See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228907215

Design of Compliant Mechanisms for Morphing Structural Shapes

Article  in  Journal of Intelligent Material Systems and Structures · June 2003

DOI: 10.1177/1045389X03035563

CITATIONS READS
160 1,276

2 authors:

Kerr-Jia Lu Sridhar Kota

16 PUBLICATIONS   716 CITATIONS   
University of Michigan
142 PUBLICATIONS   4,728 CITATIONS   
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Sridhar Kota on 20 June 2014.

The user has requested enhancement of the downloaded file.

 Design of Compliant Mechanisms for Morphing
Structural Shapes
KERR-JIA LU* AND SRIDHAR KOTA
Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109-2125, USA

ABSTRACT: Various compliant mechanism synthesis methods have been developed over the
past decade; however, very little attention has been directed towards adaptive shape change
problems. In this paper, we present a systematic method for synthesizing compliant
mechanisms to morph a given curve or profile into a target curve utilizing minimum number
of actuators (typically one). Two objective functions are formulated, using Least Square
Errors and a modified Fourier Transformation, to capture the shape differences. The topology
and dimensions of the optimal compliant mechanism are generated using Genetic Algorithms.
Applications of this synthesis approach are demonstrated through two adaptive antenna
design examples.

Key Words: compliant mechanism, shape morphing, genetic algorithm, adaptive structure,
structural optimization

INTRODUCTION of MEMS motion amplifiers and compliant grippers.

Previous research on compliant mechanism synthesis
compliant mechanism is a single-piece flexible has typically employed a two-step synthesis approach as
A structure that delivers the desired motion by
undergoing elastic deformation as opposed to the rigid
shown in Figure 1 (Joo, 2001): (a) topology synthesis;
then (b) size and geometry optimization. The topology
body motions in a conventional mechanism (Kota et al., and dimensional aspects of an elastic body is illustrated
1999). It is designed to accomplish with a one-piece in Figures 2 and 3; given a one-piece elastic body, the
device what conventional mechanisms can do with topology is defined by the number of interior holes and
multiple pieces. It is designed to be flexible enough to their locations with respect to the boundary conditions
transmit motions, yet stiff enough to withstand the (loading and support locations), while the dimensions of
external loads. The hingeless nature of compliant the remaining material is determined by the size and
mechanism eliminates the backlash error and effectively shapes of these holes. The two-step approach decom-
reduces the production and maintenance costs associated poses the interrelated topology and dimensional syn-
with the multiple piece assembly. Moreover, the dis- theses into two separate stages: the topology synthesis
tributed compliance throughout the compliant mechan- ensures the motion in the desired output direction (a
ism provides a smooth deformation field, which reduces quality aspect); the size and geometry optimization
the stress concentration. These features are particularly refines the mechanism dimensions to achieve a desired
attractive when dealing with applications such as shape objective, such as maximizing displacement (a quantity
change in aircraft wings and antenna reflectors. These aspect). Research has suggested that the decomposition
systems generally require a reduced system weight, can simplify the problem and yield successful results in
accurate and smooth surface shapes, and simple control real single-input single-output (SISO) applications
schemes that compliant mechanisms can offer. (Ananthasuresh et al., 1994; Frecker, 1997; Hetrick
Various systematic strategies for compliant mechan- and Kota, 1999; Joo et al., 2001).
isms, using a structural optimization approach, have Although various methods for synthesizing compliant
been developed in the past decade to fulfill various mechanisms have been developed in the past decade,
functional requirements, such as maximizing energy very little attention (Saggere and Kota, 1999) has been
efficiency (Hetrick and Kota, 1999) or maximizing directed to problems related to shape change. Unlike the
geometric advantage (Joo et al., 2001). These synthesis SISO problems studied in most of the previous research,
approaches have been successfully applied to the design the quality of the solutions to the shape change
problems depends greatly on the precise deformation
(direction and magnitude) of all or several discrete
*Author to whom correspondence should be addressed.
E-mail: kjlu@umich.edu. output points along the shape-changing boundary. With

JOURNAL OF INTELLIGENT MATERIAL SYSTEMS AND STRUCTURES, Vol. 14—June 2003 379
1045-389X/03/06 0379–13 $10.00/0 DOI: 10.1177/104538903035563
ß 2003 Sage Publications
380 K.-J. LU AND S. KOTA

Figure 1. Typical compliant mechanism synthesis involves a two-step approach, where the topology is optimized first, and with the optimized
topology as an input, the optimal dimensions for the compliant mechanism are determined in the second step (Joo, 2001): (a) topology
synthesis; (b) dimensional synthesis.

variables. Therefore, the objective of this research is to

develop a generalized scheme to synthesize compliant
mechanisms that can morph one shape to another
desired shape with minimum error. In addition, discrete
and continuous variables should be used simultaneously
in this unified synthesis approach to determine the
topology and dimensions of the compliant mechanism.

Figure 2. Two elastic bodies with different topologies.

