Struct Multidisc Optim (2006) 31: 470479 DOI 10.

1007/s00158-005-0582-5

R E S E A R C H PA P E R

K. Abe K. Koro

A topology optimization approach using VOF method

Received: 27 April 2005 / Revised manuscript received: 20 September 2005 / Published online: 27 January 2006 Springer-Verlag 2006

Abstract Topology optimization methods with continuous design variables obtained by the homogenization formula or the solid isotropic microstructure with penalty (SIMP) model are widely used in the layout of structures. In the implementation of these approaches, one must take into account several issues, e.g., irregularity of the problem, occurrence of the checkerboard pattern, and intermediate density. To suppress these phenomena, the employment of additional strategies such as the perimeter control or the ltering method will be required. In this paper, a topology optimization method which can eliminate these difculties is developed based on the volume of uid (VOF) method. In the method, shape design is described in terms of the VOF function. Since this function is dened by a volume fraction of material occupying each element, it can be recognized as a continuous material density in the SIMP model. Within the framework of the VOF analysis, the topology optimization procedure is reduced to a convection motion of the material density governed by a HamiltonJacobi equation as in the level set method. Through numerical examples, the validity of the proposed method is investigated. Keywords Topology optimization VOF method Material density FEM

1 Introduction Design optimization can be a helpful tool in the layout of structures and composite materials. Since the pioneering work of Bendse and Kikuchi 1988, various topology optiK. Abe (B) Department of Civil Engineering and Architecture, Niigata University, Niigata, Japan Tel.: +81-25-2627028 Fax: +81-25-2627021 e-mail: [email protected] K. Koro Graduate School of Science and Technology, Niigata University, Niigata, Japan

mization methods have been developed by many researchers (cf. Eschenauer and Olhoff 2001; Bendse and Sigmund 2003). During the past decade, application of the method was further extended to nonlinear structures (Swan and Kosaka 1997; Bruns and Tortorelli 2001; Cho and Jung 2003). In the context of the so-called material approaches (Eschenauer and Olhoff 2001) in which a nite element mesh is xed in the design domain, the topology optimization is achieved by exploring an optimal black(material)and-white(void) distribution in the given design domain. It is well known that the discrete (01) design can be an ill-posed problem and may not converge under a sequence of rening meshes. Various remedies for this irregularity have been proposed, e.g., relaxation of the design space by the introduction of microstructures (Bendse and Kikuchi 1988; Suzuki and Kikuchi 1991), restrictions of the distribution pattern such as the perimeter control (Haber et al. 1996) and the slope constraint (Eschenauer and Olhoff 2001; Bendse and Sigmund 2003), and regularization by the ltering of material distribution (Sigmund 1994; Bourdin 2001) or by the diffusion method (Wang et al. 2004). While the microstructure evaluated by means of the homogenization method (Guedes and Kikuchi 1990) enhances the design space by replacing the black-and-white distribution with continuous design variables, it may generate intermediate densities. To reduce the intermediate densities, the solid isotropic microstructure with penalty (SIMP; Bendse 1989; Bendse and Sigmund 1999; Rozvany 2001) model in which a penalty is imposed on the intermediate density has been proposed. Though the SIMP model cannot contribute to the regularization of design space, a combination with the restrictions can be effective technique. This method is thus used popularly. Besides, it is remarkable that, despite of its simplicity, in some cases, the SIMP model is realized by a microstructure (Bendse and Sigmund 1999). Another issue which should be overcome in the material approaches is the occurrence of the so-called checkerboard pattern. For the suppression of this phenomenon, the employment of the ltering method, the diffusion method, or the continuous interpolation of material density (Matsui and Terada 2004) will be effective.

A topology optimization approach using VOF method

471

The material approaches with continuous density make it possible to replace the topology optimization problem with the optimal sizing problem of design variables; as stated above, the employment of additional strategies is needed for a desirable result. The necessity of restrictions or regularizations in the continuous variable design is caused by the original optimization procedure, in which an optimal material distribution is obtained by updating the design variables of nite elements independently of each other. In this case, the geometrical connectivity of the distribution pattern cannot be taken into account in the topology optimization. A new design method which is released from this obstacle has been developed based on the level set method (Allaire et al. 2002, 2004; Wang et al. 2003). The level set function is signed distance (Sussman et al. 1994) from the material (or structural) boundaries, in which > 0 in the material domain. The motion of boundary contours results from the evolution of the level set function. The topology optimization is thus reduced to the determination of an optimal distribution of the level set function dened in the design domain. In this paper, inspired by the success of the level set method, application of the volume of uid (VOF) method (Hirt and Nichols 1981) to the capture of material boundaries is attempted. The VOF method has been originally developed to analyze uid dynamics with free surfaces, as has the level set method. The VOF function is dened as a volume fraction of material occupying each element. Hence, this function can be regarded as the material density in the SIMP method. However, in the proposed method, the topology optimization is realized through the evolution of the VOF function governed by a HamiltonJacobi equation, as in the level set method. The cause of the checkerboard pattern is the inconsistency in the nite element discretization of the density and displacement (Jog and Haber 1996). In this case, the element-wise correction of the density may induce the anomalies. Since the proposed method changes the shape by updating the boundary contour, the value of VOF function inside the domain cannot be disturbed. Therefore, owing to this strategy, the checkerboard pattern will be avoided in the VOF function. This is the advantage of the proposed method over the classical topology optimization methods and is the common benet to the VOF and the level set methods. The superiority of the VOF method over the level set method is the simplicity of the formulation and numerical implementation. The VOF function directly denes the shape of the structure by itself, while the level set function denes the shape by its contour line. In the VOF method, a nite element locating on the boundary contour has an intermediate density. Thus, the resolution of the boundary capture depends on the mesh size. It may be a drawback of the VOF method. In Section 2, topology optimization problem is dened in the context of the VOF method. Sections 3 and 4 describe the topology optimization algorithm and the numerical implementation, respectively. Numerical examples of twodimensional mean compliance problem are demonstrated and the capability of the developed method is discussed in Section 5.

2 Topology optimization by VOF method 2.1 Linear elastic problem In this paper, we consider a two-dimensional linear elastic problem given by the following fundamental equations; the equilibrium equation, + f = 0 in , (1) (2) the constitutive relation, = E:, the boundary conditions, u = u on u , n := t = on , t (3)

where and are stress and linear strain tensors dened in the structural domain , f is body force and E is the elastic tensor. u is the displacement and n = t is the traction. Here n denotes the unit outward normal vector. u and are t prescribed displacement and traction on sub-boundaries u and , respectively. The boundary is given by the union of u and . The weak form of these equations is described, by virtue of the principle of virtual work, as a(u, u) = (u), u = u on u , for u D, (4)

where D stands for a set of displacement functions satisfying the condition u=0 on u , and a(u, u) and (u) are dened by a(u, u) = (u) = (u) : E : (u) d , fud + u d. t (5)

2.2 Optimization problem We consider an optimization problem of linear elastic structures dened by min J (u) :=

F(u)d ,

subject to a(u, u) = (u), u = u on u , for u D, V := d Vmax , (6)

where the objective function J is specied by a function F(u), and V is the volume constrained by an allowable volume limit Vmax .

472

K. Abe, K. Koro

2.3 VOF function As mentioned in the preceding section, the VOF function is dened by a volume fraction of material in each element. Therefore, the VOF function has an element-wise constant distribution, in which = 1 for an element lled with the material and = 0 for a void element. In the numerical analysis, a lower limit 1 min > 0 is introduced to assure the solvability of nite element equations. The time evolution of VOF function is governed by the HamiltonJacobi equation as + v = 0, (7) t where v is a convection velocity which produces an optimal topology. The evaluation of is achieved using an interpolation of as will be stated in Section 4.1. The convection of the VOF function makes it possible to change the shape without inducing the checkerboard pattern. The derivation of v will be described in Section 3. 2.4 Optimization problem with VOF function The VOF function is dened in the design domain which includes the structure as a subset. The original optimization problem can be recast, using the Lagrange multiplier + , as min J (u; ) := J (u; ) + + (V Vmax ),

In the following, the traction boundary is assumed to be xed on the boundary irrespective of the distribution of , i.e., the integration associated with the prescribed traction is independent of the design process. t 3 Topology optimization procedure 3.1 Evaluation of increment of functional The increment of functional J due to the change of VOF function can be expressed as J = where J = V = J + + V, J u

(11)

u+

(12) (13)

d .

Here u is the displacement increment induced by . In (12), the derivatives of J are interpreted in the sense of Frchet derivative, and they are given by J u J (8) u= =

F u

ud ,

(14) (15) u ex-

F(u) d .

subject to a(u, u; ) = (u; ), u = u on u , for + (V Vmax ) = 0, + 0,

To evaluate the derivative of (14) without using plicitly, we introduce an adjoint problem a(u, u; ) =

D,

F(u) ud , u
u

(16)

u = 0 on u for

D.

where J (u; ), a(u, u; ), (u; ) and V () are given by J (u; ) = a(u, u; ) = (u; ) = V () =

F(u) d , (u) : E : (u) d , (9) + u d, t

f u d

Note that, since u + u = u on u , it can be concluded that u=0 on u . It means that, if (16) is fullled for u D, then this equation holds for u. Therefore, once we obtain the solution u which satises (16), the following relation can be derived from (14) J u = a( u, u; ). (17) u The evaluation of the right-hand side of (17) can be achieved by considering the equilibrium condition for u + u and + , i.e., a(u + a(u + u, u; + u, u; + ) = (u, + ) u+ a + , + . (19) ), (18) where the terms on both sides can be expanded as

d .

It can be regarded that the elastic tensor is extended to a function given in proportionally to as E(x) = (x)E, x . (10)

= a(u, u; ) + (u; +

Note that, since the VOF function (x) corresponds to the material density, (10) gives the SIMP model with the penalization power p = 1 (variable thickness sheet).

a u

) = (u; ) +

A topology optimization approach using VOF method

473

From (8), (18), and (19), we obtain the next relation a u u= a , (20)

Substitution of (25) into (23) yields J =

[f u (u) : E : (u) + F(u) + + ]2 (26) t 0.

where the left-hand side is expressed as a u u= (u) : E : ( u) d (21)

|| d

= a( u, u; ). Combining (17), (20), and (21) yields J u u= = = a(u, u; ) u u

(u; ) a(u, u; ) (22)

(26) implies that the functional J gives a monotone decreasing sequence through the time evolution analysis. In the analysis, the last term /|| in (24) is regarded as 0 when =0. Therefore, v=0 at almost everywhere in the design domain except in the neighborhood of the structural boundaries. The Lagrange multiplier + is determined so that the material volume is conserved when it reaches the upper limit (Wang et al. 2003). To achieve this, let us consider the change of volume V due to of (25), i.e., V =

[f u (u) : E : (u) + F(u) + + ] (27) t.

f u d (u) : E : (u) d .

|| d

Since V should be zero when V = Vmax , (27) gives the following condition + +

Substituting (13), (14), and (22) into (11), we obtain the increment J as J =

|| d [f u (u) : E : (u) + F(u)] || d (28)

[f u (u) : E : (u) (23)

+F(u) + + ] d . (23) makes it possible to evaluate the increment of the functional associated with the change of VOF function without using the increment of displacement explicitly.

= 0. From this relation, + can be expressed as

f u (u) : E : (u) + F(u) || d

+ = 3.2 Convection velocity of VOF function Distribution of the VOF function is to be updated based on (7). In this paper, the following velocity eld which can lead the material distribution to an optimal topology is proposed v = [f u (u) : E : (u) + F(u) + + ] . ||

|| d

. (29)

4 Numerical implementation (24) 4.1 Discretization of HamiltonJacobi equation by CIP method The time evolution of the VOF function is calculated based on (7). Since the VOF function is interpolated by a continuous function in (7), it is necessary to cope with the numerical diffusion in the discretization of the HamiltonJacobi equation. While the conventional upwind difference method may approximate the convection equation, it cannot necessarily reduce the numerical diffusion satisfactorily. Unless this deterioration is avoided, the domain with intermediate density will be widened, and then the recognition of topology will be difcult. To maintain the numerical accuracy during the time

(24) can be justied by evaluating J caused by the velocity v. From (7) and (24), the change of the VOF function at a time step is expressed as = v t = [f u (u) : E : (u) + F(u) + + ] || t, where t is time increment. (25)

474

K. Abe, K. Koro

4.3 Adaptive setting of

vi t

The time increment t must be small enough so that the numerical stability can be assured. On the other hand, large time increment is preferable for the time step reduction. In general, the time increment at each time step has different upper bound because the convection velocity varies every step. The time increment which ensures the stability is determined adaptively at each step by the criterion tk = h , k 2vmax (32)

Fig. 1 Convection analysis of VOF function at the ith element

stepping analysis, the employment of the cubic interpolated prole (CIP) method (Yabe 1991) is attempted. As an example, let us consider the ith element as illustrated in Fig. 1. Since the VOF function is dened at each nite element, nodes of this function are located at the central points of elements. From (7), the VOF function at t + t of the ith element is given by (xi , t + t) = (xi vi t, t), (30)

k where h denotes the element size, and t k and vmax are the time increment and the maximal velocity at the kth time step.

5 Numerical examples 5.1 Minimum compliance problem In this section, the developed topology optimization method is applied to the minimum compliance problem. The objective function J of this problem is given by the strain energy stored in the structure, i.e., J (u; ) =

where xi is the location of the central node of the ith element, and vi is the velocity at this element given by (24). (30) implies that the VOF function can be updated based on the distribution of this function at the preceding time step. Note that, while the VOF function has constant value at each element in the evaluation of integrals described in Section (3), it is interpolated continuously in the calculation of (30). The CIP method approximates the spatial distribution using cubic-polynomial interpolation expressed in terms of the function and its derivatives at each central node. The additional information concerning the derivatives contributes to the suppression of the numerical diffusion. In the following analysis, tensorial products of Hermite polynomials are used for the interpolation of rectangular elements.

(u) : E : (u) d .

(33)

In the following examples, an isotropic material with Poissons ratio =0.3 is considered for two-dimensional plane strain problems without body force. Since the func tional of (33) gives a self-adjoint problem, J in (23) has the following simple form, J=

[ (u) : E : (u) + + ] d .

(34)

The nite element analyses are carried out with uniform meshes consisting of four-node linear elements. The lower 4.2 Reinitialization of VOF function Since the present VOF method is enhanced by the CIP method, the numerical diffusion induced during the analysis of convective motion will be vanished sufciently. Nevertheless, as will be shown in a numerical example, the distribution of the VOF function may stick in a local solution at the very early stage. To prevent the local convergence, the reinitialization of the VOF function is introduced periodically. Contrary to the level set method (Allaire et al. 2004; Sussman et al. 1994), the present reinitialization is quite simple. That is, the value of VOF function at each element is replaced with = 1 ( 0.5), min ( < 0.5), (31)
40h0

P=1
Fig. 2 Initial topology of two-dimensional cantilever

where is the VOF function after the correction.

40h0

A topology optimization approach using VOF method

475

(0th step)

(200th step)

(400th step)

(600th step)

(800th step)

(1000th step)

Fig. 3 Evolution of material distribution of cantilever

limit of VOF function min is set to 1104 . The material distribution is updated iteratively through the convection analysis until the topology converges. Since the aim of this section is the validation of numerical performance of the proposed method, a simple stop criterion will be enough. The following analyses are terminated when the iteration reaches a specied step.

5.2 Topology optimization of cantilever A two-dimensional cantilever xed at the left side, as depicted in Fig. 2, is analyzed. The square design domain of side length of 40h 0 (h 0 is mesh size) is discretized into 4040 square elements. The initial structure has 55 square voids as shown in Fig. 2. A concentrated vertical load P=1 is applied at the

20 non dimensionalized objective function

volume ratio

10

0.5

0 0

step

1000

step

1000

Fig. 4 Time history of nondimensionalized objective function

Fig. 5 Time history of volume ratio

476

K. Abe, K. Koro

60h0

P=1

60h0

Fig. 8 Two-dimensional simple beam

Fig. 6 Material distribution at the 100th step given by a design process with the reinitialization

lower corner of the right side, where P = P/Eh 0 is the load nondimensionalized by h 0 and Youngs modulus E. The evolutional process of the optimization design is shown in Fig. 3. In this analysis, the VOF function is reinitialized at every 20 steps, and the allowable volume ratio to the whole designable space is set to 0.3. A two-bar truss-like structure is formed in an early process (Fig. 3b). After that, the third member is generated at the loading point and grows with time step (Fig. 3cf). Note that the checkerboard pattern is not observed in the VOF function. Furthermore, the material distribution has sharp boundaries, and the intermediate density is suppressed sufciently. Time histories of the nondimen sionalized objective function J = J/Eh 2 and the volume 0 ratio are shown in Figs. 4 and 5, respectively. The uctuation of J which is induced by the reinitialization decreases with progressing topology. While the objective function J is increasing during the process of volume reduction, it turns out that it decreases monotonically after the convergence of volume similar to the case of the level set method (Wang et al. 2003). Though the objective function uctuates, the topology is not changed by the reinitialization. Therefore, the shape will not converge to different topology due to the uctuation. After the 200th step, the change of the objective function J is quite small. Nevertheless, the present design process

consumes a large number of iterations for the development and the convergence of topology. It should be noticed that, though the numerical performance has problem dependence, the application of the conventional SIMP method with p = 3 shows similar convergence to the present method. Figures 6 and 7 are showing the material distributions at the 100th step given by design processes with and without reinitialization, respectively. In the former case, many holes have already been merged and the evolutional process to the two-bar truss is recognizable; in the latter case, as shown in Fig. 7, the material domain is eroded uniformly under the same topology and then nally frozen. This result implies that, though the cause of this is not claried, the reinitialization is essential for the avoidance of stagnation. This may be an issue peculiar to the present method. Though in Fig. 6 we can nd the material islands which are isolated from the

(0th step)

Fig. 7 Material distribution at the 100th step given by a design process without the reinitialization

(1000th step)

Fig. 9 Initial and nal layouts for 31 initial voids

40h0

A topology optimization approach using VOF method

477

(0th step)

(0th step)

(400th step)

(300th step)

Fig. 10 Initial and nal layouts for 32 initial voids

Fig. 11 Initial and nal layouts for 64 initial voids

structure, they vanished during the optimization process as shown in Fig. 3.

5.3 Topology optimization of simple beam A simple beam subjected to a concentrated load at the center of the bottom side is considered. The outline of this problem is illustrated in Fig. 8. In the analysis, due to the symmetry, the right half of the beam is discretized. The numerical results for various number of initial voids of 31, 32, and 64 are shown in Figs. 9, 10, and 11. The convergences of the objective function J and the volume ratio are also shown in Figs. 12 and 13, respectively. In these analyses, the design domain is discretized into 6040 elements. The allowable volume ratio and the period of reinitialization are 0.4 and 20 steps, respectively. From the gures, we can nd that the nal topology is independent of the initial layout, while the convergence slows down with decreasing number of voids. The number of iterations with 64 voids is about 300 steps for the nal topology layout, in contrast with the case of 31 voids which consumes more than 1,000 steps. This tendency has been perceived in the level set method (Wang et al. 2003), and thus, it seems to be common for these approaches. Although the nal topology is independent of the initial layout (voids) in the above analyses, it does not necessarily hold in every case. Figures 14 and 15 show the initial and the

nal material distributions for 31 and 32 initial voids with 3020 elements, respectively. It can be found that the 31 initial voids result in simpler nal topology than the case of 6040 elements. However, the design is realizable. In contrast with the initial layout of 31 voids, the 32 initial void arrangement leads to an unnatural structure (Fig. 15). While J converges to around 17.0 in the former, it converges to 23.0

40 non dimensionalized objective function 3x1 voids 3x2 voids 6x4 voids

20

step

1000

Fig. 12 Time histories of nondimensionalized objective function for 31, 32, and 64 initial void arrangements

478

K. Abe, K. Koro

1 3x1 voids 3x2 voids 6x4 voids

volume ratio

0.5

a
0 0 step 1000

(0th step)

Fig. 13 Time histories of volume ratio for 31, 32, and 64 initial void arrangements

in the latter. From this fact, it can be concluded that, in general, the nal topology depends on the initial layout and may lead to a wrong design. When the initial shape has no hole, voids will not be generated in two-dimensional problems as in the level set method (Allaire et al. 2004). Figure 16 shows the initial and the nal material distributions for 12080 elements. As can be seen from Figs. 911,

(300th step)

Fig. 15 Initial and nal layouts for 32 voids with 3020 elements

(0th step)

(0th step)

(300th step)

(800th step)

Fig. 14 Initial and nal layouts for 31 voids with 3020 elements

Fig. 16 Initial and nal layouts for 12080 elements

A topology optimization approach using VOF method

479

14, and 16, although the structural forms are similar to one another, the number of members of the truss-like structure increases as the mesh size decreases. The mesh dependence may be improved by introducing the perimeter control as mentioned by Matsui and Terada (2004). In spite of these dif ferent topologies, the nal objective function J ranges from 17.0 to 18.0. Considering the dependence of accuracy on the mesh resolution, we can conclude that the difference in performance of these results is negligible; hence, any of these structures will be acceptable practically.

application to various problems will also be possible by introducing adjoint problems. References
Allaire G, Jouve F, Toader A-M (2002) A level-set method for shape optimization. C R Acad Sci Paris Ser I 334:11251130 Allaire G, Jouve F, Toader A-M (2004) Structural optimization using sensitivity analysis and a level-set method. J Comput Phys 194:363393 Bendse MP, Sigmund O (2003) Topology optimization: theory, methods and application. Springer, Berlin Heidelberg New York Bendse MP, Kikuchi N (1988) Generating optimal topologies in structural design using a homogenization method. Comput Methods Appl Mech Eng 71:197224 Bendse MP (1989) Optimal shape design as a material distribution problem. Struct Optim 1:193202 Bendse MP, Sigmund O (1999) Material interpolation schemes in topology optimization. Arch Appl Mech 69:635654 Bourdin B (2001) Filters in topology optimization. Int J Numer Methods Eng 50:21432158 Bruns TE, Tortorelli DA (2001) Topology optimization of non-linear elastic structures and compliant mechanisms. Comput Methods Appl Mech Eng 190:34433459 Cho S, Jung H-S (2003) Design sensitivity analysis and topology optimization of displacement-loaded non-linear structures. Comput Methods Appl Mech Eng 192:25392553 Eschenauer HA, Olhoff N (2001) Topology optimization of continuum structures: a review. Appl Mech Rev 54(4):331390 Guedes JM, Kikuchi N (1990) Preprocessing and postprocessing for material based on the homogenization method with adaptive nite element methods. Comput Methods Appl Mech Eng 83:143198 Haber RB, Jog CS, Bendse MP (1996) A new approach to variabletopology shape design using a constraint on perimeter. Struct Optim 11:112 Hirt CW, Nichols BD (1981) Volume of uid (VOF) method for the dynamics of free boundaries. J Comput Phys 39:201225 Jog CS, Haber RB (1996) Stability of nite element models for distributed-parameter optimization and topology design. Comput Methods Appl Mech Eng 130:203226 Kwak J, Cho S (2005) Topological shape optimization of geometrically nonlinear structures using level set method. Comput Struct 83:22572268 Matsui K, Terada K (2004) Continuous approximation of material distribution for topology optimization. Int J Numer Methods Eng 59:19251944 Rozvany GIN (2001) Aims, scope, methods, history and unied terminology of computer-aided topology optimization in structural mechanics. Struct Multidiscipl Optim 21:90108 Sigmund O (1994) Materials with prescribed constitutive parameters: an inverse homogenization problem. Int J Solids Struct 17:23132329 Sussman M, Smercea P, Osher S (1994) A level set approach for computing solutions for incompressible two-phase ow. J Comput Phys 114:146159 Suzuki K, Kikuchi N (1991) A homogenization method for shape and topology optimization. Comput Methods Appl Mech Eng 93:291318 Swan CC, Kosaka I (1997) Voigt-Reuss topology optimization for structures with nonlinear material behaviors. Comput Methods Appl Mech Eng 40:37853814 Wang MY, Wang X, Guo D (2003) A level set method for structural topology optimization. Comput Methods Appl Mech Eng 192:227246 Wang MY, Zhou S, Ding H (2004) Nonlinear diffusions in topology optimization. Struct Multidiscipl Optim 28:262276 Yabe T (1991) A universal solver for hyperbolic equations by cubicpolynomial interpolation. I. One-dimensional solver, II. Two- and three-dimensional solvers. Comput Phys Commun 66:219242

6 Conclusions In this paper, we have developed a topology optimization approach using the VOF method. The structural shape is given by the distribution of VOF function which has the same role as the material density in the SIMP model. The design optimization is described as a variational problem of an objective function dened in terms of the VOF function. The motion of structural domain during the optimization process is reduced to the convection of the VOF function governed by the HamiltonJacobi equation. Here the convection velocity eld in the design domain is derived from a functional. To reduce the intermediate density due to numerical diffusion, the CIP method is employed in the time evolution analysis. Furthermore, the reinitialization of VOF function is introduced to prevent the stagnation of material topology in the design process. Owing to these two strategies, the intermediate density is no longer an issue. The developed method was applied to two-dimensional minimum compliance problems. Through numerical experiments, it has been found that the undesirable material distribution such as the checkerboard pattern and the intermediate density can be avoided by the proposed method. While the numerical results shown in this paper have mesh dependence, the nal topologies are acceptable. It should be taken into account that, in general, the nal topology depends also on the initial layout and may converge to an undesirable design. Indeed, we encountered few cases that different designs are obtained from different initial topologies. Although the VOF method allows intermediate density and thus cannot attain the black-and-white design, one can clearly recognize the topology from the nal material distribution. Contrarily, in the level set method, the structural prole is dened distinctly by contours. However, when continuous functions are introduced to approximate Heaviside and delta functions as proposed by Wang et al. (2003), the boundary will be shaded. In this case, the level set method is thought to be indistinguishable practically from the VOF method. Though the discussion of this paper is restricted to linear elastic problems, the application will be extended to nonlinear problems without any difculty, provided that the convergence is preserved (Kwak and Cho 2005). Moreover, the