Research paper Struct Multidisc Optim 25, 1–10 (2003) DOI 10.1007/s00158-003-0304-9
Effects of microscale material randomness on the attainment of optimal structural shapes
T. Liszka, M. Ostoja-Starzewski
Abstract Classical procedures of shape optimization of fectly homogeneous continuum. Consequently, the pro- engineering structures implicitly assume the existence of cess of optimization does not recognize the presence of a hypothetical perfectly homogeneous continuum – they any microscale material randomness, although the lat- do not recognize the presence of any microscale mate- ter becomes progressively more relevant as the mesh rial randomness. By contrast, the present study inves- is refined. A study of this effect conducted recently on tigates this aspect for the paradigm of a Michell truss the paradigm of the Michell truss (Michell 1904) made with minimum compliance (maximum stiffness) that has of elastic–plastic members showed that the random mi- a prescribed weight. The problem involves a stochastic crostructure prevents the attainment of the classically generalization of the topology optimization method im- predicted optimal shape using the homogeneous con- plemented in the commercial Altair’s OptiStruct com- tinuum (Ostoja-Starzewski 2001). The aforementioned puter code. In particular, this generalization allows for investigation was carried out under the assumption of the dependence of each finite element’s stiffness matrix fixed load on the structure and the search was con- on the actual microstructure contained in the given ele- ducted for a structure of optimal (i.e., minimal) weight. ment’s domain. Contrary to intuition, stochastic material The study considered the effect on the ability to at- properties may improve the compliance of optimal design. tain the classically expected minimum of a spatially dis- This is because the optimization is performed on a given ordered (non-periodic) material microstructure of which random distribution, so that the design process has an op- the Michell truss-like continuum would be manufactured: portunity to choose ‘stiffer’ cells and discard those with the denser the truss (i.e., the finer its mesh spacing), the weaker material. The paper does not aim for a robust de- more significant the effect of microstructural fluctuations sign process, but tries to answer a simpler intermediate on the plastic limit σ0. This led to the question of the question: how the random fluctuation of material prop- sensibility of the ad infinitum refinement of the Michell erties influences a structure that has been designed using truss, and hence to the open issue of the plausibility of its classical continuum-based optimization algorithms. attainment. In general, the planar material of which the truss is Key words optimal structures, random microstructure manufactured is a random medium:asetB = {B(ω); ω ∈ Ω},whereeachB(ω) is a deterministic, albeit spa- tially inhomogeneous, body with Ω being a sample space. More specifically, each B(ω) is described by a scalar- valued planar field of plastic limit k,sothatwehave 1 the random field k = k(ω,x); ω ∈ Ω, x ∈ R2 specifying Background B. The actual setup of the field k depends on a non- dimensional scale parameter δ = L/d > 1, relating the Problems of shape optimization of engineering structures mesh spacing L to the size of the microheterogeneity d implicitly assume the existence of a hypothetical per- (e.g., crystal diameter). Micromechanics of random het- erogeneous media (Jeulin and Ostoja-Starzewski 2001), Received: while only recently making inroads into plasticity (Jiang Published online: 2003 et al. 2001), suggests that the following two aspects Springer-Verlag 2003 should be included in any physically acceptable model: 1, u 2 T. Liszka , M. Ostoja-Starzewski (i) the scatter in k (as described by its standard devia- 1 tion σk) should grow as δ decreases, and (ii) the statistical Altair Engineering, Inc., 7800 Shoal Creek Blvd, Austin, TX, 78757, USA average k may be taken to be constant as δ decreases: e-mail: [email protected] 2 Department of Mechanical Engineering, McGill University, 1 k =const,σk(δ) ∼ . (1) Montreal, Que. H3A 2K6, Canada δ 2
As is typically done in mechanics of random media, it is where convenient to split the field k into its constant mean and the randomly fluctuating part. We can then write aC(u, v)= ε(u)C(x)ε(v)dx (4) V k(ω)=k + k (ω) , k =0, (2) is the bilinear form of the energy, and where k is the zero-mean noise in k. In setting up the model, we also assumed: (i) the truss spacing L of inter- L(u)= fvdx+ tvdx (5) est to us is greater than the grain size d,sothatk may be V ∂Bt treated as a field of independent random variables when entering the finite difference formulation generalized to is the linear form of the load. That is, we seek the opti- ≡ inhomogeneous media; and (ii) the underlying material mal choice of the stiffness tensor C( Cijkl)insomegiven microstructure is space-homogeneous and ergodic. set of admissible tensors Uad. C are generally fields over R2 ∈ ∞ 6 Since the reference problem of this Michell truss is ,sothatUad (L (V )) , corresponding to the six in- governed by a quasi-linear hyperbolic system (Hegemier dependent elements of the in-plane stiffness tensor. In (3), and Prager 1969), ours was described by a stochastic gen- by ‘design constraints’, we mean constraints on stresses, eralization thereof (Ostoja-Starzewski 2001). The term strains, displacements, etc., while sizing constraints, vol- ‘stochastic’ refers to an ensemble of deterministic prob- ume constraints, etc., are accounted for in the choice of lems that exist on spatially inhomogeneous realizations of Uad. Finally, U is the space of kinematically admissible the random field k. Due to the randomness of k on scales displacement fields. δ<∞, the net of characteristics display a statistical scat- In the case of optimal shape design, elements C of Uad ter that increases with decreasing δ, which translates into take the form a trend to use more structural material to carry the pre- C (x)=χ(x)C , (6) scribed loading. This, of course, is an opposing tendency ijkl ijkl to the convergence to the Michell truss with mesh refine- where Cijkl is the constant stiffness tensor for the ma- ment, and, beyond a certain mesh refinement, the solu- terial employed for the construction of the mechanical tion to the governing hyperbolic system breaks down. To element, and χ(x) is the indicator function. sum up, there is a limitation to this ideal optimality. In the case of sizing problems, like the design of sheets Therefore, the above method cannot be used for in- of variable thickness, the admissible Cijkl’s have the form vestigation of systems with too fine a mesh for a given microstructural randomness, nor can it be employed for Cijkl(x)=h(x)Cijkl , (7) elastic structures. These restrictions suggested another ∈ ∞ approach – in fact, much more in line with the conven- where again Cijkl is a constant tensor, and h(x) L (V ) tional methodology of shape optimization of engineering is the sizing function. structures (Rozvany et al. 1995; Bendsøe and Kikuchi The discretized formulation of the topology optimiza- 1988) – which seeks the shape of an elastic structure with tion problem (3) can be stated as follows: minimum compliance (maximum stiffness) that possesses min f(ρ) a prescribed weight. Note, however, that this is differ- ∗ ent from the problem considered in (Ostoja-Starzewski s.t. V = ρjνi ≤ V , (8)
2001), which dealt with a plastic structure of minimal η ≤ ρi ≤ 1,i=1,...,N. weight under a fixed load. The present paper reports where f represents the objective function, ρi and νi are a stochastic generalization of the topology optimization ∗ method implemented in a commercial Altair’s OptiStruct element densities and volumes, respectively, V is the computer code, which takes into account a material mi- target volume, N is the total number of elements, and crostructure of random chessboard type. η is a small number that prevents the stiffness matrix from being ill-conditioned. The common objective func- tion is the weighted sum of the compliance under all load 2 cases. Note that the problem in (8) is a relaxation for- Problem formulation mulation of the topology problem, in which the density should only take the values 0 or 1. To enforce the design to be close to a 0/1 solution, a penalty is introduced to re- The minimum compliance problem for a planar (respec- duce the efficiency of intermediate density elements. Here tively, 3-D) body of volume V in R2 (R3) subjected to the penalization is achieved by the following power-law body forces f and tractions t takes the form: formulation: min L(u) ∗ p Ki (ρ)=ρi Ki , (9) Cijkl ∈ Uad ∈ ∗ subject to aC(u, v)=L(v), all v U, design constraints, where Ki and Ki represent the penalized and the real (3) stiffness matrix of the i-th element, respectively, and p is 3 the penalization factor, which is bigger than 1 (Bendsøe and Kikuchi 1988). A more general formulation of the topology optimiza- tion problem can be stated as follows: min f(ρ) − ∗ ≤ s.t. gj(ρ) gj 0,j=1,...,m, (10)
0 ≤ ρi ≤ 1,i=1,...,N, ∗ where gj and gj represent the j-th constraint and its up- per bound, respectively, and m is the total number of Fig. 1 One-dimensional tension problem on a square domain constraints. The ‘minimum member size’ constraint, implemented in Altair OptiStruct is based on constraining the discrete slope of the density. To achieve a predetermined mini- mum member size of radius rmin = dmin/2, the slope con- straint can be formulated for a general irregular finite element mesh as follows:
|ρi − ρk|≤(1.0 − ρmin)dist(i, k)/rmin , (11) where dist(i, k) denotes the distance between adjacent elements i and k,andρmin,say0.1, is the threshold that is interpreted as void in the final solution (Zhou et al. 1999). This condition guarantees that whenever an element j reaches a density of 1.0, the member connected to this element has a diameter of at least dmin. This condition introduces n · N,withn bigger than 2, additional linear constraints and makes the direct solution of this for- Fig. 2 Optimal solution for the uniform unidirectional prob- mulation computationally prohibitive. Our code Altair lem (dark color represents the converged optimal shape OptiStruct implements a different algorithm (Zhou (density = 1.0), and void areas are shown as white; the few et al. 1999) that is very efficient for these types of con- remaining cells with intermediate densities are shown with straints. For a survey of historical developments and different colors on a gray scale) a summary of the theory and techniques of topology op- timization, we refer the reader to (Rozvany et al. 1995), generated for an isotropic material, but any distribu- while for a background on penalization formulation to tion of horizontal trusses will satisfy the optimization (Allaire and Kohn 1993). problem. Two predefined (non-random) distributions of mate- rial density were tested: each consisted of regular bands 3 of alternating materials parallel to the axis. For the ho- Numerical results rizontal distribution (Fig. 3a), the optimization process placed the trusses inside the stiffer material (gray and 3.1 white bands represent hard and soft material, while black Preliminary tests represents the optimal shape). For the vertical distribu- tion (Fig. 3b – material types not shown in the figure) Before proceeding with the solution of the Michell truss, the trusses were generated with varying thickness, and we start with a very simple example of a unidirectional smaller thicknesses coincided with bands of stiffer mate- stress state in a square domain of size 1×1. The boundary rial. Although both cases are clearly contrived, the solu- conditions are ux = uy =0atx =0,andux =0.1,uy =0 tion in Fig. 3a proves that it is possible to obtain a better at x = 1 (see Fig. 1). We are trying to find the optimum optimal shape (lower compliance) in the presence of non- distribution of material with total volume V ∗ =0.1. Be- uniform material properties. cause the problem is ill-posed, the results of the simula- Finally, Fig. 4 shows the optimal shape in the presence tion exhibit poor convergence or numerical artifacts (e.g. of very strong noise (s =0.9). By comparing the mate- splits at each member near the edges of the boundary). rial distribution (Fig. 4b), representing material strength However, this (as expected) provides an optimization al- (black-gray-white scale, black being the soft material) gorithm with significant “freedom” that results in a vis- with the optimal shape shown in Fig. 4a, one can verify ible tendency of the “optimal” structure toward harder that the optimization process really avoids soft cells in the cells in the design space. The solution in the Fig. 2 was model. 4
Fig. 3 Optimal solution for the unidirectional problem with “banded” material properties: a) horizontal, and b) vertical distri- bution of material properties
Fig. 4 Optimal solution for the non-uniform unidirectional problem (noise = 0.9): a) optimal shape, b) distribution of material strength, c) and d), zoom of the lower part of the domain
3.2 Michell truss problem
The formulation outlined above was applied to the clas- sical problem of finding an optimal structure supported by a circular foundation boundary F and subjected to A a loading condition at A (see Fig. 5). First, for the sake F of reference, in Fig. 6a we display the truss found by car- rying out an optimization on a 27 × 27 mesh of square- shaped finite elements, all with identical properties. Next, we considered the optimal structure problem under the same support and loading conditions, but defined on 7 7 a2 × 2 mesh of a random chessboard with single elem- Fig. 5 Michell truss problem: design space and boundary ents being isotropic and linearly elastic: conditions 5
Cijkl = λδij δkl + µ(δjkδil + δjlδik) . (12) As an example, for the noise scale factor s =0.3, the Young modulus was varied from 0.7 to 1.3. The Pois- The material randomness is introduced via a statement son’s ratio was ν =0.3. All the elements are generated ac- analogous to (4): cording to the same process, independent of their spatial location, so that we effectively have a binomial random 7 7 C(ω)= C + C (ω) C (ω)=0 (13) field on the 2 ×2 mesh. All presented examples used the same random field r with different scale factors s. ≡ and generating C (ω)( Cijkl (ω)) by multiplying each ≡ element’s mean part C ( Cijkl ) by a random num- 3.3 − ber r sampled from a uniform distribution [ 1, 1] multi- Experiment 1 plied by a constant scale factor s: