An Effective Method of Synthesizing Compliant Adaptive

Structures using Load Path Representation

KERR-JIA LU1,* AND SRIDHAR KOTA2
1
Department of Mechanical and Aerospace Engineering, George Washington University
Washington, DC 20052
2
Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109-2125, USA

ABSTRACT: The synthesis of shape morphing compliant mechanism is inherently different

from the typical single output design problems, due to the multiple output points along the
morphing boundary. We have previously developed a genetic algorithm (GA)-based synthesis
approach, incorporating a binary ground structure parameterization, to systematically design
shape morphing compliant mechanisms. However, the approach is ineffective due to issues
such as the generation of disconnected structures and the need to choose an initial mesh. In
this paper, we present the ‘load path representation,’ which is developed to overcome the
issues encountered using the binary ground structure parameterization. The performance of
the load path approach over the binary ground structure approach is demonstrated through
several design examples. The results have shown that the load path approach offers several
advantages, such as (a) eliminating the need of an initial ground structure, (b) ensuring
structural connectivity, and (c) yielding solutions that generate desired shape change
efficiently.

Key Words: compliant mechanism, adaptive shape morphing, load path representation,
binary ground structure, genetic algorithm.

INTRODUCTION In a previous paper (Lu and Kota, 2003), we had

compliant mechanism is a single-piece flexible
A structure that is designed to be flexible to transmit
motions, yet stiff enough to withstand external loads.
Various synthesis approaches (Ananthasuresh et al.,
1994; Frecker et al., 1997; Larsen et al., 1997; Hetrick
and Kota, 1999; Saggere and Kota, 1999; Tai and Chee,
2000), utilizing structural optimization techniques, have
been developed in the past decade to design practical
devices such as motion amplifiers (Hetrick and Kota,
1999) in MEMS and compliant grippers (Joo et al.,
2000). One unique feature of compliant mechanisms is
its smooth deformation field, due to the distributed
compliance, that provides a novel means to morph
structural shape. However, unlike the single output
problems studied in most of the previous research, shape
morphing involves deforming the structural boundary,
where every point is an output point, from its initial
profile to a desired target shape. Development of a new
synthesis approach is, therefore, required to address the
multiple output points presented in such a type of
problem (Lu and Kota, 2003).
*Author to whom correspondence should be addressed. E-mail: kjlu@gwu.edu
developed a genetic algorithm (GA)-based synthesis
approach for shape morphing compliant mechanisms,
using a binary ground structure parameterization. As
shown in Figures 1 and 2, the ‘binary ground structure
approach’ starts with a preprocessor to estimate
problem feasibility and identify several candidate loca-
tions along the morphing boundary to be the ‘output
points’ of the compliant mechanism. The design domain
is then parameterized using a ground structure of beam
elements (Figure 2(b)). Each beam element is described
by two design variables: one binary variable determining
the presence of an element (topology), and one real
variable defining the cross sectional area (dimension).
The use of binary topology variables distinguishes itself
from the traditional ground structure optimization
approach, hence the name ‘binary ground structure.’
Although the results have demonstrated feasibility of the
synthesis approach, it is unclear how the initial ground
structure is selected (the mesh configuration and
resolution in Figure 2(b)). Furthermore, the representa-
tion using binary variables actually includes invalid
structures that are disconnected from input or ground
support in the design space, as shown in Figure 3. An
additional verifying algorithm is thus required to ‘search
for’ and penalize the disconnected designs, leading to

JOURNAL OF INTELLIGENT MATERIAL SYSTEMS AND STRUCTURES, Vol. 16—April 2005 307
1045-389X/05/04 0307–11 $10.00/0 DOI: 10.1177/1045389X05050104
ß 2005 Sage Publications
308 K.-J. LU AND S. KOTA

inefficiency in the overall synthesis process. Moreover,
due to the direct correspondence between the design
variables and the FE elements, the number of design
variables will increase dramatically when a more refined
initial mesh is used.
In this paper, we have developed a ‘load path repre-
sentation’ to address the issues associated with the
binary ground structure. The load path representation
renders a novel parameterization to replace the binary
ground structure parameterization in Figure 1. The
overall synthesis approach is, therefore, referred to as
the ‘load path approach.’ It ensures structural con-
nectivity and allows variable mesh configurations with-
out the need to choose an initial discretization mesh. It is
largely inspired by the morphological representation
(Tai and Chee, 2000) tailored to design structures and
single output compliant mechanisms. Interested readers
can refer to Tai and Chee (2000) for more details on
the morphological representation. The load path repre-
sentation is presented in this paper, followed by several
design examples, to study the improvements over the
binary ground structure approach.

