Home Search Collections Journals About Contact us My IOPscience

A nonlinear stretching based electromagnetic energy harvester on FR4 for wideband

operation

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2015 Smart Mater. Struct. 24 015013

(http://iopscience.iop.org/0964-1726/24/1/015013)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 143.239.66.244
This content was downloaded on 30/11/2014 at 12:45

Please note that terms and conditions apply.

 Smart Materials and Structures

Smart Mater. Struct. 24 (2015) 015013 (14pp) doi:10.1088/0964-1726/24/1/015013

A nonlinear stretching based

electromagnetic energy harvester on FR4 for
wideband operation
Dhiman Mallick1, Andreas Amann2 and Saibal Roy1
1
Microsystems Center, Tyndall National Institute, University College Cork, Lee Maltings, Dyke Parade,
Cork, Ireland
2
School of Mathematical Science, University College Cork, Cork, Ireland

E-mail: saibal.roy@tyndall.ie

Received 28 August 2014, revised 17 October 2014

Accepted for publication 4 November 2014
Published 28 November 2014

Abstract
We report a nonlinear stretching-based electromagnetic energy harvester using FR4 as a
vibrating spring material due to its low Young’s modulus. We show analytically that the
nonlinearity is caused by the stretching, in addition to the bending, of the specially designed
spring arms; this gives rise to a wider half-power bandwidth of 10 Hz at 1 g acceleration, which
is almost 5 times higher than that of a comparable linear counterpart. The output spectra show
the first reported experimental evidence of a symmetry broken nonlinear secondary peak in a
single potential well system at frequencies close to the nonlinear jump frequency, which may
appear to be due to the dynamic symmetry breaking of the oscillator or to the inherent
asymmetry of the built prototype. The presence of this secondary peak is useful in generating a
significant amount of power compared to the symmetric states, producing ∼3 times more power
at the secondary peak than the nearby symmetric states. 110% of the peak power obtained for
0.5 g acceleration is achieved at the secondary peak during the frequency up-sweep. The
experimental results are compared with a deterministic numerical model based on the Duffing
oscillator, and we include a qualitative discussion on the influence of noise in an experimental
energy harvesting system.
Keywords: electromagnetic energy harvesting, nonlinear, symmetry breaking, FR4, wideband

(Some figures may appear in colour only in the online journal)

1. Introduction boost. Piezoelectric [3, 4], electrostatic [5, 6] and electro-

Lately, vibrational energy harvesting (VEH), which concerns
the efficient conversion of the ambient mechanical vibration
energy that is present in the environment into electrical
energy, has received a significant amount of attention due to
the reduced power requirements of small-scale electronic
components [1, 2]. Providing power to long-term, low-power
wireless sensing applications has been recognized as one of
the main motivations behind the research in VEH. In parti-
cular, with the growing utility of portable, implantable and
wireless sensor nodes in health monitoring, military and
industrial applications, the research on energy harvesting
systems for powering these devices has seen a tremendous
magnetic [7–10] transducers are the most widely used meth-
ods that are utilized to convert ambient vibrations to electrical
energy. Piezoelectric and electrostatic energy harvesters have
been fabricated using a MEMS batch fabrication process in a
number of reported works [4–6], whereas due to the difficulty
of integrating magnetic materials, the realization of a micro-
fabricated electromagnetic generator is still a challenge. Also,
the output power of micro-electromagnetic energy harvesters
is significantly lower compared to that of their macro-coun-
terparts due to scaling [11–14]. Consequently, many works
have been reported over the years focussing on electro-
magnetic energy harvesters using macro-sized and discrete
components [9, 15, 16].

0964-1726/15/015013+14$33.00 1 © 2015 IOP Publishing Ltd Printed in the UK

Smart Mater. Struct. 24 (2015) 015013
Figure 1. (a) Schematic of the nonlinear electromagnetic energy harvester showing different components (inset figure shows the cross-section
of the prototype). (b) Architecture of the designed resonator on FR4.
A major part of the reported works in the last decade in
the field of VEH (including electromagnetic transducers)
involved linear resonating structures. In linear systems, the
response drops significantly as the external excitation fre-
quency shifts from the natural frequency of the structure. The
mismatch of frequency is inevitable, as in the majority of
practical scenarios, the environmental vibration is either fre-
quency-varying or random in nature. As a result, a number of
strategies have been adopted to widen the response of the
energy harvesting systems. Resonant frequency tuning [17–
19], multimodal energy harvesting [20, 21] and, more
recently, nonlinear energy harvesting [6, 22–30] have
emerged as potential solutions. Nonlinear effects in
mechanical oscillators are widely utilized these days to
achieve a wideband performance. Nonlinearity could be
introduced in energy harvesters through nonlinear stiffness.
Depending upon the nature of the potential function of the
system, nonlinear energy harvesters are broadly classified as
monostable and bistable. A bistable nonlinear oscillator has
been implemented through bistable cantilever beams [22, 28],
whereas monostable nonlinear oscillators have been realized
through nonlinear magnetic levitation [23] by introducing
nonlinear stiffness though stretching strain in a doubly
clamped beam [24, 30] and by the interaction between mag-
nets [32]. Mechanical stoppers are also used to increase the
operational bandwidth [33] using the monostable non-
linearity. The monostable energy harvesters can only work in
conditions with a slow and steady frequency sweep [31]. The
main advantage of using a monostable system is that there is
no need for any special arrangement to introduce the non-
linearity, as in the case of a bistable system. Nonlinearity
appears much more inherently in monostable cases; thus, the
total volume is often not that bulky.
In this paper, a nonlinear stretching-based electro-
magnetic energy harvesting system has been studied. The
device was designed in such a way that the nonlinearity
appears due to stretching of the spring arms, in addition to the
conventional bending component in a fixed-fixed beam con-
figuration. FR4 is used as the resonator material, due to its
low Young’s modulus (Y = 21 GPa), to operate in a relatively
lower frequency region. An analytical formulation of the
2
D Mallick et al
bending and stretching forces are derived to gain insight on
the source of the nonlinearity. A dynamical model has also
been developed from the governing equation of a nonlinear
oscillator incorporating electromagnetic transduction. Later,
detailed experimental studies are performed on the fabricated
prototype, and the results are compared with the modelled
data. Additionally, the paper investigates the effect of dif-
ferent dynamical parameters on the performance of the energy
harvesting device. Dimensional and non-dimensional models
are introduced to analyse the performance and to understand
the discrepancies observed between the modelled and the
experimental results.
2. Design and fabrication of the energy harvester
An electromagnetic (EM) energy harvester has been designed
that utilizes nonlinear stretching strain to obtain a wideband
response. The magnets are placed on the top of the central
paddle (1 cm2 square) of the vibrating resonator. Under the
influence of external vibrations, the magnets move up and
down, and the flux lines are cut by a coil assembled above it.
As a result, voltage is induced into the coil. A schematic of
the overall prototype is shown in figure 1(a). We have chosen
FR4 (standard PCB material), a relatively inexpensive mate-
rial, as the structural material of the vibrating body. Nonlinear
springs on FR4 could be a novel approach for wideband, low
frequency applications, due to its low Young’s Modulus
(Y = 21 GPa), and particularly for EM energy harvesters, as
conducting coils can be easily routed on FR4 laminates [34].
A low Young’s modulus reduces the necessity of using an
additional proof mass to bring the operational frequency
down within a comparatively smaller footprint.
2.1. Spring structure
The spring structure of the designed prototype has been
fabricated using a CNC machining process. The designed
spring structure is shown in figure 1(b). A similar spring
configuration was reported by Marinkovic et al [24] for pie-
zoelectric energy harvesting at the MEMS scale. In this work,
Smart Mater. Struct. 24 (2015) 015013
z (a)
y x
z
y x
Figure 2. Eigen modes of the designed resonator from COMSOL: (a) vertical up and down movement at 170 Hz; (b) twisting movement at
384 Hz; (c) another twisting movement at 484 Hz.
Figure 3. (a) Simulation result for the magnetic field. The orientations of the magnets are shown by the arrows. (b) Prototype harvester. The
overall volume of the prototype is 0.78 cm3. The magnets act as the proof mass for the FR4 resonator.
the spring structure has been designed and optimized for
macro-scale EM harvesting in which the magnets act as the
proof mass. The thin (150 μm) cantilever springs allow the
resonator to move in the out-of-plane direction. One end of
each of the four cantilever arms is fixed, while other ends of
the cantilevers are connected pairwise. This gives a fixed-
fixed cantilever configuration. As explained analytically in the
next section, nonlinear stretching arises due to this fixed-fixed
configuration of the cantilever. The overall outer dimension of
the resonator including the frame is 1.72 × 1.72 cm2. The
COMSOL eigen frequency modes for the spring structure are
shown in figure 2. Later, it is seen from the experiment that
the out-of-plane movement occurs in the first mode of
vibration at 170 Hz, whereas twisting movements are
obtained at 384 Hz and 484 Hz, respectively.
2.2. Magnets
Two oppositely polarized Neodymium Iron Boron (NdFeB
N42H) permanent magnets were epoxy bonded at the middle
3
D Mallick et al
z
(b)
y x
(c)
of the central paddle of the resonator. Each magnet has a
dimension of 4 mm × 2 mm × 1 mm, with 1 mm in the poled
direction. To provide a strong magnetic field, sintered NdFeB
N42H magnets (Magdev Ltd, UK) were used, which have a
typical (BH)max of 40 MGOe [35]. The field strength of this
magnet arrangement was simulated using Ansoft Maxwell 2D
v12. Figure 3(a) shows the simulation result. It is found that at
0.5 mm above the magnets, the magnetic field strength is
about 0.4 T. The orientations of the magnets are also shown in
figure 3(a) by the arrow.
2.3. Coil
An enamelled copper wire wound coil (Recoil Ltd, UK) was
also epoxy bonded to a glass slide and placed at a gap of
1 mm from the top of the magnets. The gap was created using
the spacers. The coil has an outer diameter, inner diameter
and thickness of 3 mm, 1.15 mm and 0.5 mm, respectively,
and has about 450 turns. The copper wire has a diameter of
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

Table 1. Different parameter values for the designed prototype that

are used for modelling and experiments.
Parameters Value
Spring Arm Length (L) 9.2 mm
Spring Arm Width (W) 0.8 mm
Thickness of the Resonator (d) 0.15 mm
Area of the Central Paddle 1 cm2
Young’s Modulus of FR4 (Y) (COMSOL) 21 GPa
Coil Outer Diameter 3 mm
Coil Inner Diameter 1.15 mm
Coil Number of Turns 450
Coil Resistance (RC) (Measured) 69 Ohm
Magnets Dimension 4 × 2 × 1 mm3
Linear Stiffness (k) (Calculated) 185.66 N m−1
Nonlinear Stiffness (kn) (Calculated) 3.56 × 109 N m−3
Electromagnetic Coupling Coefficient (γ) 0.035 Wb m−1
(Calculated)
Inertial Mass (m) (Measured) 16 × 10−5 Kg
Mechanical Damping Coefficient (ρ) 0.015
(Measured)

The contribution to spring force due to bending of the

beam can be approximated using the Euler–Bernoulli beam
equation:

∂ 4ϕ
=0 (1)
∂x 4
where ϕ (x) is the deflection of the beam at position x. As
seen from figure 4(a), the boundary conditions for the pro-
Figure 4. (a) Simplified model of a clamped-guided beam for the blem are
calculation of spring force. (b) Variation of spring force as a function ϕ (0) = 0,
of deflection. The inset figure shows the potential energy function of
the spring arm as it gets deformed. ϕ′ (0) = 0,
ϕ (L ) = d f ,
32 μm. The resistance of the coil was found to be 69 Ω upon ϕ′ (L ) = 0
measurement. Now, as equation (1) is a fourth order equation, ϕ can be
Different parameters for the designed harvester, which expressed as
are used in the modelling and the experiment in the following
sections, are listed in the table 1. The final assembled proto- ϕ = a 0 + a1 x + a 2 x 2 + a 3 x 3 (2)
type is shown in figure 3(b). The overall volume of the pro-
totype is 0.78 cm3. with real coefficient a0 to a3.
Solving equation (1) for the given boundary conditions,
we obtain an expression for bending along the length (x) of
the beam as
3. Analytical and numerical study
d f ⎡ 2 2x 3 ⎤
ϕ (x ) = ⎢ 3x − ⎥ (3)
3.1. Static analysis L2 ⎣ L ⎦

In this section, the components of spring force due to bending where Y is the Young’s modulus of the spring arm. Now, the
and stretching are obtained separately. For analytical model- strain energy due to the bending component of the deforma-
ling of the nonlinear spring force, a simplified model of the tion is
clamped-guided cantilever beam is used, as shown in
figure 4(a). It consists of a beam of length L, width W and Y 3 L ⎛ ∂ 2ϕ ⎞2
thickness d. One end of the beam (x = 0) is clamped, and the
E [ϕ] =
24
d W ∫0 ⎜ 2 ⎟ dx
⎝ ∂x ⎠
(4)
other end (x = L) is guided. The tip deflection at the guided
end is df. In this model, the possibility of any twisting Differentiating ϕ (x) two times and using it in
movement in other directions has been neglected. equation (4), the energy stored due to the bending component

4
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

of the strain is found to be Table 2. Comparison of linear and nonlinear stiffness values
obtained from the analytical calculation and the COMSOL model.
1 d 3W
( )
Eb d f = Y 3 df2
2 L
(5) Stiffness Analytical COMSOL
−1
Linear (k) 275.5 N m 185.66 N m−1
Hence, the component of spring reaction force (Fb) arising
Nonlinear (kn) 9.32·109 N m−3 3.56·109 N m−3
due to bending is calculated as
∂E b
Fb = −
∂d f The above equation is similar to the well-known non-
d
d3 linear spring force (F = kx + k n x 3). When the ratio df is less
= −Y 3 Wd f (6)
L than one, the cubic nonlinearity is insignificant. However, as
the tip deflection (df) becomes greater than the thickness of
In the case of stretching, the beam can be modelled using the d
the arm (d), or, in other words, if df is greater than one, only
same figure (figure 4(a)). The stretched or elongated length of
the beam is the effect of nonlinearity comes into play. The device was
designed in such a way that it exploits this analytically
L
derived conclusion of cubic nonlinearity, which appears due
L + ΔL = ∫0 1 + (ϕ′ (x ) )2 dx (7)
to large deformation and results in a wider bandwidth. The
Hence, the length of the stretched portion is given by variation of the spring force with deflection is shown in
figure 4(b). The same result is validated by finite element
L
ΔL = ∫0 ( )
1 + (ϕ′ (x ) )2 − 1 dx analysis using COMSOL. However, there is a difference
between the COMSOL and the analytical calculation of
L 2 stiffness at large deflection. The analytical model is based on
(ϕ′ (x ) )
≈ ∫0 2
dx a simple clamped-guided beam, and the derivation shows that
3d 2f the spring force can be expressed in terms of the sum of a
= (8) linear term and a cubic order term of the deflection of the
5L spring arm. This is a simplified model of the designed system,
The stretching component of the strain is formulated as fol- which did not consider any higher-order terms in the calcu-
lows lation, which also contribute to the spring force. COMSOL,
on the other hand, considers the entire structure in finite-sized
ΔL 3d 2f
Strain (v) = = (9) elements and numerically calculates the spring force con-
L 5L2 sidering in every term. The calculated values of linear stiff-
The energy stored due to this stretching component of strain ness (k) and nonlinear stiffness (kn) from the analytical model
can be written as and COMSOL are given in table 2. However, there is a 262%
difference between the values of nonlinear stiffness (kn) from
⎛ ⎞
ES = ∫ ⎝ 12 YdWν2⎠ dL
⎜ ⎟ the analytical model and COMSOL. To investigate the root of
this mismatch, we looked further into our COMSOL model.
⎛ 2 ⎞2 Under 0.2 N applied force, the deformation of the entire
1 ⎜3 df ⎟
= Y⎜ dWL central platform of our structure is not uniform (figure 5). The
2 ⎝ 5 L2 ⎟⎠ deflection at the middle point of the central platform is
9 dW 4 378.2 μm, whereas that at the two corners (the connection
= Y df (10) points between the beams and the central platform) is
50 L3
290.1 μm. This mismatch could arise from the softness of the
Similarly, the component of spring reaction force (Fs) arising spring material. In effect, the deflection of the guided ends of
due to stretching can be expressed by equation (11). the beams is not as high as it should be and resulted in a
∂E s 18 d negative effect on the ‘large deflection’ of the spring arms,
Fs = − = − Y 3 Wd f 3 (11)
∂d f 25 L which thus resulted in the mismatch.
For small deflections, the spring force is mainly domi-
Combining both the components, the total spring reaction nated by the linear term, and the nonlinear stretching-based
force is obtained analytically. force becomes significant only as the deflection becomes
F = Fb + Fs large. Using the COMSOL stiffness values for 50 μm, 150 μm
Y ⎡ 18 ⎤ and 300 μm deflections, the contributions due to linear force
= − 3 W ⎢ d 3d f + dd f 3⎥ term Fb are 0.0093 N, 0.028 N and 0.056 N, whereas that due
L ⎣ 25 ⎦
to nonlinear force term Fs are 0.0004 N, 0.012 N and 0.096 N.
⎡d 3⎤
Y 4⎢ f 18 ⎛ d f ⎞ ⎥ So, for the increasing deflection of the spring arm, the con-
= − 3 Wd + ⎜ ⎟ (12)
L ⎢⎣ d 25 ⎝ d ⎠ ⎥⎦ tributions of the nonlinear force term in the total force also
increase significantly and are 4.2% for 50 μm, 30% for
150 μm and 63.2% for 300 μm. We define U as the nonlinear

5
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

f(200)=0.2 Subdomain: Total displacement [m] Deformation: Displacement equation (14), we get the equation of motion of the oscillator
as
m ẍ + 2mρω 0 ẋ + kx + k nx3 + γ I = −mω2Z 0 sin ωt (15)
where ρ is the mechanical damping coefficient, z = Z 0 sin ωt
is the sinusoidal external perturbation and ω is the external
frequency. Since EM transduction is considered, the external
conversion circuit can be modelled using the following
equation
z
y x
LI ̇ + RI − γ ẋ = 0 (16)

Max: 3.782e-4
R is the total resistance, which includes coil resistance RC
Min: 0

and load resistance RL, and L is the electromagnetic induc-

tance. At low frequencies the effect of L is neglected.
0

0.5

1.5

2.5

3.5

x10-4
Figure 5. Deformation of the designed structure under 0.2 N applied
force. Inverse deflection of the corners is marked by the arrows. The
deflection at the middle point of the central platform is 378.2 μm,
whereas that at the two corners is 290.1 μm.
Since equation (14) describes a second-order non-
autonomous system, we can rewrite this system in an auton-
omous form by considering the phase ωt as the third state-
space variable. The system can now be represented in a three-
dimensional state space form

⎧ x1̇ = x 2 ;
⎪
( 1 1
Potential U = Eb + Es = 2 kx2 + 4 k n x4 analytically and ) ⎪ ⎛
⎨ x 2̇ = − ⎜ 2ρω 0 +
γ2 ⎞ k
⎟ x 2 − x1 −
kn 3
x1 + ω2Z 0 sin x 3;(17)
plotted the variation with deflection in the inset of figure 4(b). ⎝ mR ⎠
⎪ m m
U has a monostable configuration, which implies the hard- ⎪
ening effect due to large deformations. From equation (12), ⎩ x 3̇ = ω ;

the relation between linear stiffness constant k = ( YWd3

L3 ) and Equation (17) has been solved numerically using the 4th
nonlinear stiffness constant k n = ( 18YWd
25L3 ) is order Runge–Kutta method. The power dissipated in the
electrical domain for the electromechanical oscillator is given
kn 18 1 γ2
= (13) by P(t) = R ẋ 2 ; consequently, the average power has been
k 25 d 2 1 t+T
calculated as Pavg = T ∫ P(t′)dt′. Different parameter
t
Equation (13) provides an important implication for values used for numerical studies are already listed in table 1.
designing this kind of nonlinear system. The nonlinearity that We have performed all of our numerical simulations using the
arises due to stretching has inverse square dependence on the stiffness constants found from the COMSOL modelling, and
thickness of the spring arm. For this design, the values of k the results are compared with the experimental findings and
and kn are found to be 185.66 N m−1 and 3.56·109 N m−3 from are described in the results and discussions section.
the COMSOL modelling, and the resonant frequency is
( ω
)
f0 = 2π0 171.4 Hz. Using the linear stiffness derived from
the analytical calculations (275.5 N m−1), the linear resonant 4. Experimental study
frequency comes out to be 208.8 Hz. There is a 21.8% mis-
match between the linear resonant frequency derived from The fabricated EM energy harvesting prototype was tested
COMSOL and the analytics. using an experimental test set-up, as shown in figure 6. The
prototype was excited by a Brüel & Kjær LDS V455 Per-
3.2. Dynamical analysis manent Magnet Shaker. The shaker vibrates vertically. As the
assembled device also has pre-dominant out-of-plane move-
The governing equation for an electromechanical oscillator ment, for intended vibration, it was simply mounted on the
dU
with restoring force dx , external excitation Z̈ and mechanical shaker table. The shaker was controlled by a LDS Comet
damping ratio c that includes the electromagnetic conversion USB vibration control system, which was fed to a LDS PA
mechanism [36] is given by 1000 L power amplifier. The vibration controller generates
dU and transmits a sweeping sinusoidal signal from 5 Hz to
m ẍ + cẋ + + γ I = − m z̈ (14) 1000 Hz via the power amplifier to get a constant input
dx
acceleration level ranging between 0.05 g to 2 g (g = 10 m s2).
where m is the moving mass, γ is the electromagnetic cou- A miniature piezoelectric CCLD accelerometer (LDS 4394)
pling coefficient and I is the current flowing due to EM was placed on the shaker table to monitor the acceleration
transduction. level during the experiment. The output signal from the
As discussed in the previous section, the restoring force energy harvester was recorded using an Agilent DSO1014A
dU
is of the form F = − dx = kx + k n x3. So, substituting this in oscilloscope.