SHAPE CHANGE USING COMPLIANT
MECHANISMS

The performance of many mechanical–structural

systems, such as aircraft wings and antenna reflectors,
is directly related to the geometric shapes of their
components. Such systems require different shapes for
different operating conditions, but they are generally
designed to have one fixed shape that constitutes a
compromise with respect to all the operating conditions.
Figure 3. Two elastic bodies with different topology and dimensions. To respond to varying operating conditions and external
disturbances, the component shape has to change
adaptively to maintain optimal system performance
multiple output points presented in these problems, it and enhance versatility. Various adaptive shape change
might be inappropriate to use the two-step approach (shape morphing) systems have been developed and
that focuses on the performance of only one particular incorporated into practical applications (Austin and
output point. Since the deformation of the shape- Van Nostrand, 1995; Washington, 1996; Austin et al.,
changing boundary is simultaneously influenced by the 1997; Martin et al., 1997, 1998, 2000; Yoon and
topology and dimensions of the compliant mechanism, Washington, 1998; Webb et al., 1999; Yoon et al.,
it is critical to unify the two synthesis steps and address 2000; Ameduri et al., 2001; Angelino and Washington,
them simultaneously. Furthermore, previous research 2001). Most of these shape morphing systems involve
on topology synthesis has employed continuous optimi- the use of smart actuators and materials, such as shape
zation methods to find the optimal material distribution memory alloys (SMAs) and lead zirconate titanate
by eliminating unnecessary elements from the initial (PZT). Although the smart actuators can provide
element network (Figure 1(a)-2). However, the elimina- light-weight actuation schemes, their scalability is
tion or preservation of elements is, in fact, a discrete uncertain when realistic scale problems are considered.
decision, and it should be addressed using discrete A compliant mechanism, on the other hand, changes
 Design of Compliant Mechanisms 381

its shape through the structural deformation, which is such as aircraft wings, antenna reflectors, lumbar
independent of the scale of the problem, thus offering support, or fluid flow control devices. Two antenna
an alternative means to achieve a desired shape change. shape change examples will be shown later to illustrate
In this research, we will investigate the feasibility of the synthesis approach developed in this research.
this alternative shape change method and explore its
potential benefits.
Figure 4 gives a simple illustration of how a compliant METHODOLOGY
mechanism changes its shape. The actuator provides a
displacement or force input to the system, and the This paper presents a systematic synthesis approach
compliant mechanism deforms due to the structural for compliant mechanisms that can achieve the desired
flexibility, which, in turn, changes the boundary from its shape change from an initial curve profile to a target one.
initial state (initial curve profile) into a deformed state The problem specifications include the initial and target
(deformed curve profile). In this research, it is assumed curve profiles, support locations, external loads, actuator
that the compliant mechanism is composed of frame-like type and location, available design domain, and material
elements, where the structural deformation comes properties. As can be seen from the flowchart in Figure 5,
mainly from bending of the beams. The goal is to find the synthesis approach starts with a feasibility check to
the optimal topology and dimensions for the compliant ensure the shape change is attainable. The next step is to
mechanism, so that the deformed curve profile, due to
input actuation, matches the target curve with minimum
error. Although Figure 4 only illustrates the shape
morphing between two curves using a two-dimensional
planar compliant mechanism, it can be regarded as one
cross-section of a shape morphing surface, such as an
aircraft wing cross-section along the wingspan. More
complicated three-dimensional surface shape change can
be explored in the future by expanding the synthesis
method developed in this research, but the scope of this
paper will be restricted to the design of two-dimensional
compliant mechanisms. Several additional assumptions
also include: (1) the shape-changing object will change
from its initial profile to only one target profile; (2) the
initial and target profiles are specified a priori; (3)
the shape-changing object is integrally attached to the
compliant mechanism; and (4) the compliant mechanism
has only a single external input actuator at a specified
location. The actuator can be selected from a variety of
ranges and types, such as electric motor or even the
smart actuators, as long as the required motion and
force can be provided. The shape change concept can be Figure 5. Flowchart for the compliant mechanism synthesis
applied to many engineering fields in various scales, approach.

Figure 4. An illustration for a shape change compliant mechanism: (a) shows the problem specifications and how the input actuator is
connected to the active points; (b) shows the deformed structure with its deformed curve boundary and how that compares to the target curve.
382 K.-J. LU AND S. KOTA

discretize the design domain into an initial grid, and the where
INI ðl Þ and
TAR ðl Þ are the curvature functions of
grid nodes are connected with beam elements to form an the initial and target curves; l is the normalized arc
initial finite element mesh. This initial mesh then goes length; E is the Young’s Modulus; h is the beam height
through an optimization process that searches for the (in-plane dimension).
optimal topology and dimensions of the compliant
mechanism. In this optimization process, the objective
is to minimize the error between the deformed and target Problem Setup – Determine Optimal Active Points
curve. A finite element analysis (FEA) is used to solve for
the deformed boundary shape, which is, then, compared Once the problem passes the feasibility check, the
to the target curve profile to define the shape change design domain is parameterized so that the designs can
error. Due to the mixed variable types from the topology be described in terms of the design variables that will be
(discrete) and dimensions (continuous), a genetic algori- incorporated into the optimization process in the later
thm (GA) is adopted to simultaneously determine the stage of the synthesis process. Previous research has
optimal topology and dimensions of the compliant been using an initial finite element mesh to discretize the
mechanism. The steps in this synthesis procedure are design domain either with two-dimensional quadri-
described in detail in the following sub-sections. lateral elements or with a grounded structure consists
of beam/truss element network (Figure 1(a)-2). The
Problem Setup – Check Shape Change Feasibility design variables are then defined as the dimensions
of the elements, such as the cross-section areas. The
A preprocessor first examines the given initial and discretization mesh size is typically pre-specified by the
target curves in order to evaluate if the shape change designer without any clear guideline on the size
is attainable. The initial curve is assumed to be stress selection. Although finer mesh can generate results
free before the shape change and to have a constant with higher resolution, it also increases the number of
rectangular cross-section. The nominal dimensions of the design variables required to represent a design. In this
cross-section are the minimum feature sizes from the research, we want to gain some insight into the required
manufacturing constraints. The feasibility check then complexity of the initial discretization element network
estimates the maximum stress along the curve when it is by studying the curve profiles before and after the shape
deformed into the target shape. The maximum stress is change.
checked against yielding with two criteria: (1) the stress It is assumed that the shape change boundary is
due to axial tension or compression must be smaller than connected to the interior elements of the compliant
yielding stress; and (2) the stress due to bending must mechanism through several ‘active points,’ such as P1,
be smaller than yielding stress. For the first criterion, P2, and Pn in Figure 4. The compliant mechanism can
the preprocessor calculates the length of both curves then transmit the input actuation motion through these
(initial and target), and estimates the axial stress that is active points to control the motion/shape of the
required to stretch or compress the curve from its initial morphing boundary (the adaptive surface). Hence, the
length to the target one. For the second criterion, the active points can also be regarded as the ‘output points’
preprocessor first calculates the curvature functions of of the (interior) compliant mechanism. The shape
both curves,
INI ðlÞ and
TAR ðlÞ, where l is the normalized change can be achieved with minimum error if we
arc length varying from 0 to 1. The curvature difference have full control of every point along the curve, but
function, d
ðlÞ, is then defined in Equation (1) as the designing a compliant mechanism with so many output
difference between the two curvature functions. This can points (active points) might involve a tremendous
be considered as the local curvature change along the number of design variables. It is, thus, important to
curve. Since the bending stress along a beam is identify the minimum number of required active points,
proportional to its local curvature change, according to such that, when they are connected to the interior
Euler–Bernoulli beam theory, the maximum bending elements, the whole compliant mechanism can achieve
stress can be calculated from the curvature difference the desired shape change with acceptable error.
function as shown in Equation (2). If both of the To determine the active point locations, a piecewise
maximum stresses due to axial tension and bending stay linear function is used to approximate the curvature
within the yielding limit of the selected material, the difference function, d
ðlÞ, and the active points are
shape change is considered feasible, and the algorithm defined as the end points of each linear section. The
invokes the next step in the synthesis procedure. curve fitting function is piecewise linear because the
curvature distribution is linear for a straight beam with
d
ðlÞ ¼
INI ðlÞ 
TAR ðl Þ ð1Þ
constant cross-section, subjected to uniform pressure
loads or end loads. The morphing boundary can
Eh be regarded as several beam segments connected in
max ¼ maxðd
ðl ÞÞ ð2Þ
2 series with the interior compliant mechanism applying
 Design of Compliant Mechanisms 383

(transmitting) forces/moments to the end points of each structure approach seen in previous research (Figure
segment (active points). An optimization process is then 1(a)-2), but instead of discretizing the design domain
used to determine the minimum number of active points with arbitrary grid size, the grid size in one direction
while constraining the curve fitting error below an is now determined by the active points. The grid size in
acceptable tolerance, ". They are shown in Equations (3) the other direction and the configuration of the initial
and (4) as the objective function and the constraint. The network are selected based on the problem specification
binary design variable, activePti, represents the ith given and engineering experience to form a variety of available
data point along the shape change boundary; the ith data configurations. It is shown later that the selection of
point becomes an active point only if activePti ¼ 1. A initial element network is critical to the resulting optimal
GA (Goldberg, 1989) is employed to find the optimal solution.
number of design variables, because the binary variables After creating the initial element network, two design
can be easily incorporated in the GA. The optimal variables are assigned to each beam element to describe
locations of the active points can be obtained from the the compliant mechanism topology and the dimensions
optimal activePt vector; the j th data point on the shape of each beam segments. In this research, each beam
change boundary is considered an optimal active point element connecting any two grid nodes is assumed to
location if activePtj ¼ 1. More details on GA will be have a constant rectangular cross-section. The in-plane
described later in this paper. beam dimensions (beam heights) are considered the
dimensional design variables, while the out-of-plane
!
X
n dimensions (beam widths) are prescribed to be constant
min activePti Minimize number for all elements. As shown in Equations (5) and (6), for
activePti ð3Þ
i¼1 the ith beam element, it is assigned a binary variable
of active points (hTopi) and a continuous real variable (hDimi) to
represent the topology and dimension respectively. The
Subject to final beam heights, hi, are simply the multiplication of
the two variables, expressed in Equation (7). It is also
  assumed that the shape change boundary always ‘exists’
g1 : d
ðli Þ  d
 ðli Þ  " Curve fitting
ð4Þ and has a constant in-plane dimension, hBoundary.
error constraint
Therefore, for an initial discretization mesh with n
elements, there will be a total of 2n þ 1 design variables,
where n is the number of given data points along the including n binary topology variables, n continuous
shape change boundary; activePti 2 f0, 1g; d
ðlÞ is the dimensional variables, and one continuous variable for
curvature difference function; d
 ðl Þ is the piecewise the in-plane dimension of the shape change boundary.
linear function connecting ðlj , d
ðlj ÞÞ 2 f j j activePtj ¼ 1g; An optimization process is then incorporated to find the
and " is the acceptable fitting error. optimal values for these design variables, so that the
desired shape change can be achieved with minimum
Problem Setup – Discretize Design Domain and error.
Define Design Variables
hTopi 2 binary values ð0: element eliminated,
After obtaining the locations of the active points, the
1: element remainsÞ ð5Þ
preprocessor proceeds to discretize the design domain
and define the topology and dimensional design vari-
hDimi 2 positive real numbers ð6Þ
ables for the compliant mechanism optimization in the
next step. As shown in Figure 4, the design domain is
hi ¼ hTopi  hDimi ,
defined by the initial curve and the input actuator ð7Þ
location, considering space constraints and boundary where i ¼ 1, 2, . . . , number of elements
condition specifications as well. Based on the active
points and boundary conditions, the design domain is
then discretized into an initial grid, where grid nodes are Optimization Procedure – Optimization Method
connected with beam elements to form an initial
discretization network. With a given initial grid, the An optimization problem is formulated at this point
initial element network provides many possible topolo- in order to find the optimal compliant mechanism that
gies for the final solution simply by removing or changes the initial curve into a target shape with
including different elements within this network. An minimum error. In this work, we use a GA to search
optimization procedure can then be used to determine for the optimal design, because it can search for a wider
the best topology among them in the later stage of this range of solution space (compared to local search), and
synthesis approach. This is similar to the grounded it is capable of handling mixed design variables
384 K.-J. LU AND S. KOTA

(discrete þ continuous). It also has a simple binary the chromosome code, including the topology and
encoding data structure that enables the binary topology dimensional variables. The mutation for each new
variables to be incorporated in GA naturally. design, following the crossover, is simply defined by
The GA is a heuristic optimization method that replacing an arbitrary value (any hTop, hDim, or
simulates the selection scheme found in nature, based on hBoundary) in the chromosome string with a randomly
the principle of ‘survival of the fittest.’ The design generated value.
variables are encoded into a chromosome string
analogous to the genes on the DNA. It starts with a 2
randomly generated population of individuals (designs),
6
and subsequent new generations are created through a designðkÞ ¼ 4 hk Top1 , hk Top2 , hk Topn ,
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
reproduction process. The chance of survival for one Topology
individual depends on its fitness value evaluated from 3
the objective function defined in the optimization
hk Dim1 , hk Dimn , hk Boundary5
problem. The individual with higher fitness value has a |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Dimensions
higher chance to reproduce, but inferior ones can still
survive with a lower probability to preserve the variety ð8Þ
in each generation. Reproduction repeats until reaching
the maximum number of generations, and the fittest where k ¼ 1,2, . . . , number of individuals in one genera-
design in the final generation is then considered the tion; n ¼ total number of beam elements; hTopi and
‘optimal’ solution. GA allows solutions to jump from hDimi are the topology and dimensional values for the
one local region to another in the solution space through ith beam element.
the genetic operations, thus preventing the solution
from being trapped in a local optimum. This also ½h1 Top1 , h1 Topn , h1 Dim1 , h1 Dimk ,
p1 ¼
implies that the solution obtained from GA is indepen- jh1 Dimkþ1 , h1 Dimn , h1 Boundary
dent of starting point, as opposed to that obtained ð9Þ
from gradient-based continuous optimization methods.
Although global optimum is still not guaranteed, GA ½h2 Top1 , h2 Topn , h2 Dim1 , h2 Dimk ,
p2 ¼
gives a wider range of possible optimal solutions and jh2 Dimkþ1 , h2 Dimn , h2 Boundary
is ‘more likely’ to find the global optimum. Details ð10Þ
regarding GA can be found in the reference (Goldberg,
1989). In this paper, we will only highlight some ½h1 Top1 , h1 Topn , h1 Dim1 , h1 Dimk ,
important issues in GA, such as the encoding of k1 ¼
h2 Dimkþ1 , h2 Dimn , h2 Boundary
chromosome string and the reproduction operation.
To utilize GA in this research, the topology and ð11Þ
dimensional design variables are encoded into a
chromosome string shown in Equation (8). The GA ½h2 Top1 , h2 Topn , h2 Dim1 , h2 Dimk ,
randomly creates a population of designs in the first k2 ¼
h1 Dimkþ1 , h1 Dimn , h1 Boundary 
generation, and successive generations are generated
ð12Þ
through the reproduction scheme, including selection,
crossover, and mutation. The selection scheme is based
on the fitness value of each design, which is the
‘deviation’ or ‘shape difference’ between the deformed Optimization Procedure – Objective Function for Curve
and target curves. The deviation evaluation will be Comparison
described in more detail in the next section when
defining the objective function. As shown in Equations To evaluate the performance of a design, the
(9)–(12), using a roulette wheel selection scheme, two associated fitness value is evaluated using a curve
parent chromosomes ( p1 and p2) are selected from the comparison scheme that captures the shape differences
parent generation, and two new chromosomes (k1 and between the deformed and target curves. For each
k2) are then created through genetic operations on p1 compliant mechanism defined by a chromosome string,
and p2. The crossover operation is done through a FEA is used to solve for the structural deformation.
exchanging part of the chromosome strings of the The deformed curve profile is then extracted from the
parents; the vertical lines in Equations (9) and (10) deformed structural boundary and compared to the
indicate the point of crossover, while the underline desired target curve. The deformed and target curves are
portions in Equations (11) and (12) show the segments expressed in terms of two sets of sampling points that
of chromosome strings coming from the other parent. are evenly distributed along the curve lengths, as shown
Note that the crossover point can be anywhere within in Figure 6.
 Design of Compliant Mechanisms 385

Figure 7. (a) A shape change example: bending a symmetric curve

Figure 6. Sampling points on the deformed and target curves. (about y-axis) towards the right; (b) an example solution (not the
actual solution) to this problem; (c) a mirror image solution which
could be discarded during the optimization process using LSE.
Previous research (Saggere and Kota, 1999) defined
the ‘deviation’ between the deformed and target curves
as the average Euclidian distance between the two
curves at the sampling points. This is termed as the
least square error (LSE) deviation and shown in
Equation (13).

deviationLSE ¼
n qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X ð13Þ
ðxDEF, i  xTAR, i Þ2 þ ðyDEF, i  yTAR, i Þ2
n i¼1 Figure 8. (a) A shape change example that changes a circle into an
ellipse; (b) a design with deformed curve that matches the target
curve exactly (shape and locations); (c) a design that can achieve
Since LSE deviation compares the differences between shape change from a circle to an ellipse at an angle.
all sampling points along the curve, the scale, orienta-
tion, and shape information are captured and can
potentially lead to a solution that matches the target
curve exactly. But this might discard a portion of
possible solutions when there is symmetry in the
problem. Figure 7(a) shows an example that changes a
symmetric curve (about y-axis) into a curve that bends
towards the right. Assuming that Figure 7(b) is a
solution to this problem, it can be seen that Figure 7(c)
can also be a valid solution simply by mirroring the
Figure 9. A one-dimensional periodic signal and its frequency
solution about y-axis, but this solution can be discarded spectrum from the Fourier Transformation. Lower frequencies dictate
using LSE deviation. Figure 8 shows another shape the overall shape, while higher ones describe the finer details.
change example where LSE deviation is inappropriate
to describe the ‘shape’ difference. The goal is to deform features. This is especially useful for the curve compar-
a circle into an ellipse, but the ellipse can be in any ison task not only because the information is purely
orientation because the circle is axisymmetric. Using about the shape, but it also provides shape information
LSE deviation will prevent the algorithm from finding of different importance level.
the solution in Figure 8(c). Therefore, for symmetric
problems like these, it is necessary to use a curve Z 1
comparison scheme that focus purely on the differences Fð!Þ ¼ F f f ðtÞg ¼ f ðtÞejwt dt ð14Þ
1
in ‘shape’ instead of the sampling point locations.
For situations emphasizing the shape difference, we
use a modified Fourier Transformation (FT) to char- To utilize FT, the curve has to be represented as a
acterize and compare the curves. Standard Fourier one-dimensional periodic signal. One way to achieve this
Transformation is shown in Equation (14), where f (t) is setting is to express the curve shape in terms of its
a periodic function in time-domain. It transforms the location in Y-direction as a function of its arc length.
periodic function into its frequency content in terms of Since the start and end points of the curve do not
harmonic amplitudes and their corresponding frequen- necessarily have the same value in global Y-direction,
cies, as shown in Figure 9. Lower frequency information the sampling points on the curve have to go through a
dominates the overall signal shape, while higher coordinate transformation from the original global X–Y
frequency information usually contributes to finer coordinates to the x0 –y0 coordinates, where the x0 -axis is
386 K.-J. LU AND S. KOTA

Figure 10. The coordinate transformation used in the modified FT

method, so that the deformed and target curves can have the same y
value (in the y0 -direction) at the starting and end points.

Figure 11. An example of the differences in harmonic amplitudes for

determined by connecting the start and end points of the two curves.
curve, as shown in Figure 10. The curve is then
expressed as a function of the arc length, hence, a
one-dimensional periodic signal. Instead of using FT Equation (21) is also applied to all elements to prevent
directly, the curve in the new coordinates and para- structural failure. The optimization model can, thus, be
meterization is approximated using Discrete Fourier summarized in the following.
Transformation (DFT), because the exact function of
Objective Function
the curve is unknown. The DFT takes discrete sampling
information on the curve and approximates the curve
min ðdeviationLSE Þ or min ðdeviationFT Þ
with discrete frequencies and their corresponding hTopi , hDimi hTopi , hDimi
harmonic amplitudes, that is, the Fourier Descriptors hBoundary hBoundary

(FDs). To further accelerate the calculation speed, Fast ð16Þ

Fourier Transformation (FFT) can be used instead of
DFT by resampling the modified curve (arc length vs. Subject to
y0 -direction value) with 2N points, where N is an integer.
To evaluate the deviation between the deformed and g1 : d ¼ K1 F FEA equilibrium ð17Þ
target curves, the modified FT (coordinate transforma-
tion þ FFT) is used to describe both of the curves g2 : hDimmin < hDimi  hDimmax size constraint
separately, resulting in two sets of FDs. As shown ð18Þ
in Figure 11 and Equation (15), the sum of amplitude
differences at corresponding frequencies is then defined g3 : hDimmin < hBoundaryi  hDimmax size constraint
as the FT deviation between the two curves.
ð19Þ
X
nAmp
g4 : maxðdj Þ  dallow stiffness constraint ð20Þ
deviationFT ¼ jAmpTARk  AmpDEFk j ð15Þ
k¼1
g5 : i  allow stress constraint ð21Þ
th
Where AmpTARk and AmpDEFk are the k harmonic
amplitudes for the target and deformed curves g6 : hi ¼ hTopi  hDimi mixed variables
respectively, and nAmp is number of amplitudes
(number of resampling points). hTopi 2 f0, 1g, hDimi 2 Rþ , i 2 all beam elements, and
As shown in Equation (16), either of the two deviation j 2 degrees of freedom of nodes on the shape change
measures (LSE or FT) can be used as the objective boundary
function in the optimization procedure, depending on
the nature of the problem. The structural deformation is For each chromosome string, the fitness value is
obtained from FEA, as indicated in Equation (17). All evaluated using the objective function, and the best
the beam elements are subjected to size constraints, design in each generation is recorded throughout the
shown in Equations (18) and (19). The stiffness evolution process. The reproduction process repeats
constraint shown in Equation (20) limits the maximum until reaching the maximum number of generations,
deformation on the shape change boundary before hence, terminating the evolution process. The optimal
activating the input actuator. A stress constraint solution will be the best solution across all generations.
 Design of Compliant Mechanisms 387

Table 1. Several test shapes and their corresponding LSE and FT deviation values.
Test1 Test2 Test3 Test4

LSE dev: 2.57 mm LSE dev: 51.97 mm LSE dev: 13.72 mm LSE dev: 11.52 mm
FT dev: 5.01 mm FT dev: 5.01 mm FT dev: 0 mm FT dev: 51.33 mm

VALIDATION OF OBJECTIVE FUNCTION

To examine how the LSE and FT deviations explore

different aspects of shapes, the deviation measures
defined in Equations (13) and (15) are applied to several
test shapes, shown in Table 1. The curve set in Tests 1
and 2 are the same, but the dash curve is translated and
rotated with respect to the solid curve. Thus, FT
deviation has the same values for the two cases (same
shape), but LSE gives a large deviation value in Test 2
due to the location difference. Test 3 shows two curves
that are mirror images of each other, so the FT
deviation is zero while LSE deviation gives a large Figure 12. An antenna reflector that changes from a parabolic
value. Test 4 shows that when two curves have very shape into a circular shape.
different shapes, both LSE and FT have large deviation
values, but FT captures the shape difference better.
listed in Table 2. As shown in Figure 12, displacement
input of 5 mm in the y-direction is located at (0, 30),
DESIGN EXAMPLES
and a fixed point is located at (0, 0). Multiple external
loads of 1 N are applied to the reflector surface. The
Recent studies (Washington, 1996; Martin et al.,
design domain is then defined by connecting the input,
1998, 2000; Yoon and Washington, 1998; Yoon et al.,
2000; Angelino and Washington, 2001) have shown that fixed point, and the initial curve. Using ABS plastic
antenna reflector adaptation can potentially enhance (Acrylonitrile–butadiene–styrene: Young’s Modulus ¼
system performance and increase flexibility, such as 2480 MPa, Yield Stress ¼ 34.45 MPa) as the material,
changing the signal pattern or coverage area. Two the estimated maximum axial stress and bending
examples similar to the ones shown in Washington moment are both below the yielding limit of ABS,
(1996) are presented in this section to illustrate the thus the shape change is feasible. The curvature
synthesis approach described in this paper and demon- functions (
INI ðlÞ and
TAR ðlÞ) and curvature difference
strate the feasibility of shape change using compliant function (d
ðlÞ) are shown in Figure 13. After the active
mechanisms. points are determined using GA, an initial discretization
network are created as in Figure 14 by connecting the
Antenna Reflector – Beam Shaping actuation location to the optimal active points along the
boundary with a beam element mesh. The horizontal
Figure 12 is an example of antenna reflector in its element members are included to provide stiffness or
beam-shaping mode that changes the focus of the alternative paths between input and the active points.
reflector to vary the radiation pattern. In this example, To search for the optimal topology and dimensions
we design a compliant mechanism that is capable of for the compliant mechanism, the LSE deviation
changing the parabolic curve (initial curve) into a objective function is used in this example, because
circular curve (target curve). Due to symmetry about both the initial and target curves are asymmetric in the
the y-axis, only the left half of the reflector is modeled. half-model. The out-of-plane beam dimension is speci-
The data points along the initial and target curves are fied as 4 mm, and the in-plane dimension is allowed to
388 K.-J. LU AND S. KOTA

Table 2. Information of the data points along the initial

and target curves in the beam shaping example.
Data point Initial Curve (x, y) Target Curve (x, y)
No. (mm) (mm)

1 100.0000 10.0000 100.0000 17.1573

2 90.3321 8.1599 90.3321 13.9229
3 80.5598 6.4899 80.5598 11.0188
4 70.6946 4.9977 70.6946 8.4485
5 60.7477 3.6903 60.7477 6.2148
6 50.7306 2.5736 50.7306 4.3204
7 40.6550 1.6528 40.6550 2.7675
8 30.5324 0.9322 30.5324 1.5578
9 20.3745 0.4151 20.3745 0.6927
10 10.1931 0.1039 10.1931 0.1732
11 0.0000 0.0000 0.0000 0.0000 Figure 14. Initial discretization element network and boundary
conditions are shown. Only half of the antenna is modeled due to
symmetry.

Figure 15. The optimized compliant mechanism with the deformed

and target curves.

Figure 13. (a) The curvature function for the initial curve; (b) the
curvature function for the target curve; (c) the curvature difference
function for the beam shaping example shown in Figure 12.

vary between 1.5 and 4 mm. The GA starts with an

initial population of 200 individuals and allows 50
generations, while the crossover and mutation prob-
abilities are 0.8 and 0.4, respectively. The total CPU Figure 16. Two ends of the antenna reflector are required to bend
time for the optimization process is less than 15 min. As upwards and downwards to simulate a change in orientation of 3
clockwise.
can be seen from the result in Figure 15, the optimized
compliant mechanism can indeed achieve the desired
target curves are listed in Table 3. As shown in Figure 16,
shape change with the LSE error of only 0.4689 mm. It is
a displacement input of 1 mm is located at (0, 15), and
also shown that the optimized topology is different from
a fixed point is located at (0, 0). Uniform external loads
the initial element network in Figure 14, which implies
of 1 N downwards are again applied to the reflector
that the topology and dimensions of the compliant
surface. Following the same procedure in the previous
mechanism are determined simultaneously.
example, the initial discretization network created by
connecting the input and active points are shown in
Antenna Reflector – Beam Steering Figure 17.
Since the initial curve is symmetric about y-axis, FT
Figure 16 is an example of antenna reflector in its deviation objective function is used in this example. In
beam-steering mode that changes the orientation of this example, the out-of-plane beam dimension is 4 mm
the reflector to vary the coverage area. In this example, and the in-plane dimension is allowed to vary between
we design a compliant mechanism that is capable 2 and 4 mm. The GA starts with an initial population of
of simulating a rotation of 2.86 (0.05 rad) clockwise 60 individuals and allows 20 generations, while the
about the center. The data points along the initial and crossover and mutation probabilities are 0.8 and 0.5,
 Design of Compliant Mechanisms 389

Table 3. Information of the data points along the initial

and target curves in the beam steering example.
Data point Initial Curve (x, y) Target Curve (x, y)
No. (mm) (mm)

1 70.6946 8.4485 98.9249 22.1573

2 60.7477 6.2148 89.4997 18.1974
3 50.7306 4.3204 79.9410 14.5743
4 40.6550 2.7675 70.2573 11.3074
5 30.5324 1.5578 60.4582 8.4156
Figure 18. The optimized compliant mechanism for the beam
6 20.3745 0.6927 50.5551 5.9183
steering mode.
7 10.1931 0.1732 40.5614 3.8345
8 0.0000 0.0000 30.4916 2.1827
9 10.1931 0.1732 20.3623 0.9814
10 20.3745 0.6927 10.1918 0.2481
11 30.5324 1.5578 0.0000 0.0000
12 40.6550 2.7675 10.1944 0.0983
13 50.7306 4.3204 20.3867 0.4040
14 60.7477 6.2148 30.5732 0.9329
15 70.6946 8.4485 40.7486 1.7005
16 80.5598 11.0188 50.9061 2.7225
17 90.3321 13.9229 61.0372 4.0140
18 100.0000 17.1573 71.1319 5.5896 Figure 19. The mirror image design of the compliant mechanism
19 70.6946 8.4485 81.1786 7.4633 shown in Figure 18.
20 60.7477 6.2148 91.1645 9.6484
21 50.7306 4.3204 101.0751 12.1573
deviation can be further reduced by adding more
subsequent generations.

DISCUSSIONS

Determination of Active Points

In the problem setup step, the number and locations of

active points are determined by fitting a piecewise linear
curve through the curvature difference function, d
ðlÞ,
assuming that all beam elements have uniform cross-
section areas and that only uniform pressure loads or
nodal loads are applied. However, if nonuniform
Figure 17. Initial discretization element network and input actuation.
distributed loads and cross-sections are used, d
ðlÞ will
no longer be piecewise linear; thus, a curve fitting process
should be applied using a piecewise curve with a different
order. For example, if linear varying distributed loads
respectively. The CPU time is less than 10 min. As are presented, the moment along the element would be
shown in Figure 18, an interesting result that seemed to quadratic. Therefore, the active points will then be
rotate in the opposite direction (counterclockwise) was determined by finding a piecewise quadratic curve with
obtained from the GA with the FT deviation of minimum segments that fits d
ðlÞ with minimum error.
1.1758 mm (LSE deviation of 6.0869 mm). This simply When using nonuniform cross-section elements, the
implies that the desired shape change can be achieved same rule still applies, but it would involve a more
by mirroring the current design. The mirror image design complicated calculation.
is shown in Figure 19 with the same FT deviation value,
but the LSE deviation is reduced to 0.6801 mm. This
example suggests another useful aspect of the modified Selection of Initial Discretization Mesh
FT: when the actuation direction specified by the user
does not allow the shape change in the ‘right’ direction, In the problem setup step, the initial discretization
unlike the LSE which might prevent GA from finding a mesh is created by interconnecting the input point, fixed
feasible solution, using the modified FT can help points, active points, and some intermediate grid points
identify the mirror solution. Note that the FT deviation within the design domain with a selected mesh config-
of 1.1758 mm is obtained after 20 generations. The uration. In the later stage of the synthesis process, the
390 K.-J. LU AND S. KOTA

optimization toolbox searches for the optimal topology, conditions and is more likely to explore a larger variety
which is the best connectivity among all possible of solutions.
solution (from the fully connected initial mesh). The
resulting elements can be regarded as paths through Convergence of Genetic Algorithm
which the input motion transfers to the output
boundary (shape morphing boundary). The initial The GA used in this synthesis approach is a heuristic
mesh lays the framework for many possible paths, optimization technique that is independent of starting
while the GA searches for the optimal and necessary point and does not require sensitivity information.
ones within all available paths. However, the grid size However, the computation time can be expensive when
and mesh configuration (how grid points are connected), the number of design variables increases (with smaller
actually controls the size and complexity of the solution grid size and more complex mesh type). If the
space, because selecting the initial mesh establishes the neighborhood of the optimal solution is relatively
scope of all possible combinations of paths. Therefore, small, compared to the entire solution space, it might
the initial mesh configuration is critical to the final encounter difficulties in even getting close into this
result. Although the solution space of a more compli- neighborhood, let alone zeroing on the optimal solution.
cated initial mesh could include that of a simpler one, One way to accelerate the convergence speed is to
the computation time increases dramatically as the employ a local search following the GA, or use a hybrid
solution space gets larger. Thus, it is important to GA that includes some gradient information within the
consider the trade-off between the complexity of the evolution process. The local search or gradient informa-
solution space and the available computation time. The tion can help converging to the local optimum, instead
majority of the previous research has been discretizing of jumping between different regions when using GA
the design domain into arbitrary mesh size. In our work, alone. On the other hand, the crossover and mutation
the mesh is created by, first, connecting the input to the operations in GA can be modified to better preserve the
active points to create load paths from input to the structure features than randomly switching elements
output points. Some intermediate cross members, such between two designs.
as that shown in Figure 14, are then introduced by
dividing the design domain between the input and the
active points into several sections to provide structural CONCLUSION
stiffness. More complicated mesh configuration, such as
including an ‘X’ in each grid box as shown in Figure 17, We have developed a systematic synthesis approach
can also be used to exploit different solution space. for compliant mechanisms that are particularly focused
Although the determination of cross members and mesh to achieve adaptive shape change. The approach
configuration is still somewhat arbitrary, we believe the employs a discrete optimization scheme, GA, to explore
use of active points is a definite improvement over the discrete nature of the topology, while simultaneously
existing approaches that discretize the entire design determining the dimensions for the compliant mechan-
domain arbitrarily. ism. In the optimization process, we have also included
two different objective functions: LSE for asymmetric
Specification of Boundary Conditions problems, and modified FT for symmetric ones. Two
antenna shape change examples are then presented to
The boundary conditions specified in this synthesis demonstrate the feasibility of this approach. The results
problem are critical to the final solution because show that we can indeed achieve the desired shape
the location/number of fixed points (grounded points) change using compliant mechanisms and thereby realize
and location/direction of input can also change the inherent benefits associated with them. More shape
the solution space. With the same mesh configuration, change examples are currently being studied to explore
the solution space complexity might be the same, but the limitation of this approach. Issues regarding the
a different boundary condition (different locations discretization mesh and GA convergence are also being
of fixed points) would imply a completely different investigated to refine the synthesis approach.
solution space (kinematically). Although the bound-
ary conditions are specified at the beginning of the
synthesis procedure, they can are not completely fixed ACKNOWLEDGMENTS
throughout the optimization process in this research.
When the element(s) connecting to a particular fixed Authors gratefully acknowledge the funding support
point is(are) eliminated, the number and locations of the of U.S. Air Force Office of Scientific Research for
fixed points are changed. Thus, the synthesis problem this work under the research contract number
is not confined to one particular set of boundary F49620-01-1-0160.
 Design of Compliant Mechanisms 391

REFERENCES Kota, S., Hetrick, J., Li, Z. and Saggere, L. 1999. ‘‘Tailoring
Unconventional Actuators Using Compliant Transmissions:
Design Methods and Applications,’’ IEEE/ASME Transactions
Ameduri, S., Esposito, C. and Concilio, A. 2001. ‘‘Active Shape Airfoil on Mechatronics, 4(4):396–408.
Control through Composite Piezoceramic Actuators,’’ SPIE
Martin, C.A., Jasmin, L., Flanagan, J., Appa, K. and Kudva, J.N.
Proceedings, 4327:641–650.
1997. ‘‘Smart Wing Wind Tunnel Model Design,’’ SPIE
Ananthasuresh, G.K., Kota, S. and Kikuchi, K. 1994. ‘‘Strategies for Proceedings, 3044:41–47.
Systematic Synthesis of Compliant Mems,’’ ASME Winter Annual
Martin, J.W., Main, J.A. and Nelson, G.C. 1998. ‘‘Shape Control of
Meeting, 55:677–686.
Deployable Membrane Mirrors,’’ ASME Adaptive Structures and
Angelino, M. and Washington, G. 2001. ‘‘Point Actuated Aperature Materials Systems Conference, ad57/md83:217–223.
Antenna Development,’’ Proceedings of SPIE, 4334:147–155.
Martin, J.W., Redmond, J.M., Barney, P.S., Henson, T.D., Wehlburg,
Austin, F., Siclari, M.J., Van Nostrand, W., Weisensel, G.N., J.C. and Main, J.A. 2000. ‘‘Distributed Sensing and Shape Control
Kottamasu, V. and Volpe, G. 1997. ‘‘Comparison of Smart Wing of Piezoelectric Bimorph Mirrors,’’ Journal of Intelligent Materials
Concepts for Transonic Cruise Drag Reduction,’’ SPIE Systems and Structures, 11:744–757.
Proceedings, 3044:33–40.
Saggere, L. and Kota, S. 1999. ‘‘Static Shape Control of Smart
Austin, F. and Van Nostrand, W. 1995. ‘‘Shape Control of an Structures Using Compliant Mechanisms,’’ AIAA Journal,
Adaptive Wing for Transonic Drag Reduction,’’ SPIE Proceedings, 37(5):572–578.
2447:45–55.
Washington, G. 1996. ‘‘Smart Aperature Antennas,’’ Journal of Smart
Frecker, M. 1997. ‘‘Optimal Design of Compliant Mechanisms,’’ PhD Materials and Structures, 5(6):801–805.
Dissertation, University of Michigan, Ann Arbor.
Webb, G.V., Lagoudas, D.C. and Kulkarini, M. 1999. ‘‘Adaptive
Goldberg, D. 1989. Genetic Algorithms in Search, Optimization, and Shape Control for an Sme-Actuated Aerofoil Rib Structure,’’
Machine Learning, Addison-Wesley, 1–57. Proceedings of IMECE, 59:205–212.
Hetrick, J. and Kota, S. 1999. ‘‘An Energy Formulation for Parametric Yoon, H.S. and Washington, G. 1998. ‘‘Piezoceramic Actuated
Size and Shape Optimization of Compliant Mechanisms,’’ ASME Aperature Antenae,’’ Journal of Smart Materials and Structures,
Journal of Mechanical Design, 121:229–234. 7(4):537–542.
Joo, J. 2001. ‘‘Nonlinear Synthesis of Compliant Mechanisms: Yoon, H.S., Washington, G. and Theunissen, W.H. 2000. ‘‘Analysis
Topology and Size and Shape Design,’’ PhD Dissertation, and Design of Doubly Curved Piezoelectric Strip-Actuated
University of Michigan, Ann Arbor. Aperature Antennas,’’ IEEE Transactions on Antennas and
Joo, J., Kota, S. and Kikuchi, N. 2001. ‘‘Nonlinear Synthesis of Propagation, 48(5):755–763.
Compliant Mechanisms: Topology Design,’’ 2001 ASME Design
Engineering Technical Conferences, DETC2001:MECH–14141.

View publication stats