Topology optimization of multi-scale structures: a review

A starting point in the history of structural optimization is the classical work by Michell (1904) where an optimal truss design is represented by a continuum description. Arguably, this is the first work on multi-scale optimization since the truss-like continuum description is actually a limit case that can only be realized by using an infinite number of discrete truss members; see Prager and Rozvany (1977) for more details. Similarly, Cheng and Olhoff (1981) found that for the case of optimizing the stiffness of a plate by varying the thickness, more stiffener-like reinforcement appeared when the design space was refined. They made the important conclusion that in the limit of an infinitely fine mesh an infinite number of stiffener-like reinforcing members would occur (Cheng 1981). Only by restricting the variation in shape can existence of a solution be proven (Niordson 1983). For compliance minimization in 2 or 3 dimensions, Kohn and Strang (1986) proved that the optimal material distribution can only be found by relaxing the design space. This means the use of a continuous composite material description using a characteristic unit-cell length \(\epsilon \rightarrow 0\), allowing for much more detail than a single-scale discretized point-wise material or void description. Here, the theory of homogenization (Bensoussan et al. 1978) enters the field of structural optimization, since this is the perfect tool to bridge the scale between the microscopic periodic composite material and its effective (homogenized) properties on the macroscale. Using this theory, several research groups independently and more or less simultaneously realized that there exists a class of sequential laminates, the so-called rank-N laminates that can achieve the theoretical upper bounds for maximum strain energy density (Lurie and Cherkaev 1984; Norris 1985; Francfort and Murat 1986; Milton 1986). These composite materials consist of several scales in themselves, which is a necessary condition to achieve ultimate stiffness (Allaire and Aubry 1999). With this knowledge, Bendsøe and Kikuchi (1988) came up with the landmark paper that provided the computational framework to do homogenization-based topology optimization of continuum structures. Contrary to using rank-2 laminates that are optimal for a single load case in 2D, they came up with a single-scale interpretation, the square microstructure with a rectangular hole, shown in Fig. 3. By changing the widths μ₁ and μ₂ as well as the orientation 𝜃 throughout the design domain, near-optimal structures could be achieved. The use of rank-2 laminates as a microstructure was included in later works (Bendsøe 1989; Allaire and Kohn 1993).

Shortly after the homogenization-based topology optimization approach was introduced, an alternative known as the SIMP (Solid Isotropic Material with Penalization) or power-law approach was suggested (Bendsøe 1989; Zhou and Rozvany 1991; Mlejnek 1992). Here, the material distribution in a design domain is represented by a scalar field, with a relative density per element in the discretized domain. Here, ρ = 1 means solid and ρ = 0 means empty. This integer-programming problem is relaxed to allow for intermediate densities during optimization. Regardless of the value of the relative density, the material within each element is assumed to be isotropic and homogeneous. Material properties (e.g., Young’s modulus) are related to relative densities by a power-law interpolation. This seemingly artificial relationship was later proved physically permissible for a properly chosen penalty power, e.g., with a power p ≥ 3 for Poisson’s ratio ν = 1/3 (Bendsøe and Sigmund 1999). Based on this, a heterogeneous material distribution within the element domain, i.e., a microstructure of solid base material and void, can be found to match the expected material properties of an intermediate density. Hence, even the simple “mono-scale” density approaches may be interpreted as multi-scale approaches albeit with suboptimal microstructures. With a justified penalization power, the SIMP approach converges to near 0-1 solutions, representing mono-scale structures. In a number of multi-scale approaches to be discussed later, a smaller p is chosen for obtaining solutions with a large portion of intermediate densities, which provide a basis for filling porous microstructures. However, this may not be mechanically valid, c.f. if p < 3 for ν = 1/3, this power-law scheme violates the Hashin-Strikhman bounds (Hashin and Shtrikman 1963). The SIMP approach is easy to implement; see educational codes (Sigmund 2001; Andreassen et al. 2011; Ferrari and Sigmund 2020). Here, the three-field projection approach is recommended (Guest et al. 2004; Xu et al. 2010; Wang et al. 2011).

Mono-scale structures can also be optimized by methods based on level sets (Wang et al. 2003; Allaire et al. 2004), evolutionary procedures (Xie and Steven 1993), and geometric morphing and projection (Norato et al. 2004; Guo et al. 2014). For an overview of these methods, we refer to the reviews by Sigmund and Maute (2013) and Deaton and Grandhi (2014). The use of non-gradient approaches for solving topology optimization problems with many variables is not recommended (Sigmund 2011).