LOAD PATH REPRESENTATION

In every compliant mechanism, three classes of points,

Figure 1. Flowchart for both GA-based synthesis approaches
using binary ground structure and load path representation.
termed as ‘essential ports,’ must exist to ensure proper
functionality of the system: (1) input points for
actuation; (2) fixed points to ground the structure; and
(3) output points to interact with the external environ-
ment. In the shape morphing problem, there are
multiple output points, while the number of input and

Figure 2. Compliant mechanism synthesis using the binary ground structure shown in Figure 1: (a) problem specification; (b) design domain
parameterization; and (c) the hypothetical design obtained from the GA.

Figure 3. Three types of invalid structures that can be generated using the binary ground structure shown in Figure 2(b).
 Synthesis of Compliant Adaptive Structures 309

Figure 4. The essential ports should be connected directly or

indirectly by load paths (nondirectional).

fixed points are specified by the designer. In this

research, only one actuator is employed. As can be seen
in Figure 4, these essential ports should be connected
directly or indirectly to each other to form a valid Figure 5. An example design with 6 intermediate connection ports.
This design is considered fully connected because pTopi ¼ 1 for all
compliant mechanism. The physical connections are the paths (i ¼ 1–7).
‘load paths’ in a structure to transmit force between the
input actuator and the output points. In the load path
Table 1. The pathSeq and pTop for the designs in
representation, it is assumed that ‘every essential port is
Figures 5 and 6. Note that the pathSeqs are identical for
connected to every other essential port of a different the two designs, and the pDim is only for Figure 6.
class via one load path.’ Thus, three types of load paths
are generated: (1) paths from Input ! Output, (2) paths Path Path Path pTopi pTopi pDimi
Type No. Sequence Figure 5 Figure 6 Figure 6
from Input ! Fixed points, and (3) paths from Fixed
point ! Output. In addition, a set of intermediate In ! Out 1 {1, 6, 3} 1 1 {1, 0.75}
connection ports (interconnect ports) are introduced to 2 {1, 7, 11, 4} 1 0 {1, 1, 2}
3 {1, 7, 11, 5} 1 1 {4.5, 0.75, 3}
allow connections between different load paths, thus
In ! Fix 4 {1, 6, 2} 1 1 {2.5, 5}
achieving ‘indirect’ connections between input/output/
fixed points. Fix ! Out 5 {2, 3} 1 1 {0.75}
6 {2, 6, 7, 4} 1 1 {5, 1.5, 8}
7 {2, 7, 11, 5} 1 0 {1.8, 1, 1}
Topology and Connectivity

Based on the layout shown in Figure 4, each design

can be represented by a graph with vertices as the
essential and interconnect ports. Figure 5 shows an
example design that is ‘fully connected’ (defined later).
As seen in Table 1, each load path is represented in
terms of the ‘path sequence’ ( pathSeq) containing
vertices along the path. For each load path, there is
also a corresponding binary variable, pTop, which
indicates the presence or elimination of this path:
when all pTop ¼ 1, the graph is considered ‘fully
connected’ with all paths present; when some pTop ¼ 0
in a graph, the pTop ¼ 0 paths are eliminated from the
design. For example, the design in Figure 6 has the same
path sequences as Figure 5, but different pTop values
create a different topology. The pTop values for Figure 6
are shown in the fifth column in Table 1, while the Figure 6. An example design with identical pathSeq as in Figure 5,
pathSeq’s (third column) are identical to those in but different pTopi values (fifth column, Table 1) lead to a different
Figure 5. The pathSeq and pTop are, therefore, consid- topology.
ered the topology design variables. By changing the
pathSeq and pTop, various topologies can be created. rest of the structure. These rules can easily be
To ensure that all designs are well-connected, each incorporated in the load path approach by monitoring
design must satisfy two connectivity requirements: the pTop values of paths from input and fixed point(s)
(1) the structure must be grounded at one or more to the output points. At least one of the pTop
fixed points, and (2) the input must be connected to the values in each path type has to be 1 to satisfy the
310 K.-J. LU AND S. KOTA
connectivity requirement, thus the structural connectiv-
ity is now shown explicitly in the design variables. This
is a huge advantage over the binary ground structure
approach, where a topology variable only determines
the existence of one element and the connectivity has
to be ‘searched for.’
Size and Geometry Aspects
It is assumed that the load paths are comprised of
rectangular beam elements, thus the size design varia-
bles are then defined as the cross section dimensions
of each beam section between two vertices. In this
research, the out-of-plane beam dimension is pre-
scribed as a constant, thus only in-plane dimensions
are considered design variables. Take Figure 6 for
example, each path has a corresponding sequence,
pDim, to describe the section dimensions, shown in
the sixth column in Table 1. The upper and lower
bound of the section dimensions are based on
manufacturing capability. When two load paths have
overlapping sections, such as in paths #1 and #4
between vertices 1 and 6 (bold in Table 1), one of the
two available pDim values (1 and 2.5 from pDim1 and
pDim4) is randomly selected to describe the section
dimension. Since only one value is required to define
one section, this random selection can be seen as the
selected value dominating the recessive one. An
additional variable, hBoundary, is also included in
each design to define the in-plane dimension of the
morphing boundary. Finally, to control the geometry
of the compliant mechanism, the locations of the
interconnect ports are allowed to move within the
design domain and are considered the geometry design
variables ( portLocation).
Using the design variables defined in the load
path representation, the topology, size, and geometry
of the compliant mechanism can be explored simulta-
neously. This parameterization scheme is incorporated
into the same synthesis procedure shown in Figure 1
in place of the binary ground structure, and is referred
to as the ‘load path approach’ in this paper. One crucial
component in this GA-based synthesis approach is
the reproduction scheme, including selection, crossover,
and mutation. Due to the difference in data structure
from the binary ground structure, a new set of genetic
operation strategies are required to work with load path
representation. The genetic operations need to maintain
sufficient diversity in each generation to prevent pre-
matured convergence to a dominant design. In the
following section, we present a brief overview of the GA
synthesis approach (Lu and Kota, 2003) and introduce
the genetic operation schemes incorporated in the
load path approach. Additional information on GAs,
in general, can be found in Holland (1975) and
Goldberg (1989).
GENETIC OPERATION STRATEGIES IN GA
Genetic algorithm simulates the selection scheme seen
in nature and evolves a population of designs based on
the principle of ‘survival of the fittest.’ The ‘fitness’ of
each design is defined based on the objective function.
In the shape morphing problem, the objective is to
minimize the ‘shape deviation’ between the deformed
shape and the desired target shape. Two shape devia-
tion objective functions have been developed based on
least square errors and Fourier transformation, respec-
tively (Lu and Kota, 2003). The smaller the deformed
shape deviates from the target shape, the better the
shape morphing is achieved, hence, the fitter the design.
As shown in Figure 1, in the GA, the designs in the first
generation are randomly generated to somewhat sample
the whole solution space. This parent generation then
produces a new generation from a reproduction process.
The load path approach incorporates a ‘roulette wheel’
selection scheme, where fitter individuals have higher
chances to be selected for reproduction, and inferior
ones have lower probability to reproduce. The selected
parent designs then produce new offspring designs
through the genetic operations: crossover and mutation.
The genetic operations create diversity within each
generation, which, in fact, provides the power behind
GA to improve designs as generations evolve. The
crossover and mutation strategies are, therefore, essen-
tial to the performance of GA.
Crossover Strategies
The crossover strategy in the load path approach
is to ‘exchange’ randomly selected paths between two
parent designs. More than one path can be selected
for crossover. The pathSeq and pTop of the parent
designs are exchanged during this process while preserv-
ing the original pDim and portLocation. Since new
pathSeq leads to new connectivity in the offspring
design, the number of segments (length of pDim) along
each path may be different from that in the parent
design. The pDim must be modified by inserting
additional values or removing extra ones to maintain
compatibility with the pathSeq. In addition, the bound-
ary dimension (hBoundary) of the two parent designs
can be exchanged according to the crossover probability.
Figure 7 shows two example designs (P1 and P2) with
their load path information listed in Table 2. Paths #1
and #5 are selected, for example, to illustrate the
crossover operation. The two new designs are shown
in Figure 8 and Table 3 as K1 and K2. As shown in
Table 3, the original pathSeq and pTop of Paths #1 and
#5 (asterisked) are exchanged between P1 and P2,
but pDim and interconnect port locations remain
the same. Since, the length of pDim changes after
the crossover. Therefore, a random value (within the
 Synthesis of Compliant Adaptive Structures 311

Figure 7. Two example parent designs with load path

information listed in Table 2.

Table 2. The load path information for the parent designs shown in Figure 7.
Figure 7(a) P1 Figure 7(b) P2

Path Type Path # pathSeq pTop pDim pathSeq pTop pDim

In ! Out 1 {1, 6, 3} 1 {1, 0.75} {1, 8, 11, 3} 0 {2, 5, 3}

2 {1, 7, 11, 4} 0 {1, 1, 2} {1, 6, 4} 1 {3, 1.5}
3 {1, 7, 11, 5} 1 {4.5, 0.75, 3} {1, 5} 1 {4}
In ! Fix 4 {1, 6, 2} 1 {2.5, 5} {1, 10, 2} 0 {2, 4}
Fix ! Out 5 {2, 3} 1 {0.75} {2, 9, 3} 1 {1, 2}
6 {2, 6, 7, 4} 1 {5, 1.5, 8} {2, 8, 6, 4} 1 {1, 5, 2}
7 {2, 7, 11, 5} 0 {1.8, 1, 1} {2, 8, 9, 5} 0 {2, 1.2, 3}
hBoundary {3, 4, 5} 1 {5} {3, 4, 5} 1 {2}

Figure 8. Two offspring designs obtained from the

parent designs in Figure 7 by exchanging paths #1
and #5. Their load path information is listed in
Table 3.

Table 3. The load path information of the offspring designs shown in Figure 8.
Figure 8(a) K1 Figure 8(b) K2

Path Type Path # pathSeq pTop pDim pathSeq pTop pDim

In ! Out 1* {1, 8, 11, 3} 0 ! {1, 2, 0.75} {1, 6, 3} 1 ! {5, 3}

2 {1, 7, 11, 4} 0 {1, 1, 2} {1, 6, 4} 1 {3, 1.5}
3 {1, 7, 11, 5} 1 {4.5, 0.75, 3} {1, 5} 1 {4}
In ! Fix 4 {1, 6, 2} 1 {2.5, 5} {1, 10, 2} 0 {2, 4}
Fix ! Out 5* {2, 9, 3} 1 ! {0.75, 3} {2, 3} 1 ! {1}
6 {2, 6, 7, 4} 1 {5, 1.5, 8} {2, 8, 6, 4} 1 {1, 5, 2}
7 {2, 7, 11, 5} 0 {1.8, 1, 1} {2, 8, 9, 5} 0 {2, 1.2, 3}
hBoundary * {3, 4, 5} 1 {2} {3, 4, 5} 1 {5}
312 K.-J. LU AND S. KOTA

Table 4. The load paths for the original K1 in Figure 8(a) and its mutated version in Figure 9.
Figure 8(a) K1 – original Figure 9 K1 – mutated

Path Type Path # pathSeq pTop pDim pathSeq pTop pDim

In ! Out 1* {1, 8, 11, 3} 0 {1, 2, 0.75} {1, 8, 11, 3} !1 {1, 2, 0.75}

2 {1, 7, 11, 4} 0 {1, 1, 2} {1, 7, 11, 4} 0 {1, 1, 2}
3 {1, 7, 11, 5} 1 {4.5, 0.75, 3} {1, 7, 11, 5} 1 {4.5, 0.75, 3}
In ! Fix 4 {1, 6, 2} 1 {2.5, 5} {1, 6, 2} 1 {2.5, 5}
Fix ! Out 5* {2, 9, 3} 1 {0.75, 3} {2, 9, 3} !0 {0.75, 3}
6* {2, 6, 7, 4} 1 {5, 1.5, 8} {2, 6, 7, 3} 1 {5, 1.5, 8}
7 {2, 7, 11, 5} 0 {1.8, 1, 1} {2, 7, 11, 5} 0 {1.8, 1, 1}
hBoundary * {3, 4, 5} 1 {2} {3, 4, 5} 1 {4.2}

pDim bounds) is inserted into pDim if the new pathSeq

is longer; a randomly selected segment cross section
value is removed from pDim, if shorter. In addition, the
boundary information is exchanged during crossover;
since the connectivity of the morphing boundary is
invariant, only the cross sectional dimension is changed
(hBoundary, also indicated by * in Table 3).

Mutation Strategies

Four options are offered in the mutation process:
(1) mutation of hBoundary, (2) mutation of path desti-
nation, (3) mutation of pTop, and (4) mutation of
portLocations. In hBoundary mutation, the boundary
dimension is replaced by a randomly generated value
within the upper and lower bounds. To mutate the
path destination, the end vertices of some randomly
selected paths can mutate to a different one within
the same class. For example, a path originally connect-
ing the input to one of the output points can be mutated
into a path connecting the input to another output
point, simply by changing the last vertex in the pathSeq.
The binary topology variable ( pTop) is also allowed
to mutate from 0 to 1 and vice versa for one randomly
selected load path, thus changing the topology. The
connection port location can also be mutated to a
different location within the design domain. Figure 9
shows an example design mutated from K1 in
Figure 8(a). The mutation in hBoundary, the destination
change in Path #6, and mutation of pTop1 and pTop5
are shown in Table 4, while the location change of
interconnect port 7 can be seen in Figure 9. All mutated
paths are indicated by * in Table 4.
It is noted that when the two parent designs are
identical, the ‘exchanging paths’ (crossover) strategy
fails to produce any new design. In fact, the offspring
designs will be identical to the parent designs. The
mutation probability is, therefore, higher in this
approach to enhance diversity in each generation.
Higher mutation also helps improve the crossover
performance, because the more diverse a generation
is, the less likely it is to select two identical parent
designs.
Figure 9. The new K1 (Figure 8(a)) after mutation shown in Table 4.
Note that interconnect port 7 is also mutated to a different location.
Convergence to Local Optimum
Owing to the heuristic nature of GA, the algorithm
is capable of searching the whole solution space more
extensively without being trapped in a local region.
Although GA is more efficient in locating a region close
to a local optimum, finding the exact optimum may be
quite difficult. If the GA can indeed explore the entire
solution space thoroughly, performing a local search
following the GA can accelerate the convergence to the
nearest local optimum, which is very likely to be the
global optimum. However, there is no guarantee that
GA can explore or sample the solution space evenly, so
adding a local search after GA can only lead to a local
optimum. In order to enhance the chance of finding
the global optimum, a global search, DIRECT optimi-
zation algorithm (Jones et al., 1993), is adopted to help
investigate the global optimality. DIRECT optimization
algorithm is a sampling algorithm that requires no
knowledge of the objective function gradient. The
algorithm samples points in the solution space and
uses the information it has obtained to decide where to
search next. It operates at both the global and local
levels. Once the global part of the algorithm finds the
basin of convergence of the optimum, the local part of
the algorithm quickly and automatically exploits it
(Jones et al., 1993). The topology of the optimal solution
obtained from GA is, therefore, used as a basic
layout for the DIRECT algorithm to perform
 Synthesis of Compliant Adaptive Structures 313

Figure 11. The morphing leading edge design obtained from ten
Figure 10. The GA-based load path approach followed by additional trials of the binary ground structure approach.
global/local search to improve the convergence.

additional iterations on the connection port locations

and beam section dimensions. However, the sampling
nature of DIRECT algorithm implies that the obtained
solution depends greatly on the number of iterations
(sampling points). Therefore, a local search algorithm is
utilized to accelerate the convergence to the nearby local
optimum, after a prescribed number of iterations are
carried out using the DIRECT algorithm. In this paper,
the optimization toolbox in Matlab is used to perform
the local search. More information on the optimization
toolbox can be found in the Matlab documentation. The
load path approach using GA, global search, and local
search, is shown in Figure 10.
Figure 12. The morphing leading edge design obtained from ten
trial runs of the load path approach.
EXAMPLES AND DISCUSSIONS

Shape morphing can be seen useful in many areas,
such as changing the aircraft wing shape to reduce drag,
or changing the lumbar support shape in chairs to
enhance comfort. To study the performance of the
synthesis approach for compliant mechanisms, the
binary ground structure approach and the improved
load path approach are applied to several examples to
understand their capabilities and limitations. All of the
examples use the Least Square Error (LSE) deviation
shown in Equation (1) as the objective function in GA
to evaluate the difference between the achieved shape
(deformed curve) and the desired target curve shape,
subject to size, node locations, stress, stiffness, and
connectivity constraints. Interested readers can refer to
our previous paper (Lu and Kota, 2003) for more details
regarding the problem formulation.
n qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X
LSE dev ¼ ðxDEF,i
xTAR,i Þ2 þ ð yDEF,i
yTAR, j Þ2
n i¼1
ð1Þ
where, n is the total number of data points along the
morphing boundary; (xDEF,i, yDEF,i) and (xTAR,i, yTAR,i)
are the ith data point on the deformed and target
morphing boundaries.
Morphing Aircraft Leading Edge
Most aircraft wings are optimized to produce mini-
mum drag under a particular flying speed, at which the
largest proportion of fuel is expended. However, in
reality, flying speed varies continuously throughout the
flight. Hence, to obtain optimal fuel efficiency, the wing
shape should be able to change in response to the
change in flying speed (Lu and Kota, 2002). The shape
morphing of a hypothetical airfoil leading edge is
investigated here to compare the performance of the
binary ground structure and load path approaches.
Figures 11 and 12 are the results obtained from the
binary ground structure and load path approaches,
respectively. The structural topologies are shown in
314 K.-J. LU AND S. KOTA

Table 5. The LSE deviation value and computation time

from ten trials of both approaches.
Aircraft Leading Binary Ground Load Path
Edge Example Structure Approach Approach

LSE dev. of best design 8.9193 mm 3.7189 mm

Average LSE dev. 9.4540 mm 6.7385 mm
Average CPU time 533 s (8.88 min) 274 s (4.57 min)

solid lines, while the dark dash lines show the target
shape, and the light dash lines represent the actual shape
(deformed curve) achieved due to input actuation. Both
solutions are obtained using the same number of
population (150), number of generation (50), crossover Figure 13. The optimal solution for antenna reflector beam steering
probability (0.8), and mutation probability (0.5). obtained from the binary ground structure approach.
In addition, the output points along the morphing
boundary are determined in a preprocessor (Lu and
Kota, 2003). The overall dimension is 260 mm (10.24 in.)
by 230 mm (9.06 in.) by 20 mm (0.79 in.) (out-of-plane),
and the material is aluminum.
Owing to the heuristic nature, GA can provide
a different result for the same problem in each run.
The designs shown in Figures 11 and 12 are the best
solutions from ten trial runs of each approach, with LSE
deviations of 8.92 mm (0.35 in.) and 3.72 mm (0.15 in.),
respectively. The average computation time and LSE
deviation for the ten trials are shown in Table 5. As can
be seen, the binary ground structure approach requires
almost twice the computation time of the load path
approach. This may result from the larger number of
design variables used in the binary ground structure Figure 14. The optimal solution for antenna reflector beam steering
approach. Moreover, the verifying algorithm for struc- obtained from the load path approach.
tural connectivity may also lead to excessive computa-
tion time. Since the mesh is fixed in this approach, the
optimal solution is always a subset of all possible Flexible Antenna Reflector
designs embedded in the initial discretization mesh.
However, the true optimal solution might not be Recent studies (Washington, 1996; Yoon and
included in the initial mesh. Therefore, the selection of Washington, 1998; Martin et al., 2000; Angelino and
the initial mesh is critical to the quality of the final Washington, 2001) have shown that antenna reflector
solution. Note that there are several ‘trivial’ elements adaptation can potentially enhance system performance
that have one ‘free end’ as shown in Figure 11. These and increase flexibility, such as changing the signal pat-
elements have no strain/stress in them (they undergo tern or coverage area. In this example, a flexible antenna
rigid body motion), so they can be removed without reflector changes its shape to direct the radiation signal to
affecting the compliant mechanism performance. a different direction. As shown in Figures 13 and 14, the
The load path approach, on the other hand, can two tips of the cylindrical reflector move in opposite
generate various structural topologies, because the directions to redirect the signal to the right. Figures 13
locations of the connection ports are part of the design and 14 are the solutions obtained from the binary ground
variables. The use of load path representation also structure approach and load path approach, respectively
eliminates the need of an additional verifying algorithm with the corresponding LSE deviation values of 0.53
for connectivity. Therefore, the computation time is (0.02 in.) and 0.51 mm (0.02 in.). The overall dimension
reduced and the desired shape morphing can be achieved of the reflector is approximately 200 mm (7.87 in.) by
with smaller deviation. As seen in Figure 12, all ele- 40 mm (1.57 in.) by 4 mm (0.16 in.) (out-of-plane) and the
ments are connected at both ends because the topology material is ABS (Acrylonitrile Butadiene Styrene) plastic.
is now represented in terms of the load path, thus there As can be seen, both designs are able to achieve the
are no more trivial loose ends. desired shape morphing of less than 2% of the shorter
 Synthesis of Compliant Adaptive Structures
Figure 15. Natural sitting spinal model in an ideal driver’s seat
(Harrison et al., 2000).
overall dimensions. Note that both designs were obtained
from only one trial run of each algorithm, while multiple
trials are sometimes necessary for more complicated
shape morphing. This suggests that the binary ground
structure and load path approaches are equally effective
in finding a design that can achieve the desired shape
morphing when the problem is less complex. As will be
described later, the complexity is typically related to
the number of inflection points in the problem. Since
the shape change required in the reflector is similar to
cantilever beam bending without generating any inflec-
tion point, the problem is thus considered ‘simple’
shape change.
Shape Morphing Lumbar Support
Lower back pain occurs frequently and is one of the
most costly health problems affecting industry and
society. Lifetime prevalences of 60–90% have been
reported (Andersson, 1991). Lumbar support is one of
the commonly used preventive strategies (Lahad et al.,
1994). This example is inspired by the lumbar support
system that is commonly used in car seats and office
chairs to prevent lower back pain. The downward
‘seating’ motion from a person is used as an input to
actuate the lumbar support that changes the initially
straight back support shape into a curved profile. The
curved profile should match the natural profile of
human spine, which typically includes an inflection
point as shown in Figure 15.
To study the performance of the binary ground
structure and load path approaches, we created a
curve to roughly approximate the spinal shape. The
initial (straight) and target (spine) curves are shown
315
Figure 16. The optimal lumbar support design obtained from the
binary ground structure approach.
Figure 17. The optimal lumbar support design obtained from the
load path approach. Note that relative motions are allowed between
overlapping paths (physical element connections are denoted by *).
in Figures 16 and 17, utilizing the downward ‘seating’
motion as the input. Figures 16 and 17 are the best
designs obtained from 10 trial runs of the binary ground
structure approach and load path approach each.
Asterisk symbols are placed beside physical element
connections to distinguish them from the visual inter-
sections created by overlapping elements. The corre-
sponding LSE deviation values are 11.24 (0.44 in.) and
10.55 mm (0.42 in.), respectively. The overall dimension
is 200 mm (7.87 in.) by 500 mm (19.69 in.) by 5 mm
(0.2 in.) (out-of-plane) and the material is ABS plastic.
The average LSE deviation and required computation
time are listed in Table 6.
As can be seen from Table 6, the binary ground struc-
ture approach requires almost twice as much computa-
tion time as the load path approach. Although the LSE
deviation values are close in Figures 16 and 17, the
316 K.-J. LU AND S. KOTA
Table 6. The LSE deviation value and computation time
from ten trials of both approaches.
Lumbar Support Binary Ground Load Path
Example Structure Approach Approach
LSE dev. of best design 11.2738 mm 10.5463 mm
Average LSE dev. 12.3716 mm 10.6284 mm
Average CPU time 460 s (7.67 min) 243 s (4.05 min)
shape morphing does look different (visually). This may
result from the insufficient numbers of data points along
the initial and target curves. There are currently 11 data
points along the curve, and the Euclidian distance
between the deformed and target locations of each point
are used to measure the LSE deviation in Equation (1).
Additional data points along the morphing boundary
can potentially lead to more significant difference
between the LSE deviations obtained from the two
approaches.
Simple cantilever beam bending can be seen as the
whole beam ‘pivoting’ about the fixed end (Howell and
Midha, 1995). However, creating an inflection point
requires changing the center of curvature from one
side to the opposite side of the beam. Since it requires
a moment or some opposite (push/pull) motions to
generate a couple at the inflection point, shape morph-
ing involving creating an inflection point, such as this
lumbar support example, is considered more compli-
cated than simple cantilever beam bending (antenna
reflector example). It is observed that, from the 10 trial
runs using binary ground structure approach, the
resulting designs can typically achieve the shape
morphing on the top portion (flatter side) in Figure 16,
but most of them fail to match the portion below the
inflection point. In other words, it is relatively straight-
forward for the algorithm to find a solution without an
inflection point, but generating an inflection point can
be quite difficult. On the other hand, the load path
approach appears to capture the inflection point better.
As can be seen in Figure 17, the overlapping paths from
the input and fixed point to the opposite sides of the
inflection point seem to provide a push/pull motion,
leading to a moment about the inflection point. This
overlapping ‘X ’ configuration can also be vaguely seen
in the result in Figure 16. However, if this X configura-
tion is not a subset of the initial discretization mesh, the
binary ground structure approach may be difficult to
find a topology that can produce an inflection point on
the morphing boundary. When selecting an initial mesh
for Figure 16, we intentionally included the X config-
uration in the initial mesh. The decision was based on
our understanding of the problem and some results
obtained from the load path approach, but this kind of
information is generally unavailable beforehand.
Therefore, we believe the load path approach is a
better means to systematically design shape morphing
compliant mechanisms without the need of intuition
or prior experience to select an initial discretization
mesh.
CONCLUSIONS
In this paper, we introduced the load path representa-
tion method and its incorporation into the GA-based
synthesis approach to design shape morphing compliant
mechanisms. Several examples are included to study the
performance of the load path approach and the binary
ground structure approach. The results showed that the
load path approach requires almost half of the com-
putation time required by the binary ground structure
approach, due to the absence of additional verifying
algorithm for structural connectivity. Generally, both
approaches are able to achieve simple shape morphing
when no inflection points are involved. However, even
for more complicated problems with inflection points,
load path approach consistently yields solutions that
successfully achieve the desired shape change. More
importantly, unlike the binary ground structure
approach, the load path approach does not have an
initial discretization network, which typically requires
intuition or prior experience when determining the
complexity and configuration of the mesh. Therefore,
the load path approach can potentially lead to a fully
automated synthesis approach for shape morphing.
Although the load path representation was developed
for synthesis of shape morphing compliant mechanisms
in this context, it is not limited to multiple output prob-
lems. For structural optimization or single output com-
pliant mechanism synthesis, the load path representation
can be used by simply replacing the curve comparison
objective function with other objective functions, such
as minimizing strain energy or maximizing geometric
advantages. In other words, the load path representation
is a general method to parameterize the design domain
into appropriate design variables to simultaneously
optimize the compliant mechanism topology and dimen-
sions. Future and on-going research include the study
of boundary conditions since the locations of the input
actuator and ground supports also play significant roles
in the structural topology. The load path representation
is also being applied to structural optimization and other
compliant mechanism design problems to understand
the generality of this approach.
ACKNOWLEDGMENTS
Authors gratefully acknowledge the funding support
of U.S. Air Force Office of Scientific Research for
this work under the research contract number F49620-
96-1-0205.
 Synthesis of Compliant Adaptive Structures
REFERENCES
Ananthasuresh, G.K., Kota, S. and Kikuchi, K. 1994. ‘‘Strategies
for Systematic Synthesis of Compliant MEMS,’’ International
Mechanical Engineering Congress and Exposition, 55:677–686.
Andersson, G. 1991. The Epidemiology of Spinal Disorders,
pp. 107–146, The Adult Spine: Principles and Practice, New York,
Raven Press.
Angelino, M. and Washington, G. 2001. ‘‘Point Actuated Aperture
Antenna Development,’’ Proceedings of SPIE, 4334:147–155.
Frecker, M.I., Ananthasuresh, G.K., Nishiwaki, S., Kikuchi, N. and
Kota, S. 1997. ‘‘Topological Synthesis of Compliant Mechanisms
Using Multi-Criteria Optimization,’’ ASME Journal of Mechanical
Design, 119(2):238–245.
Goldberg, D. 1989. Genetic Algorithms in Search, Optimization, and
Machine Learning, Addison-Wesley, Boston, MA.
Harrison, D.D., Harrison, S.O., Croft, A.C., Harrison, D.E. and
Troyanovich, S.J. 2000. ‘‘Sitting Biomechanics, Part Ii: Optimal
Car Driver’s Seat and Optimal Driver’s Spinal Model,’’ Journal of
Manipulative and Physiological Therapeutics, 23(1): 37–47.
Hetrick, J. and Kota, S. 1999. ‘‘An Energy Formulation for Parametric
Size and Shape Optimization of Compliant Mechanisms,’’ ASME
Journal of Mechanical Design, 121:229–234.
Holland, J.H. 1975. Adaptation in Natural and Artificial Systems, The
University of Michigan Press, Ann Arbor, MI.
Howell, L.L. and Midha, A. 1995. ‘‘Parametric Deflection
Approximations for End-Loaded, Large-Deflection Beams in
Compliant Mechanisms,’’ ASME Journal of Mechanical Design,
117(1):156–165.
Jones, D.R., Perttunen, C.D. and Stuckman, B.E. 1993. ‘‘Lipschitzian
Optimization without the Lipschitz Constant,’’ Journal of
Optimization Theory and Applications, 79(1):157–181.
317
Joo, J., Kota, S. and Kikuchi, N. 2000. ‘‘Topological Synthesis of
Compliant Mechanisms Using Linear Beam Elements,’’ Mechanics
Based Design of Structures and Machines, 28(4):245–280.
Lahad, A., Malter, A.D., Berg, A.O. and Deyo, R.A. 1994. ‘‘The
Effectiveness of Four Interventions for the Prevention of Low
Back Pain,’’ Journal of the American Medical Association, 272(16):
1286–1291.
Larsen, U.D., Sigmund, O. and Bouwstra, S. 1997. ‘‘Design and
Fabrication of Compliant Micromechanisms and Structures with
Negative Poisson’s Ratio,’’ Journal of Microelectromechanical
Systems, 6(2):99–106.
Lu, K.J. and Kota, S. 2002. ‘‘Compliant Mechanism Synthesis for
Shape-Change Applications: Preliminary Results,’’ SPIE Modeling,
Signal Processing, and Control Conference, 4693:161–172.
Lu, K.J. and Kota, S. 2003. ‘‘Design of Compliant Mechanisms for
Morphing Structural Shapes,’’ Journal of Intelligent Material
Systems and Structures, 14(6):379–391.
Martin, J.W., Redmond, J.M., Barney, P.S., Henson, T.D., Wehlburg,
J.C. and Main, J.A. 2000. ‘‘Distributed Sensing and Shape Control
of Piezoelectric Bimorph Mirrors,’’ Journal of Intelligent Material
Systems and Structures, 11:744–757.
Saggere, L. and Kota, S. 1999. ‘‘Static Shape Control of Smart
Structures Using Compliant Mechanisms,’’ AIAA Journal,
37(5):572–578.
Tai, K. and Chee, T.H. 2000. ‘‘Design of Structures and Compliant
Mechanisms by Evolutionary Optimization of Morphological
Representations of Topology,’’ ASME Journal of Mechanical
Design, 122:560–566.
Washington, G. 1996. ‘‘Smart Aperture Antennas,’’ Journal of Smart
Materials and Structures, 5(6):801–805.
Yoon, H.S. and Washington, G. 1998. ‘‘Piezoceramic Actuated
Aperture Antennae,’’ Journal of Smart Materials and Structures,
7(4):537–542.