe discontinuity of the optimized orientation field (Figure 7d) that still hurts the precision of de-homogenization, though it does not necessarily lead to failure in the de-homogenization process.

From Figure 7i and Figure 1f, we further see that the proposed de-homogenization is stable with respect to changes of the triangular mesh resolution. By comparing Figure 7c, Figure 7f, and Figure 7i, one can observe that the homogenization-based topology optimization combined with proposed Φstress\Phi_{stress} and orientation regularization directs to a more desirable structural design, which achieves better stiffness and geometric regularity concurrently. Thus, this setting is used in the following experiments.

Comparison analysis. We first compare our results to the uniform triangular lattice structure and porous infill structure. The latter is produced via density-based topology optimization under a local volume constraint [34]. Figure 8 and Figure 9 show that the uniform triangular lattice structure always gives rise to the highest compliance. The triangular lattice structure is not significantly stiffer than the porous infill structure. However, density-based topology optimization is a full-scale method and requires a sufficiently high simulation resolution to obtain a fine structural design. This significantly increases the computational cost. In addition, the method does not always converge [31]. This is shown in Figure 8b, where a binary layout is still unavailable even after 1000 iterations.

Image not extracted

Figure 8. (a) The uniform triangular lattice structure. (b) The porous infill structure from density-based topology optimization under the local volume constraint. (c) The proposed de-homogenization result. All examples share the same problem description.

Image not extracted

Figure 9. (a) A triangular design domain with 2 different loading cases, the normal force FnF_n and shear force FtF_t. (b) Result by our method. (c) Uniform triangular lattice structure. (d) Porous infill structure after 2000 iterations.

While our method is primarily developed to generate infill structures that spread across the entire design domain, it can be adapted to generate lattice structures that cover only part of the design domain, by setting Lmin=1.0×10−6L_{min}=1.0\times10^{-6}, Lmax=1.0L_{max}=1.0. The obtained structural layout is a subset of the initial design domain. With this adaptation, we evaluate the trade-off between the structural performance and regularity of the lattice, taking the results from the image-based de-homogenization method [16] as a reference. The design domain and load conditions are illustrated in Figure 10a. The results from our method and the image-based approach are shown in Figure 10b and Figure 10c, respectively. The structure from our method consists of equilateral triangles with varying thicknesses, while that from the image-based approach exhibits a large number of intersections of differently oriented beam-like substructures. The improved regularity in lattice structures is achieved with a compromise in the stiffness. The difference in stiffness is 8.2% measured by Cours−CalternativeCalternative\frac{C_{\mathit{ours}}-C_{\mathit{alternative}}}{C_{\mathit{alternative}}} when the volume fraction is the same, with Cours=2.78C_{\mathit{ours}}=2.78 and Calternative=2.57C_{\mathit{alternative}}=2.57.

Image not extracted

Figure 10. (a) Beam with 5 different loads. (b) De-homogenized structural design by our method. (c) The corresponding result from the image-based de-homogenization, reproduced from [16].

Mechanical parts. With ‘Bracket-1’ and ‘Bracket-2’ we consider two mechanical parts for emphasizing the basic features of the proposed de-homogenization scheme (see Figure 11). In particular, the results are in line with those obtained for ‘femur’, concerning compliance and geometric regularity.

Problem descriptions, density layouts, and de-homogenized structural designs for bracket 1 and bracket 2

Figure 11. (a), (d) Problem descriptions of bracket 1 and bracket 2, respectively. (b), (e) The corresponding density layouts generated by homogenization-based topology optimization. (c), (f) The resulting de-homogenized structural designs.

Multiple boundary conditions. We further demonstrate the adaptivity of the proposed method to the optimization problems with totally different boundary conditions, i.e., the fixations and loads vary concurrently. Here we consider an equilateral triangular design domain for infill optimization (Figure 12a).

Equilateral triangular design domain, optimized density layout with layer orientations, and de-homogenized triangular lattice structure

Figure 12. (a) The equilateral triangular design domain and the corresponding boundary conditions. Here the thick lines in black indicate the fixation positions. The loads are shown in different colors, each load is outward of the design domain and perpendicular to the corresponding domain boundary. (b) The optimized density layout and the corresponding layer orientations shown by streamlines. (c) The de-homogenized triangular lattice structure.