The theory of numerical homogenization, which approximates effective properties of a periodic unit cell, can also be exploited to design a microstructure with desired properties. This method was introduced by Sigmund (1994) and is generally referred to as inverse homogenization. With compliance minimization in mind, it is common practice to formulate an optimization problem to maximize the microstructure stiffness with respect to the applied stresses or strains. For the difference in stiffness between such microstructures and the optimal energy bounds, see, e.g., Guedes et al. (2003) and Träff et al. (2019). Furthermore, microstructures with negative Poisson’s ratio (Sigmund 1994; Andreassen et al. 2014; Clausen et al. 2015b), materials with maximum shear and bulk moduli (Sigmund 2000), or materials with increased buckling strength (Thomsen et al. 2018; Wang and Sigmund 2020) can be designed using inverse homogenization.

Besides problems in elasticity, inverse homogenization can be applied to many other types of physics. Sigmund and Torquato (1996) considered thermal expansion problems, while Torquato et al. (2002) included both thermal and electrical conductivity. Sigmund (1999) and Challis et al. (2012) considered combined stiffness and fluid permeability. Inverse homogenization for the design of phononic and photonic materials has been considered by Sigmund and Jensen (2003) and Jensen and Sigmund (2011), respectively. Lately, inverse homogenization has been applied to the design of photonic isolaters by Christiansen et al. (2019). For more details on material design using inverse homogenization, we refer interested readers to the review papers by Cadman et al. (2013) and Osanov and Guest (2016) and the Ph.D. thesis by Andreassen (2015).

In the context of elasticity, the minimum number of sequential layers that is required to parameterize the optimal solution using rank-N microstructures depends on both the number of load cases and the dimension of the problem (Avellaneda 1987; Francfort and Murat 1986): Hence, at most, 7 different length scales are needed to describe an optimal material parameterization in 3D, i.e., six length scales for the microstructure and one for the macroscopic parameterization. The elasticity tensors of these multi-scale laminates can be derived analytically, allowing the local microstructure optimization problems to be solved in a very efficient manner. For example, for a rank-2 laminate shown in Fig. 5 (right) the first layering is constructed on the infinitesimal length scale x/𝜖² shown in Fig. 5 (left). This layering can be described by the relative layer width μ₁ that describes the ratio of the isotropic stiff material (+) compared to the weak void material (-) and the interface normal n¹ and tangent t¹. Subsequently, the rank-2 laminate is constructed using a second layering on length scale x/𝜖 using relative layer width μ₂ that describes the ratio of the stiff material to the first layering. By ensuring orthogonal interface vectors, the elasticity tensor is thus defined by only three parameters, μ₁, μ₂, and 𝜃₂, besides the material properties of the stiff (+) and weak (-) phase respectively.

The first work in which topology optimization using rank-2 laminates was considered is the work by Bendsøe (1989). Subsequently, Allaire and Kohn (1993) proved that the optimal design for plane problem subject to a single load case can be described purely by the stress distribution in the macroscopic domain. Multiple load case problems, where the design is parameterized using optimal rank-3 laminates, have been considered for both plane problems (Allaire et al. 1996; Cherkaev et al. 1998) and plate/shell problems (Díaz et al. 1995; Hammer et al. 1997; Krog and Olhoff 1999). By choosing different plate layouts, e.g., using a fixed core thickness, Soto and Díaz (1993) and Krog and Olhoff (1999) showed the effect of using several density restrictions on the performance. The effect of restricting the amount of unique microstructures in the domain has been investigated by Cherkaev et al. (1998), who demonstrated that the performance can be at least twice as bad if only 1 or a few unique micostructures are used in the design domain.

The method has been applied to 3D topology optimization problems, where orthogonal rank-3 laminates are used for single load case problems (Cherkaev and Palais 1996; Allaire et al. 1997; Díaz and Lipton 1997; Olhoff et al. 1998; Czarnecki and Lewiński 2006). Multiple load cases are considered by Díaz and Lipton (2000). A nice feature of rank-N laminates is that they provide a lower bound on the theoretical compliance that can be reached. It is thus recommended to include this comparison when presenting a multi-scale approach. To help doing so, we supply a code for the optimization of rank-2 microstructures, which is discussed in Appendix.

A downside of structures optimized using rank-N laminates is that they consist of several length scales. This poses two challenges. Firstly, although technology is rapidly developing and lab work has demonstrated multi-scale manufacturing capability (e.g., Zheng et al. (2016)), a majority of manufacturing techniques do not yet support precise production of structures spanning multiple length scales. Hence, single scale interpretation of multi-scale rank-N laminates is required. Träff et al. (2019) have shown that the multi-scale rank-N microstructures can be approximated on a single scale with only a small loss in performance. Secondly, it is necessary to compile globally consistent structures from locally defined rank-N laminates (or their simplified versions), i.e., to de-homogenize the results. This problem was first addressed by Pantz and Trabelsi (2008), who used an implicit geometry description to interpret the optimized multi-scale designs on a single length scale with little loss in performance. More details on de-homogenization will be presented later in Section 5.3.1.

Another difficulty with the interpretation of rank-N microstructures is that, for multiple load cases, there exist an infinite number of different microstructure parameterizations that can reach the same elasticity tensor. As an example, consider the planar unit-cell optimization problem using four independent load cases shown in Fig. 6 (top). If we optimize for minimum complementary energy and find the corresponding rank-3 laminates using the method of Lipton (1994a), we obtain the four distinct microstructure realizations shown in Fig. 6 (bottom). These microstructures and an infinite number of combinations thereof can achieve the same minimum complementary energy, which is both an advantage and a difficulty when finding practical interpretations of rank-N designs.

Finally, it should be mentioned again that for many problems beyond elasticity and heat conduction, a full parameterization of the optimal set of constitutive properties is yet to be found. Finding these bounds on the properties and geometries that reach these bounds is a research field on its own. The interested reader is referred to the books on the theory of composites by Milton (2002, 2016).

Predefined cellular structures such as those shown in Fig. 7 have been used to efficiently design multi-scale components. Common in many of these multi-scale approaches is the use of homogenization to evaluate the effective material properties of predefined unit cells. Since the cells are predefined, homogenization can be performed off-line prior to optimization. This allows to generate structures with fine geometric details at a run-time computational complexity comparable to mono-scale approaches (Section 2.2). This is attractive, as evidenced by the large number of publications and industrial examples falling into this category. Predefined unit cells reduce the solution space of the microstructures to a dimension of 1 (or a few, see Section 5.3). This distinguishes this approach from hierarchical or concurrent approaches (Sections 5.4 and 5.5), where microstructures are represented by multi-variable density fields. This distinction is meaningful for implementation. Since the solution space is reduced, it is convenient to construct a differentiable function that maps the reduced parameters to homogenized properties, and use this function in macroscale optimization.

A typical application of these approaches is to design 3D printed components with uniform infill patterns. From the optimization perspective, this requires no or little adaption of mono-scale optimization approaches. The repeating unit cell can be interpreted as a material, the distribution of which is then optimized by mono-scale approaches. Note that the unit cells interpreted as a material can be isotropic or anisotropic. An example of the former is a 2D beam-like lattice structure constructed from a regular triangular tiling (Clausen et al. 2016). In contrast, the 2D lattice structure from a regular square tiling is anisotropic. Most mono-scale optimization approaches in their standard form deal with isotropic materials. To handle anisotropic materials, the stiffness matrices in finite element analysis must be adapted to reflect the homogenized elasticity tensor of the anisotropic material. Also note that if not optimally oriented along local principal stress directions, anisotropic microstructures may result in highly suboptimal results.

There has been a growing interest in designing components with graded cellular structures. Here, the input to optimization extends from a single unit cell to a set of cells with gradation in porosity and effective material properties (e.g., elasticity tensor). This set can thus be referred to as functionally graded cells. They are typically parameterized by the fraction of solid material within each cell (ρ), naturally serving as the design variable in optimization. The homogenized elasticity tensor (E^H) is a function of ρ, E^H(ρ), which can be constructed by interpolating elasticity tensors of sample cells with equally spaced material fractions. E^H(ρ) thus depends on the specific functionally graded cells, and in general it deviates from the power-law relation. This function for X-shaped lattice structures with varying thicknesses is visualized in Fig. 8a, where the power-law curves with p = 1 (i.e., unpenalized), 2 and 3 are also plotted. For accurately representing the corresponding mechanical properties, the cell-specific E^H(ρ) shall be used in lieu of the generic power-law in density-based topology optimization. The cell-specific E^H(ρ) is commonly constructed using numerical homogenization. Wu et al. (2019) proposed an alternative surrogate model for mapping the effective properties of density parameterized unit cells. It used static condensation to reduce the degrees of freedom of unit cells, followed by proper orthogonal decomposition and diffuse approximation for mapping the density to unit cell stiffness matrix.

