Mechanism and Machine Theory 117 (2017) 56–77

Contents lists available at ScienceDirect

Mechanism and Machine Theory

journal homepage: www.elsevier.com/locate/mechmachtheory

Research paper

Design and optimization of a general planar zero free length

spring
Arnoud A.T.M. Delissen∗, Giuseppe Radaelli, Just L. Herder
Precision and Microsystems Engineering, Delft University of Technology, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: A zero free length (ZFL) spring is a spring with special properties, which is commonly used
Received 4 January 2017 in static balancing. Existing methods to create ZFL springs all have their specific drawbacks,
Revised 17 June 2017
which rises to the need of a new method to create such a spring. A method is proposed to
Accepted 3 July 2017
design planar ZFL springs with specified stiffness (250–750 N/m) within a certain range (up
to 20 mm of displacement). Geometric non-linearities of a curved leaf spring are exploited
Keywords: by changing its shape. The shape is determined by a non-linear least squares algorithm,
Zero free length spring minimizing the force residuals from a non-linear numerical analysis. Constraints are in-
Null-length spring troduced to help in preventing the spring from intersecting itself during deformation. For
Ideal spring three types of springs with different boundary conditions, designs are found with charac-
Curved flexure teristic shapes and maximum force errors less than 1%. A trend is observed between spring
Shape optimization
size, maximum stress and desired stiffness. New type of ZFL springs can now be designed,
Desired force-displacement
which can not only be used in existing applications, but also enables the use of ZFL springs
in micro mechanisms.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction

The spring is one of the most commonly used physical elements in engineering. A special sub-category in springs is the
zero free length spring (ZFL spring), also null-length spring or ideal spring, which is - as the name implies - a spring with
zero physical length when no forces are applied. This property results in a spring element which enables a range of special
mechanisms to be realized, predominantly in the field of static balancing. To mention a few: Slow wave seismometer [1,2],
Anglepoise suspension [3] which are basically balanced arms with a mass [4], zero stiffness mechanisms [5,6], mobile arm
support for humans [7,8], camera stabilizer apparatus [9] and in (robotic) manipulators [10,11]. In all these applications any
unwanted potential energy differences (resulting from gravity or elastic deformations) are counteracted by one or more ZFL
springs, delivering the necessary forces to neutralize the undesired loads.
Before explaining the emerging implementation difficulties of a ZFL spring, it is important to note the properties which
define the unique behavior of a ZFL spring. A ZFL spring gets its unique properties from the sole fact that the unstretched
(free) length of a linear spring is zero. Observing Fig. 1, a ZFL spring is shown on the left. The length being zero results in
the spring pivot and endpoint being coincident. The spring force is now directly proportional to the displacement vector, i.e.
the force is in the same direction and its magnitude is proportional to the extension length of the spring (F = ku). This in
contrast to a normal (non-zero free length linear) spring, where the force is not only dependent on the displacement, but
also on the initial position L0 of the spring (F = k(L0 + u )). This is shown on the right in Fig. 1. In one dimension there is

∗
Corresponding author at: Delft University of Technology, Mekelweg 2, 2628 CD Delft, Netherlands.
E-mail address: [email protected] (A.A.T.M. Delissen).

http://dx.doi.org/10.1016/j.mechmachtheory.2017.07.002
0094-114X/© 2017 Elsevier Ltd. All rights reserved.
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 57

List of variables (in order of appearance)

F Reaction/spring force
k Spring stiffness
u Displacement
L0 Initial length of spring
xi Control points
N Number of segments
li Segment length
αi Segment relative angle
q Design vector
w Beam width
h Beam height
Fe External force vector
K Stiffness matrix
Kt Tangent stiffness matrix
Fi Internal force vector
Fx , Fy Reaction force component in x and y-direction
FA Axial (reaction) force component (aligned with displacement direction)
FT Transverse (reaction) force component (force perpendicular to displacement direction)
M Reaction moment
nu Number of displacement steps in one track
nθ Number of tracks in different directions
umin Minimum displacement (displacement at first sample point)
umax Maximum displacement range of the spring
F˜ Desired force
r Force residual
fobj Objective function
α max Maximum relative segment angle
lmin Minimum segment length
lmax Maximum segment length
r0 Direction vector from the light source node to the shadow segment
rs Orientation vector of the shadow segment
R90 90° angle rotation matrix
d0 Normal vector of the shadow segment
nˆ 0 Corrected normal vector of the shadow segment
r1 , r2 Direction vectors from the light source node to the corners of the shadow segment
d1 , d2 Normal vectors of the corners
n1 , n2 Corrected normal vectors of the corners
μ Direction parameter
R(θ ) Rotation matrix of angle θ
θe Extra outwards rotation of normal vectors
nˆ 1 , nˆ 2 Normal vectors of the shadow segment corners
f0 , f1 , f2 Distances to the border of the shadow
gI Constraint equation for the shadow-method constraint to prevent self-intersection
Nc Number of shadow constraint equations
gII Constraint equation for contact during displacement
rg Penalty function for constraint equations
cI , cII Constraint constants for penalty functions
fobj, c Constrained objective function
J Jacobian matrix
p Refined control points
ɛA Maximum relative error of forces in axial direction
ɛT Maximum relative error of forces in transverse direction
ɛshape Error on the shape of the spring

no difference between the two springs, as the reference (zero) point can be chosen freely. But in two or three dimensions
the zero point will always be the pivot point on which the spring is fixed.
58 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 1. Comparison of a ZFL spring (left) and a ‘normal’ spring (right). In the ZFL spring F = ku, while in the normal spring F = k(L0 + u ). The undeformed
spring is shown in black, the deformed in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version
of this article.)

Fig. 2. Various existing methods to create ZFL springs. By using a cantilever spring (a), a pre-stressed spring (b), a spring made from non-linear rubber (c)
or a compound spring (d).

Currently, a general ZFL spring is not an existing product, although various methods exist to approach the behavior of
a ZFL spring at least in a certain range. The simplest ZFL spring is created by using a cantilever bar on its perpendicular
plane. In all directions there is a linear proportional force from the origin. The plane of ZFL action for this spring is shown
in Fig. 2a. For large deformations however, it is not linear anymore, plus for some designs an out-of plane structure may not
be desirable. Cawardine [3] proposed to create a pre-stressed spring using a special coiling process. By applying a pre-stress
the effective undeformed length of the spring would become zero. In Fig. 2b this effect is shown. A third method to create
a ZFL spring is to use a material which exhibits a linear elastic behavior over a limited range, for instance implemented
by a rubber band [8,12]. The non-linear material properties of the material cause a linear proportional part in the force
profile of the spring. This effect is schematically shown in Fig. 2c. Another method to get a ZFL spring, would be to hide
the length of a normal linear spring behind a guiding point. For instance, the spring can be connected to a wire, which is
guided along a pulley. This guiding point then acts as a new zero point for the ZFL spring, which is shown in Fig. 2d. Also
other possible compound structures simulating ZFL springs exist [13,14]. A recently explored method is to use an in-plane
compliant mechanism. By the arrangement of leaf springs and links a desirable force-displacement curve can be obtained.
It has been shown that it is possible to obtain ZFL behavior in one direction over large deflections [15].
The existing methods of creating ZFL springs have big disadvantages. For instance, the range of operation is limited by
the range of linear material behavior in rubber or by the physical space a (complex) compound structure occupies. It is hard
to produce a precise ZFL spring from a compound structure, as effects of friction and backlash are introduced by a guiding
mechanism. Furthermore, it is not possible to create a spring on micro scale because fully three-dimensional structures,
assembly methods and material choice are limited. The existing ZFL spring based on a mechanism [15] is unable to obtain
ZFL behavior close to zero and is limited to one direction.
The goal of this research is to provide the designer with a method to create their own ZFL spring, tailored for any
chosen spring stiffness. The spring is realized by a beam which is formed into a yet unknown shape. When displacing one
end of the beam over a predetermined area of operation, the beam shape inherently causes a linear force-displacement
proportionality in all in-plane directions. This method is applied on three cases with different boundary conditions, each
behaving differently and being suitable for different applications.
The method starts off with describing the numerical model which is used for simulation, followed by an explanation how
of the desired and actual force profiles are translated to a mathematical objective function. Furthermore, the optimization
process is treated. The results section shows different designs and their performance, obtained from various initial designs
and desired stiffness values. Finally the performance is analyzed and the optimization process, shape, possible applications
and difficulties are discussed.

2. Method

In this section, a method is proposed which is able to obtain designs for ZFL springs. This research is limited to planar
designs of curved flexures with constant thickness and height. The first step is to obtain a suitable numerical model. Then,
the reaction forces can be calculated by numerically analyzing this model. An optimization problem is constructed, for which
an objective function needs to be formulated from the reaction forces. Next, some constraints on the shape are introduced
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 59

Fig. 3. An arbitrary curved spline with design variables being angles and lengths. The control point scaffold (red) and the resulting spline (black). (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

and incorporated into the objective function, to help prevent infeasible designs. Finally, an optimization algorithm is used
to obtain designs for the ZFL spring.

2.1. Analysis

Before anything can be analyzed, the spring geometry needs to be parameterized. By using a (quadratic) B-spline [16] a
smooth shape can be generated, dependent on a set of coordinates or control points xi which form the scaffold of the curve.
In this way complex shapes can be described by using relatively few parameters. As design parameters during optimization,
the Cartesian control points will not directly be used, but instead they will be generated from N segments with lengths li
and a relative angle α i with respect to each other (Fig. 3). This is to have better control on the relative positioning of the
control points. The design parameters α i and li are captured in the design vector q. The first control point is chosen to be
x0 = 0, any further coordinates can be calculated using Eq. (1). Furthermore, the cross section of the beam is uniform over
the whole curve and is also kept constant during optimization. The out-of-plane width is called w and the in-plane beam
height h.
⎡ ⎤
p
  α
xi i ⎢cos q ⎥
xi = = x0 + l p⎢  q=1 ⎥ (1)
yi ⎣ p ⎦
p=1 sin αq
q=1

The values of interest are the reaction forces. These are calculated using the non-linear equation
Fe = K (u )u (2)
The beam discretization is based on the Isogeometric Analysis (IGA) framework, developed by Hughes et al. [17,18]. Being
very similar to Finite Element Analysis (FEA), it distinguishes itself by the fact that it is based on the exact design geometry
as defined with splines. This means that a coarse (unrefined) mesh already exactly represents the shape of the beam. In FEA,
the discretization into finite elements results in a poor geometrical continuity for coarse meshes, causing inaccuracies. With
IGA this problem is overcome. Additionally, the IGA method is able to perform refinements to the discretization without
changing the geometry. A Kirchoff-Love rod formulation capable of finite strain, large rotations and large deformations was
implemented into IGA by Nagy [19].
A linear elastic material model with isotropic behavior can be used since a metal will be chosen as the base material,
operating in the linear regime. On account of occurring geometrical non-linearities resulting from large deformations, the
system of equations is solved using an iterative solver with full Newton steps to find the static solution. Eq. (2) is iteratively
solved by additive displacements resulting from the force imbalance:
Kt (u )u = Fe − Fi (u ) (3)
Boundary conditions are applied on both ends of the beam by use of Lagrange multipliers [20]. Both rotations and
displacements in two directions can independently be chosen constrained, displaced, or free. For different combinations
of boundary conditions, distinct types of springs are distinguised. Three spring versions are identified as pinned-pinned,
clamped-pinned and clamped-clamped, also seen in Fig. 4. There is a fundamental difference between these types of springs.
Because of the rotational freedom of the pinned-pinned spring, the force is always aligned along the two pinned points,
shown in Fig. 4a. This is comparable to a simple spring element as commonly used in engineering where the zero point or
rotation point is at the base of the spring. The next type introduces a clamp on the base side, causing a reaction moment to
be added. This allows for the direction of the force to be in any direction, as seen in Fig. 4b. The result of this is that there is
60 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 4. Different boundary conditions give springs with different behaviors. The undeformed spring (black) and arbitrarily deformed state (blue). Distin-
guished are the pinned-pinned spring (a), the clamped-pinned (b) and the clamped-clamped spring (c). (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

Fig. 5. A curved beam which is displaced over a straight line at discrete steps (sample points). At each of these locations the reaction loads are calculated.

no rotation point, but only a zero force point at the end of the spring. The begin and endpoint do not need to coincide to get
zero free length properties. Another clamp on the displaced side also introduces a reaction moment M at the displacement
side in addition to the reaction force (Fig. 4c). The force components for the clamped-pinned and clamped-clamped spring
can be described as Cartesian (Fx and Fy ) or as the axial and transverse force (FA and FT ). The axial force is the component
aligned with the displacement direction, while the transverse force is perpendicular to the displacement.
Last in the analysis phase is to displace the endpoint over a certain track and calculate the corresponding reaction forces
at several sample points on this track. Starting with the pinned-pinned spring, it can already be deduced that due to its
rotational freedom, the smallest spring will be the one that exploits this freedom. The force-displacement profile will thus be
axial symmetric around the rotation point. Therefore, the pinned-pinned spring will be displaced over a straight line in only
one single direction. Because the spring’s endpoint is not always near the base point, the spring’s end first is pre-stressed to
the first sample point, close to the spring base joint. Then, at a number (nu ) of sample points ui on this displacement track
the reaction loads are calculated. In the case of the pinned-pinned spring, displacement actually means distance of the end
point to the base point.
The clamped springs do not have this rotational freedom, so being dependent on direction, multiple of straight tracks
are made in different directions from the initial endpoint. A track is made in nθ different directions, where on each line
nu sample points are taken between umin and umax . An example of multiple straight tracks can be seen in Fig. 5. For the
clamped-clamped spring, it is chosen that the endpoint additionally keeps the same orientation - always staying parallel to
its original angle. In some applications this angle might be chosen differently, resulting in a different force profile.

2.2. Objective function

Now being able to calculate the reaction loads (Fij and Mij ) at the set of displacement sample points uij , the next step
will be to compare these values with a desired force value (F˜i j ). In order to obtain a ZFL spring, the reaction force has to
be linear proportional to the displacement, which means F˜i j = kui j . These Cartesian forces are used during the optimization.
For the pinned-pinned spring and the clamped-pinned spring only forces are acting on the endpoint. The clamped-clamped
version also generates a torque at its endpoint. This moment is not taken into account during the optimization in this
work. By taking the difference between the actual loads and the desired loads, residuals are formed as in Eqs. (4) and (5),
illustrated in Fig. 6. Since the displacements are taken constant during optimization, the loads are only dependent on the
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 61

Fig. 6. An arbitrary force displacement plot corresponding with the displacement points on one track. The residuals indicated should be minimized.

design q. Only the spline geometry will thus be optimized.

rx,i j (q ) = Fx,i j (q ) − F˜x,i j (4)

ry,i j (q ) = Fy,i j (q ) − F˜y,i j (5)

For optimization purposes, it is favorable to have a single scalar component to optimize. This is done by combining all the
residuals using the sum of squares formulation. Note that this step is done internally by the chosen optimization algorithm,
which will be explained further in Section 2.5.
 
nθ nu

fobj (q ) = rx,i j (q )2 + ry,i j (q )2 (6)
i=1 j=1

By changing the design variables q, the objective function is minimized by the algorithm.
min fobj (q ) (7)
q

2.3. Constraints

To help in generating feasible designs, three different constraint types are applied. The simplest constraints in the op-
timization are the bounds on the design parameters. By placing a lower bound on the lengths li of the segments, they
are prevented from having negative or zero length. That could otherwise cause problems in the analysis. The upper length
bound prevents the spring from being extremely large compared to the displacement range. Also the relative angles α i are
bounded, because numerically the use of angles introduces a periodicity in the design space. Additionally it is not favorable
to have spring segments which are folded flat onto each other due to the beam thickness. The first angle α 1 is bounded
between −2π < α1 ≤ 4π , since it only changes the orientation of the spring instead of an angle between two segments. The
initial orientation is chosen between 0 < α 1, init ≤ 2π , to give enough freedom in orientation, not to be blocked immedi-
ately by the bounds. The optimizer now effectively has the possibility to go from 2π to 0 by moving towards 3π , preventing
the optimizer getting stuck at the boundary. The only exception to this is the pinned-pinned type of spring for which the
orientation is fixed at α1 = π /2 (since it is unaffected by orientation). Below, the bounds are summarized in equation form.
They are enforced by the optimization algorithm to be discussed.
−2π < α1 ≤ 4π (for PP : α1 = π /2 ) (8)

−αmax ≤ αi ≤ αmax for i = 2, . . . , N (9)

lmin ≤ li ≤ lmax for i = 1, . . . , N (10)

Even with bounds on the relative angles, a combination of design parameters could still cause the planar spring to
intersect with itself. This is undesirable as it would be impossible to produce in a planar design. Not intersections of the
spline itself, but crossings of the constructing scaffold lines are observed. Because the spline line is interpolating the scaffold
lines, some cases may exist where the segments are not intersecting whereas the spline is, or the reverse. However, using
the scaffolds leads to a simple and fast intersection prevention. Two subsequent segments are by definition already unable
to intersect. All remaining unique combinations of the N segments make pairs that are able to intersect. No off the shelve
solution exists for this constraint problem, so a new method is proposed.
62 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 7. The location of x j+1 results in an infeasible design if placed in the area shown by the filled contours. Constraint border normal are shown in black
(unrotated) and blue (rotated). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Taking two random oriented segments i and j (with j > i + 1), their nodes xi , xi+1 , xj and x j+1 are placed, as illustrated
in Fig. 7. The infeasible area for point x j+1 looks like a shadow behind segment i cast by a light source at xj . The constraint
value in the infeasible domain will be defined as the shortest distance from the location of x j+1 to the feasible domain. First,
the normal of segment i (d0 ) is determined, of which it is important it points towards the infeasible domain. The vector rs
aligned with segment i, is rotated 90° by using rotation matrix R90 . The direction of this normal vector is corrected such
that it points in the infeasible direction by aligning it with the vector pointing from xj to the middle of the segment (r0 ).
Doing this leads to the correct normal vector nˆ 0 .
xi + x j+1
r0 = − xj rs = xi − xi+1 (11)
2
 
0 −1
R90 = (12)
1 0

rs (r0 · d0 )d0
d0 = R90 nˆ 0 = (13)
r s  (r0 · d0 )d0 
The next two normals are the ones which define the edges of the shadow. First, the direction vector towards the seg-
ment’s corners r1 and r2 are determined. By rotating these 90°, the normal vector is obtained. Since the shadow is behind
the segment, the normals need to point inward. This is done by aligning the normals along the segment, which leads to the
normal directions n1 and n2 .
r1 = xi − x j r2 = xi+1 − x j (14)

r1 r2
d1 = R90 d2 = R90 (15)
r 1  r 2 

n1 = −sgn(d1 · rs )d1 n2 = sgn(d2 · rs )d2 (16)

An extra rotation of the normals is added to the shadow to prevent segment j to be placed directly through the nodes of
i. The constraint function value would be exactly zero, so no penalty would be given for that case. The normals need to be
turned away from the source xj . By using the determinant, the orientation of segment i with respect to the source (xj ) can
be determined. A rotation of θe = 10◦ is added, acting as a buffer.
 
  cos (θ ) − sin (θ )
μ = det r0 rs R (θ ) = (17)
sin (θ ) cos (θ )

nˆ 1 = R (sgn(μ )θe )n1 nˆ 2 = R (−sgn(μ )θe )n2 (18)

Three individual shortest distances from point x j+1 to the feasible domain can be calculated as f0 , f1 and f2 .
 
f0 = nˆ 0 · x j+1 − xi (19)

 
f1 = nˆ 1 · x j+1 − xi (20)

 
f2 = nˆ 2 · x j+1 − xi+1 (21)
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 63

The shortest of these three distances gives the constraint value:

gI,i j = min ( f0 , f1 , f2 ) ≤ 0 (22)
For each unique combination of two different segments, not being adjacent to each other, this constraint equation will
be calculated. A curve with N segments thus gives Nc unique constraints.
(N − 1 ) (N − 2 )
Nc = (23)
2
With the previous constraints, only the undeformed design has been subjected to intersection-prevention. During the
displacement intersections could still occur, which are called displacement-contacts. It is very cumbersome to include in-
tersections in the deformed spring: the model is refined for the analysis, which causes the (few) segments to refine into
many smaller segments. Due to the quadratic relation (Eq. (23)) for the number of constraints, this approach would lead
to an explosive amount of constraint equations. Therefore, a simple measure to prevent self-contact is proposed. The dis-
placement of the endpoint is known prior to the analysis, of which the maximum is umax in any direction. Although almost
the whole spring is deforming, the base point x0 is defined at standstill. Therefore, x0 and the endpoint xN have to be at
least a distance of umax apart from each other. Although this method does not prevent all cases of displacement-contact, it
certainly helps. The constraint condition is mathematically written as:
gII = umax − x0 − xN  ≤ 0 (24)
Note that this constraint is only added for the clamped-clamped and clamped-pinned spring. For the pinned-pinned
spring the distance between these two points needs to be smaller than the distance from the base point to the sample
point, in order to realize a positive (pulling) force. Therefore, the constraint is reversed for pinned-pinned situations:
gII = x0 − xN  − umin ≤ 0 (25)
The non-linear constraints are incorporated into the objective function as a residual by a penalty formula rg = c max(g, 0 ).
The constant c is used for scaling or weighing of the constraint equations. Using all constraint equations, the constrained
objective function can be written as
 
 
N N−i +1  2
fobj,c (q ) = fobj (q ) + cI2 max gI,i j (q ), 0 + cII2 max (gII (q ), 0 )2 (26)
i=3 j=1

2.4. Design sensitivities

Design sensitivities describe how sensitive an objective value or residual is to a change of its design parameters. They
give a big advantage in the optimization procedure and computation time if they are derived analytically. For every residual
(either force or constraint) the sensitivity can be calculated analytically, resulting in the Jacobian matrix J = ddqr . For the
residuals they can be calculated using the chain rule:
dr dF d p dx
= (27)
dq d p dx dq
dF dp
In which dp
are obtained using the direct sensitivity method. The refinement sensitivities dx
describe the relation be-
dx
tween changes of unrefined control points and refined coordinates [21]. Lastly, the term dq
are the sensitivities between the
Cartesian control points and the design variables - the derivatives of Equation 1.

2.5. Optimization algorithm

The final stage is to obtain designs using an optimization algorithm, which minimizes the constrained objective function
(Eq. (26)). The Matlab function for non-linear least squares optimization “lsqnonlin” is used, which minimizes the sum of
squares value of all the residuals. This optimizer is able to use the Jacobian sensitivity matrix. In this optimization function
the trust-region-reflective algorithm is used [22]. This optimization algorithm needs an initial design vector to start with.
Due to the possible existence of local minima, multiple different initial designs are used. These are randomly generated in
the entire design space, as defined by the bounds1 , by using a Latin hypercube [23]. This ensures the initial designs are
spread randomly but evenly distributed over the design space (in terms of angles and lengths).

2.6. Properties and constants

The material properties and geometry settings used for the optimization are listed in Table 1. The material properties
used, are those of titanium alloy Ti-6Al-4V. This material has a high yield strength compared to its Young’s modulus, making
it suitable for the large deformations and high internal stresses of the spring designs. Furthermore, the parameters which
define the sample points and optimization settings for the different designs can be found in Table 2. The desired stiffness
values k are chosen from the range {250, 375, 500, 625, 750} N/m.
64 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Table 1
Geometry, bound limits and material constants.

Height h 1.00 mm

Width w 10.0 mm
Maximum segment angle α max 0.9π rad
Minimum segment length lmin 0.1 mm
Maximum segment length lmax 200.0 mm
Number of segments N 4
Young’s modulus E 113.8 GPa
Yield strength σ yield 880 MPa

Table 2
Design and optimization parame-
ters.

Design PP CP & CC

umin 5 mm 0 mm
umax 20 mm 20 mm
nu 10 10
nθ 1 8
c 10 10

3. Results

In this section the resulting designs from optimizations of the different spring types will be shown and analyzed. Starting
with the pinned-pinned designs for various stiffness values, their performances are quantified using an error measure to
enable comparison. Followed by clamped-pinned and clamped-clamped designs which have to cope with extra parasitic
loads (forces in transverse direction of the displacement, which should be zero) and had to be optimized for more than one
direction. Finally some general observations about the designs and their limitations are made.
For human interpretation it is more intuitive to speak in terms of axial and transverse forces instead of the Cartesian
forces. The desired axial and transverse force respectively are F˜A,i j = kuA,i j and F˜T,i j = 0, where the transverse force is called
a parasitic force since it should be zero. The term k is the required stiffness and uA, ij the axial displacement.
The performances of the springs are quantified using two error measures. The first one giving information about the
forces in axial direction - the maximum relative error ɛA - is introduced as the maximum error between the actual force
in axial direction FA, ij and the desired force F˜A,i j . This measure gives the error with respect to the desired axial force in
percentages:
   
 FA,i j   FA,i j 
εA = 100 max  − 1 = 100 max  − 1 (28)
F˜ A,i j
uA,i j k
Similarly, the performance of parasitic forces in transverse direction is quantified (if applicable). Since the desired trans-
verse force is zero, the desired axial force will be used as a reference. This is now the error of the (parasitic) transverse
force with respect to the desired axial force in percentages:
 
 F 
εT = 100 max  T,i j − 1 (29)
uA,i j k
Note that in obtaining these performance parameters, the final design is re-evaluated with more sample points than
used in the optimization phase (twice the steps in axial direction and thrice the number of directions). This is to prevent
(possibly large) errors at points between the sample points to go unnoticed.
All three type of springs were optimized using the same initial design vectors. The four initial designs used are shown in
Fig. 8 and were generated using a Latin hypercube as described in the method. The results from each of these initial designs
are given to see the influence of differing initial design. For ease of indicating the various designs, they are indicated using
a special notation. The first two letters indicate the spring type - PP for pinned-pinned, CP for clamped-pinned and CC
for clamped-clamped. Secondly, the initial design number is given and finally, the desired stiffness (separated with a dot).
For instance, a clamped-pinned spring originating from the third initial design and optimized for a stiffness of 250 N/m is
indicated with CP3.250.

3.1. Pinned-pinned spring

The pinned-pinned type of spring only had to be optimized for axial force as parasitic forces were absent. Optimizing for
five different values of spring stiffness (k = 250, ..., 750 N/m) resulted in a range of designs - of which three will be shown
every time, for overview in the figures. The designs from initial design 2 are shown in Fig. 9a, which generally perform the
best of all initial designs. In Fig. 9b the generally worst performing designs from initial design 4 are shown.
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 65

Fig. 8. The four initial designs used during optimization, generated by random Latin hypercube sampling.

Fig. 9. The axial force profiles of the (un-optimized) initial designs. The same spring designs (indicated by differently colored lines, result in different
behaviors for (a) the pinned-pinned spring, (b) the clamped-pinned spring and (c) the clamped-clamped spring. The maximum relative axial error ranges
from 70% up to 800%. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
66 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 10. Pinned-pinned spring designs performing best PP2.250, PP2.500 and PP2.750 (a) and worst PP4.250, PP4.500 and PP4.750 (b). The designs are rotated
to align with each other. (For interpretation of the references to color in the text, the reader is referred to the web version of this article.)

Fig. 11. The axial force profiles for (a) the best (PP2.250, PP2.500 and PP2.750) and (b) the worst designs (PP4.250, PP4.500 and PP4.750) compared to the
desired force profile. The dashed lines are the desired forces, with circles at the sample points. (For interpretation of the references to color in the text,
the reader is referred to the web version of this article.)

The resulting shapes tend to a drop-like shape, with the begin and endpoint of the spring close to each other. For the
different stiffness values the shapes resulting from the same initial design are comparable to each other. Only the size is
different (Fig. 10a). The springs with lowest stiffness were the largest, while stiffer springs were bigger. The axial force-
profiles of the best and worst springs are shown in Fig. 11. It can be seen that the best designs are seemingly spot on with
the desired profile, while the worst set of designs are off more. No springs were found to be intersecting themselves during
displacement.
The axial errors were calculated for all the obtained designs on the refined number of sample points (Table 3). It can be
seen that the results from the 3rd and 4th initial designs are all performing bad compared to the other designs, with errors
over 100%. The designs resulting from initial design 2 were performing the best, all scoring ɛA < 1%. Also the designs from
initial design 1 perform well, with scores errors around 1%.

3.2. Clamped-pinned spring

Two new difficulties in the spring design are introduced by adding a clamp to create a clamped-pinned spring. Paths in
every direction had to be optimized and parasitic transverse forces were introduced. Again, the optimization was run with
4 different initial designs and for 5 stiffness values of which the best and worst performing designs are respectively shown
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 67

Table 3
The relative axial error ɛA for all optimized pinned-pinned de-
signs. The colored values are the designs shown in Figs. 10 and
11, the best performing spring in bold and the worst in italic.

ɛA (%)

Stiffness (N/m) Initial 1 Initial 2 Initial 3 Initial 4

250 0.954 0.711 55.507 114.063

375 1.515 0.423 69.968 95.668
500 1.174 0.157 79.592 107.907
625 1.091 0.314 84.284 116.491
750 0.815 0.279 100.916 102.962

Fig. 12. Clamped-pinned spring designs performing best CP3.250, CP3.500 and CP3.750 (a) and worst CP2.250, CP2.500 and CP2.750 (b). Designs are rotated
to align with each other.

Fig. 13. The axial force-displacement profile for the best (CP2.500) and worst (CP3.250) clamped-pinned designs compared to the desired force profile. The
desired forces at the sample points are indicated by circles.

in Fig. 12a and b. The results are (more or less) shaped like a spiral, with the pinned endpoint at the center of the spiral.
By increasing the required stiffness, the spring design becomes smaller. The implemented constraints were not violated in
any design. However, during displacement intersection occurs in two springs (CP2.500 as shown in Fig. 12b and CP2.375).
Just like for the pinned-pinned springs, the axial force-displacement profiles in different directions of the best and worst
designs is shown in Fig. 13. The best design shows axial force profiles very close to each other, while the worst design has
force profiles scattered over a larger range.
68 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 14. Axial force profiles for the best CP3.250 (a) and worst CP2.500 (b) performing springs. The graphs represent force profile in circles around the zero
point with different constant radii. The direction of displacement is shown on the horizontal axis.

Fig. 15. The reaction moments at the base for the best CP3.250 (a) and worst CP2.500 (b) performing springs for displacements in different directions. The
dashed lines indicate the direction in which the endpoint is displaced towards or away from the base point.

Table 4
The relative axial and transverse error ɛA and ɛT for all optimized clamped-pinned springs. The
best and worst performing designs are highlighted in respectively bold and italic.

Initial 1 Initial 2 Initial 3 Initial 4

Stiffness (N/m) ɛA (%) ɛT (%) ɛA (%) ɛT (%) ɛA (%) ɛT (%) ɛA (%) ɛT (%)

250 1.821 0.961 10.742 10.125 1.817 0.957 1.822 0.961

375 2.355 1.237 24.328 14.597 2.350 1.233 2.355 1.237
500 2.816 1.473 33.820 21.261 2.816 1.473 2.816 1.473
625 3.242 1.693 3.235 1.687 3.235 1.687 3.242 1.693
750 3.623 1.886 3.629 1.891 3.623 1.886 3.624 1.887

Since the springs were optimized in 8 different track directions, it is interesting to see the force behavior at points in
between the tracks. Therefore, the axial force-profile is shown for constant displacement length uA . Each line shows the
axial forces at locations on a circle around the undeformed pinned endpoint. These forces for the best spring CP3.250 and
worst spring CP2.500 are shown in Fig. 14. In the same manner the reaction moments in the base are shown in Fig. 15.
In Table 4 both the axial and transverse errors can be seen for all designs. Only designs for stiffness k = 250 N/m man-
aged to get a relative transverse error of ɛT < 1%. No designs had an axial error of ɛA < 1%. The results from initial designs
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 69

Fig. 16. Clamped-clamped spring designs performing best CC1.250, CC1.625 and CC1.750 (a) and worst CC2.250, CC2.625 and CC2.750 (b). Designs are rotated
to align with each other.

Fig. 17. The axial force-displacement profile for the best (CC1.250) and worst (CC2.625) clamped-clamped designs compared to the desired force profile.
The desired forces at the sample points are indicated by circles.

1, 3 and 4 result in good springs (ɛA < 5%) compared to initial design 2 (with most designs ɛA > 10%). Designs CP2.625 and
CP2.750 do come to a good design, similar to the other springs of the same stiffness. It is noticed that the results found
from initial designs 1, 3 and 4 were similar in shape and also have the same performance. The transverse errors follow the
results of the axial errors - if the axial error is low, the transverse is consistently somewhat lower.

3.3. Clamped-clamped spring

The last new difficulty was introduced by also clamping the endpoint of the spring, which introduces extra reaction
moments at the tip. Once more the optimization was run for 4 different initial designs and the range of stiffness values, of
which the best and worst resulting designs are shown in Fig. 16.
The resulting shapes either look like the shape of a horseshoe (CC1.625), a spiral which is less spiraling than the clamped-
pinned springs (CC1.250 and CC1.750) or an S-shape (CC2.250 and CC2.750). Also here the design becomes smaller as a higher
stiffness is required. Only in one case intersection during displacement occurs, which is for spring CC2.625.
Again, for the best and worst performing spring, the axial force-displacement profiles and the circular axial force profiles
are respectively shown in Figs. 17 and 18.
The two kinds of errors (ɛA and ɛT ) and the maximum reaction moments M are presented in Table 5. Most designs had a
relative axial error and relative transverse error of ɛA < 1% and ɛT < 1%. The results from initial design 1 are nearly identical
70 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 18. The axial force profiles for the best CC1.250 (a) and worst CC2.625 (b) performing springs. The graphs represent force profile in circles around the
zero point with different constant radii. The direction of displacement is shown on the horizontal axis.

Table 5
The relative axial and transverse errors ɛA and ɛT plus moments M for all opti-
mized clamped-clamped springs. Best and worst values are indicated with respec-
tively bold and italic.

Initial 1 Initial 2

Stiffness (N/m) ɛA (%) ɛT (%) M (Nm) ɛA (%) ɛT (%) M (Nm)

250 0.406 0.102 0.211 0.660 0.287 0.368

375 0.723 0.199 0.414 0.808 0.361 0.482
500 0.859 0.242 0.507 0.953 0.430 0.587
625 0.979 0.282 0.594 13.998 10.007 0.906
750 0.839 0.270 0.375 1.164 0.547 0.770

Initial 3 Initial 4

Stiffness (N/m) ɛA (%) ɛT (%) M (Nm) ɛA (%) ɛT (%) M (Nm)

250 0.670 0.278 0.369 0.406 0.102 0.211

375 0.825 0.361 0.484 0.723 0.199 0.414
500 0.953 0.430 0.587 0.859 0.242 0.508
625 1.064 0.491 0.681 0.979 0.282 0.595
750 1.164 0.547 0.770 1.093 0.320 0.680

to those of 4 (except CC1.750 compared to CC4.750). Also initial design 2 and 3 are almost the same, with exception of
CC2.625, which did not come to a good design. The reaction moments are in the range of 0.2 Nm to 0.9 Nm for all designs.

3.4. Results | general design observations

To examine the deterioration of performance resulting from small deviations in the design, Monte-Carlo simulations were
run. Small errors were introduced in each control point of the best clamped-clamped design (CC1.250). The beam design was
refined by knot insertion to obtain 17 instead of 5 adjustable control points, resulting in  a greater shape freedom. For each
of these control points a random error in x and y-direction is added of maximum size x2 + y2 < εshape (see Fig. 19).
Two different sizes of shape errors are chosen as εshape = {1.0, 5.0} mm and for each of these, 10 0 0 random variations are
analyzed. This resulted in the data shown in Fig. 20. In Table 6 a summary of the errors is given.
The maximum size measure of the spring is defined as the diameter of the smallest circle totally encompassing the
spring design. This length measure is normalized by dividing with the maximum displacement range umax and is shown
in Fig. 21a with the variation of normalized spring stiffness horizontally. The stiffness normalization is obtained by dividing
with the Young’s modulus E and the beam width w. A trend of smaller springs for a higher stiffness is observed for all kinds
of springs. It is also noticed that the pinned-pinned springs are smaller than their clamped counterparts.
For each design in the maximum deformed position, the stresses were calculated. All optimizations resulted in springs
which did not reach yield strength during deformation with the chosen stiffness values. Extracting the maximum occurring
stresses, normalizing (by dividing with the yield strength σ yield ) and plotting them versus the normalized spring stiffness
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 71

Fig. 19. The control points are randomly varied within the areas indicated in (a), of which several examples are shown in (b) for the different shape error
sizes.

Fig. 20. The uncertainty range of spring design CC1.250, with the original best performance in a black dashed line. This uncertainty results in the maximum
axial error distribution (a) and the maximum transverse error distribution (b) for 10 0 0 samples in two different shape error regimes. The smaller shape
error results are shown in red, the larger is shown in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the
web version of this article.)

Table 6
Summary of errors resulting from shape errors of the spring.

Error size Minimum error Maximum error Mean error

ɛA (%) ±5.0 mm 0.823 12.64 5.146

±1.0 mm 0.465 3.253 1.287
Original 0.406
ɛT (%) ±5.0 mm 0.283 9.615 1.287
±1.0 mm 0.149 1.485 0.610
Original 0.102
72 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

Fig. 21. (a) Different normalized length measures for different designs and (b) maximum normalized stresses for different designs plotted versus the
normalized stiffness. The different line styles represent the different spring types, while the markers represent the used initial condition.

Fig. 22. The normalized length plotted against the normalized maximum stress. The different line styles indicate the different spring types and the markers
represent the used initial design.

results in Fig. 21b. The maximum stresses show an increasing trend when required stiffness is increased. This trend is valid
for all spring versions. Finally, both the normalized maximum stresses are plotted against the normalized length measures
in Fig. 22.
From a practical point of view it might be attractive to have a spring with a finite number of bends instead of a con-
tinuously curved beam, possibly leading to easier fabrication. Optimizations using straight beams were performed to create
CP250 and CC250 springs. Linear interpolations with a set number of straight edges were made from the best resulting
splines (CP3.250 and CC1.250). The initial shapes are shown in Fig. 23a and the final shapes after optimization in Fig. 23b.
Also their performance values are given in Table 7.

4. Discussion

Following the structure of the Results section, the findings are discussed in order of the springs defined by the three
different types of boundary conditions. First, the optimization phase is treated for each spring type and secondly, the per-
formance of the final designs. Next, the general shape of each type of spring is qualitatively analyzed and finally specific
applications and restrictions of the different spring types are elaborated. The section is closed with some remarks about the
general design of these springs.
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 73

Fig. 23 . The initial designs compared to the original spline shape and the optimized straight shapes for clamped-pinned springs, respectively (a) and (b).
And for the clamped-clamped springs (c) and (d). (For interpretation of the references to color in this figure legend, the reader is referred to the web
version of this article.)

Table 7
Overview of performance values for original design and optimized
straight shapes.

Initial design No. of segments ɛA (%) ɛT (%) M (Nm)

CP3.250 3 24.609 15.140 –

CP3.250 4 2.579 1.425 –
CP3.250 5 2.152 1.205 –
CP3.250 Original spline 1.817 0.957 –
CC1.250 3 1.824 0.612 0.402
CC1.250 4 0.637 0.409 0.242
CC1.250 5 0.453 0.113 0.229
CC1.250 Original spline 0.406 0.102 0.211
74 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

4.1. Pinned-pinned spring

Because the pinned-pinned spring has a rotational freedom, the forces only need to be optimized over one track. In this
case, the forces already point to the point of rotation, making the transverse forces automatically zero. This makes it the
spring with least residuals to optimize. When looking at the performance values (Table 3) and comparing to the other spring
types (Tables 4 and 5) this appears to result in lower errors (for the best performing springs). It also results in a smaller
spring compared to the others (Fig. 21).
For this spring there is great variation in the solutions found. Not every initial design gives good ZFL springs. Because this
spring needs to move to the first sample point, a pre-stress was applied before starting the analysis. This gives difficulties
in the cases of initial design 3 and 4 (Table 3). The spring starts as the initial design in a stretched configuration. To get to
the first starting point a negative (pushing) force is applied, and this force remains negative over the entire displacement
range from there (Fig. 9). Now the optimizer seeks to minimize the difference between the negative actual force and positive
desired force. One option to do this is to move the endpoint closer to the begin point - helped by the additional constraint.
The other is to lower the stiffness of the entire spring, thus lowering the overall forces (and hereby obtaining a lower
objective value). Once the spring forces cross over to the positive side, the spring’s compliance needs to be reduced again.
By this time however, the optimizer also succeeded in bringing the two endpoints together, which introduces trouble. The
fact that these two rotationally free points are (almost) coinciding causes a rigid body motion to occur, i.e. the solver cannot
find one single solution anymore as all orientations are a valid solution, the stiffness matrix will become singular. Close to
this point the design is also very sensitive in its force-response. A small change in design will cause the spring to behave
entirely different. All in all, this means that the initial designs need to be chosen with care for this type of spring to prevent
the optimizer to get stuck in a local minimum (10 out of 20 designs have bad performance). A possible solution for this
is to choose an already designed zero free length spring with another stiffness close to the new desired stiffness as initial
design.
Fig. 11 shows that the force profile is smooth in between the sample points. This proves that the number of sample
points is enough for this application. A sudden drop in force is observed close to a displacement of zero. This is caused by
the fact that the spring is forced towards its base point instead of extending. In the final designs a tiny gap is left between
the begin and endpoint of the springs. If the end point is displaced towards the begin point, the force will rapidly grow in
negative direction.
The general shape of the springs is a droplet shape, where begin and endpoint are close together. To get linearity and a
low enough stiffness, the spring needs length, which it generates by making the droplet shape. The short distance between
begin and endpoint are a direct result of the rotational freedom of the spring and extra enforced by the helping constraint of
the distance between begin and endpoint (which speeds up the optimization considerably). The shape of the best performing
springs in Fig. 10a has only three segments of considerable length, the first segment being very short (shorter than the
width of the spring). This implies a redundancy in design variables, where the same spring shapes and performances might
be obtained using less segments.
Possible applications for this spring are in the macro scale, where joints can be made such as weight balancing [13,24],
robotics [10,11] and arm support [7,8]. Especially for systems with low stiffness in transverse direction or rotation this would
be a good spring, since there are no parasitic transverse forces or torques which could cause errors in the movement of the
system.

4.2. Clamped-pinned spring

For the clamped-pinned springs multiple track directions were added, furthermore a second type of error was introduced
here. Initial design 1, 3 and 4 result in good springs for all chosen stiffness values, while initial design 2 only produces two
well performing springs. This shows the clamped-pinned is less inclined to get stuck in a local minimum than a pinned-
pinned spring (only 6 out of 20 designs have bad performance compared to the others). Compared to the pinned-pinned
springs the clamped-pinned are performing worse, which might be attributed to the added force requirements. On the other
hand no convergence issues were encountered, since the newly introduced clamp prevents rigid-body motion.
Most of the designs found are around an error of 1–5% (Table 4), but three have errors beyond 10%. Fig. 14a shows that
the forces in between the sample points vary a little, however enough sample directions were used to capture the behavior
in all directions as no strange peaks occur in between sample points. In Fig. 14b the variation is bigger, but the error is still
captured well by the chosen sample points. For larger displacements however it might be the case that the sample points
are starting to move too far from each other. In this case, more sample directions or other sampling patterns might be
used.
The general spring shape can be characterized as being spiral shaped (Fig. 12a and b). The low design stiffness gives
need of generating enough flexure length. Combined with the fact that the stiffness needs to be equal in every direction,
the material is distributed in a uniform shape with the same characteristics in every direction. To compensate the reaction
moment in the base of the spring, the reaction forces at the endpoint are used. However, in the displacement direction
towards and away from the base point this gives a problem. A perfect ZFL spring has in this direction only a force component
aligned with the base point, thus having no moment arm to create a couple. The graph in Fig. 15 shows that the best design
has negligible reaction moment in the direction aligned with the base point, while the worst design has a considerable
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 75

moment. This is reflected in the spring design, which minimizes the bending moment at the base when displacing towards
the base point. Also the spring becomes bigger in size due to the added directions (Fig. 21) - in fact they are almost twice
as big as the pinned-pinned springs, but also the desired force profile range is twice its size (one in forward direction and
one in reverse direction).
The distance constraint (Eq. (24)) helps keeping the begin and endpoint apart from each other, to prevent contact during
deformation. However, not all cases of contact during deformation are prevented by this. For instance, if the endpoint of
design CP2.500 is displaced to the left a contact point will occur. This constraint is only applied between the last and first
node so it does not prevent any of the other nodes to make contact.
In practice this spring would also be used in macro applications due to the need of a joint like the pinned-pinned spring.
In systems sensitive to transverse errors this spring type is less desirable. Still there are no parasitic torques possibly twisting
the structure.