In Figure 12, we show the optimization and de-homogenization results, where the design deviation between the de-homogenized result and the corresponding homogenization-based topology optimization result is at ξ=10.15%\xi= 10.15\% and close to other examples. In addition, the triangular lattice structure exhibits good regularity.

Conclusion and Future Work

In this paper, we have introduced an innovative geometry-based de-homogenization technique for designing and optimizing functionally graded triangular lattices capable of withstanding various loading conditions. By conducting simulations and optimizations at a low-resolution discretization of the design domain, our de-homogenization approach, akin to other methods in this domain, generates highly detailed structures with significantly reduced computational demands compared to traditional density-based topology optimization methods.

Unique to our approach is its ability to explicitly represent the resulting structural design in a concise geometric format, i.e., a global triangular mesh where each edge is assigned a distinct thickness. This explicit geometric representation facilitates downstream processes such as editing and additive manufacturing. In terms of mechanical performance, the de-homogenized structure exhibits stiffness comparable to that of the optimal design achieved through homogenization-based topology optimization. Additionally, when compared to porous infill generated using local volume constraints, our de-homogenized structure consistently demonstrates higher stiffness.

In future work, we aim to extend this geometry-based de-homogenization method to address 3D problems. Just as we simplified a general rank-3 laminate to an equilateral triangle, we speculate that a general rank-6 laminate in 3D could potentially be approximated by an equilateral tetrahedron. We plan to verify this hypothesis and devise geometric techniques for the de-homogenization process.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Availability of code

The demonstration code for homogenization-based topology optimization can be accessed through https://github.com/3.4

CRediT authorship contribution statement

Junpeng Wang: Conceptualization, Methodology, Software, Writing - Original Draft. Rüdiger Westermann: Conceptualization, Methodology, Writing - Review & Editing, Supervision, Funding acquisition. Xifeng Gao: Methodology, Writing - Review & Editing. Jun Wu: Conceptualization, Methodology, Writing - Review & Editing, Supervision.

Acknowledgment

We thank Peter Dørffler Ladegaard Jensen from the Technical University of Denmark for the helpful discussion. This work was supported in part by a grant from German Research Foundation (DFG) under grant number WE 2754/10-1.

Elasticity Tensor of Rank-3

In a variety of related works [17, 12, 16], the general description of the elasticity tensor (S\mathbf{S}) of rank-3 is given as below

Sijkl=Sijkl+−(1−ρ)[(Sijkl+−Sijkl−)−1−ρ(1−v02)E+(∑n=13PnΛijkln)]−1(15)S_{ijkl}=S^{+}_{ijkl}-(1-\rho)\left[\left(S^{+}_{ijkl}-S^{-}_{ijkl}\right)^{-1}-\frac{\rho\left(1-v_{0}^{2}\right)}{E^{+}}\left(\sum_{n=1}^{3}P_{n}\Lambda^{n}_{ijkl}\right)\right]^{-1} \tag*{(15)}

Here, same as the elasticity tensor of the plane problem, SijklS_{ijkl} is a fourth-order tensor. Sijkl+S^{+}_{ijkl} and Sijkl−S^{-}_{ijkl} are the elasticity tensors of the isotropic solid and ’void’ materials and are given by Hooke’s law

[S1111+/−S1122+/−S1112+/−S1122+/−S2222+/−S2212+/−symsymS1212+/−]=E+/−1−v02[1v0v10symsym2(1−v0)](16)\begin{bmatrix} S^{+/-}_{1111} & S^{+/-}_{1122} & S^{+/-}_{1112}\\ S^{+/-}_{1122} & S^{+/-}_{2222} & S^{+/-}_{2212}\\ sym & sym & S^{+/-}_{1212} \end{bmatrix} = \frac{E^{+/-}}{1-v_{0}^{2}} \begin{bmatrix} 1 & v & 0\\ v & 1 & 0\\ sym & sym & 2(1-v_{0}) \tag*{(16)} \end{bmatrix}

E+E^{+} and v0v_{0} are Young’s modulus and Poisson’s ratio of the isotropic solid material. Pn>0P_{n}>0 depicts the relative contribution of layer nn to the stiffness, which subjects to ∑n=13Pn=1\sum_{n=1}^{3}P_{n}=1, and Pn=ρnρP_{n}=\frac{\rho_{n}}{\rho}.

Combining PnP_{n} and Λijkln\Lambda^{n}_{ijkl}, the item ∑n=13PnΛijkln\sum_{n=1}^{3}P_{n}\Lambda^{n}_{ijkl} parametrizes all of the possible properties of rank-3 laminate in the entire design space with a fixed volume fraction. Λ\Lambda is defined as

Λijkln=ninnjnnknnln+tinnjntknnln+nintjnnkntln+tinnjnnkntln+nintjntknnln2(1−v0)(17)\Lambda^{n}_{ijkl}=n^{n}_{i}n^{n}_{j}n^{n}_{k}n^{n}_{l}+\frac{t^{n}_{i}n^{n}_{j}t^{n}_{k}n^{n}_{l}+n^{n}_{i}t^{n}_{j}n^{n}_{k}t^{n}_{l}+t^{n}_{i}n^{n}_{j}n^{n}_{k}t^{n}_{l}+n^{n}_{i}t^{n}_{j}t^{n}_{k}n^{n}_{l}}{2(1-v_{0})} \tag*{(17)}

Here, nnn^{n} and tnt^{n} are the components of normal and tangent of layer nn. Let nn=[cos⁡θn sin⁡θn]n^{n}=[\cos\theta_{n}\ \sin\theta_{n}] and tn=[−sin⁡θn cos⁡θn]t^{n}=[-\sin\theta_{n}\ \cos\theta_{n}], here θn\theta_{n} is the orientation angle of the normal of layer nn to the xx-axis in a standard coordinate system, all the components of Λijkln\Lambda^{n}_{ijkl} are respectively expressed by

Λ1111n=cos⁡(4θn)+4cos⁡(2θn)+38+1−cos⁡(4θn)4(1−v0);Λ2222n=cos⁡(4θn)−4cos⁡(2θn)+38+1−cos⁡(4θn)4(1−v0);Λ1122n=Λ2211n=1−cos⁡(4θn)8−1−cos⁡(4θn)4(1−v0);Λ1112n=Λ1121n=Λ2111n=2sin⁡(2θn)+sin⁡(4θn)8−sin⁡(4θn)4(1−v0);Λ1222n=Λ2221n=Λ2122n=Λ2212n=2sin⁡(2θn)−sin⁡(4θn)8+sin⁡(4θn)4(1−v0);Λ1212n=Λ1221n=Λ2112n=Λ2121n=1−cos⁡(4θn)8+1+cos⁡(4θn)4(1−v0).(18)\begin{aligned} \Lambda^{n}_{1111} &= \frac{\cos(4\theta_n)+4\cos(2\theta_n)+3}{8}+\frac{1-\cos(4\theta_n)}{4(1-v_0)};\\ \Lambda^{n}_{2222} &= \frac{\cos(4\theta_n)-4\cos(2\theta_n)+3}{8}+\frac{1-\cos(4\theta_n)}{4(1-v_0)};\\ \Lambda^{n}_{1122} &= \Lambda^{n}_{2211}=\frac{1-\cos(4\theta_n)}{8}-\frac{1-\cos(4\theta_n)}{4(1-v_0)};\\ \Lambda^{n}_{1112} &= \Lambda^{n}_{1121}=\Lambda^{n}_{2111}=\frac{2\sin(2\theta_n)+\sin(4\theta_n)}{8}-\frac{\sin(4\theta_n)}{4(1-v_0)};\\ \Lambda^{n}_{1222} &= \Lambda^{n}_{2221}=\Lambda^{n}_{2122}=\Lambda^{n}_{2212}=\frac{2\sin(2\theta_n)-\sin(4\theta_n)}{8}+\frac{\sin(4\theta_n)}{4(1-v_0)};\\ \Lambda^{n}_{1212} &= \Lambda^{n}_{1221}=\Lambda^{n}_{2112}=\Lambda^{n}_{2121}=\frac{1-\cos(4\theta_n)}{8}+\frac{1+\cos(4\theta_n)}{4(1-v_0)}. \tag*{(18)} \end{aligned}

In practical computation, Equation (15) is usually given in a convenient matrix form [17]

S[2]=S+−(1−ρ)[(S+−S−)−1−ρ(1−v02)E+M]−1(19)S^{[2]}=S^{+}-(1-\rho)\left[(S^{+}-S^{-})^{-1}-\frac{\rho(1-v_0^{2})}{E^{+}}\mathcal{M}\right]^{-1} \tag*{(19)}

Here, M=∑n=13pnΛn\mathcal{M}=\sum_{n=1}^{3}p_n\Lambda^n, and the matrices S+S^{+}, S−S^{-} and Λn\Lambda^n are obtained from the corresponding fourth-order tensor components through the transformation below

A=[12(A1111+A2222)−A1122A1112−A222112(A1111−A2222)2A121212(A1112+A2221)sym12(A1111+A2222)+A1122](20)A=\begin{bmatrix} \frac{1}{2}(A_{1111}+A_{2222})-A_{1122} & A_{1112}-A_{2221} & \frac{1}{2}(A_{1111}-A_{2222})\\ & 2A_{1212} & \frac{1}{2}(A_{1112}+A_{2221})\\ \text{sym} & & \frac{1}{2}(A_{1111}+A_{2222})+A_{1122} \end{bmatrix} \tag*{(20)}

Let

m1=∑n=13pncos⁡(2θn),m2=∑n=13pnsin⁡(2θn),m3=∑n=13pncos⁡(4θn),m4=∑n=13pnsin⁡(4θn).(21)\begin{aligned} m_1&=\sum_{n=1}^{3}p_n\cos(2\theta_n),\qquad m_2=\sum_{n=1}^{3}p_n\sin(2\theta_n),\\ m_3&=\sum_{n=1}^{3}p_n\cos(4\theta_n),\qquad m_4=\sum_{n=1}^{3}p_n\sin(4\theta_n). \tag*{(21)} \end{aligned}

Then, the matrix M\mathcal{M} is rearranged into a simpler form

M=[3−v0−(1+v0)m34(1−v0)−(1+v0)m44(1−v0)m123−v0+(1+v0)m34(1−v0)m22sym12](22)\mathcal{M}= \begin{bmatrix} \frac{3-v_0-(1+v_0)m_3}{4(1-v_0)} & -\frac{(1+v_0)m_4}{4(1-v_0)} & \frac{m_1}{2}\\ & \frac{3-v_0+(1+v_0)m_3}{4(1-v_0)} & \frac{m_2}{2}\\ \text{sym} & & \frac{1}{2} \end{bmatrix} \tag*{(22)}

Upon obtaining SS, one can get the components of SijklS_{ijkl} in the following way

S=[S1111S1122S1112S2222S2221symS1212]=[12(S11[2]+S33[2])+S13[2]−12(S11[2]−S33[2])12(S11[2]+S33[2])12(S11[2]+S33[2])−S13[2]−12(S11[2]−S33[2])sym12S22[2]](23)S = \begin{bmatrix} S_{1111} & S_{1122} & S_{1112} \\ & S_{2222} & S_{2221} \\ \mathit{sym} & & S_{1212} \end{bmatrix} = \begin{bmatrix} \frac{1}{2}(S_{11}^{[2]}+S_{33}^{[2]})+S_{13}^{[2]} & -\frac{1}{2}(S_{11}^{[2]}-S_{33}^{[2]}) & \frac{1}{2}(S_{11}^{[2]}+S_{33}^{[2]}) \\ & \frac{1}{2}(S_{11}^{[2]}+S_{33}^{[2]})-S_{13}^{[2]} & -\frac{1}{2}(S_{11}^{[2]}-S_{33}^{[2]}) \\ \mathit{sym} & & \frac{1}{2}S_{22}^{[2]} \end{bmatrix} \tag*{(23)}

For details of the derivation process above, we refer to [17].

References

  1. [1]Grégoire Allaire, Perle Geoffroy-Donders, and Olivier Pantz. Topology optimization of modulated and oriented periodic microstructures by the homogenization method. Computers & Mathematics with Applications, 78(7):2197–2229, 2019. doi:10.1016/j.camwa.2018.08.007.DOI
  2. [2]Martin P. Bendsøe. Optimization of structural topology, shape, and material, volume 414. Springer, 1995. doi:10.1007/978-3-662-03115-5.
  3. [3]Martin Philip Bendsøe and Noboru Kikuchi. Generating optimal topologies in structural design using a homogenization method. Computer Methods in Applied Mechanics and Engineering, 71(2):197–224, 1988. doi:10.1016/0045-7825(88)90086-2.
  4. [4]Blaise Bourdin. Filters in topology optimization. International journal for numerical methods in engineering, 50(9):2143–2158, 2001. doi:10.1002/nme.116.DOI
  5. [5]Anders Clausen, Niels Aage, and Ole Sigmund. Exploiting additive manufacturing infill in topology optimization for improved buckling load. Engineering, 2(2):250–257, 2016. doi:10.1016/J.ENG.2016.02.006.DOI
  6. [6]Stephen Daynes, Stefanie Feih, Wen Feng Lu, and Jun Wei. Optimisation of functionally graded lattice structures using isostatic lines. Materials & Design, 127:215–223, 2017. doi:10.1016/j.matdes.2017.04.082.DOI
  7. [7]Martin Ohrt Elngaard, Niels Aage, Jakob Andreas Bærentzen, and Ole Sigmund. De-homogenization using convolutional neural networks. Computer Methods in Applied Mechanics and Engineering, 388:114197, 2022. doi:10.1016/j.cma.2021.114197.
  8. [8]Perle Geoffroy-Donders, Grégoire Allaire, and Olivier Pantz. 3-d topology optimization of modulated and oriented periodic microstructures by the homogenization method. Journal of Computational Physics, 401:108994, jan 2020. doi:10.1016/j.jcp.2019.108994.
  9. [9]Jeroen P. Groen and Ole Sigmund. Homogenization-based topology optimization for high-resolution manufacturable microstructures. International Journal for Numerical Methods in Engineering, 113(8):1148–1163, 2018. doi:10.1002/nme.5575.DOI
  10. [10]Jeroen P. Groen, Florian C. Stutz, Niels Aage, Jakob A. Bærentzen, and Ole Sigmund. De-homogenization of optimal multi-scale 3D topologies. Computer Methods in Applied Mechanics and Engineering, 364:112979, jun 2020. doi:10.1016/j.cma.2020.112979.DOI
  11. [11]Jeroen P Groen, Jun Wu, and Ole Sigmund. Homogenization-based stiffness optimization and projection of 2D coated structures with orthotropic infill. Computer Methods in Applied Mechanics and Engineering, 349:722–742, 2019. doi:10.1016/j.cma.2019.02.031.DOI
  12. [12]Jeroen Peter Groen. Multi-scale design methods for topology optimization. DTU Mechanical Engineering, 2018.
  13. [13]B. Hassani and E. Hinton. A review of homogenization and topology optimization I—homogenization theory for media with periodic structure. Computers & Structures, 69(6):707–717, 1998. doi:https://doi.org/10.1016/S0045-7949(98)00131-X.
  14. [14]Yuming Huang, Guoxin Fang, Tianyu Zhang, and Charlie CL Wang. Turning-angle optimized printing path of continuous carbon fiber for cellular structures. Additive Manufacturing, 68:103501, 2023. doi:10.1016/j.addma.2023.103501.DOI
  15. [15]Jakob Wenzel, Marco Tarini, Daniele Panozzo, Olga Sorkine-Hornung, et al. Instant field-aligned meshes. ACM Trans. Graph., 34(6):1–15, 2015. doi:10.1145/2816795.2818078.DOI
  16. [16]Peter Dørrfler Ladegaard Jensen, Ole Sigmund, and Jeroen P. Groen. De-homogenization of optimal 2D topologies for multiple loading cases. Computer Methods in Applied Mechanics and Engineering, 399:115426, 2022. doi: 10.1016/j.cma.2022.115426.
  17. [17]LA Krog and Niels Olhoff. Topology and reinforcement layout optimization of disk, plate, and shell structures. Springer, 1997. doi:10.1007/978-3-7091-2566-3_9.DOI
  18. [18]Tsz-Ho Kwok, Yongqiang Li, and Yong Chen. A structural topology design method based on principal stress line. Computer-Aided Design, 80:19–31, 2016. doi:10.1016/j.cad.2016.07.005.DOI
  19. [19]Jonathan Palacios and Eugene Zhang. Rotational symmetry field design on surfaces. ACM Transactions on Graphics (TOG), 26(3):55–es, 2007. doi:10.1145/1276377.1276446.DOI
  20. [20]Ajit Panesar, Meisam Abdi, Duncan Hickman, and Ian Ashcroft. Strategies for functionally graded lattice structures derived using topology optimisation for additive manufacturing. Additive Manufacturing, 19:81–94, 2018. doi: 10.1016/j.addma.2017.11.008.
  21. [21]O. Pantz and K. Trabelsi. A post-treatment of the homogenization method for shape optimization. SIAM Journal on Control and Optimization, 47(3):1380–1398, 2008. doi:10.1137/070688900.
  22. [22]Pauli Pedersen. On optimal orientation of orthotropic materials. Structural optimization, 1(2):101–106, 1989. doi: 10.1007/BF01637666.DOI
  23. [23]Nicolas Ray, Bruno Vallet, Wan Chiu Li, and Bruno Lévy. N-symmetry direction field design. ACM Transactions on Graphics (TOG), 27(2):1–13, 2008. doi:10.1145/1356682.1356683.
  24. [24]F. C. Stutz, J. P. Groen, O. Sigmund, and J. A. Bærentzen. Singularity aware de-homogenization for high-resolution topology optimized structures. Structural and Multidisciplinary Optimization, aug 2020. doi:10.1007/s00158-020-02681-6.DOI
  25. [25]Florian Cyril Stutz, Tim Felle Olsen, Jeroen Peter Groen, Niels Aage, Ole Sigmund, Justin Solomon, and Jakob Andreas Bærentzen. Synthesis of frame field-aligned multi-laminar structures. ACM Transactions on Graphics, 2022. doi: https://doi.org/10.1145/3516522.
  26. [26]Krister Svanberg. The method of moving asymptotes—a new method for structural optimization. International journal for numerical methods in engineering, 24(2):359–373, 1987. doi:10.1002/nme.1620240207.DOI
  27. [27]Erik Träff, Ole Sigmund, and JP Groen. Simple single-scale microstructures based on optimal rank-3 laminates. Structural and Multidisciplinary Optimization, 59(4):1021–1031, 2019. doi:10.1007/s00158-018-2180-3.DOI
  28. [28]Fengwen Wang, Boyan Stefanov Lazarov, and Ole Sigmund. On projection methods, convergence and robust formulations in topology optimization. Structural and Multidisciplinary Optimization, 43(6):767–784, 2011. doi:10.1007/s00158-010-0602-y.DOI
  29. [29]Junpeng Wang, Christoph Neuhauser, Jun Wu, Xifeng Gao, and Rüdiger Westermann. 3D-TSV: The 3D trajectory-based stress visualizer. Advances in Engineering Software, 170:103144, 2022. doi:10.1016/j.advengsoft.2022.103144.DOI
  30. [30]Junpeng Wang, Rüdiger Westermann, and Jun Wu. A streamline-guided dehomogenization approach for structural design. Journal of Mechanical Design, 145(2):021702, 2023. doi:10.1115/1.4056148.DOI
  31. [31]Junpeng Wang, Jun Wu, and Rüdiger Westermann. Stress topology analysis for porous infill optimization. Structural and Multidisciplinary Optimization, 65(3):1–13, 2022. doi:10.1007/s00158-022-03186-0.DOI
  32. [32]Junpeng Wang, Jun Wu, and Rüdiger Westermann. Stress trajectory guided structural design and topology optimization. In International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, volume 1, page 1. American Society of Mechanical Engineers, 2022. doi:10.1115/DETC2022-89030.DOI
  33. [33]Rebekka V. Woldseth, J. Andreas Bærentzen, and Ole Sigmund. Phase noise for dehomogenisation in 2d multiscale topology optimisation. Computer Methods in Applied Mechanics and Engineering, 418:116551, 2024. doi:10.1016/j.cma.2023.116551.DOI
  34. [34]Jun Wu, Niels Aage, Rüdiger Westermann, and Ole Sigmund. Infill optimization for additive manufacturing – approaching bone-like porous structures. IEEE Transactions on Visualization and Computer Graphics, 24(2):1127–1140, 2018. doi:10.1109/TVCG.2017.2655523.DOI
  35. [35]Jun Wu, Christian Dick, and Rüdiger Westermann. A system for high-resolution topology optimization. IEEE transactions on visualization and computer graphics, 22(3):1195–1208, 2015. doi:10.1109/TVCG.2015.2502588.DOI
  36. [36]Jun Wu, Ole Sigmund, and Jeroen P Groen. Topology optimization of multi-scale structures: a review. Structural and Multidisciplinary Optimization, pages 1–26, 2021. doi:10.1007/s00158-021-02881-8.DOI
  37. [37]Jun Wu, Weiming Wang, and Xifeng Gao. Design and optimization of conforming lattice structures. IEEE Transactions on Visualization and Computer Graphics, 27(1):43–56, 2021. doi:10.1109/TVCG.2019.2938946.

Paper details

vlnc:W_mtpfhr2h
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents