Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

0

Hybrid Cellular Automata in Structural Design Optimization


Andrs Tovar and John E. Renaud
University of Notre dame Indiana, USA

1. Introduction
Structural optimization is an active eld of research that combines computational mechanics and mathematical programming. A general formulation of a structural optimization problem can be dened as nding the shape or topology of a structure that maximizes a performance objective and/or minimizes a cost function, subject to functional and side constraints, while satisfying a set of governing equations. An archetypical problem is to maximize the stiffness of a structure subject to a mass constraint. The development of numerical techniques for structural analysis, particularly the nite element method, has resulted in the conception of numerical design algorithms exclusively devoted to structural optimization (Arora, 1989). Finite element-based optimization techniques were rst developed by UCLA Professor Lucien Schmit in the 1960s (Schmit, 1960). He recognized the potential of combining optimization techniques with nite element analysis for structural design. Depending on the nature of the design variables, there are ve common ways to state a structural optimization problem: sizing, shape, topometry, topography, and topology optimization. Sizing optimization focuses on well-dened geometric structural features. This approach is largely used in truss structures in which the design variables correspond to cross-sectional parameters such as area or moment of inertia. In sizing optimization, the structure preserves its mesh during the optimization process. Shape optimization involves determining the optimal prole or boundary of the structure. Two of the most common approaches for shape optimization are the basis vector and the grid perturbation approach. The basis vector approach requires the denition of different trial designs called basis vectors. The design variables are the weighting parameters that dene the participation of each basis vector in the design process. On the other hand, the grid perturbation approach requires the denition of perturbation vectors. These vectors perturb or deform the boundary of the design domain. The design variables are the values that determine the amount of the perturbation during the optimization process. In shape optimization, the structures mesh varies as the design evolves; however, the genus of the structure, which roughly speaking corresponds to the number of holes, is preserved. Topometry optimization and topography optimization are applications of sizing an shape optimization to shell-type structures, respectively. In topometry optimization, the thickness distribution is optimized along the shell so the mesh is maintained. On the other hand, in topography optimization, the mesh is deformed along the perpendicular plane to optimize the shape or contour of the

shell. Topology optimization is being used more often in recent studies to nd preliminary, sometimes completely innovative, structural congurations. Finally, topology optimization involves the optimal distribution of material within a design domain where there is no prescribed structural shape. In topology optimization, the design domain is comprised of a large number of nite elements. The optimization algorithm selectively alters the material in each element resulting in an optimal distribution of solids and voids. Topology optimization is the most relaxed formulation as it does not impose any constraints on the geometry of the structure. For this reason, optimum topology designs achieve better performance measures than any other feasible structure. However, the use of conventional optimization methods is impractical to solve this integer programming problem due to the size of the design space and the computational cost of the structural analysis. Even the design of a simple two-dimensional structure may involve thousands of design variables, which correspond to the number of elements in the nite element model. In general, a nal design is obtained by combining different techniques (e.g., topology optimization followed by shape optimization). This is referred to as a generalized shape optimization problem. The ve structural optimization approaches described in this section are depicted in Fig. 1.
Initial design

Topometry

Sizing

Structural optimization

Topography Topology

Shape

Fig. 1. Five structural optimization approaches: size, shape, topometry, topography, and topology optimization. The purpose of this chapter is to present the application of a cellular automaton algorithm to solve structural optimization problems. This algorithm combines cellular automaton (CA) principles, control theory, and nite element analysis. For this reason, it is referred to as a hybrid cellular automaton (HCA) algorithm. Special emphasis is given to topology optimization problems.

2. Topology optimization
2.1 Problem formulation

A topology optimization problem can be dened as a binary programming problem in which the objective is to nd the distribution of material in a prescribed design domain. A classic formulation, referred to as the binary compliance problem, is to nd the 0-1 structure (i.e., solids and voids) that minimizes the structures end compliance C subject to a constraint in mass M. This is optimization problem can be expressed as
min x{0,1}n

C ( x) M( x) = M,

subject to

(1)

where M is the mass limit and x is an n-dimensional binary vector that represents the solid and void distribution in the design domain. As other topology optimization problems, the binary compliance problem (1) is known to be ill-posed (Kohn & Strang, 1986a;b;c). In particular, by increasing the number of design variables it is possible to obtain a non-convergent sequence of feasible 0-1 designs that monotonically reduce the structures compliance. Consequently, the design will progress towards a chattering design with innite number of holes of innitesimal size. That makes the compliance problem unbounded and, therefore, ill-posed. In order to make the problem well-posed, at least two problem formulations have been proposed: perimeter control and the homogenization method. The perimeter control refers to the addition of a constraint on the perimeter of the structure (Haber et al., 1996; Jog, 2002). This method effectively avoids chattering congurations, but its implementation is not free of complications. These complications include uctuations during the iterative process (Duysinx, 1997) and drastic changes in the nal topology due to small variations in the algorithms parameters (Jog, 2002). The homogenization method for topology optimization involves the relaxation of the binary problem to include intermediate material density values, that is
min 0 x 1

C ( x) M( x) = M.

subject to

(2)

In this way, the chattering congurations become part of the problem statement by assuming a periodically perforated microstructure. The mechanical properties of the material are determined using homogenization theory (Allaire, 2001; Bendse, 1995). The main drawback of this approach is that the optimal microstructure, which is required in the derivation of the relaxed problem, is not always known. This can be alleviated by restricting the method to a subclass of microstructures, possibly suboptimal, but fully explicit. This approach, referred to as partial relaxation (Allaire & Kohn, 1993; Bendse & Kikuchi, 1988). However, another problem with the homogenization methods is the manufacturability of the optimized structure. The gray areas of intermediate bulk density found in the nal designs contain microscopic length scale holes that are impossible to fabricate. This can be solved by penalizing intermediate density values (Bendse, 1995), but doing so reverts the problem back to the original ill-posed character with respect to the mesh renement.

An alternative that avoids the application of the homogenization method is to relax the binary problem using a continuous density value with no microstructure. The mechanical properties of the material are determined using a power-law interpolation function between void and solid (Bendse, 1989; Mlejnek, 1992). The power law implicitly penalizes elements with intermediate density by assigning a high compliance value, which drives the structure to a 0-1 conguration. This approach is usually referred to as the solid isotropic material with penalization (SIMP) method (Zhou & Rozvany, 1991). The SIMP method does not solve the problems ill-posedness, but it is simpler to implement than other penalization methods. Therefore, the SIMP method is used in this work.
2.2 Solution methods

Several optimization techniques have been utilized to solve topology optimization problems. For instance, the binary compliance problem (1) has been tackled with stochastic programming methods such as genetic algorithms (Chapman et al., 1994; Jakiela et al., 2000), simulated annealing (Anagnostou et al., 1992), and ant colony optimization (Kaveh et al., 2008). The main problem of the stochastic methods is the large number of function evaluations required for convergence. However, this problem can addressed by combining them with sensitivity analysis (Chan & Wong, 2008; Svanberg & Werme, 2006; 2007). A well-referenced topology optimization approach is the so-called evolutionary structural optimization (ESO) method proposed by Xie & Stevens (1993). The ESO method generates stiff structures via sequential element rejections and admissions, without the use of sensitivity analysis. As explained by Rozvany (2009), the ESO method has several drawbacks including the use of a greater number of iterations with respect to gradient-based methods and the possible convergence to an entirely non-optimal solution. On the other hand, the relaxed compliance problem (2) can be solved by a greater variety of optimization methods. Furthermore, when the sensitivities can be expressed in a closed form, virtually any gradient-based optimization method can be effectively used. General optimization approaches such as sequential linear programming (SLP) and sequential quadratic programming (SQP) have been extensively used (Calvel & Mongeau, 2007; Yang & Chuang, 1994). Some of the most cited approaches include general structural optimization algorithms such as the method of moving asymptotes (MMA) (Svanberg, 1987) and the convex linearization method (CONLIN) (Fleury, 1989). Besides these general approaches, there are several specialized algorithms tailored to solve particular topology optimization problems. One example is the optimality criteria (OC), which is presented as a 99-line Matlab code (Sigmund, 2001). Additionally, there is a family of heuristic methods in which the iterative scheme for topology optimization can be dened in terms of local interactions between neighboring elements. These methods are referred to as cellular automaton (CA) methods for topology optimization (Chia et al., 2006; Inou et al., 1994; Kita & Toyoda, 2000; Missoum et al., 2005; Tovar et al., 2007; 2006). The SAND-CA method presented by Missoum et al. (2005) avoids the use of global nite element analysis as it locally solves equilibrium equations in a simultaneous analysis and design (SAND) framework. The hybrid cellular automaton (HCA) method presented by Tovar et al. (2006) and Tovar et al. (2007) combines traditional nite element analysis with a local element iteration in a multivariate control system scheme.

The HCA algorithm was originally developed to solve the relaxed compliance problem. To this end, two formulations were implemented. The rst makes use of a heuristic approach in which element strain energy is uniformly distributed along the design domain (Tovar et al., 2006). The nal structures are free of numerical instabilities and convergence is achieved after a few function evaluations. The second formulation employs Karush-Kuhn-Tucker conditions with the SIMP model (Tovar et al., 2007). Convergence is achieved after considerably more iterations when compared to the previous formulation but, for some test problems, its rate of convergence is still superior to the one of the OC and MMA methods (Patel et al., 2008). The HCA method has been recently implemented by Livermore Software Technology Corporation as LS-OPT/Topology for LS-DYNA users (Goel et al., 2009). This chapter summarizes the principles and the state of development of the HCA method.
2.3 Numerical complications

The implementation of optimization algorithms in topology optimization problems is not free of numerical complications (Sigmund & Petersson, 1998). Some of these complications include: mesh-dependency, checkerboard patterns, local minima, and the presence of intermediate densities in the nal topology. Mesh-dependency naturally arises from to the ill-posedness of the topology optimization problem. This issue can be mitigated with the use of perimeter control (Jog, 2002), density slope constraints (Petersson & Sigmund, 1998), and lters (Sigmund, 1997). Filters dene a size in the elements neighborhood and, consequently, a length scale in the model. Furthermore, a slope control mitigating the mesh-dependency is provided since the ltered variables are approximately the same in a given neighborhood. A checkerboard pattern refers to alternating solid and void elements that resemble a checkerboard. This issue is usually reported in two-dimensional, displacement-based nite element analysis (Bendse & Sigmund, 2003). Checkerboard patterns are caused by numerical approximations in the nite element analysis that overestimate their stiffness (Daz & Sigmund, 1995). Some procedures to address checkerboard patterns are reviewed by Sigmund & Petersson (1998). These procedures include image ltering, higher-order nite elements, and patches (superelements on the nite element mesh), ltering being the simplest and more efcient approach (Sigmund, 2007). Intermediate densities refer to elements in the nal design that are neither solid nor void. The presence of such elements draws the structure away from a binary solution. Causes of intermediate densities include a decient penalization method and the use of ltering schemes. Penalization methods can be implicit such as the SIMP model (Bendse, 1989) or explicit such as the addition of constraints in the problem formulation (Fuchs et al., 2005). The latter method does not offer any signicant advantage with respect to the simpler implicit penalization. For this reason, the SIMP model is commonly used in topology optimization. Finally, most topology optimization problems are non-convex and, therefore, they have many local minima. When gradient-based optimization methods are used, the solution may be affected by the initial design, the type of optimization algorithm, and even the set of parameters within the same algorithm. Numerical experience has shown that the continuation method mitigates this complication. The idea of the continuation method is to gradually change the

optimization problem from an articial convex formulation to the nal non-convex original problem. At each step, a gradient-based algorithm is used until convergence (Bendse & Sigmund, 2003). In the context of the SIMP method, the continuation can be performed by successively increasing the value of the penalization power.
2.4 Sensitivity analysis

The sensitivities of an optimization problem can be numerically obtained using the nite difference method, which is the simplest procedure, but computationally unaffordable for many engineering problems. On the other hand, analytical methods can be used to derive a closed form for the sensitivity coefcients; however, they have three major problems: the derivation process is lengthy, they require several simplications and assumptions, and the implementation in a nite element code is cumbersome. Nevertheless, let us now present the main elements to obtain an analytic expression for the sensitivities of a nonlinear compliance problem (2). The mass M of the structure can be dened as M( x) =

i =1

xi m0i ,

(3)

where m0i is the mass of the i-th solid element and xi is its relative density. Hereafter, the subindex 0 refers to the properties of the bulk, solid material. The sensitivity of M with respect to xi can be expressed as M = m0i , (4) xi for i = 1, . . . , n. The compliance C of the structure is the scalar quantity dened by C ( x ) = F T U ( x ), ext (5)

where F ext is the applied force vector and U is the nodal displacement vector. The sensitivity of C with respect to xi can be determined using the adjoint method and expressed by F C = F T K 1 int , ext T xi xi (6)

for i = 1, . . . , n, where K T is the tangent stiffness matrix and F int is the internal force vector. Therefore, the sensitivity analysis of the compliance is reduced to nding the derivative of F int with respect to xi . In literature, this sensitivity for nonlinear problems has been approximated for nonlinear structures by authors like Maute et al. (1998) and Schwarz & Ramm (2001). In summary, an analytical expression for the sensitivity of the compliance in nonlinear analysis can be stated as C 1 = ci , (7) xi xi for i = 1, . . . , n, where ci is a complex integral function. Under several approximations, ci can be expressed in terms the elastic material tensor and additive components of a consistent elastoplatic material tensor (Maute et al., 1998). These sort of approximations are not only complex and difcult to implement in nite element analysis but also restricted to a particular

set of material models. On the other hand, using the SIMP method for linear analysis, the elastic material tensor C i can be interpolated as

C i = xi C 0i ,

(8)

where the exponent p is carefully chosen to penalize intermediate densities. In this way, the integral function ci can be expressed in terms of the element compliance ci as C 1 = pci , xi xi (9)

where ci = f T ui , where f i and ui are the element nodal force and displacement vectors, i p respectively. In this linear expression, f i = k i ui and k i = xi k 0i , where k i represents the element stiffness tensor. Therefore, the sensitivity in linear problems is readily obtained from a single nite element analysis at no extra computational cost.

3. Hybrid cellular automata


3.1 Biological inspiration

Bones have been considered optimal structures in terms of light weight and high stiffness. As a living tissue, bones continually adapt to changes in the physical environment. Throughout an individuals lifetime, old bone is removed (resorption) and new bone is added (formation). The functional adaptation process driven by resorption and formation is called remodeling. In human adults, 5% of cortical bone and 25% of trabecular bone is replaced every year by remodeling (Martin et al., 1998, p. 64). Bone remodeling is considered a control mechanism by which bone tissue is resorbed in regions exposed to low mechanical stimulus, whereas new bone is formed where the stimulus is high (Frost, 2003; Huiskes et al., 2000). Sensor cells within the bone, called osteocytes, are believed to translate mechanical stimuli into chemical signals that activate the remodeling process (Cowin et al., 1991). Osteocytes that lay on the bone surface, called bone lining cells, are also thought to initiate bone remodeling in response to various chemical and mechanical stimuli (Miller & Jee, 1992). The actuators in this control mechanism are referred to as basic multicellular units or BMUs. These BMUs are formed by bone resorbing cells or osteoclasts and bone forming cells or osteoblasts. The control system that regulates bone remodeling is depicted in Fig.2. Numerically, bone can be modeled by cellular automata that mimic the behavior of bone cells as a control system (Tovar, 2004). The mechanical stimulus and set point are determined according to the structural optimization process. The nite elements method is used to obtain the mechanical stimulus that eventually activates the remodeling process. The hybrid cellular automaton (HCA) algorithm incorporates control theory and nite element analysis in a CA framework.

Mechanical set point

Mechanical signal

Load

Mechanical properties

Osteocytes
Lining cells

Osteoclasts
Osteoblasts BMUs

' bone mass

Fig. 2. Control system in bone remodeling. The mechanical load is sensed by the osteocytes (mechanical signal). Based on an equilibrium condition or set-point, the osteocytes send a remodeling signal which regulates net bone resorption or formation by basic multicellular units (BMU) of osteoclasts and osteoblasts. The effect of the structural change is translated into changes of mechanical properties and, therefore, changes in the mechanical stimulus. The adaptation process continues until equilibrium is reached. Figure after Huiskes (2000).

3.2 Control-based optimization

To implement the HCA model, let us determine the expression of the mechanical stimulus and the set point for each CA in the model. Assuming that bone remodeling drives the structure towards an optimal design of minimum mass and maximum stiffness, let us restate the optimization problem (2) as an unconstrained, multi-objective problem. This optimization problem can be written as
min 0 x 1

C ( x) + (1 ) M( x),

(10)

where is a weighting factor in the range [0, 1]. The optimality conditions of (10) for an interior point 0 < xi < 1 yield M C + (1 ) = 0. xi xi (11)

Using the sensitivity coefcients stated in (4) and (9), the optimality condition (11) can be expressed as 1 pci + (1 )m0i = 0. (12) xi This condition can be restated as a control problem in which the error signal ei is obtained from the mechanical stimulus yi and the set point si dened as ei = yi si , where yi = 1 1 (1 ) c , and si = m0i , xi i p (13)

for every CA in the discrete location i. The constant weighting factor is dened in the objective function (10). The mass of the solid element m0i depends on the material density and

element volume. The exponent p is dened by (8). A value of p 2.5 effectively penalizes intermediate densities in a topology optimization problem. The function yi can be obtained from the nite element analysis. In order to achieve the optimality conditions, the algorithm has to drive the error signal ei to zero. It can be shown that for a void element xi = 0, the optimality conditions are satised when ei 0. Similarly, for a solid element xi = 1, the optimality conditions are satised when ei 0. From these conditions, it is possible to state a control-based algorithm described by
k k k xi +1 = min max 0, xi + xi , 1 ,

(14)

where the local change in the design variable xi is a function of ei . This iterative scheme can be reconstructed as a closed-loops control acting on a discrete dynamic system. Several control strategies can be incorporated into this system. For example, using a two-position controller, the change in the design variable can be expressed as k k +KT if ei > 0, k k xi = (15) 0 if ei = 0, k KT if eik < 0, where KT is a constant value in the range [0, 1]. Using a proportional-integral-derivative (PID) controller, this change can be expressed as
k k k k 0 1 k k k k xi = KPi ei + KIi ei + ei + + ei + KDi ei ei 1 ,

(16)

where KPi , KIi , and KDi are positive scalars referred to as the proportional, integral, and derivative gains, respectively. These control rules are presented and utilized by Tovar et al. (2006). The convergence of this algorithm is studied by Penninger et al. (2010).

Controller

Plant

Analysis

Fig. 3. Embedded controlled-based optimization in the HCA algorithm. This multivariable feedback control system simultaneously operates on every discrete element of the design domain. This system includes the controller, plant or structure, and the nite element analysis (sensor). The discrete variable k denotes the iteration of the algorithm.

3.3 Numeric model and implementation

The numerical implementation of the HCA algorithm requires the discretization of the structure into CAs. Each CA represents a sensor cell and the surrounding material. In bones, the osteocyte density is estimated between 12, 000 and 20, 000 cells per cubic millimeter (Boyde,

1972; Frost, 1960; Mullender et al., 1996). Assuming a density of 16, 000 cells per cubic millimeter, a three-dimensional lattice of 25 25 25 = 15, 625 CAs approximately represents one cubic millimeter of bone. Each osteocyte is enclosed in a mineralized lacuna and probably has as many as 80 cytoplasmic processes resembling axons. These cytoplasmic processes are 15 m long. They are three-dimensionally arrayed in a manner that permits them to connect with processes from about 12 neighboring cells (Cowin & Moss, 2001). In order to reproduce the inter cellular communication among sensor cells, a radially symmetric three-dimensional CA neighborhood needs to be dened. This layout contains the immediate CAs arranged on the three orthogonal planes for a total of 19 CAs. The CA lattice and the CA neighborhood are shown in Fig. 4.

(a) CA lattice (b) CA neighborhood Fig. 4. Cellular automaton model of bone. The CA lattice (a) represents one cubic millimeter (1 mm3 ) of bone tissue. This Cartesian three-dimensional lattice is composed of 25 25 25 = 15, 625 CAs. The CA neighborhood (b) in a bone model is formed by 19 CAs which models inter-cellular communication among osteocytes. In order to model the cellular communication between sensor cells and the actuators, the effective error signal perceived by the BMUs at a discrete location is an average value of the error in the neighborhood. The effective error signal ei can be modeled by k ei = ik
k j Ni j ek j

| Ni |

(17)

where j represents the efciency of the transmission of the error signal from the location j to central location i and i is a strength factor of that signal. The parameters j and i are used to model the microcracks in the neighborhood and osteocytes health, respectively (Tovar, 2004). The set Ni represents the neighborhood of the cell i. In this application, the cardinality of this set is | Ni | = 19. With the use of this method, trabecular bone structure can be predicted (Fig. 5). These principles can be extended to solve structural optimization problems.

Fig. 5. Trabecular bone structure predicted with HCA (Tovar et al., 2004). The discrete times k correspond to the ones used by the algorithm.

4. Applications
4.1 Topology optimization

The compliance problem (2) can be solved using the control system of bone remodeling in a CA framework. In this case, the optimality conditions for an interior point yield M C + = 0, xi xi (18)

where is the Lagrange multiplier associated with the mass constraint. Comparing (18) with (11) yields (1 ) . (19) = Using the result from (19) in (13), the new expressions for the error ei , mechanical stimulus yi , and set point si are ei = yi si , where yi = 1 1 c , and si = m0i . xi i p (20)

Since is unknown, the HCA algorithm implements an additional loop to determine its value at every iteration. This addition to the control-based algorithm is presented in Fig. (6). A complete formulation for linear and nonlinear compliance problems using a control-based algorithm is presented by Tovar & Khandelwal (2010).

Controller

Plant

Analysis

Yes

Constraint satisfied? No

Update

Fig. 6. Embedded controlled-based optimization in the HCA algorithm. This is a multivariable feedback control system simultaneously operating in every discrete element of the design domain. This system includes the controller, plant or structure, and nite element analysis (sensor). The discrete variable k denotes the iteration of the algorithm.

The use of lters in topology optimization effectively mitigates numerical complications such as mesh-dependency and the formation of checkerboard patterns. The effective error signal as modeled by (17) can be expressed in terms of a lter imposed on ei . Traditionally, the HCA incorporates a box lter described by the average error in the neighborhood, i.e., k ei = j Ni ek j , (21)

| Ni |

where Ni is the neighborhood of the element i within a radius R (Sigmund, 2007). This can be expressed as Ni = { j : d(i , j ) R}, (22) where d() denotes distance and i is the spatial location of the i-th CAs centroid. The shape of the neighborhood depends on the lter radius R. Figure 7 shows the neighborhood layouts for three lter radii. R = 0.0 2D R = 1.2 R = 1.7 R = 2.3

3D

Fig. 7. CA neighborhoods in 2D and 3D lattices. The layouts result from different lter radii for a unit size element. To dene the neighborhood at the structures boundary, the design domain can be extended in different ways to dene the boundary condition (Chopard & Droz, 1998). A xed boundary is

dened so that the neighborhood is completed with cells having a pre-assigned xed state. An adiabatic boundary is obtained by duplicating the value of the cell in an extra virtual neighbor. In a reecting boundary, the state of the opposite neighbor is replicated by the virtual cell. With a periodic boundary, the design domain is assumed to be wrapped in a torus-like shape. Figure 8 shows the topologies of a microstructure obtained by using the HCA algorithm for different lter radii with periodic boundary conditions.

(a) Initial design

(b) R = 0.0 (c) R = 1.2 (d) R = 1.7 (e) R = 2.3 Fig. 8. Topology optimization of a microstructure. The design domain is uniformly loaded by traction forces (a). The optimization problem is to minimize compliance with a mass limit M = 0.5M0 . The results where obtained using periodic boundary conditions for different lter radii. When no lter is used (b), the nal structure depicts checkerboard patterns. The complexity of the nal structure decreases when the lter size increases. In the context of the compliance problem (2) under the SIMP model (8), the use of p = 1 makes the problem convex and the grey solution is unique. This problem is referred to as the thickness problem. On the other hand, the problem looses its convexity when p increases and intermediate density values are penalized. The solution is a 0-1 structure. Since the solution is not unique, the use of the continuation method is advised (Bendse & Sigmund, 2003). Figure 9 shows the nal designs obtained with the HCA algorithm for p = 1 and p = 3. Filter and the continuation method are implemented in the latter case.
4.2 Shape optimization

Shape optimization involves determining the optimal prole or boundary of the structure. Two of the most common approaches for shape optimization are: the basis vector approach and the grid perturbation approach. The basis vector approach requires the denition of different trial designs called basis vectors. The design variables are the weight parameters that dene the participation of each basis vector in the design process. On the other hand, the grid perturbation approach requires the denition of perturbation vectors. These vectors perturb or deform the boundary of the design domain. The design variables are the values that determine the amount of the perturbation that occurs during the optimization process.

(a) Initial design (b) p = 1 (c) p = 3 Fig. 9. Thickness and topology optimization of a Michell-type structure referred to as the wheel problem. The design domain is discretized in 50 25 CAs of unit size. One of the lower edges is restrained to prevent vertical and horizontal displacements, while the opposite lower edge is restrained in the vertical direction. A sufciently small vertical load is uniformly distributed along the center of the lower surface as depicted in (a). The compliance problem is constrained with a mass limit M = 0.5M0 . The unique solution of the convex thickness optimization problem for p = 1 is shown in (b). A 0-1 optimal design of the non-convex problem for p = 3 is shown in (c).

With the addition of appropriate CA conditions, the HCA algorithm can be applied to shapelike optimization. In the process of bone remodeling, the structural changes are localized on the surfaces of the mineralized tissue. In the context of HCA, this restriction is referred to as the surface condition (Tovar, 2004). The surface condition strives to avoid the spontaneous creation of voids within solid neighborhood and the spontaneous creation of solids within a void neighborhood. This basic surface condition can be stated as 0< j Ni x k j

| Ni |

< 1.

(23)

Additionally, since remodeling occurs only around the solid matrix, the neighborhood of a void element should contain at least one solid element. On the other hand, remodeling requires direct contact with bone marrow, then the neighborhood of a solid element should contain at least one void element. These conditions can be expressed as if if xi = 0 xi = 1 then then max{ x j Ni } = 1, min{ x j Ni } = 0. (24)

Any gray element is assumed to contain enough marrow to completely remodel. The set of conditions stated by (23) and (24) are referred to as the strict surface conditions. The application of surface conditions in topology optimization constraints the changes to the surface of the structure, in a way similar to the one of the shape optimization. This process is shown in Fig. 10. In this example, the genus of the structure is maintained.

(a) Initial design (b) Final design Fig. 10. Shape-like optimization of the wheel problem described in Fig. 9. During this process, the surfaces of the structure are reshaped as in shape optimization. The initial design (a) and the nal design (b) have the same genus.

4.3 Compliant mechanisms

Compliant mechanisms are exible single-piece structures that transform motion, force, or energy by way of elastic deformation of exible components, rather than from movable joints in traditional rigid-body mechanisms. In a conventional rigid-linked mechanism, multiple parts exist and deformations occur at hinged joints. Compliant mechanisms can transform the same energy by elastically deforming throughout the entire structure. While structures are exible enough to be able to transmit motions, they are also stiff enough to support the input load. The elastic deformation of compliant mechanisms can either be concentrated in the regions of the exible hinges as in the case of lumped compliant mechanisms, or the deformation can be distributed throughout the entire mechanism. Compliant mechanisms are especially well suited for micro-scale models, where manufacturability of hinges is an issue. The use of topology optimization to generate compliant mechanisms was rst proposed by Ananthasuresh et al. (1994). Their work proposes a multi-criteria formulation that accounts for both exibility and stiffness requirements. To accomplish this, an input load-case and an output load-case are considered. The input load-case accounts for the boundary conditions of the input load at the point of actuation. The output load-case includes a dummy load at the output deection point and the output reaction force using a linear spring of stiffness that represents a workpiece that the mechanism is acting against. The topology optimization can be expressed as maximizing mutual potential energy W while minimizing compliance C with a mass constraint. This is
min 0 x 1

W ( x) + (1 )C ( x)
M( x) = M.

subject to

(25)

A compliant mechanism is a structure between these two conditions. The mutual potential energy is a measure of the output deformation of the mechanism shared between the two loading conditions (Shield & Prager, 2001). In a sense, it is a measure of the exibility of the structure. When = 0, this problem becomes the compliance problem and the structure has maximum stiffness, at least for linear analysis. On the other hand, when = 1, the structure is of maximum exibility.

Similarly to the denition of compliance (5) and its sensitivity (9), the mutual potential energy and its corresponding sensitivity can be dened by W ( x) = F in U out and ext
T

1 W = pwi , xi xi
T

(26)

where element mutual potential energy is wi = f in uout , and in and out refer to the i i input and output loading conditions, respectively. In the context of HCA, using the optimality conditions of (25), the error signal, the eld variable, and the set point for every CA can be written as (1 ) 1 ei = yi si , where yi = wi + ci and si = m0i , (27) xi xi p where is the Lagrange multiplier associated with the mass constraint. In the design of compliant mechanisms, the HCA algorithm makes use lter. An example of an elevation mechanism is presented in Fig. (11). Although, the nal design can not be claimed to be optimal for non-linear cases, they present themselves as good candidate designs.
Output port

Input port

Input port

(a) Initial design (b) Final design Fig. 11. Topology optimization of a compliant mechanism. In this problem, input forces are applied at the prescribed input ports on the sides of the simply supported structure. The nal displacement is prescribed at the output port in the z direction on the upper free surface.

4.4 Crashworthiness

Crashworthiness designers seek to tailor a vehicles structural design to maximize energy absorption while maintaining structural stiffness for occupant protection (Patel, 2007). To avoid impact or deceleration injury to passengers, the reaction force during a crash event needs to be driven to permissible values. Long exposures to high deceleration forces can cause uid displacement, tissue deformation, and internal-organ rupture among other injuries (Horgan & Gilchrist, 2004). At the same time, a crashworthy design must ensure that the energy is absorbed within a prescribed allowable displacement to prevent the passengers from being penetrated by the structure. To prevent excessive deceleration levels, the available deformation distance in front of the passenger compartment must be optimally utilized for crashworthiness. This implies that the structural design of a given vehicle must balance stiffness and deformation for a given crash velocity. One way to pose the structural optimization problem is maximizing the energy absorption U during the dynamic event, while keeping the maximum deceleration a and maximum pene-

tration d within reasonable limits a and d. This problem can be written as


max 0 x 1

U ( x, t) max{ a( x, t)} a(t) t max{d( x, t)} d(t),


t

subject to

(28)

where t is time involved in the dynamic event. Over the past decade, there has been an array of advancements to improve motor vehicle safety by designing structures that trade-off the two main objectives: deceleration (energy absorption) and penetration (stiffness). For mildly nonlinear problems, investigators (Mayer et al., 1996; Min et al., 1999; Pedersen, 2003), have simplied the complex interactions in the problem, so that they can use approximated analytical sensitivities instead of requiring numerical computation. These applications are limited to a very small class of nonlinear phenomena. When the transient behavior is combined with contact forces, large displacements, large (and permanent) deformations, strain rate, plastic hardening, and other phenomena for 3D structures, the analysis is both complex and numerically expensive. Heuristic methods have been proposed in order to avoid these numerical difculties for simplied material models (Forsberg & Nilsson, 2007; Ma & Kikuchi, 2006; Soto, 2004). Currently, there is not one publication where such sensitivities, considering all phenomena simultaneously, are analytically derived or numerically computed. In design for crashworthiness, the HCA algorithm adopts a concept similar to fully stressed design (Haftka et al., 1990) and uniform strain energy density (Pedersen, 2000). A fully stressed design formulation states that for the optimum design, each member of the structure that is not at its minimum gage is fully stressed under at least one of the design load conditions (Patnaik & Hopkins, 1998). In the context of HCA, the goal is to absorb energy efciently, i.e., with the minimum possible mass, while constraining force/displacement. In a discretized structure, the energy absorbed by each element can be measured by integrating the load transmitted to the element over the resulting displacement. If the elements internal energy is driven by the algorithm to a uniform set point value, then the resulting structure is of minimum mass for the given set point. In the context of HCA, the error signal, the eld variable, and the set point for every CA can be written as ei = yi si , where yi = ui and si = si ( a, d), (29) where ui is the internal energy density at the i-th CA. The function si ( a, d) seeks to satisfy the functional constraints imposed by deceleration and penetration. For complex design requirements, the design domain is decomposed into sub-domains with particular set points (Mozumder, 2007). However, when the conditions of the problem impose, for example, a penetration function d(t), then only one constraint needs to be satised. Figure 12 depicts the results of applying the HCA algorithm to the design a bumper-like structure under static and dynamic loading condition and linear and nonlinear analysis (Patel, Kang, Renaud & Tovar, 2009). Using CA rules, extrusion constraints can be applied to reduce the manufacturability cost of the optimized topology (Patel, Penninger & Renaud, 2009).

5. Conclusion
The HCA algorithm is a structural optimization tool used in topology optimization. It is derived from the process of bone functional adaptation via remodeling. In that context, the bone

(a) Initial design

(b) Linear/static (c) Nonlinear/static (d) Nonlinear/dynamic Fig. 12. Crashworthiness design of a bumper structure. The design domain (a) is discretized into 80 20 20 identical CAs of unit size. A rigid pole impacts the design. The structure is fully supported at both sides. The problem is rst solved for static load under linear analysis (b). Then geometric and material nonlinearities are applied under static load (c). Finally, nonlinear analysis and dynamic loading conditions are used (c). LS-DYNA is used as the nite element analysis tool.

cells are represented as CAs. Their evolution according to a design rule iteratively distributes the material in each CA. Communication among the CAs is modeled with different neighborhood layouts and boundary conditions (e.g., periodic and xed). The complexity of the nal design is inversely proportional to the size of the neighborhood. This chapter demonstrates the use of the HCA algorithm in topology optimization for problems such as the compliance problem, compliant mechanism design, and crashworthiness. With the addition of surface conditions, the HCA algorithm operates on the free structures surfaces within the design domain. The HCA algorithm makes use of nite element analysis in a closed-loop control system. It requires three fundamental quantities: the design variable, the eld variable, and the set point. The design variable corresponds to each CAs relative density. When sensitivity coefcients are available, the eld variable and the set are obtained from the optimality conditions of the optimization problem. In general, the eld variable is a function of the objectives sensitivity, while the set point is a function of the constraints sensitivity. When the sensitivities are not presented, these quantities are dened according to heuristics, e.g., internal energy density for crashworthiness design.

6. Acknowledgements
The authors would like to acknowledge Honda R&D Americas and the National Science Foundation (NSF: CMMI 0700730) that have sponsored the development of the HCA algorithm and its use in automotive applications. Also, the former and current graduate students Neal Patel, Chandan Mozumder, and Punit Bandi for their contribution to compliant mechanism and

crashworthiness optimization, and Charlie Penninger, Huade Tan, and Jack Goetz for the extension of the HCA algorithm to other areas of Engineering. Finally, Prof. Kapil Khandelwal for the long hours of discussion on control-based optimization.

7. References
Allaire, G. (2001). Shape Optimization by the Homogenization Method, Springer, New York. Allaire, G. & Kohn, R. (1993). Optimal design for minimum weight and compliance in planestress using extremal microstructures, European Journal of Mechanics 12(6): 839878. Anagnostou, G., Ronquist, E. & Patera, A. (1992). A computational procedure for part design, Computer Methods in Applied Mechanics and Engineering 97(1): 3348. Ananthasuresh, G., Kota, S. & Kikuchi, N. (1994). Strategies for systematic synthesis of compliant mems, dynamic systems and control, 1994 ASME Winter Annual Meeting, Vol. DSC-Vol 55-2, Chicago, IL, pp. 677686. Arora, J. (1989). Introduction to Optimum Design, The McGraw-Hill Companies, Inc., New York. Bendse, M. (1989). Optimal shape design as a material distribution problem, Structural Optimization 1(4): 193202. Bendse, M. (1995). Optimization of structural topology, shape, and material, Springer-Verlag, Berlin. Bendse, M. & Kikuchi, N. (1988). Generating optimal topologies in structural design using a homogenization method, Computer Methods in Applied Mechanics and Engineering 71(2): 197224. Bendse, M. & Sigmund, O. (2003). Topology Optimization: Theory, Methods and Applications, second edn, Springer Verlag. Boyde, A. (1972). The Biochemestry and Physiology of Bone, Academic Press, chapter Scanning electron microscope studeis of bone. Calvel, S. & Mongeau, M. (2007). Black-box structural optimization of a mechanical component, Computers & Industrial Engineering 53(3): 514530. Chan, C. & Wong, K. (2008). Structural topology and element sizing design optimisation of tall steel frameworks using a hybrid oc-ga method, Structural and Multidisciplinary Optimization 35(5): 473488. Chapman, C., Saitou, K. & M.J., J. (1994). Genetic algorithms as an approach to conguration and topology design, Journal of Mechanical Design 116(4): 10051012. Chia, C., Rongong, J. & Worden, K. (2006). Structural optimisation using a hybrid cellular automata (HCA) algorithm, Applied Mechanics and Materials 5-6: 93100. Chopard, B. & Droz, M. (1998). Cellular automaton modeling of physical system, Cambrige University Press. Cowin, S. C. & Moss, M. L. (2001). Mechanosensory mechanisms in bone, in S. C. Cowin (ed.), Bone Mechanics Handbook, CRC Press. Cowin, S., Moss-Salentijn, L. & Moss, M. (1991). Candidates for the mechanosensory system in bone, Journal of Biomechics 113(2): 191197. Daz, A. & Sigmund, O. (1995). Checkerboard patterns in layout optimization, Structural Optimization 10(1): 4045. Duysinx, P. (1997). Layout optimization: A mathematical programming approach, DCAMM report, Technical report, Danish Center of Applied Mathematics and Mechanics, Technical University of Denmark, DK-2800 Lyngby.

Fleury, C. (1989). CONLIN: an efcient dual optimizer based on convex approximation concepts, Structural Optimization 1(2): 8189. Forsberg, J. & Nilsson, L. (2007). Topology optimization in crashworthiness design, Structural and Multidisciplinary Optimization 33(1): 112. Frost, H. (1960). Measurements of osteocytes per unit volume and volume components of osteocytes and caliculae in man, Henry Ford Hospital Medical Bulletin 8: 208211. Frost, H. (2003). Bones mechanostat: A 2003 update, The Anatomical Record. Part A 2: 1081 1101. Fuchs, M., Jiny, S. & Peleg, N. (2005). The SRV constraint for 0/1 topological design, Structural and Multidisciplinary Optimization 30(4): 320326. Goel, T., Roux, W. & Stander, N. (2009). A topology optimization tool for LS-DYNA users: LS-OPT/Topology, 7th European LS-DYNA Conference, number AIAA 2008-6046, Salzburg, Austria. Haber, R., Jog, C. & Bendse, M. (1996). A new approach to variable-topology shape design using a constraint on perimeter, Structural Optimization 11(1): 112. Haftka, R., Grdal, Z. & Kamat, M. (1990). Elements of Structural Optimization, Kluwer Academic Publishers, Dordrecht, The Netherlands, 2nd ed. Horgan, T. & Gilchrist, M. (2004). Inuence of fe model variability in predicting brain motion and intracranial pressure changes in head impact simulations, International Journal of Crashworthiness 9(4): 401418. Huiskes, R. (2000). If bone is the answer, then what is the question?, Journal of Anatomy 197: 145156. Huiskes, R., Ruimerman, R., van Lenthe, G. . H. & Janssen, J. D. (2000). Effects of mechanical forces on maintenance and adaptation of form in trabecular bone, Nature 405(6787): 704706. Inou, N., Shimotai, N. & Uesugi, T. (1994). Cellular automaton generating topological structures, 2nd European Conference on Smart Structures and Materials, number 2361-08, Glasgow, United Kingdom, pp. 4750. Jakiela, M., Chapman, C., Duda, J., Adewuya, A. & Saitou, K. (2000). Continuum structural topology design with genetic algorithms, Computer Methods in Applied Mechanics and Engineering 186(2-4): 339356. Jog, C. (2002). Topology design of structures using a dual algorithm and a constraint on the perimeter, International Journal for Numerical Methods in Engineering 54(7): 10071019. Kaveh, A., Hassani, B., Shojaee, S. & Tavakkoli, S. (2008). Structural topology optimization using ant colony methodology, Engineering Structures 30(9): 25592565. Kita, E. & Toyoda, T. (2000). Structural design using cellular automata, Structural and Multidisciplinary Optimization 19(1): 6473. Kohn, R. & Strang, G. (1986a). Optimal design and relaxation of variational problems (part I), Communications of Pure and Applied Mathematics 39(1): 113137. Kohn, R. & Strang, G. (1986b). Optimal design and relaxation of variational problems (part II), Communications of Pure and Applied Mathematics 39(2): 139182. Kohn, R. & Strang, G. (1986c). Optimal design and relaxation of variational problems (part III), Communications of Pure and Applied Mathematics 39(3): 353377. Ma, Z. & Kikuchi, N. (2006). Multidomain topology optimization for structural and material designs, ASME Journal of Applied Mechanics 73(4): 565573. Martin, R., Burr, D. & Sharkey, N. (1998). Skeletal Tissue Mechanics, Spriger-Verlag, New York.

Maute, K., Schwartz, S. & Ramm, E. (1998). Adaptive topology optimization of elastoplastic structures, Structural Optimization 15(2): 8191. Mayer, R., Kikuchi, N. & Scott, R. (1996). Applications of topology optimization techniques to structural crashworthiness, International Journal for Numerical Methods in Engineering 39: 13831403. Miller, S. & Jee, W. (1992). Bone lining cells, Bone 4: 119. Min, S., Kikuchi, N., Park, Y., Kim, S. & Chang, S. (1999). Optimal topology design of structures under dynamic loads, Structural and Multidisciplinary Optimization 17(2-3): 208 218. Missoum, S., Grdal, Z. & Setoodeh, S. (2005). Study of a new local update scheme for cellular automata in structural design, Structural and Multidisciplinary Optimization 29(2): 103 112. Mlejnek, H. (1992). Some aspects of the genesis of structures, Structural Optimization 5(1-2): 64 69. Mozumder, C. (2007). Topometry Optimization of Sheet Metal Structures for Crashworthiness Design using Hybrid Cellular Automata, PhD thesis, University of Notre Dame. Mullender, M., Meer, D., Huiskes, R. & Lips, P. (1996). Osteocyte density changes in aging and osteoporosis, Bone 18(2): 109113. Patel, N. (2007). Crashworthiness Design using Topology Optimization, PhD thesis, University of Notre Dame. Patel, N., Kang, B., Renaud, J. & Tovar, A. (2009). Crashworthiness design using topology optimization, ASME Journal of Mechanical Design 131(6): 061013. Patel, N., Penninger, C. & Renaud, J. (2009). Topology synthesis of extrusion-based nonlinear transient designs, ASME Journal of Mechanical Design 131(6): 061003. Patel, N., Tillotson, D., Renaud, J., Tovar, A. & Izui, K. (2008). Comparative study of topology optimization techniques, Structural and Multidisciplinary Optimization 46(8): 1963 1975. Patnaik, S. & Hopkins, D. (1998). Optimality of fully stressed design, Computer Methods in Applied Mechanics and Engineering 165(1-4): 215221. Pedersen, C. (2003). Topology optimization design of crushed 2d-frames for desired energy absorption, Structural and Multidisciplinary Optimization 25(5-6): 368382. Pedersen, P. (2000). On optimal shapes in materials and structures, Structural and Multidisciplinary Optimization 19(3): 169182. Penninger, C., Watson, L., Tovar, A. & Renaud, J. (2010). Convergence analysis of hybrid cellular automata for topology optimization, Structural Multidisciplinary Optimization 40(1-6): 271282. Petersson, J. & Sigmund, O. (1998). Slope constrained topology optimization, International Journal of Numerical Methods in Engineering 41(8): 14171434. Rozvany, G. (2009). A critical review of established methods of structural topology optimization, Structural and Multidisciplinary Optimization 37(3): 217237. Schmit, L. (1960). Structural design by systematic synthesis, Proceedings of the 2nd Conference on Electronic Computation, number ASCE, New York, NY, pp. 105122. Schwarz, S. & Ramm, E. (2001). Sensitivity analysis and optimization for non-linear structural response, Engineering Computations 18(3-4): 610641. Shield, R. T. & Prager, W. (2001). Optimal structural design for given deection, Journal of Applied Mathematics and Physics, ZAMP 21: 513523.

Sigmund, O. (1997). On the design of compliant mechanisms using topology optimization, Mechanics of Structures and Machines 25(4): 493524. Sigmund, O. (2001). A 99 line topology optimization code written in Matlab, Structural and Multidisciplinary Optimization 21(2): 120127. Sigmund, O. (2007). Morphology-based black and white lters for topology optimization, Structural and Multidisciplinary Optimization 33(4-5): 401424. Sigmund, O. & Petersson, J. (1998). Numerical instabilities in topology optimization: A survey on procedures dealing with checkerboards, mesh-dependencies and local minima, Structural and Multidisciplinary Optimization 16(1): 6875. Soto, C. (2004). Structural topology optimization for crashworthiness, International Journal for Crashworthiness 9(3): 277283. Svanberg, K. (1987). The method of moving asymptotesa new method for structural optimization, International Journal of Numerical Methods in Engineering 24(2): 359373. Svanberg, K. & Werme, M. (2006). Topology optimization by a neighbourhood search method based on efcient sensitivity calculations, International Journal for Numerical Methods in Engineering 67(12): 16701699. Svanberg, K. & Werme, M. (2007). Sequential integer programming methods for stress constrained topology optimization, Structural and Multidisciplinary Optimization 34(4): 277299. Tovar, A. (2004). Bone Remodeling as a Hybrid Cellular Automaton Optimization Process, PhD thesis, University of Notre Dame. Tovar, A. & Khandelwal, K. (2010). Control-based topology optimization and applications to linear and nonlinear compliance problems and approximate formulations, Submitted to Structural and Multidisciplinary Optimization . Tovar, A., Patel, N., Kaushik, A. & Renaud, J. (2007). Optimality conditions of the hybrid cellular automata for structural optimization, AIAA Journal 45(3): 673683. Tovar, A., Patel, N., Letona, G., Niebur, G., Sen, M. & Renaud, J. (2004). Evolutionary model for bone adaptation using cellular automata, 14th Conference of the European Society of Biomechanics, Hertogenbosch, The Netherlands. Tovar, A., Patel, N., Niebur, G., Sen, M. & Renaud, J. (2006). Topology optimization using a hybrid cellular automaton method with local control rules, Journal of Mechanical Design 128(6): 12051216. Xie, Y. & Stevens, G. (1993). A simple evolutionary procedure for structural optimization, Computers & Structures 49(5): 885896. Yang, R. & Chuang, C. (1994). Optimal topology design using linear programming, Computers & Structures 52(2): 265275. Zhou, M. & Rozvany, G. (1991). The COC algorithm, part II: Topological, geometrical and generalized shape optimization, Computer Methods in Applied Mechanics and Engineering 89: 309336.

You might also like