Cell-specific material interpolation models have been used in the optimization for components consisting of gyroid-based cellular structures (Li et al. 2018a) and beam-like lattice structures (Wang et al. 2018; Watts et al. 2019), as well as topology optimized microstructures (Garner et al. 2019). Graded cellular structures can also be obtained by post-processing an unmodified SIMP-based mono-scale topology optimization approach, e.g., by replacing the intermediate densities by unit cells of the corresponding material fractions (Brackett et al. 2011; Panesar et al. 2018). Figure 8 compares interpolation schemes for two different unit cells (c.f., (b) and (c)) with varying member thicknesses. The comparison is performed on the 2D cantilever problem, thoroughly studied by Sigmund et al. (2016). The domain is fixed at its left edge and loaded with a unit vertical load distributed over the central 10% of the right edge. Young’s moduli for the solid and void material are E₀ = 1 and \(E_{\min \limits }=0.001\), respectively¹. Poisson’s ratio is ν₀ = 0.3 for both materials². The compliances of the designs using the cell-specific interpolation functions (b and c) are much smaller than the design using generic (unphysical) power-law with p = 1 (d) and post-processed by interpreting gray elements with square- or X-shaped cells, demonstrating the significance of using an accurate material interpolation model in optimization. It simply does not make mechanical sense to use a non-physical macroscale model like SIMP with p = 1 to control the macroscopic density in multi-scale approaches. In (d), the post-process of replacing gray elements by square-shaped cells leads to a large compliance of c^s = 1017.03, since low-density square-shaped cells are weak in shear. The designs using cell-specific interpolations (b and c) have higher compliance values than design (e) using the power law with p = 3 (p starts from 1 and is increased by 0.2 every 50 iterations). This suggests the optimization using cell-specific interpolations gets stuck in local minima. The number of iterations for obtaining both (b) and (c) is 400. Design (b) has a density distribution close to black and white, while design (c) has large gray regions. X-Shaped cells have large shear moduli, making it economical to place gray elements in the middle of the beam which is predominately under shear stress. Rank-2 material achieves optimal stiffness (f) due to the alignment of material anisotropy with principal stress directions, whereas the fixed orientation of the square- and X-shaped lattice cells restricts the adaptation of material anisotropy.

The curves for \({E}_{ij}^{H}(\rho )\) for the square- and X-shaped cells are lower than the theoretical limit according to the Hashin–Shtrikman bounds (Hashin and Shtrikman 1963). This means a suboptimal use of material using cellular structures with a simple geometry. To improve the properties, Zhou and Li (2008) applied inverse homogenization (see Section 2.3) to design a series of unit cells with varying material fractions. Figure 9 (top) shows unit cells independently optimized for an increasing material fraction, using the code provided by Xia and Breitkopf (2015). These optimized cells have distinct topology, owing to the large solution space in topology optimization. It is also noticeable that neighboring cells, despite being close in terms of material fractions, may have poor connectivity across the interface (see Fig. 9 (top), between the first and second and between the third and fourth cells). The connectivity issue is also recognizable in hierarchical approaches (Rodrigues et al. 2002), to be discussed in Section 5.4. Note that this is less of a problem in cellular structures parameterized by thickness since the topology there is constant, as long as the thickness does not vanish to 0.

Zhou and Li (2008) proposed three methods to address the connectivity issue between topology optimized microstructures: kinematic connectivity constraint, pseudo load, and unified formulation with non-linear diffusion. In the first two, unit cells are optimized individually, while solid non-design regions or pseudo loads are prescribed along the domain boundary to stimulate connectivity. In the last, unit cells are optimized altogether, while a nonlinear diffusion term defined on the domain covering all cells is integrated in the objective function to penalize disconnection and suppress checkerboard patterns. Radman (2013a, 2013b) improved the computational efficiency of the unified formulation by successively optimizing new unit cells, while considering connection to already optimized cells. Du et al. (2018) proposed a physics-independent connectivity index, which measures the amount of overlap in adjacent cells across the shared interface. Garner et al. (2019) proposed to optimize the connectivity, quantified by the physical properties of interest (e.g., bulk modulus) of an extended domain that covers adjacent cells. Poor connectivity between adjacent cells leads to an inferior bulk modulus for the extended domain as a compound cell, and thus is effectively suppressed in optimization. Figure 9 (bottom) shows results from this compound formulation. Mapping these compatibility optimized cells to a compliance-minimized macroscale density distribution, and evaluating structural compliance using both full-scale and homogenization-based analyses, it was found that the discrepancy is small (a relative error of 2%). This small error is also attributed to the limited and slowly varying nature of the microstructure which satisfies basic assumptions of the homogenization-based multi-scale modelling. This method has been demonstrated for optimizing up to 100 varying unit cells in 2D. Optimizing a large number of unit cells in 3D can be computationally expensive. Cramer et al. (2016) proposed geometric interpolation to obtain transitioning microstructures between individually optimized unit cells. This geometric approach works for microstructures of similar topology.

Cellular structures, both assemblies of geometric primitives and topology optimized microstructures, are typically defined within a square or cube domain. Such a fixed design domain may limit the achievable properties (c.f., Träff et al. (2019)). These cellular structures may be square symmetric (2D) or cubic symmetric (3D) but not (necessarily) isotropic. If orthotropic directions are oriented along regular finite element grid directions and not along principal stress directions as recommended by Pedersen (1989), inferior stiffness may be the result.

Functionally graded cells can be obtained by uniformly varying the thickness of the geometric primitives in the cell domain, e.g., the square- and X-shaped 2D lattices in the previous subsection. Consequently, each element in the macroscale optimization is assigned a single design variable, the material volume fraction, ρ(d), where d is the thickness. To enlarge the solution space, the parameterization of unit cells can be extended. The unit cell in Fig. 10 (middle) has two superimposed geometric patterns, i.e., an X shape and a plus shape, each with an independent thickness, leading to more variations in the attainable elasticity tensor, E^H(d, t). The number of design variables per element can be further increased, e.g., by assigning an individual thickness per geometric primitive, indicated by different colors in Fig. 10 (right). Graded lattice structures in 2D and 3D optimized with multiple design variables per macro element have been demonstrated by Wang (2018, 2020). Imediegwu et al. (2019) used a lattice cell with seven independent parameters for 3D optimization. As the number of independent parameters for describing the unit cell increases, a large number of numerical homogenization operations is required. White et al. (2019) proposed a neural network surrogate model to provide the mapping from parameters to resulting metamaterial properties. Zhu et al. (2017) pre-computed the space of attainable Young’s modulus and Poisson’s ratio for certain types of microstructures, and used this space as a constraint for macroscale topology optimization with microstructures. Zhu et al. (2019) presented an approach for concurrent optimization of the microscale material distribution and a macroscale mapping function which transforms periodic microstructures into graded ones. This was extended and demonstrated for 3D problems (Xue et al. 2020).

As the number of parameters increases, the potential to generate a variety in the anisotropic directions of the microstructures increases, and this, to some extent, allows an adaption of microstructural anisotropy to the local stress directions during optimization. To ultimately align the anisotropy of microstructures to stress directions, it is critical to release the rotational freedom in the unit cell parameterization, as pioneered by Bendsøe and Kikuchi (1988).

Predefined cellular structures are beneficial for computational efficiency. However, they may significantly restrict the solution space. To access the full solution space, while avoiding intensive full-scale analysis, an idea is to use a hierarchical formulation. Hierarchical optimization of material and structure dates back to Rodrigues et al. (2002) which was later extended to 3D (Coelho et al. 2008), both using SIMP. Comparable results were reproduced with level sets (Sivapuram et al. 2016). Hierarchical optimization combined with inverse homogenization is also referred to as concurrent (or simultaneous) optimization of structure and material (or microstructure). Like in approaches in other categories, here, a two-scale discretization of the domain is employed. The formulation involves one problem at the global (or macro) scale and many problems at the local (or micro) scale. The global problem determines the macroscopic spatial distribution of homogenized material, and local problems determine microscopic spatial distribution of solid and void phases by optimizing for homogenized properties. The structural equilibrium in the macroscale is in general nonlinear due to the microstructure adaptation (Jog and Haber 1996). A nonlinear resolution framework based on FE² scheme was developed to address this nonlinearity (Xia and Breitkopf 2014). This subsection discusses approaches that prescribe the domain shape of microstructures and/or their orientation.

The microstructural optimization problem (inverse homogenization) should result in single-scale approximations of theoretically optimal rank-N composites. In a recent work, it was found that inverse homogenization at low volume fractions is prone to producing local optima, and a simple mapping approach was suggested to approximate rank-3 laminates (Träff et al. 2019). These mapped microstructures perform relatively close to theoretical energy bounds, and can serve as starting guesses for inverse homogenization problems to achieve performances even closer to the bounds.

In each iteration of a hierarchical solution process, following a solved global problem, the local problems become independent from each other. On the positive side, the independent problems can be solved in parallel by sending sets of local problems to different processors (Coelho et al. 2011). This gains a computational speedup and thus allows solving two-scale problems in 3D. On the other hand, the independent nature of the local problems creates a critical challenge regarding the compatibility of microstructures across the shared boundary. We emphasize that the problem of concern is related to structural properties beyond the disconnected geometry, and thus choose to use compatibility in lieu of connectivity. The compatibility problem arises since disconnections between adjacent microstructures are not captured in the global analysis using homogenized properties (separation of scales). A mechanical indication of compatibility is the discrepancy between the objective (e.g., compliance) evaluated by a full-scale analysis and by an analysis using the homogenized properties (Garner et al. 2019). Some approaches summarized in Section 5.2 for improving compatibility in functionally graded microstructures are also applicable to the hierarchical optimization. For instance, by using extended domains that overlap in local optimizations, the compatibility can be significantly improved, reducing the discrepancy in compliance values between full-scale and homogenization analyses from six orders of magnitude to two (Garner et al. 2019) (see Fig. 14). While this is good progress, a discrepancy of two orders of magnitude is still alarming. To examine the optimality, we can visually compare hierarchically optimized structures with those from full-scale approaches with local volume constraints (Section 3) and de-homogenization approaches (Section 5.3). For a single load, orthogonal microstructures that are individually aligned with principal stress directions are known to be close to optimal. Such orthogonal microstructures are distinct in full-scale approaches with local volume constraints as well as de-homogenization approaches, but are difficult to discover in large regions of hierarchically optimized structures. A reason behind the poor compatibility is that the fixed, axis-aligned rectangular domain used in inverse homogenization is incapable of accommodating rotation of these orthogonal microstructures.

Other strategies have been proposed to enhance connectivity. Wang et al. (2017) proposed a shape metamorphosis method based on level-set representations to interpolate a prototype microstructure to generate a family of graded microstructures. The interpolated microstructures are connectable to each other in a natural way since they present similar topological features and material distribution patterns at their edges. Li et al. (2018b) adopted the kinematic constraint approach (Zhou and Li 2008) for level-set–based topology optimization of functionally graded cellular composites hosting auxetic metamaterials. Zhou et al. (2019) proposed a geometric connectivity index upon which constraints were defined and included in the macroscale optimization to improve connectivity. Liu et al. (2020) recently proposed to ensure connectivity between any two types of the microstructures by introducing pre-defined connective regions in the microstructural unit cells and keeping these regions sharing the same topology. In these approaches, the connectivity was often visually assessed, and a mechanical assessment was unfortunately missing.

The compatibility issue is circumvented if the optimization problem is reformulated to design structures consisting of repetitive microstructures, at the cost of reduced structural performance. Such a formulation was presented by Fujii et al. (2001) for designing repetitive microstructures in the entire design space. Liu et al. (2008) incorporated a macroscale variable to concurrently design the microstructure and its distribution. Here, the solution space is reduced to a single microstructure, a strategy similar to pattern repetition in full-scale approaches (Section 3). The global analysis is performed using homogenized properties rather than on the full scale, assisted with an interpolation scheme called Porous Anisotropic Material with Penalization (PAMP). It was first demonstrated for compliance minimization, and extended for maximizing fundamental frequency (Niu et al. 2009), for considering load uncertainties (Guo et al. 2015) and more. Similar formulations and extensions using evolutionary procedures can be found in Huang et al. (2013) and more.

In between the spectrum from a single microstructure to the full solution space, approaches have been developed to design the structural layout of a few unique, unprescribed microstructures (Deng and Chen 2017; Zhang et al. 2018; Liu et al. 2020). Pizzolato et al. (2019) built on the PAMP framework (Liu et al. 2008) and developed a level-set approach to optimize the distribution of multiple concurrently optimized microstructures. It was studied in the context of heat transfer problems. When the location of unique microstructures is not prescribed, it is often cast as a multi-material optimization problem, for which discrete material optimization (Stegmann and Lund 2005) and ordered SIMP interpolation (Zuo and Saitou 2017) are applicable.

As discussed in Section 3, periodic and graded microstructures can also be designed using full-scale approaches. When the structural analysis is performed on the full resolution, a poor connectivity is reflected by a suboptimal objective. Thus, full-scale approaches naturally ensure good connectivity between microstructures or subdomains, at the price of intensive full-scale analyses. Therefore, results from relevant full-scale approaches may serve as a reference for multi-scale approaches of graded microstructures. From the comparison in Fig. 4, it is clear that a logical (but often underreported) consequence of using periodic microstructures with a fixed orientation is a large reduction in stiffness.

All approaches in the previous section have one thing in common; the shape of the unit-cell domain is rectangular. However, it is very important to acknowledge the effect of the unit-cell domain on the performance. In a large amount of works that consider a single load case, the numerical examples show that the optimized unit cells resemble a rotated version of the microstructure by Bendsøe and Kikuchi (1988) (Fig. 3). However, since the unit-cell domain cannot be rotated, the shape of the microstructures is modified to account for a periodic shape, in turn reducing performance. In other words, if the unit-cell design domain was allowed to rotate, a simpler and close to optimal microstructure would have been found.

To investigate the effect of the unit-cell shape and orientation on the performance, Träff et al. (2019) did an extensive comparison of the effect of the starting guess, when inverse homogenization is used to optimize the microstructure performance. They observed that the material design problem is non-unique and highly non-convex, i.e., different starting guesses result in completely different unit-cell designs and performances. Furthermore, a starting guess of the unit-cell shape and topology based on an optimal rank-3 microstructure resulted in a significantly better performance compared to the use of a rectangular domain. The parallelogram shape of the unit cell, which was optimized as well, allowed for periodicity patterns that cannot be described by a rectangular domain as seen in Fig. 15. Especially for lower volume fractions, Träff et al. (2019) observed that a unit cell with a starting guess and shape based on an optimal rank-3 laminate significantly outperformed (e.g., up to 30% more efficient) the designs using a traditional rectangular domain. Hence, the choice of the unit-cell domain significantly influences the performance and researchers have to be aware of this when choosing a multi-scale topology optimization algorithm.

Recently, Wang et al. (2019) showed that near-optimal and periodic truss lattice structures could be obtained for multiple load cases by distorting simple Bravais-like lattice structures to a parallelepiped. Besides using a parallelogram in 2D or a parallelepiped in 3D, one can use many more different types of polygons to solve the homogenization equations (Barbarosie et al. 2017; Podestá et al. 2019). For example, in 2D, a hexagon can be used to describe a periodic isotropic hexagonal microstructure (Sigmund 2000).

From all studies discussed above, we can conclude that performing multi-scale optimization with unit cells that are optimized using inverse homogenization on a rectangular domain reduces the optimality. This effect has been acknowledged by Barbarosie and Toader (2014) who combined the optimization of the microstructure topology and shape. This approach can achieve structural designs that are close to optimal; however, an extensive comparison unfortunately has not been performed. To address the large computational cost included in solving the problem, the approach has been parallelized. To the authors’ knowledge (and surprise), this is the only work on multi-scale topology optimization that simultaneously addresses the macroscopic design and both unit-cell topology and shape to get as close to the optimal design that can be achieved on a single length scale. Nevertheless, this method does not address the connectivity of the spatially varying unit cells over the design domain. More research has to be done to efficiently blend unit cells of different shapes together without a loss in performance to further advance this method.

Two important questions that people working on multi-scale structures should ask themselves are: These two questions are related but apparently independent from each other. While multi-scale structures are most often optimized, not surprisingly, by multi-scale approaches, they can also be designed using full-scale approaches with some additional constraints (Section 3). One important motivation of multi-scale topology optimization is to accelerate the computation for optimizing structures at high resolution. Here, it shall be reminded that theoretical stiffness-optimal structures span multiple scales. A higher resolution discretization enables the appearance of fine geometric details, which may bring the performance of optimized structures closer to theoretical limit. This, however, implies significant computational cost. Thus, multi-scale approaches are introduced for means of acceleration. These include the original hierarchical approach by Rodrigues et al. (2002), de-homogenization approaches (Pantz and Trabelsi 2008; Groen and Sigmund 2018), and works along these directions.

Many multi-scale approaches, while being formulated as an optimization problem (e.g., for maximizing stiffness), restrict the solution space by enforcing one or a few repetitive microstructures that are either predefined or concurrently optimized, with a fixed orientation and/or limited density range. This restriction may be beneficial under various considerations, from structural such as buckling strength (Clausen et al. 2016), robustness (Wu et al. 2018), and multi-functionality to operational such as inspection and repair, and from manufacturability over aesthetics to sustainability. In this regard, the microstructures play the role of accounting for these considerations. From a mathematical perspective, it would be interesting to explicitly model these requirements and integrate them in an optimization problem with a high-resolution discretization. This is challenging, since a mathematical model of some of the requirements is not available yet or comes with computational complications. Thus, restricting the solution space by a reduced parameterization can be understood as a strategy to balance the defined objective and various other considerations. A general recommendation is to always motivate use of stiffness suboptimal microstructures. Too many works seem to forget this aspect and show optimized designs that are clearly not optimal with respect to stiffness.

Analogous to technical optimization, nature has been an advocate for multi-scale structures. Hierarchically organized, functionally graded structures can be found in plant and animal bodies such as bamboo and bone (Lakes 1993; Fratzl and Weinkamer 2007). These bio-structures are remarkable from a mechanical perspective while supporting biological functionalities.

Most (if not all) multi-scale approaches make use of homogenization, which assumes separation of scales, i.e., microstructure should be much smaller than the macrostructure. This assumption often becomes invalid when considering the finite resolution of manufacturing processes. A general rule of thumb is that cells should be repeated 5 to 10 times before effective properties can be trusted. This certainly means that commonly seen approaches where one macroscale finite element corresponds to one microstructure cell, where cells may vary between each element, should be used with extreme caution. Interestingly, however, multi-scale optimization, where microstructure is appropriately adapted to local stress fields, unimpeded by cell geometry or orientation (c.f., de-homogenization approaches) provides extremely good performances even for quite large periodicities, i.e., with lack of scale separation (see Groen and Sigmund (2018) and Wu et al. (2021)). The reason is that the optimal microstructures ensure purely tension-compression-dominated deformations at both micro- and macroscale, which homogenization-wise need fewer cell repetitions for accuracy than required for bending or shear dominated deformations. Under all circumstances, it is strongly recommended to verify any assumptions of scale separation with subsequent full-scale analyses. Full-scale analysis of course implies heavy computation. Fortunately, however, recent adoption of advanced linear solvers and parallel computing has partially alleviated this problem (Amir et al. 2014; Aage et al. 2015; Wu et al. 2016a). Validation by full-scale analysis in 2D starts to appear in a handful of recent papers (Groen and Sigmund 2018; Garner et al. 2019; Wu et al. 2021), and should be an integral part of the validation of all multi-scale works.

Optimization in general and multi-scale approaches in particular are about making a delicate balance between conflicting objectives and constraints, and between the quality of results and computational efficiency. In this regard, gains in one aspect are often, understandably, accompanied by losses in other aspects. Both sides are valuable for inspiring future development. These are especially important for researchers who may be less familiar with the topic, e.g., students, application engineers, and colleagues from other disciplines. Therefore, we strongly recommend to include a quantitative comparison (or a discussion if a quantitative analysis is difficult to perform, e.g., regarding aesthetics) when a new approach is introduced. Comparisons can be made on different levels, with non-optimized designs, with optimized mono-scale structures, with designs from alternative multi-scale approaches, etc. In the case of a 2D compliance design subject to a single load case, we recommend to perform a quantitative comparison with what is theoretically possible using rank-2 microstructures. To do this, we provide a Matlab script in Appendix.

Our review has been focusing on design parameterizations, which are typically demonstrated in compliance minimization under the assumption of small deformations. Accommodating stress constraints, buckling constraints, and geometric and material nonlinearities represents an important and non-trivial next step. Some recent works have started to tackle these challenges.

Going beyond structural mechanics, design of multi-scale structures involving other physics and even multi-physics is another important venue to explore. Some of the works along this direction have been mentioned in previous sections when specific multi-scale approaches were introduced. For instance, Das and Sutradhar (2020) presented an extension of the full-scale approach with the local volume constraints (Wu et al. 2018) to optimize heat-dissipating structures considering structural and thermal performance. Fluid flow through the pores in lattice structures is another interesting topic, relevant for design of actuators (Andreasen and Sigmund 2011) and biomedical implants (Challis et al. 2012). Furthermore, as discussed in Section 2.3, inverse homogenization has been used to design metamaterials with extreme or counter-intuitive physical properties such as a negative Poisson’s ratio and negative thermal expansion (Sigmund and Torquato 1996). Designing multi-scale structures composed of such metamaterials is fascinating. It can potentially open innovative application areas in shape-morphing products, soft robots, and 4D printing (i.e., with an extra dimension of transformation over time).

Optimized structures from multi-scale approaches typically have two scales. It is worth mentioning that buckling-enhanced microstructures themselves exhibit two (or more) hierarchical scales (Thomsen et al. 2018). Integrating them in macroscale structural optimization would effectively lead to three-scale structures. The microstructures in two-scale approaches are normally defined in a uniform discretization of the macro domain, and thus have the same spatial extent. Wu et al. (2016b) proposed a full-scale approach to design octree-tree like structures, where the cells exhibit spatially varying sizes. This allows to achieve a large range of variations in porosity and pore sizes. The discrete problem of hierarchical subdivision was later reformulated by continuous variables to facilitate gradient-based optimization (Wu 2018).

Microstructures are typically defined and analyzed based on a Cartesian grid discretization with a fixed orientation. Approximating free-form macroscale shape by square or cubic microstructures leads to staircasing on the boundaries. This artifact may be less of a concern when the macroscale domain is orders of magnitude larger than the microstructure size. However, with the limited resolution of manufacturing processes, the effects of this artifact may be non-negligible, both visually and mechanically. Conforming quad/hex meshing can be an alternative for constructing a boundary-aligned grid (Wu et al. 2021). However, mapping square or cubic microstructures into irregular grids may introduce error in mechanical properties. Note that homogenization assumes a repeatable domain, and thus direct homogenization over irregular cells seems not possible.

As stated in the Introduction section, much of the recent interest in designing multi-scale structures is triggered by contemporary advances in additive manufacturing (AM). While being able to produce highly complex shapes and even structural features spanning seven orders of magnitude (Zheng et al. 2016), existing AM processes are not free from manufacturability and post-processing requirements. Structural design thus needs to consider, e.g., a minimum feature size, self-supportingness (i.e., free of critical overhang), accessibility for removing unsolidified powder or resin, and auxiliary support structure (Liu et al. 2018a). In the context of multi-scale design, these requirements (or some of them) can be satisfied by carefully selecting unit cells, e.g., beam-like lattice cells (see prints in Fig. 2) or self-supporting rhombic cells (Wu et al. 2016b). Stiffness optimal rank-N laminates and closed-walled cells with stiffness close to theoretical bounds (Sigmund et al. 2016) may, depending on process, be less favorable than open-walled cells regarding the removal of unsolidified powder or resin. Along with developments of advanced manufacturing technologies, design for manufacturability will continue to be an important topic in (multi-scale) topology optimization.

The design of multi-scale structures is a topic of interest in multiple other disciplines apart from applied mathematics, computational mechanics, and mechanical engineering. We envisage that cross-pollination with material science (e.g., multi-scale materials modelling (van der Giessen et al. 2020)) as well as geometry modelling and processing (e.g., texture synthesis (Dumas et al. 2015)) will further advance optimization and inverse design of multi-scale structures. The cross-pollination is actually already benefiting the field. For instance, the conformal-like mapping and field-aligned meshing have been used for de-homogenization.