4.3. Clamped-clamped spring

The last spring type added an extra clamp, introducing a second reaction moment. Surprisingly by the addition of this
clamp, the found designs performed better in axial and transverse force in comparison to their clamped-pinned counterparts
(Table 5 compared to Table 4) when comparing designs resulting from the same initial design and stiffness value. This could
be explained by the extra reaction moment. To satisfy moment equilibrium, the moment in the clamped base has to be
counter-acted. In the case of the clamped-pinned spring, there is only a reaction force counter-acting in the tip of the
spring, while in the clamped-clamped spring a reaction moment is added. This gives more freedom in the optimization
since the new reaction moment can help the forces in counter-acting the reaction moment in the base.
It is difficult to say whether the reaction moments (Table 5) are big or small since there is no reference. If the connected
structure is stiff in rotation direction this influence will of course be smaller than a very compliant structure in rotational
direction. This causes the need of the spring to be designed in conjunction with the entire system it will be incorporated
in. For instance, a balancer with a mass on an arm [25] makes a rotation during operation. The designed spring was opti-
mized for no rotations, so if a rotation were to be imposed, larger errors would occur. In such a case the (usually known)
rotation would need to be used in the optimization procedure, possibly resulting in different designs than presented in this
\ work.
Compared to the other spring types, the clamped-clamped spring results are very close together, with only one spring
design (out of 20) which has much higher errors than the others. This indicates very little influence of local minima far
away from the optimum. There are three main shapes found (Fig. 16), which can be identified as three local minima. The
first being the shape of a horse-shoe (CC1.625), the second a spiral (CC1.250 and CC1.750) - which is less spiraling than the
clamped-pinned springs - or the third, an S-shape (CC2.250 and CC2.750). Similar to previous designs also here the length
in the loop of the spring is used to gain a low stiffness in every direction. The symmetry in the designs can be explained
by the fact that the boundary conditions are identical on both ends. A symmetric design has equal, but opposite reaction
moments in both ends (if displaced in a symmetric manner). When the ends are moving directly from and to each other,
thus keeping symmetry, there is no need for the spring reaction forces to generate a couple.
This kind of spring is very useful for applications where friction is a big issue as no joints are needed. This also results
in application possibilities in systems on micro-scale.

4.4. General

In all cases spring designs were found, having a large improvement in performance on the initial design (compare
Fig. 9 to Figs. 11, 13 and 17). From the optimization results and previous discussion, it can be noted that the pinned-pinned
spring is hardest to optimize, followed by the clamped-pinned and the clamped-clamped spring. For future research, (ana-
lytical) methods should be developed which can provide better initial designs. For instance, the linear stiffness in different
directions can be optimized first before optimizing the non-linear force-displacement response.
The process of designing a spring starts with the requirements given on spring stiffness k and displacement range umax .
Then, depending on application, the spring type can be chosen. After that, a suitable material is chosen. The parameters
of the beam cross-section are chosen by technical capabilities (e.g. laser cutting), material type (e.g. sheet metal) and/or
available design space. Furthermore, the maximum segment length can be adjusted to the available design space. With this
information the optimization can be executed. If the results are not as desired, iterations can be done by changing the
beam cross-section. Thicker beams will result in springs with higher stiffness, but also higher stresses. Fig. 21a shows that
the spring size is increasing for lower values of stiffness. The stiffness the designer is able to obtain is bounded below
by limitations in space within or around the structure and the upper bound is determined by the minimal design area
(e.g. minimum feature size). The yield strength also determines the upper bound for the stiffness, since the stresses are
increasing for springs with higher stiffness (Fig. 21b). In future research, focus should be on determining suitable cross-
section parameters before starting optimization. If approximations can be made for maximum stresses or spring size, this
would reduce the number of design iterations to be made.
Constraints were added to prevent intersections from happening. One constraint focused on preventing intersections in
the initial (undeformed) state, while another focused on preventing the end-point from being too close (or too far away in
76 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77

the case of the pinned-pinned spring). From the results, only a few spring designs were found to intersect during displace-
ment despite the constraints used. In the pinned-pinned case no intersections were detected, for the clamped-pinned spring
only two (out of 20) springs and for the clamped-clamped spring only one. These were coincidentally also bad-performing
springs compared to the others. This leads to the thought that the constraints might obstruct the design from getting better
by blocking certain directions. This is a problem that has to be looked into in future research. Also, some designs might be
excluded from the feasible design set due to the constraints. Cases exist where the spline scaffold is intersecting, but the
spline itself is not.
No spring can be manufactured in exactly the dimensions given. So from a manufacturing point of view it is interesting
to see the performance decay as the spring dimensions are affected by slight random variations. From the Monte-Carlo
simulations, a shape error of ±1.0 mm gave axial errors of average 1.29%. An error of ±5.0 mm averaged to an axial error
of 5.15%, compared to the axial error of the original spline being 0.41%. Although small errors in the design do not directly
yield large errors in performance, depending on the application, the shape needs to be controlled to the right amount to
prevent the errors from getting too large.
A way which might make production easier is to use a finite number of bends instead of a continuously curved spring.
The effect of this was analyzed using a second optimization using straight segments, starting from a linear interpolation of
the continuous spline. The straight beams tend to keep the global form of the spline, with performance getting closer to
the original spring as the number of segments is increased. This behavior was showed for both the clamped-pinned and the
clamped-clamped spring. It might well be that these final straight springs have a sub-optimal behavior in their entire shape
freedom, but using the spline design is a reliable way to start the optimization. By using enough bends, fabrication might
become a lot easier while still maintaining good performance.
Fabrication proved to be very difficult, as different approaches did not yield a satisfactory result. Three methods were in-
vestigated, the first being laser cutting the spring out of PTFE. Creating the curved shapes using laser cutting was successful,
however the material properties of PTFE are less than optimal for a spring. Effects such as hysteresis and creep dramatically
distorted the force-displacement profile that no linear behavior could be observed. The second method was by bending a
strip of spring steel. However, due to the high yield strength of this material (which is needed to obtain a spring capable of
large deflection) the springback was very large and we did not succeed in creating the proper form needed. Finally we tried
to produce a spring by wire EDM out of Titanium Ti6Al4V. The long and slender shape introduced too much vibrations for
the shape to be cut out properly.
Not in all applications a full circular range will be needed. Some systems only require one path or an oddly shaped area
to have zero free length properties. Using this fact, application specific spring could be designed, behaving more precise in
the required area. This way for systems which are more critical with respect to transverse errors or reaction moments, a
customized spring can be made.

5. Conclusions

A zero free length spring is an engineering element which is useful, yet difficult to create. A new method of designing
in-plane compliant zero free length springs is proposed, based on the given desired force characteristics within a certain
range (up to 20 mm displacement). This is done for three different types of springs, which differ in boundary conditions;
the pinned-pinned, clamped-pinned and clamped-clamped springs. The first two are most suitable for macro applications.
The latter for micro applications, since it has no need for joints. From a set of initial designs, non-linear least squares
optimization is applied to obtain optimized designs for different values of desired stiffness (250–750 N/m). Constraints are
introduced to prevent the spring from intersecting with itself (during deformation).
The clamped-clamped spring is least affected by the choice of initial design, since 19 out of 20 designs yield good re-
sults compared to each other. For the clamped-pinned this is 14 and for the pinned-pinned 10, making it more sensitive to
the choice of initial design. The constraints work well, since no constraints are violated. However, some designs still have
intersections during deformation (2 of type CP and 1 for CC). Still some improvement on the constraint functions can be
made, because some feasible solutions are considered infeasible by the constraints, and vice versa. For each type of spring,
characteristic shapes are identified from the resulting designs and their shape is analyzed. The pinned-pinned spring is
shaped like a droplet, with the begin and endpoint close together. The clamped-pinned spring prefers a spiral shape and
the clamped-clamped spring a symmetric horseshoe shape. Considering only the well performing springs, the pinned-pinned
and clamped-clamped springs perform best in general with maximum axial force errors of less than 1%. The worst perform-
ing springs are the clamped-pinned springs with axial force errors of 1–2%. The clamped-pinned springs have problems
compensating the reaction moment, which the clamped-clamped springs can counter-act with a moment in the other end
of the spring, causing better performance. The pinned-pinned springs have good performance because of the small number
of residuals to optimize. They lack transverse errors, which the other two types have.
The feasible values of stiffness are bounded by design area, minimum feature size and yield strength of the material.
For increasing stiffness with constant beam cross-section, the maximum stresses increase, while the spring size becomes
smaller. More defined relations between cross-section, maximum stresses and spring size are to be established in further
research. Finally, it is observed that small errors in the design yield errors in the obtained force profile. Shape errors in
the design of ±1.0 mm and ±5.0 mm in the best optimized clamped-clamped spring, result in an average performance de-
terioration of maximum axial relative error of respectively 1.3% and 5.1%, compared to an original error of 0.4%. Also the
 A.A.T.M. Delissen et al. / Mechanism and Machine Theory 117 (2017) 56–77 77

introduction of straight segments instead of a continuously curved beam, introduces a deterioration of error, diminishing as
more segments are used.

References

[1] L.J.B. LaCoste, A new type long period vertical seismograph, J. Appl. Phys. 5 (7) (1934) 178, doi:10.1063/1.1745248.
[2] L.J.B. LaCoste and A. Romberg, “Force Measuring Device,” U.S. Patent US2293437, August 1942.
[3] G. Cawardine, “Improvements in Elastic Equipoising Mechanisms,” U.K. Patent GB404615, January 1934.
[4] D.A. Streit, E. Shin, Equilibrators for planar linkages, J. Mech. Des. 115 (3) (1993) 604–611, doi:10.1115/1.2919233.
[5] J.L. Herder, Design of spring force compensation systems, Mech. Mach. Theory 33 (1-2) (1998) 151–161, doi:10.1016/S0 094-114X(97)0 0 027-X.
[6] M. Schenk, S. Guest, J. Herder, Zero stiffness tensegrity structures, Int. J. Solids Struct. 44 (October) (2007) 6569–6583, doi:10.1016/j.ijsolstr.2007.02.041.
[7] B. Mastenbroek, E. de Haan, M. van den Berg, J.L. Herder, Development of a mobile arm support (Armon): design evolution and preliminary user
experience, in: 2007 IEEE 10th International Conference on Rehabilitation Robotics, IEEE, June 2007, pp. 1114–1120, doi:10.1109/ICORR.2007.4428563.
[8] A.G. Dunning, Slender Spring Systems, Delft University of Technology, 2016, doi:10.4233/uuid:50404d39- d6aa- 4fd4- bb20- 04794ec01b7b.
[9] G. Brown and A.O. DiGuilio, “Support Apparatus,” U.S. Patent US4208028, June 17 1980.
[10] C.M. Gosselin, J. Wang, Static balancing of spatial six-degree-of-freedom parallel mechanisms with revolute actuators, J. Robotic Syst. 17 (3) (20 0 0)
159–170, doi:10.1002/(SICI)1097-4563(200003)17:3159::AID- ROB33.0.CO;2- J.
[11] M. Vermeulen, M. Wisse, Intrinsically safe robot arm: adjustable static balancing and low power actuation, Int. J. Social Robotics 2 (September) (2010)
275–288, doi:10.1007/s12369-010-0048-9.
[12] P.N. Kooren, A.G. Dunning, M.M.H.P. Janssen, J. Lobo-Prat, B.F.J.M. Koopman, M.I. Paalman, I.J.M. de Groot, J.L. Herder, Design and pilot validation of
A-gear: a novel wearable dynamic arm support, J. Neuro Eng. Rehab. 12 (December) (2015) 83, doi:10.1186/s12984- 015- 0072- y.
[13] J. Herder, Energy-Free Systems: Theory, Conception, and Design of Statically Balanced Spring Mechanisms, Delft University of Technology, 2001, doi:10.
13140/RG.2.1.3942.8966.
[14] R. Barents, M. Schenk, W.D. van Dorsser, B.M. Wisse, J.L. Herder, Spring-to-spring balancing as energy-free adjustment method in gravity equilibrators,
J. Mech. Des. 133 (6) (2011), doi:10.1115/1.4004101.
[15] L.C. Leishman, M.B. Colton, A pseudo rigid-body model approach for the design of compliant mechanism springs for prescribed force deflections,
ASME 2011 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference, October 2011, doi:10.
1115/DETC2011-47590.
[16] L. Piegl, W. Tiller, The NURBS Book, second ed., Springer-Verlag, Berlin Heidelberg, 1996.
[17] T.J.R. Hughes, J.A. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement, Comput. Meth. Appl.
Mech. Eng. 194 (39-41) (2005) 4135–4195, doi:10.1016/j.cma.2004.10.008.
[18] J.A. Cottrell, T.J.R. Hughes, Y. Bazilevs, Isogeometric Analysis: Toward Integration of CAD and FEA, Chichester: John Wiley & Sons, Ltd, 2009, doi:10.
1002/9780470749081.
[19] A.P. Nagy, Isogeometric Design Optimisation, Delft University of Technology, 2011.
[20] G. Radaelli, J.L. Herder, Isogeometric shape optimization for compliant mechanisms with prescribed load paths, ASME 2014 International Design Engi-
neering Technical Conferences and Computers and Information in Engineering Conference IDETC/CIE 2014, no. August, (Buffalo, New York, USA), 2014,
doi:10.1115/DETC2014-35373.
[21] X. Qian, Full analytical sensitivities in NURBS based isogeometric shape optimization, Comput. Meth. Appl. Mech. Eng. 199 (29-32) (2010) 2059–2071,
doi:10.1016/j.cma.2010.03.005.
[22] MathWorks, Least-Squares (Model Fitting) Algorithms, 2016 [Online].
[23] K.-T. Fang, R. Li, A. Susjianto, Design and Modeling for Computer Experiments, Taylor & Francis Group, LLC, Boca Raton, 2006.
[24] S.R. Deepak, G.K. Ananthasuresh, Perfect static balance of linkages by addition of springs but not auxiliary bodies, J. Mech. Robotics 4 (2) (2012)
021014, doi:10.1115/1.4006521.
[25] J.L. Herder, R. Barents, B.M. Wisse, W.D.V. Dorsser, Efficiently variable zero stiffness mechanisms, in: 4th International Workshop on Human-Friendly
Robotics, 2011, pp. 2–3.