Structural and Muliscptinary Optimization bieps/dol org/10.1007/00158-019-02255-3 RESEARCH PAPER, Revisiting topology optimization with buckling constraints Federico Ferrari! ® . Ole Sigmund? ® Cheek for Updates, Received: 6 September 2018 Revsed:29 January 2019 / Accepted: 28 February 2019 {© Springer-Verlag GmbH Germany, part of Springer Nature 2019, Abstract We review some features of topology optimization with a lower bound on the critical load factor, as computed by linearized buckling analysis. The change of the optimized design, the competition between stiffness and stability requirements and the activation of several buckling modes, depending on the value of such lower bound, are studied. We also discuss some specific issues which are of particular interest for this problem, as the use of non-conforming finite elements for the analysis, the use of inconsistent sensitivity that may lead to wrong signs of sensitivities and the replacement of the single eigenvalue constraints with an aggregated measure, We discuss the influence of these practices on the optimization result, giving some recommendations. Keywords Topology optimization - Eigenvalue optimization Finite elements - Sensitivity analysis 1 Introduction Stability and buckling have attracted considerable attention since early times of structural optimization, due to their importance in the design of structural elements, Moreover, the optimal design according to weight or compliance minimization may naturally lead to structural configurations showing poor stability (Rozvany 1996; Ohsaki and Ikeda 2007), Optimal design with respect to buckling has been thoroughly studied for beam models, where even a closed form expression for the sufficient optimality condition can be found (Olhoff and Rasmussen 1977; Cox and Overton 1992; Kinmser and Hu 1995; Cox and McCarthy 1998), Such a goal has tuned out to be much more challenging already for plate models (Armand and Lodier 1978). In this case, better results have been obtained by formulating a Responsible Bitor: nj Kato 1 Federico Ferrari feferr@mek da de Ole Siground sigmund@ mek du dk Deparment of Mechanical Enginesring, Techsial University of Denmark, Nils Koppels Alé 408, 2800 Kongens Lyngby. Denmark Published online: 15 March 2019 inearized buckling - Aggregation functions reinforcement problem (Simitses 1973; Haftka and Prasad 1981), rather than allowing for only a continuously varying thickness (Frauenthal 1972), Much research effort has been devoted to the design of trusses, where stability was at first imposed on the local level, constraining the maximum stress or displacement of individual members (Berke 1970; ‘Achtziger 1999), Then, several methods have been proposed to account for overall stability of structures (Khot ct al 1976; Szyszkowski and Watson 1988; Rozvany 1996). The interaction between local and global buckling modes, of particular concern for building frames stiffened with bracing, systems, was studied by, e-g., Hall etal. (1988). Relatively few works have appeared concerning buckling in topology optimization of continuum models. In this setting, the problem becomes much more complicated, both for the less intuitive definition of the buckling mechanism and because of several issues hampering the optimization process. On one hand, intricate spatial distribution of tension ‘and compression makes it more difficult to identify simple Duckling mechanisms. On the other hand, typical issues ‘encountered in eigenvalue optimization, such as eigenvalue ‘multiplicity (Seyranian et al. 1994), artificial buckling modes (Neves et al. 1995, 2002), and existence of many local minima, are encountered, posing serious convergence issues. A large number of buckling modes are likely to become active and therefore must be considered (Bruyncel et al. 2008; Dunning et al. 2016), making the problem also very challenging from the computational point of view. DspringerF-Ferrati and 0 Sigmund Homogenization-based topology optimization aimed at the maximization of the linearized buckling load was first. addressed by Neves et al. (1995), while Rodrigues et al, (1995) provided a comprehensive study of the optimality conditions for both single and multiple eigenvalues. The later was accounted for by Falgado and Rodrigues (1998), While solving the reinforcement problem for a plate. ‘Topology optimization (TO) is constantly spreading to ‘more fields of structural engineering as a preliminary design tool, also for large-scale, complex structures (Aage et al 2017). Due tothe specific and often counter-acting character ‘of buckling design, as opposed to compliance or stress designs, it is apparent that the former must be carefully taken into account in the optimization of such complex structures. ‘Therefore, a renewed interest in topology design with regards to buckling is rising. Some works have focused ‘on alleviating the issues due to the clashing character of stiffness and stability (Gao and Ma 2015; Gao et al. 2017) or on the use of effective iterative methods for solving the large eigenvalue problem (Dunning et al. 2016; Bian and Feng 2017), Recently, Thomsen et al. (2018) presented works fon the design of periodic microstructures with respect t0 multi-scale buckling conditions, laying the foundations for plenty of future applications within multi-scale structural ‘and material design. Other recent works and applications ccan be found where we wish to mention Zhou (2004), Lund (2009), Bochenek and Tajs-Zieliiska (2015), Cheng and Xu (2016), and Chin and Kennedy (2016). ‘The goal of this note is to provide a discussion about the influence of stability requirements, here posed in the shape ‘of a constraint on the linearized fundamental buckling load factor, in a minimum compliance TO problem. A simple bat illustrative example is studied, focusing on the modification of the design and its performance as this constraint becomes ‘more demanding, and on the progressive activation of more buckling modes with either global or local character We also discuss the use of aggregation functions for replacing the constraint on the, possibly non-smooth, lowest eigenvalue with a smooth approximation (Chen et al. 2004), Aggregation functions have been extensively used in the context of TO, mainly for stress-constrained problems (Yang and Chen 1996; Duysinx and Sigmund 1998; Le et al. 2010, 2012, 2016; Verbart et al. 2017) . Relatively few works discuss aggregation of eigenvalues, for example in the context of dynamic problems (Manh et al, 2011; Torii and Faria 2017), and we have been able to locate only the ‘work from Chin and Kennedy (2016) concerning buckling constraints ‘The influence of some other issues, such as the use of non-conforming finite elements for the buckling analysis, as opposed to the popular conforming four node elements, or the adoption of a simplified but inconsistent sensitivity D Springer expression i also investigated, An ultimate and exhaustive discussion of these topics is clearly outside the aims ofthis paper. Nonetheless, we believe thatthe insight provided hese right serve as useful guidelines to researchers working in the field Finally, it is fair to point out that the linearized analysis addressed here has always raised considerable criticism in the engineering community as, for many structures, it gives an overly simplified description of a complex phenomenon as buckling (Kerr and Soifer 1968; Brantman 1977). Therefore, in all but very few cases, the estimation of buckling should be carried out by the more appropriate, yet much more computationally expensive procedure based on non-linear equilibrium (Bathe and Dvorkin 1983). Sensitivity analysis for non-linear buckling has been addressed by, e-g., Wo and Arora (1988) and Reitinger and Ramm (1995) and TO based on it has been carted out by several other authors, (ef. Rahmatalla and Swan (2003), Kemmler et al. (2005), and Lindgsard and Dabl (2013)), A critical review and comparison of the different approaches to buckling estimation for use in ‘optimal design is given inthe recent work of Pedersen and Pedersen (2018), Despite the above, linearized pre-buckling analysis is still very popular in TO, mainly due to simpler implementation and its computational cheapness. Therefore, we consider @ discussion on ito be worthwhile ‘The outline of the paper is as follows: in Section 2, wwe introduce the basic setings and recall the theoretical lements used in the further diseussion. In Section 3, we present a numerical application, discussing issues touched upon carlier Finally, conclusions are drawn and some recommendations are given in Section 4 2 Setting and optimization problem Let us consider discretized mechanical system, repre- sented by m finite clements and n degrees of freedom (DOFS), We address linearized (eigenvalue) buckling analysis (Cook et al. 2001). This requites the selection of a reference load fo € R" and the solution of the linear system Kup = fo w whore K © RY" is the linear, symmettic, and positive definite stiffness matrix and to € IR" isthe equilibrium dis- placement vector. Given this, the symmetric but indefinite sires stifiness matrix K, (ug) ¢ R"* can be setup and the generalized eigenvalue problem tobe solved is IK42K, (a)lg=0, 9 40 @Revising topology optimization with buckling constraints ‘The cigenpairs (2. g,), 1 € Bp consist ofthe critical load factors i; and the associated buckling modes g,, normalized such that 97K. Tn this work, we are particularly concerned with the fundamental buckling load factor 2), associated with the critical load fey = fo. Therefore, from a computational point of view, itis more convenient to refer tothe eigenvalue equation UK (to) — wK] 9 = 0 aay ozo Which is equivalent to (2) given the relationship & = —1/1, such that Ay is associated with the minimum algebraic value of p, say 1 For densty-based topology optimization (Bendsse and ‘Sigmund 2004) performed on a regular grid, we consider the design variables x = (se)? belonging to the set pe feetoum % where f is the allowed volume fraction, and the design- dependent matrices K (x) and Ky (x, Uo) are assembled ‘yom the element ones, say Ke (¥-) and Be (Xe, Woe). These are parametrized by interpolation functions a8 Ke (te) = ‘hy Gre) ko and ge (Xe Woe) = hz (e) Bo (Woe), Where Ko and 40 (Woe) are independent of x, and te denotes the restric- tion of the global displacement vector to the element level For details and implementation of the element stress stiff- ness matrix go (Use) we refer the reader to de Borst etal. 2012) ‘Considering an isotropic base material, in the following hy (ie) = Fo+ xf (Ei Fo) ha (te) = APE @ where £} is the Young modulus of the solid, Ko that of the void, and p is the penalization factor. Equation (4) has proven to be an effective choice against autficial buckling modes (Gao and Ma 2015; Thomsen etal. 2018), atleast forthe material contrast of interest (Ei/Eo = 10%). ‘The problem of minimizing the linear compliance with an imposed lower bound 7. on the fundamental buckling load factor reads mip J = uf = uf Kuy xpos ° st mip > FE me where BC By is the subset of eigenvalues considered, ‘hich should be large enovgh to consider all the relevant buckling modes and to preduce a smooth behavior of the cptimizer Brayneel etal. 2008; Dunning tal. 2016). The sie constrsinin (5) can be replaced by the st Fal t1e0, 1eB o where the number a > 1 serves to introduce small gaps between the eigenvalues, preventing them from completely coalescing (Bendsse and Sigmund 2004). ‘The solution of (5) with an iterative, two-level, gradient- based approach (Hlaftka and Gurdal 2012) requites the solution of (1) and (2) and the computation of sensitivities used for updating the design variables, For a simple eigenvalue 4), the sensitivity with respect to the design variable x, reads (Rodrigues etal. 1995) a 7 (# 1H) _ yt K, ae where vis obtained by solving the adjoin system Ky = @7 [VaKol ® ‘The first term in (7) is formally the same as the sensitivity for a dynamic eigenvalue, and in the following, we will refer to it as the “requency-like” term. The second term is the adjoint term, accounting for the dependence of the stress stiffness matrix on the stress level in the prebuckling solution, and the variation of this as the design is changed (Rodrigues et al, 1995). cis sometimes claimed that the latter term is of minor importance and can be neglected (Ye et al, 2016; Munk etal. 2017). However, ignoring this erm results in the use of inconsistent gradients and may potentially lead to incorrect designs, as wil be discussed later. Equation (7) does not give univocal value for a repeated eigenvalue, as more than one eigenvector becomes associated with it In this case, the definition of the sensitivity becomes more involved, calling for the use of subgradients, and we refer to Rodrigues et al. (1995) for details. In this work, we do not consider such mote difficult ‘teatment, either because setting a > 1 in (6) will ensure a certain separation of the eigenvalues or a an alternative, we ‘will make use of the aggregation functions discussed below. For completeness, we also report the sensitivity for the eigenvalues 1, an, or (Ke | aK ak ig, ) OZ 9) tA rometaree ke 0 ° which is the one actually adopted when referring to (6). The adjoint vector v is again obtained by (8). 2.1 The finite element approximation Low-order elements, often prefered for density-based TO, may result in an inaccurate representation of stresses, which are of primary importance in buckling analysis. In this regard, conforming elements are known to suffer from DspringerF-Ferrati and 0 Sigmund shear locking and therefore to poorly represent a bending dominated behavior. Although this is usually neglected in compliance design, since the optimized topology mainly consists of tension and compression members, it might play a role for buckling. A number of accurate low-order elements have been developed in the past, mainly relying on the enrichment of the compatible displacement field with incompatible modes (Tumer et al, 1956; Wilson et al. 1973), or derived from mixed formulations, as hybrid stress models (Pian 1964; Pian and Sumibara 1984) or enhanced assumed strains ‘models (Simo and Rifai 1990). The main advantage of all these models is to substantially increase the accuracy of the siress approximation on coarse meshes without introducing new nodes or explicit DOFS. ‘The classic hybrid stress clement of Pian and Sumihara (1984) has been applied to buckling TO by Gao and Ma (2015) and Gao et al. (2017), however without discussing its advantages over the conforming Q4 element. Here, we shortly analyze the behavior of a popular incompatible element, the so-called Qs clement, proposed by Wilson et al, (1973), We remark that the conclusions we draw also apply to the Pian element, since the two have proven to be equivalent by Fraier etal. (1974), Let us consider @ fournode isoparametric element rapped to the bnnit square (-1,1}* with two DOES at each node (ie., we = (W, vi, u,v}, for a Qs clement the displacements are discretized as {i}en=daenft] 0-e) {nf -0-2) [af where NG.) = £46) nin), i = 1. 4 ane the shape functions of the bilinear Q, element. The last two terms are the so-called incompatible modes, allowing Fig. 1 Accuracy of angle approximations obtained by using Q, or Of diseretizaions forthe comprested column 220] example, Reels obtained with the Oselement ae identical (up 22 vo numerical precision) to those ‘obtained with the Pian-Sumhaza PS) clement D Springer wale to exactly represent a constant bending deformation, and depending on nodeless DOFs a!) and bY? “To test the influence of the incompatible modes on the accuracy of buckling analysis, let us consider the following example, recently studied by Pedersen and Pedersen (2018) ‘cantilever beam with dimensions Ly = 10, Ly = 1 and thickness ¢ = 0.005 (all measures non-timensionalized), subjected to uniform compression at the tip. The elastic modulus is £ = 2 10! and we set v = 0, Due to this parameter choice, the 2D plane stress model is equivalent to the 1D beam model, and the fundamental buckling load can bbe computed by the closed form expression PEL 2 = Ely x 295617108 10) aE 0 where fy = 113/12 is the modulus of inertia for in- plane bending (see the sketch in Fig. 1). We choose a reference load vector with magnitude |[fo|] = 2.5 - 10°. The convergence behavior and relative errors associated withthe Qs and the Qs element discretizations are reported in Fig. | “The use of Ox conforming elements results in consider able errors on coats discretizations with an overestimation of Pe by almost 50% for a column described by 10 x 4 ele- ments. The error is reduced to the order of 10-* and 10-> ‘when increasing the numberof elements by 8 and 32 times, respectively Use ofthe Qs element sensibiy improves the accuracy on coarse discretizations, Moreover, iti important to observe that now the convergence is attained from below, since with the incompatible finite clement model, the discretized stractureis softer than the real one. For the Qs element, we acknowledge avery mild convergence asthe mesh is refined and therefore the performance of the two elements becomes very similar on fine meshes. From the shove discussion, it should be clear that twckting modes might be inaccurately predicted. for ‘optimized designs consisting of thin slender bats, described by few Oy elements, The performance of the two elements Txt 1x4 0493 mox2 0119 mx ols 4ox8 002 80x18 5.0-10-* 10x32 8.6-10- Tass?Revising topology optimization with buckling constraints for the analysis of optimized designs and their influence ‘on the optimization itself will be further discussed in Section 3.3, 2.2 Formulation by aggregation functions ‘We can replave (6) with a single constraint FM (wil+1=0 an where M [-]is a suitable aggregation function (Chen et al, 2004). Various functions can be adopted in this regard, and ‘the most popular ones in the context of TO are the norm, tbe M, (ui) = (So) a2) i ‘where we use that the j4; of interest are strictly negative, and the KS function (Kreisselmeier and Steinhauser 1979) Mrs (ui) = 1 + ce (geen y ay iB Both (12) and (13) produce an upper bound, smooth approximation t© 1 = maxlurl where the degree of smootimess is governed by the aggregation parameter ‘Therefore, we indirectly obtain a smooth lower bound to Ideally, 1 is recovered as > 20. ‘The KS function is usually preferred if «high p value has tobe used, due to its superior numerical stability. For a more ‘complete discussion about the properties of these functions and for an overview on some more general aggregation strategies, we refer, for example, to Raspanti etal. (2000) and Kennedy and Hicken (2015). a) ()) =o buckling modes (€) DD ‘The sensitivity of the aggregated constraint (11) can be easily obtained as aM, (1 1) 3a 9-9 Dyer a4 ou ane Bie when the p-norm is used and aMxs aus te Sapo x On as) ‘when using the KS function, In the above, the sensitivities 8; /@xe are given by (9), One of the advantages of using the aggregation functions is the uniqueness of the gradient defined through (14) (viz, ((15)), even when eigenvalues are repeated. A proof of this, based on the argument of Gravesen et al. (2011), can be found in Torit and Faria (2017), for the p-norm function ‘The same argument trivially extends also to the KS function, due to the continuity of the exponential and logarithm functions involved, 3 Numerical example We refer to the configuration of Fig. 2a, where the gray regions are fixed to be solid in the optimization. The load, having total magnitude F = 0.02, is distributed over the cedge of the gray region on the right side, centered in the ‘id point c, and the pinned points a and b on the left side are on the 45° line from c, We here assume the load to be unidirectional and pointing downward, and therefore a single load case needs to be considered and only positive eigenvalues of (2) are accounted for. “The discretization is set up with = 90 x 210 elements and the solid regions consist of 9x 9 elements. We consider Ey = Land Eo = 10-8 in (4), and the Poisson ratio is v =033. (©) 10g (6¢/4mex) 2 Geometrical seting forthe considered example (a) compliance design (b) an distribution ofthe strain energy density forthe lowest four D springerF-Ferrati and 0 Sigmund ‘The penalization parameter is initialized with p = 1 and then raised in steps of Ap = 0.25 each 25 optimization steps, up to the value p = 6, With this continuation approach, we do not start with a too weak design, and we increase the chances of ending in a good local minimum, More advanced, adaptive continuation strategies can be used and for this, we refer, for example, to Rojas-Labanda and Stolpe (2015) and Gao et al. (2017). A projection with threshold n = 0.5 and sharpness parameter 8 = 6, based on a filter radius of rain = 2 elements (Sigmund 2007; Wang et al. 2011), is applied to control the structural features size. ‘We consider the optimization problem (5), for a volume fraction f = 0.2 and several values of the imposed minimum buckling load factor P;, At each step, we solve (2) and compute the first 24 cigenpairs from (2). A criterion based on the strain energy density, as described by Gao and Ma (2015), is used for assessing the physical ‘meaningfulness of the modes, ic. for filtering out the ‘ones corresponding to artificial deformations of low-density regions. However, we point out that in the following testing, ‘we never experienced this issue. The lowest 12 buckling Toad factors are then included as constraints, in the fashion of (6). For all the following examples, we set a = 1.01, for imposing the eigenvalue gaps. The design variables are updated by the method of moving asymptotes (MMA) (Svanberg_ 1987) and the optimization is run for a fixed ‘number of 700 iterations. ‘The compliance and the fundamental buckling load for the design with uniform material distribution are Jo 4.71410" and 31,0 = 0.662, respectively. The compliance design (P; = 0) consists of two identical bars, the upper one mainly in tension and the lower one in compression (see Fig. 2). The strain energy density (defined on the clement level as ¢- = 97k.) corresponding to the lowest four buckling displacements is shown in Fig. 2c and we recognize global bending modes in the lower bar with progressively shorter wave lengths. 3.1 Influence of the buckling constraint ‘The designs obtained for increasing values of, ae shown in Fig. 1. As the buckling constraint value is increased, the upper tension bar becomes thinner and the material is relocated to the lower part of the domain, where a progressively wider trss-like configuration develops, Such ‘configuration improves the resistance of the structure with respect to global bucking ‘We also noice the appearance of a bar connecting the original upper and lower bars near point c. This element. becoming progressively thicker, hasan important stiffening IIDD @F=0s — (b)R=o75 ()Fe=10 (a) B= 1.25 @ Peis DID D (QF). (e) = 18 () = 20 Fig.3. 2 Optimized designs obtained forthe example described in ig. 22 for increasing values of the lower bound on the fundamen- tal buckling load factor, Designs marked with an asters and double D Springer ( Fe=20" () Fe= 20" asterisks have been obtained starting from non-uniform inital guesses {is obtained from the compliance design of Fig. 2 and j from the design coresponding to Pe = 1 (se panel e)Revising topology optimization with buckling constraints effect, ensuring a global behavior of the structure against the rotation round point c, whichis highly strained inthe fundamental buckling mode (se Fig. 2). ‘The qualitative behavior of an optimized structure, e.. the one corresponding to Fr = 18, can be studied by looking at Fig. 4, Here, the plots on the first row show the distribution ofthe stain energy densities forthe eastic displacement and for the lowest four buckling modes. The plots in the second row display a subdivision ofthe regions where the stess energy density associated with a given displacement vector » and defined on the element level as Ye = oF gop, is either positive or negative. Positive (red) Indicates stiffened domains and negative (blue) indicates softened domains. For the pre-buckling displacement, the deformation energy is highest in the upper bar, which is in traction, While the lowest part, mainly in compression, is softened However, some of the inner bars are actually in tension, as they counteract the spreading of the two outer elements. As the load multiplier increases, the stiffness of the lower part is reduced, up to the occurrence of buckling. We can see from the strain energy distribution, and from the distribution of traction and compression regions, thatthe fundamental buckling mode consists of a global bending of the lower pat. Moreover, the short connection bar near point « is highly stained, which confi its usefulness in restraining this instability mode, whereas the thin bars of the infill are only marginally affected by this deformation. Forth higher buckling modes, the deformation progressively affects aso these and the modes become more localized, Interestingly, some bending waves similar to those appearing for the log (6e/ 40x) Fig. 4 Logarithmic plot of the rlaive stain energy density (rst to) and qualitative plot of the fegions subjected to stiffening and softening (Second row) forthe design of Fig. 3g. atefers to the elastic Aisplacement and b-e tothe lowest fur bucking modes compliance design can still be qualitatively recognized, especialy if we look at the distribution ofthe sess energy ve For high values ofthe buckling constraint another linking bar is prone to appear (see Fig. if, hi). This bar, however, tums out to provide an inconvenient deployment of material and this design hence clearly constitutes a. sub-optimal solution. Indeed, we observe that th is bar is not strictly necessary in order to improve the stability of the system, 48 the design of Fig. 3g, without this bar, has a higher 7c compared to that of (), and the design of Fig. 3j bas the same P, as those in (h, i). “The intoduction ofthe buckling constraint considerably increases the complexity of the feasible set and the chances cof ending ina local minimum are increased as this constraint becomes more influential, A natural precaution can be to start witha non-uniform material distribution (Rendsse and Sigmund 2004), closer to the design looked for The designs in Fig. 3, j for example, ae obtained for the buckling constraint P, = 2.0 starting from the designs comresponding to F: = Oand F = 1.0, respectively. We see that the longer linking clement is first made shorter and then temoved, to the benefit ofa slightly thicker upper bar. It is also intresting to noice that although the short bars in the infill have different configurations forthe three last designs, this seems not to affect significantly the buckling response, since the critical load and the corresponding mode remain practically unchanged. “The performance of the obtained designs is quantitatively assessed by looking a the ratio between the final and inital compliance Jy = Je/Jp and atthe parameters 8 = ala) i> 2 a6) which approach zero as the i-th buckling mode tends to coalesce on the fondamental one (with @ separation of wy Numerical values of these parameters are collected in Table 1 and their “at-a-glance” representation is given by Fig. 5. Here, the be continuous curve, ploted agsnst the left axis, shows the rato hetwecn the sifness of a design obtained for a given buckling constraint and that for Pe = 0, namely x = Jy7c/,7220: From the trend of «, we clearly see thatthe desighs optimized for a fixed volume fraction cannot retain the same stifiness as 7 is increased Moreover, « becomes generally steer as Pr increases, as also seen from the values of the slope 4/47. reported in Table 1. Therefore, for high values of P,, a very compliant structure emerges during the optimization, and even convergence to a well-connected design can be hampered ‘The behavior of the eigenvalue separation parameters defined in (16) is shown by the dasbed curves ploted against the right axis and we acknowledge the progressive aetvaton of DspringerF-Ferrati and 0 Sigmund Table 1 Compliance reduction factors, local slope ofthe Rh Anse ‘ompliance-bucking load ee felationship, snd eigenvalue 09 0561 0.0326 as 657 163 separation factors (as defined 5.5727 0.712 189 aus 329 ty 6) for te ond 075 05903352 065 142 1m aaa e 10 asst 0.2632 02s ost 093 125 0.6899 0.2628 Liao 2sio $3407 150.7556 0.3020 85.10% rsa? 1.440- 175 o8811 0.7080 2010-5 so. 2340-9 18 08664 0.5325 2910-5 s610-+ —as10-* 20 09m = L2a0- so10t 20° 09688 36.10 1540-3 52.10% 5.0% 20 ogg = 2 ‘an increasing number of buckling constraints. Already for Py = 0.75, we have a bi-modal Ay and for Fe = 1.25, the lowest eight eigenvalues are all very close. Then, they become all simultaneously active for the highest value Pz = 2.0. ‘The general trend is that the finer the bar distribution in the infill, the more the buckling modes coalesce. From this simple example, we can see that a large number of buckling ‘modes may contribute to the optimization, increasing the complexity of the problem, and this is bound to get even ‘mote pronounced for higher mesh resolutions and lower volume fractions. 3.2 Designs obtained on a finer discretization Figure 6 displays some designs oblained on a finer discretization 2) = 180 x 420, for decreasing values of the ‘volume constraint Fig.5 Performance of the designs optimized for inczeasing values of. he lower bound on the fundamental buckling load factor The ble curves show the tend of the siliness ratio r when using Q, (oid ceurve) and Qs (dashed eurve) elements forthe discretization D Springer ‘The designs in the first row refer to a filter size scaled according to the mesh size (1¢., rig = 4 elements), while ‘those in the second row are obtained for the same filter size (Compared to element size) used for the coarser mesh 21 Ge. rig = 2 elements) In all the cases, the comtesponding compliance design is used as initial guess, to facilitate the ‘good solutions. Now, 32 eigenpairs are computed and the lowest 24 buckling loads are included as constraints, since ‘we expect a higher number of modes to coalesce. f=o2 f=018 °. ® » DIDI ©) 0 ®) w 16 f=0at o @ Fig. 6 ah Designs obtained on the mesh 2) = 180 x 420 for decreasing values of the volume constraint and forthe fier sizes of ‘yn = 4 elements (ist row) and rnin = 2 elements (Second 10). The initial guess i, forall the eases, the corresponding compliance designRevising topology optimization with buckling constraints Fig.7 a The modification ofthe 10! compliance (refered to left 4 xis) and ofthe fundamental buckling load factor (ferred to the right axis) forthe uniform sand optimized design onthe finer mes 23, Solid lines refer {0 the case Fig = 4 and dashed ones othe cast Fain = 2B The trend of the eigenvalue gap parameters by defined i (16) for She designs with low fer size and decreasing allowed volume All these designs ideally relate tothe one coresponding to Be = 18, shown in Fig. 3g. However, since the approximate buckling load reduces, beth due to the mesh refinement and as the volume is reduced, it becomes extremely hard forthe optimizer to find a design coping with such a high buckling constraint. Therefore, P, is scaled according to the trend shown in Fig. 74, and its numerical values are 1.55, 13, 1.148, and 1002, for the four decreasing volume fractions (found by numerical experiments) From this figure, we se tat the buckling load factor (41.0) and compliance (Ja) of the uniform material Aistibation ae linearly reduced and increased, respectively, as f decreases. On the other hand, the compliance of the optimized design (2) seems to increase more than proportionally, The same tend is recognized for both the cases of filter rit Aste volume fraction f is reduced, a more complicated infill develops in the lower part, with the bats becoming thinner, The connection bar near the right end also becomes thinner and moves progressively away from point c. These rodifications are further emphasized for designs obtained with the smaller filter size, where finer structural features axe allowed. Figure 7b shows the trend ofthe 8; parameters, as defined by (U6), corresponding to the constrained eigenvalues, We see that mote nodes ate becoming active, as expected. For the volume fraction f= 0.16, the first ten i are below 1072, which means that do is only 12% larger than 2. For the case f = 0.14, the same ten go even below 10-?, and all the 4; are below 5-107, which means that 424 < 1.31. 3.3 Analysis and design with Q¢ elements All the previous designs have been obtained by using (Q; clements for the discretization, We now analyze their response also by using Qs elements rom this analysis, we find thatthe computed compliance and fundamental buckling load factor are very similar —y=02 —y-o18 j=a18 sou 0 with differences of the order of few %c and below 1%, respectively. Nevertheless, important differences in the nature of the buckling modes can be seen by looking at Fig. 8, where the strain energy density associated with the (wo lowest buckling modes, as predicted by Qy and Qs discretizations, is displayed for three representative designs For the design corresponding to F = 1.8 and the mesh £2, the buckling load predicted by the Qs discretization is, 24 = 1.791, ic, only 0.5% lower than the one given by Qt clements, However, the leftmost st of plots in Fig. 8 shows that the strain energy is much more localized in the infill, where one of the thin bars undergoes local buckling, For the second mode, the maximum stain energy is still oceurring in the same bat. “The central set of plots concerns the response of the design from Fig. 6g, and similar considerations can be rade, Again, some of the bars in the infill appear to be highly strained for the buckling mode predicted by the Qs clements, though the numeiieal value ofthe buckling Toad factor is only 0.65 smaller. ‘As an extreme case, we include the lefimost set of plots in Fig. 8, concerning the design from Fig. 6h, This fone shows some very thin bars and we acknowledge the completely different nature of the buckling modes, which become extremely local as predicted by the Oy elements ‘The fundamental mode involves only a thin bas, whicl cleatly indicates a non-optimal feature, described by very few elements. Also, the second buckling mode involves the failure of a small bar of the infill, which is subjected to a high compressive stess, These mechanisms are completely missed when using Qj elements, where such regions seem to be not mich strained, At this point, it should become apparent that even though the differences in the predicted buckling load factors appear to be negligible, the substantial discrepancy of the buckling modes is expected to definitely play a role in the sensitivity, and therefore in the optimization outcomes. DspringerF-Ferrati and 0 Sigmund 9, $002, eis f= 016, FE on ‘Compasison ofthe stain energy issibuton (og($e/ous) of he two lowest buckling modes for three optimized design, analyzed with Q, and Qe elements. We observe that inal the cass, the buck ling mode predicted by Qf elements is more localized. involving some In order to investigate this, the previous examples have been solved again by using Qs elements. Design variables are referred to the element center and therefore filtering is still needed for the stability of the approximation (Jog and Haber 1996). Some of the resulting designs are shown in Fig. 9, and we notice that the main differences to those collected in Fig. 3 concer, as expected, the bars of the infill, ‘which now appear to be generally thicker. Looking back at Fig. 5, we see the difference in the performance of these designs, compared to those obtained by using Qs elements. As we already pointed out, the discretized structure is now softer and the buckling load fof the uniform design is about 2% lower than the one PIID ()R=15 0) = Fig.9 Designs obtained by using Qs elements forthe state discretization. a-e refers to the mesh £2 and compares with design of F 38 a0 9d d0 )F=20 4, f= 044, Fe= 102 fine ba. As an extreme case, forthe design obtained on the finer mesh 1; and for f = 0.14 and rig = 2, the fondamental bueking mode involves an isolated member predicted by conforming Qs elements. However, referring to the compliance ratio Jy and to the stiffness parameter x, the comparison is still consistent. For low values of P., the differences between the two sets of results are negligible. Then, the curves start to drift as Pr is increased and the stiftness ratio « corresponding to the structures optimized with Q, elements is systematically lower. This is again expected, as the softening of the discretization implies that the buckling constraint is becoming more demanding. The difference in the « value attains, for Pl) = 2, a maximum of 3%. For the designs on the finer mesh $2, we even recognize more evident changes of the topology, with the removal oa @ms compares withthe design of Fig hand e shows the distribution of the tain energy density comresponding tots fundamental bueking mode D SpringerRevising topology optimization with buckling constraints ‘of weak thin bars. For example, Fig. 9d shows the result corresponding to f’ = 0.14 and we notice the removal of the thin bar responsible for the buckling. From (e), we then see that the fundamental buckling mode now involves more globally the lower part. Again, the cross check of all these designs, ic., their analysis by means of Q, elements results in very litle numerical differences for both the compliance and the buckling load. In summary, from our experience, the use of Qs in stead of Qj elements seems to affect only litle the value of the response parameters: compliance and buckling load factors. Also, only minor differences appear in optimized designs However, the predicted buckling modes can be different from those given by Qy discretizations, where the stiffness ‘of small features may be overestimated, This is particularly pronounced for higher order buckling modes. Therefore, ‘considering that the use of Og elements neither increases computational cost nor over-complicates implementation, Qe elements are in general recommended but do not seem imperative for linear buckling topology optimization, 3.4 Influence of the adjoint sensitivity term ‘We now claborateon the use of inconsistent sensitivities for the buckling constraint. “The set of examples discussed in Section 3.1 have been solved again considering ony the frequency-like term of the sensitivity (7), te, (i), <0 (ae) “ and neglecting the adjoin term —2,¥7 [9K /2%¢] Wo For_low values of the buckling constraint (PF; = 0.5 and Fe = 075), this causes only marginal differences in the optimized design, mainly localized along the boundaries (Gee Fig, 10a, b). However, we point out that already for these cases, we have a (small) worsening of the OP Fe Sey O)F=07 (jR=10 () F=20 Fig. 10. Designs obtained by using an inconsistent sensitivity a and b “The diference with respect to the corresponding designs of ig. 3. For hese, we have the compliance ratios Jy = 0.5758 and Jy = 0.5963. For the design of ¢, we have J, = 0.6597 and for the material Aistibution ind, the opinsization breaks iteration 150 performance, as Jy increases by 0.4% and 1% respectively, compared to what i reported in Table 1 For the buckling constraint 7 = 1.0, the differences become much more evident (see Fig, 10e) and the increase in Jy ix about 69. For higher values of the buckling constraint, even the obttinment of a design becomes dificult, As an extreme situation, we can look at Fig, 104, showing the material distribution at iteration 150 for the case P, = 20, Here, the optimization clearly breaks down, ‘To further clarify the importance of the adjoint term in driving the optimization towards a meaningful design, wwe refer to Fig. 11, showing the distibution of the total sensitivity (a) and of the frequency like (b) and adjoint (c) contributions for the optimized designs corresponding Fe =10and Fe = 18 of Fig. 3 The distibution ofthe frequency-like term qualitatively follows that of the strain energy associated with . We notice that this term becomes strongly negative near the Toad and support, and therefore this contribution alone would remove material from thete. However, in the pre buckling state, these regions show high stress gradients This is precisely reflected by the adjoint term, which there aktains ite maximum values (see Fig. 1c). It is thus clear by comparing the consistent sensitivity (plots (a) with the inconsistent one (plots (b) that neglecting the adjoint term, ‘which magnitude is comparable to that ofthe fequency-like ‘one, results in gradients with wrong sign in highly stressed ones. This at best leads to very unstable convergence Dot oo @ eo 6 Fig. 11 Plot ofthe sensitivity 3% /8x (a, of the frequency-like (b) and adjoint conbutions(€ forthe designs comesponding to P; = 1.0 (rst row) and F, = 1.8 (second row). The color scales refer to the Same range for plots ae DspringerF-Ferrati and 0 Sigmund Table 2 Results obtained with the KS function for teveal values of seeregation parameter ’ Jn & 16 0964 019 ONDA} ORS 20 os941 0.057.146.0243, 0.308 so oss 0047 0.076 =~ 009 0a 100 08376 ©0018 0.000037 16128 087030005 0.006. lsat 24192 0861 0.005 0.005 0.006.007 ‘The last two rows refer tothe application ofa continuation scheme on he aggregation parameter, eg staring from p = 16 and doubling t each 100 erations upto p = 128, for high values of P-, as this tendency becomes more prominent In our numerical testing, we have considered several other design problems from the literture and we have clearly seen that the adjoint sensitivity term always plays an important role in driving the design towards meaningful and efficient configurations. Only for very special cases, and low values of the buckling load (i.e, when moving not too far from the compliance design), neglecting this term may ive, to some extent, acceptable results Our conclusion is therefore that the use of an inconsistent sensitivity for topology optimization of continuum models invalving linearized buckling, sometimes appearing in the literature (Munk et al. 2017), must be avoided 3.5 Results obtained with the KS function We focus onthe case FE = 18 and we resolve this using the KS function to aggregate the eigenvalue constraints as in (0). 1 3 F F Hg 1 2 a} § & F F 1F 4 5 6 ‘We first observe that, due tothe conservative character of (3), replacing (6) with (11) amounts to solving a problem ‘with a slightly higher buckling constraint. This can cause convergence issues, because ofthe delicate balance between stiffness and stability requirements. A first remedy to this is to scale the constraint (11) according to the relative weight of in the aggregated measure, Therefore, we define the scale factor = fl Mal Which is updated as the material penalization parameter p is raised and jumps in the eigenvalues occur. This parameter is not considered in the sensitivity relation (15), as itis updated ‘only occasionally. This approach is conceptually similar to that adopted by Le et al. (2010) for stress-constrained problems, ‘The results obtained for some choices of the aggregation parameter p arc reported in Table 2 and the corresponding designs are shown in Fig. 12. From this figure, we can also see the stiffness of the designs, compared to that of the design in Fig. 3g. ‘The aggregation parameter clearly plays an important role but, as a rule of thumb, designs reasonably close to those of Fig. 3g ate obtained when p is set larger than 20. With = 50 or above, we obtain designs which also ‘outperform the ones obtained with the original approach, even starting with a uniform material distribution. We recognize a twofold effect of the aggregation parameter: as) 1. With a high value of p, we have an envelope which is closer to 11, and therefore the overshooting of the ‘constraint is reduced, 2. ‘The gaps between the eigenvalues will be reduced by increasing p (sce 8; values in Table 2). os 6 vee. 095} Fy of %e @ p fixed ww ° Fig. 12. Designs obtained by using the KS function for aggregating the single constraint (6), forthe case FE — 1.8 and for several values ofthe Agetegation paramster The graph on the right plots the ratio between the sifiess ofthese designs and tht ofthe design in Fg. 3 D SpringerRevising topology optimization with buckling constraints ‘The importance of choosing a relatively high value for the aggregation parameter, even for application to stress constraints, has recently been pointed out also by Zhou and Sigmund (2017) ‘Regarding the second point, performing a continuation fon the aggregation parameter gives very good results (see also Fig, 12). Here, the continuation strategy adopted was quite heuristic, as the parameter p was simply doubled ceach 50 steps, For a more systematic continuation scheme, tuned on the progresses of the optimization, one can refer to Poon and Martins (2007). However, we remark that also when employing continuation, the starting value of the aggregation parameter still has great importance. The choice ‘of a low starting value (¢.g., ¢ = 16) results ina bad design, whereas a very good one can be achieved by starting from p= 24 (see Fig. 12) Similar considerations hold for the use of the e-norm function. From our experience, a design equivalent to that of Fig. 3g is obtained upon setting p > 8. 4.Concluding iscussion We have covered several issues related to topology ‘optimization with linearized buckling constraints. ‘With the help of an illustrative 2D example, we have demonstrated the influence of the buckling constraint on the optimized design, Furthermore, we have discussed the effect of some practices, as the use of non-conforming finite element approximations, the use of a simplified but inconsistent sensitivity expression, and the replacement of the separate eigenvalue constraints with an aggregated one, ‘The following main points are highlighted: = We recognized the balance between stiffness and stability inherent to this problem, as well as the ‘coalescence of several buckling modes contributing (0 the optimization, The implications of these features in complicating the optimization for high values of the ‘buckling constraints have been discussed. = Incompatible (or equivalently mixed) finite elements improve the accuracy of the buckling analysis on coarse meshes, compared to conforming ones. Regarding the ‘optimization, using these elements is advisable when the design is likely to consist of many thin features, ‘which are prone to undergo local buckling. = The use of inconsistent sensitivities, obtained by neglecting the adjoint term, results in gradients having ‘wrong sign in regions which are highly stressed in the pre-buckling state Ithas been clearly shown that this is at best the source of sub-optimal results and, for high values of the buckling constraints, leads to completely wrong of non converged results Aggregation functions have proven to be very effec Live, leading to results which are essentially equivalent to those obtained by ditectly constraining the mini- ‘mum eigenvalue, However, the aggregation parameter should be selected sufficiently high to obtain a close approximation to the extremal value, and not to over- shoot too much the constraint. ‘As a final remark, we conclude that several issues concerning TO with stability are far from being resolved, and a proper treatment of buckling, even in its linearized form, is relatively complicated. An issue not discussed here is the large computational cost due to the eigenvalue analyses, and this will be a topic for future investigations, aimed at the application and extension of efficient solution ‘methods Dunning et al, (2016) and Ferrari et al. (2018). The present paper should be a helpful discussion for such further researches, aimed at including buckling in large-scale TO. Acknowledgements The current project is supported by the Villum Fonden through the Villum Investigator Project "InnoTop” The futhore aze_gritefol to Prof, Pauli Pedersen for several fui Aiseusions on te topie ofthe paper, References ‘Aage N, Andreassen E, Lazaroy BS, Sigmund © (2017) Gige ‘voxel computational morphogenesis for structural design. Nature 550(7674)'84-86 Achiziger W (1999) Local stability of trusses in the context of topology optimization, part I exact modelling, Strvet Optim 1235-246 ‘Armand IL, Lodie B (1978) Optimal design of bending clement, Int ‘Numer Methods Eng 13:373-384 Bathe KJ, Dvorkia E (1983) On the automatic seltion of nonlinear finite element equations. Comput Struct 17(5-6):871-879 endsge MP, Sigmund O (2008) Topology optimization: theory ‘methods and applications. Spinge, Berlin ‘Berke L (1970) An efficient approach tothe minimum weight design ‘of dellestion lined stuetires, AFFDDL-TM-10-4 Bian X, Feag Y (2017) Large-scale buekling-constained topology ‘optimization based on essembly-fie finite clement analysis. AdY Mesh Eng 9): 1-12 ochenek B, Tajs-Ziclisks K (2015) Minimal compliance topologies {or maximal buckling load of columas. Suuet Mulidiseip Optim 5115) 1149-1159 Braniman R (I977) On the ute of linearized instability analyses to investigate the buckling of nonsymmetical systems. Acta Mech 2641) 75-88 Broyneel M, Colson B, Remouchamps A (2008) Discussion on ome convergence problems in tickling optimisation. Struct Multidiscip Optim 352) 181-186 (Chen X, QT, QiL, Teo KL. (2004) Smooth convex approximation to the maximum eigenvalue function. J Gioh Optim 30(2):253-270, G. Xu L 2016) Two-scale topology design optimization of stifened or porous plate subject to out-of-plane buckling constraint. Struct Muliiseip Optim 54(5):1283-1296 Chin TW, Kennedy G1 (2016) Large-scale compliance—minimization sand bucking topology opimization of the undeformed common research model wing. In: AIAA SeiTechForum ch DspringerF.Ferarand 0, Sigmund Cook RD, Malis DS, Pleshs ME, Witt RI (2001) Concepts and applications of finite element analysis, As edn, Wiley, New York Cox 8, MeCarthy C (1998) The shape of the tallest column. SIAM J ‘Math Anal 29:547-554 (Cox S, Overton M (1992) On the optimal design of columns against ‘buckling. STAM J Math Anal 28 287-325 de Borst. Crisfield MA, Remamers JIC, Verhoosel CV (2012) Non- linear nite element analysis of solids and stuctres, 2nd eda Wiley Dunning PD, Ovtchinnikov B, Scot J, Kim A (2016) Level-set topology optimization with many linear buckling constraints using and efficent and robust eigensolver, Intemational Journal for famerical Methods, in Engineering Duysinx P, Sigmund 0 (1998) New developments in handling opal stress constraint in optimal material distributions, pp 1501-1509 Fema P, Lazarov BS, Sigmund 0 (2018) Eigeavalue topology ‘oplimization vis efficient multilevel solution of the Frequency Response, Int Numer Methods Eng 115(7)$72-892 Folgado J, Rodrigues Ht (1998) Structural optimization with & non “smooth buckling load enterion. Contzol Cybern 27:235-253, Frauenthal JC (1972) Constrained optimal design of circular plates ‘against bucling J Struct Mech 1:159-186 Freier M, Nilsson, Samuclson A (1974) The rectangular plane stress clement by Ture, Pan and Wilson. Iat J Namer Methods Eng 435187 Gao X, Ma H (2015) Topology optimization of continaum structures ‘under buckling constraints. Comput Stet 157:142-152 Gao X, LiL, Ma H (2017) An adaptive continuation method {or \opology optimization of continuum strictures considering tnukling constrains. Int 7 Appl Math 9(7):24 Gravesen J, Evgrafov A, Nguyen DM (2011) On the sensitivities of multiple eigenvalues. Struct Mulidiscip Optim 44()583-587 Haftka R, Gordl Z (2012) Elemeats of structural optimization. Solid smechanis and its applications. Springer, Netherlands Hafika R, Prasad (1981) Optimum structural with plate bending elements —a survey. ALAA 19:517-522 Hall SK_ Cameron GE, Grierson DE (1988) Least-weight design of steal frameframe aceounting for P ~ A effets, St Bog ASCE 115(6) 1463-1475 Jog CS, Haber RE (1996) Stability of finite element models for distibated-parameter optimization and topology design. Comput Methods Appl Mech Eng 130(3)203-226 KKemmler R, Lipka A, Ramm E (2005) Large deformations and stabibiy in topology optimization. Struct Multdscip Optim 30:459-476, Kennedy GJ, Hicken JE (2015) Improved constraint-ageregation methods. Comput Methods Appl Mesh Tng 289(Supplement 332354 Kerr ADT, Soifer MT (1968) Te linavizaton ofthe prebucking state ‘aad ie effect on the determined instability loads. Appl Mech 3o775-785 ‘hot NS, Venkayya VE, Berke L (1976) Optimum structural design with stability constants Int} Numer Methods Bog 10(3)1097— m4) Kirmser PG, Hu KK (1995) The shape of reconsidered, Math Intell 15(3):52-57 Kesselmeier G, Stenhauser R (1979) Systematic contol design by optimizing a vector performance index. IFAC Proceedings 12(7 13-117. FAC Symposium on Computer Aided Design of Control Systems, Zaich, Switzerland, 29-31 August Le C, Norao J. Bruns T, Ha C, Tortorelli D (2010) Stess-based topology optimization for continua, Struct Muliiseip Optim 4114):605-620 Lee B, James K, Martins J (2012) Siess-constrained topology optimization with design dependent loads, Struct Mulidiscip Opti 46(5}:647-661 he ideal column D Springer Lee K, Ahn K, Yoo J (2016) A novel p-norm conection method for lightweight topology optimization under maximum stress constraints, Comput Strut 171(Supplement C):18-30 Lindgeard E, Dahl (2013) Oa complisace and buckling objective fonctions in topology optimization of snap-through problems Struct Muliiseip Opti 47:409-421 Lund E (2009) Buskling topology optimization of laminated malti~ ‘material composite shel structutes. Compos Struct 91(2):158— 197 ‘Manh ND, Evgrafov A, Getsborg AR, Gravesen J (2011) Ieogeometrc shape optimization of vibrating membranes. Comput Methods ‘Appl Mech Bing 200(13): 1343-1353 ‘Munk DI Vio GA. Steven GP (2017) A simple alternative formlation {or strututl optimisation with dynamic and buckling objectives Struct Muliisep Optim $5(3) 969-986 [Neves MM, Rodrigues H, Guedes JM (1995) Generalized topology design of structures with a buckling load criterion. Struct Optim 1007178 [Neves MM, Sigmund 0, Bendsge MP (2002) Topology optimization ‘of periodic microstructures with a penalization of highly localized buckling modes. Int ¥ Numer Methods Eng 54(6)80— 84 ‘Onsaki M, Theda K (2007) Stability and optimization of structures: generalized seasvity analysis. Mechanical Engineering Series Springer, Belin ‘Olholf N, Rasmussen SII (1977) On single and bimodal optimism ‘buckling loads of clamped columns, Int J Solis Stuet13(7):605— ols Pedersen NL, Pedersen P (2018) Buckling load optimization for 2D continuum models with alternative formulation for buckling load ‘estimation, Stract Multissip Opti 58(5) 2163-2172 Pian TH (1968) Derivation of element slinss matrices by aseumed ‘Hees disteibutions, ATA J2:1383-1336 Plan THH, Sumibara K (1984) Rational approsch for assumed ‘Heese finite elements, Int T Numer Methods Bng 20(9) 1685 1695 Poon NMK, Mating IRA (2007) An adaptive approach to conttaint ‘sgzregation using adjoint sensitivity analysis. Suuct Muliiscip Optim 34:61-73, RahmatallaS, Swan C (2003) Continuum topology optimization of buckling-senstve structures. ALAA J 41(6): 1180-1189 Raspanti CG, Bandoni JA, Biegler LT (2000) New statepies for Aexbility analysis and design under uncertainties. Compat Chem ng 24:2193-2209 Reitinger R, Ramm E (1995) Buckling and imperfection seastivity in the optimization of shell structure, Thin-Walled Struc 23:15 im Rodeigues HC, Guedes IM, Bendsse MP (1995) Necessary conditions for optimal design of stuctures with a nonsmooth eigenvalue based erterion. Stuct Optim 9:52-56 ‘RojasLabanda S, Stolpe M (2015) Automatic penalty continuation ‘in srictral topology optimization, Steet Mulidiseip Optim 52:1205-1221 Roxvany G (1996) Difficulties in topology optimization with stress, local buckling and eystem stability constrsnts, Struct Optim 1121-317 Seyranian AP, Lund B, Olhotf N (1994) Mulsple eigenvalues in structural optimization problems. Struct Optim 8(4):207-227 Sigmund © (2007) Morphology-based black and white filters for Topology optimization, Struct Multdiscip Optim 33(4)401— 404 ‘Simitses G} (1973) Optimal versus the stiffened circular plate. AIA, TIL 1408-1412 ‘Simo JC, Rifai MS (1990) A class of mixed assumed stain metbods and the method of incompatible modes. Int} Numer Methods Eng 29:1595-1638Revising topology optimization with buckling constraints Svanberg K (1987) The method of moving arymplotes - a new ‘method for structural optimization. Int T Numer Methods Eng 24(2):358-375 Siyszkowski W. Watson T. (1988) Optimization of the buckling load of columns and frames. Eng Struct 10(4):249-256 ‘Thomsen CR, Wang F, Sigmund O (2018) Buckling strengt topology ‘opumization of 2D periodic mater based on linearized bifrcation analysis, Comput Methods Appl Mech Eng 339:115-136 “Torii AJ, Faia JR (2017) Structural optimization considering smallest ‘magnitude eigenvalues: a smooth approximation J Braz Soe Mech Sei Eng 3965): 1745-1754 ‘Turner MI, Clough RJ, Martin HC, Topp LI (1956) Stine and defle tion analysis of complex sutures. Aerosol Sct 28:305-23, Vesbart A, Langelaar M, van Keulen F (2017) A unified agere- {gation and telanation approach for suess-constained topology optimization. Struct Mulidiseip Optim 55°663-579 ‘Wang F, Lazatov B, Sigmund O (2011) On projection methods, ‘onvergence and robust formulations in topology optimization, Servet Mulidisip Optio 43(6) 757-788 ‘Wilton EL, Taylor RE, Doherty W, Glaboutsi J (1973) Incompatible ‘isplacement models, Academie Press, pp 41-57 Wu CC, Arora JS (1988) Design sensitivity analysis of non-linear buckling foad. Comput Mech 3129-140 Yang RJ, Chen C3 (1996) Stress-based topology optimization. Stuct ‘Opsin 122), 98-105 Ye HL, Wang WW. Chen N, Sui YK (2016) Plattsell topological ‘optimization subjected ta linear buckling constrains by adopting composite exponential filtering functions. Acta Mech Sinica 32(0):619-658 ‘Zhou M (2004) Topology optimization of shell structures with linear ‘bekling responses, In: Takis © (ed) WCCM VI in Beijing. Cina p 795-800 hos M, Sigmund O (2017) On fully sessed design and p-norm measures in structural optimization. Suuct Muliiscip Optim 5631-736) Publishers note Springer Nature remains neuceal with regard to junsdictonal claims in published maps and instttionalafiiations Dspringer