6
Smart Mater. Struct. 24 (2015) 015013
Figure 6. (a) Schematic of the experimental set-up. (b) The actual test set-up for experimental study of the system.
5. Results and discussions
5.1. Measurements at different acceleration levels
The fabricated prototype was tested under different peak
acceleration levels such as 0.05 g, 0.1 g, 0.5 g, 1 g, 1.5 g and
2 g (1 g = 10 m s2). The optimum load resistance (at which
maximum power is generated) is found to be 70 Ω (which is
almost equal to the coil resistance of the device) at 244 Hz for
a 1 g input acceleration level. The average load power is
0.35 μW, the RMS load voltage is 9.45 mV across the optimal
load (RL = 70 Ω) and the half-power bandwidth is 10 Hz at
this acceleration level. The RMS load voltages and load
powers for different input accelerations are shown in
figures 7(a) and (b), respectively. For all of these measure-
ments, the driving frequency of the input sinusoidal signal
was swept from 160 Hz to 360 Hz for 480 s in both the up and
down directions. The experimentally obtained data has been
compared with the result obtained from numerical modelling
of the system according to the theoretical framework, as
discussed in the previous section. For a very low input
acceleration of 0.05 g, the response is similar to that of a
linear resonant system. The linear resonant frequency is found
to be 170 Hz, which matches well with the value obtained
from the COMSOL model. As the response is the near linear
regime for an input acceleration of 0.05 g, a slow sweep of
8 mins has been performed around the resonance point, and
the open circuit quality factor (QOC) is calculated from the
resultant curve (inset of figure 7(a)); this factor is found to be
33.33. The mechanical damping factor QOC = ( 1
2ρ ) is cal-
culated (ρ = 0.015) using Qoc and is used subsequently for the
modelling purpose. With increasing input acceleration, the
system response becomes more and more non-resonant in
nature, and a wideband operational region is obtained.
However, this comes at the expense of the nonlinear hyster-
esis phenomenon in which two stable states (one high energy
7
D Mallick et al
state and the other low energy state) coexist. The jump phe-
nomenon is a characteristic feature of nonlinear oscillators.
When the external frequency is slowly varied, the steady state
amplitude of the output response changes discontinuously and
drastically at one point. This frequency is known as the jump
frequency or saddle node point. It is seen that for low
amplitude of the input accelerations, the voltage and power
spectra are quite well matched for experimental and modelled
curves. However, for high amplitudes of the input accelera-
tions (1.5 g, 2 g), the experimental output responses differ
from the modelled results. For 1.5 g and 2 g input accelera-
tions, the jump point during the up-sweep are at 259.6 Hz and
279 Hz, respectively, whereas those values in the experiment
are at 254 Hz and 260 Hz.
5.2. Comparison of saddle node bifurcation between modelling
and the experiment
The saddle node bifurcation point or the jump point during
the up-sweep of the frequency differs for the comparing
curves. The reason for this inconsistency can be twofold.
First, in our system the nonlinearity arises due to the
stretching strain of the spring arms. Certainly, some energy
is dissipated due to the effect of stretching, and dissipation
rises for large values of force (high levels of base excita-
tion). However, our model does not include such complex
dissipation terms. Additionally, for high base excitation
levels, the effect of nonlinearity is quite high; as a result, the
system becomes very sensitive to any kind of small per-
turbation. Such small perturbations are always present
within the real experimental system due to noise, whereas
numerically, we considered a deterministic system. As a
result of this, the system losses increase much more quickly
than expected for those values of the base accelerations, and
the saddle node point appears much earlier. Similar obser-
vations also have been made by Sebald et al [37] and Barton
et al [38]. To understand further the influence of noise, a
Smart Mater. Struct. 24 (2015) 015013
Figure 7. (a) Load RMS voltage; the inset shows the calculation of the Q factor around the linear resonant point and the (b) Average load
power. For all of the plots, RL was kept constant at 70 Ω. The solid lines are for the up-sweep, and the dotted lines are for the down-sweep.
basin of attractions for the system is plotted for four dif-
ferent frequencies in figure 8 using the numerical model.
The basin of attraction shows that when starting from a
different set of initial conditions at the vanishing driving
force, the solution leads to different branches in which the
high energy state is shown by red and the low energy state
is shown by blue. The fixed points of the high energy and
low energy states are also marked as FPH and FPL in the
plots. If noise is added to the system, it can be easily per-
turbed from FPH in the red region into the blue region and
ends up in FPL. At 190 Hz (figure 8(a)), i.e. at the start of
the hysteresis loop, the probability of staying in the high
energy state is much higher. This probability, however,
reduces as we increase the external frequency, as shown in
figure 8(b) for 205 Hz and in figure 8(c) for 215 Hz. At
225 Hz figure 8(d)), just before the jump occurs, there is
very little probability of staying in the high energy branch
of the response.
However, to justify our claim, we have included additive
Gaussian white noise, along with the harmonic excitation, to
our model and have performed a qualitative discussion using
the numerical simulations to exhibit the effect of dynamical
noise that becomes prominent in the experimental system
during high acceleration inputs. We have incorporated the
noise in our numerical simulation by using the Euler–Mar-
uyama method. The modified equation of motion can now be
8
D Mallick et al
written as
mẍ + 2mρω 0 ẋ + kx + k nx3 + γ I = −mω2 Z 0 sin ωt
+ Dξ (18)
where D is the noise parameter, ξ is the noise function with zero
expected value and the auto-correlation equals the delta func-
tion. The above equation is solved numerically for various noise
parameter values at 2 g input acceleration. Figure 9 shows that
the jump point in the up-sweep is sensitive to the noise, and it
becomes smaller as the noise parameter value is increased. For
zero noise the jump point is 279 Hz, whereas it becomes
264.6 Hz, 226.6 Hz and 211.5 Hz for D = 1 × 10−8 m2 / s3,
2.25 × 10−8 m2 / s3 and 4 × 10−8 m2 / s3, respectively. These
noise levels are very small compared to other reported noise
levels that are used to harvest energy [39]. Thus, the mismatch
between the saddle node bifurcation points between the
experiment and simulation could be due to the dynamical noise
that is present in experimental systems, possibly due to the
excitation of any other degrees of freedom.
For the 2 g acceleration level, the peak power that is
generated at the jump-point is 0.78 μW, whereas that value
from the simulation is 1.22 μW. However, the trend of power
variation with the frequency is similar to the experimentally
observed curves. It should also be noted that in the experi-
ment the jump-point comes earlier at 260 Hz as compared to
the 279 Hz of the model, and the generated power is 0.86 μW
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

Figure 8. Basin of attraction plots from the numerical model at (a) 190 Hz, (b) 205 Hz, (c) 215 Hz and (d) 225 Hz. The basin of attraction for
the low energy state is shown in blue, and that in the high energy state is shown in red. FPL and FPH are fixed points for the low energy and
high energy attractors, respectively.

Figure 10. Variation of jump frequency on the up-sweep and half-

Figure 9. Numerical simulation on the effect of noise in the jump
point at the 2 g acceleration level. With the increase in noise, the
saddle node bifurcation point during the up-sweep becomes lower.
at 260 Hz from the model, which is close to the experimental
value. The presence of noise cuts off the generated power at a
lower value than what could have been achieved in a noise-
free environment. The above-mentioned values for the 1 g
acceleration level are 0.35 μW and 0.38 μW, respectively.
Two important observations in this analysis are the sad-
dle node point (or jump point) during the up-sweep of the
frequency, as maximum power is normally generated at this
power bandwidth with input acceleration.
point, and the half-power bandwidth. It is seen from figure 10
that for higher values of accelerations, the half-power band-
width also deviates from the modelled values due to the
above-mentioned reasons.
5.3. Non-characteristic secondary peaks
Additionally, a non-characteristic secondary peak was
observed experimentally at around 196 Hz. To check the
possibility of an additional eigen mode due to the unbalance
of mass due to misalignment of the magnets on the

9
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

Figure 11. (a) Frequency response of the open circuit voltage for the extended frequency range is shown for an input acceleration of 1 g. Solid
rounds and hollow rounds are open circuit voltages during the up- and down-sweep, respectively. Time trajectories of open circuit voltage at
170 Hz, 185 Hz, 196 Hz, 210 Hz, 222 Hz, 244 Hz and 250 Hz during the (b) up- and (c) down-sweep are shown.

resonator, we shifted the location of the magnets relatively
by 0.5 mm / 1 mm / 2mm in either direction in the COM-
SOL model and simulated. The change in configuration of
the magnets made a small change in the frequency, but no
additional eigen mode was obtained from our simulation,
hence ruled out this possibility. To investigate further, we
recorded the time domain data of the open circuit voltage at
different frequencies for a constant input acceleration level
of 1 g. In figure 11(a), we have shown the open circuit
voltage up to an 800 Hz frequency sweep. Two small linear
resonant peaks appear at 384 and 484 Hz, respectively. The
time responses were recorded during both the up-
(figure 11(b)) and down-sweep (figure 11(c)) of the
frequency around the hysteresis region. The frequency was
swept at a very slow rate (for 16 mins in each sweep
direction), and the oscilloscope was paused to record the
signal when the required frequency was reached. Our
interest is the temporal voltage variation at 196 Hz where
the additional peak appears in the frequency sweep. It can
be seen that the time trajectory loses its symmetry around
the origin at this point, i.e. all the other time trajectories
remain symmetric if we flip them with respect to the zero
line on the voltage axis, but the trajectory at 196 Hz does
not remain so. Similar dynamic symmetry breaking phe-
nomena have been observed theoretically for the Duffing
oscillator by a number of researchers such as Parlitz et al

10
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

Figure 12. Chronological evolution of the symmetry breaking

observed near 196 Hz. The time trajectories of the open circuit
voltage from 192 Hz to 200 Hz are shown here.

[40] and Olson et al [41]. As we increase the frequency

(from 210 to 244 Hz) the peak to peak value of large
Figure 13. (a) Non-dimensional velocity at different input accelera-
amplitude oscillation increases until 250 Hz; then, the
tion levels. (b) Frequency response from the non-dimensional model
amplitude drops to the low energy branch. From the steady for F = 3, d = 0.2. In this case the SBB appears around ω = 1. The
state time traces during the up- and down-sweep of the evolution of the symmetry broken states are shown on the inset for
frequency, the coexistence of two stable states (high energy ω = 0.90, 0.95, 1.00 and 1.05, respectively, for a qualitative
and low energy) can be easily seen. In figure 12, a close comparison with our experimental observation.
examination of this symmetry broken state was performed
for a further qualitative analysis. Open circuit voltage tra- We consider a non-dimensional time τ = ωt , where
jectories show in detail the evolution of the symmetry k
ω= m
. Then,
broken state from its inception at 192 Hz to its dis-
appearance at 200 Hz. It is remarkable that a transition to a dx dx dt m m
x′ = = . = ẋ and x″ = x¨
dynamically symmetry broken state occurs in our mono- dτ dt dτ k k
stable (single valley potential well) device, which, accord- m m
ing to the best of the author’s knowledge, has not been Then, ωt = ωτ k = ω1 τ Where, ω1 = ω k we assume the
reported in the experimental literature so far. non-dimensionalized amplitude X = ax while considering the
k
value of a in such a way that a 2 = k3 . Hence, the non-
dimensional form of equation (14) can be written as
5.4. Non-dimensional model
X ″ + dX ′ + X + X 3 = f cos ( ω1τ ) (19)
However, our one degree of freedom dimensional model of c′
the Duffing oscillator did not produce this kind of symmetry where d = is the total damping and f = Z 0 a is the
mk
broken behaviour. To investigate it further, we derived a non- external force. Here, X’ refers to the differentiation with
dimensionalized model for the Duffing oscillator following respect to the non-dimensional time τ. Equation (19) is solved
the work of Parlitz et al [40]. Instead of considering the numerically, and the non-dimensional velocities, which are
mechanical and electrical damping separately, for our non- proportional to the induced voltage, are plotted in figure 13(a)
dimensional model, total damping has been taken into against the non-dimensional frequency (ω1 = 1 stands for the
account and represented as c′. natural frequency of the system ω 0 = 170 ) for various input

11
Smart Mater. Struct. 24 (2015) 015013
accelerations from f = 0.0019 (=0.05 g) to f = 0.0755 (=2 g).
The non-dimensional damping was found to be d = 0.031 for
the open circuit case. However, we carried out a further study
for the fundamental understanding of the observed symmetry
breaking in our system. For the higher-order force and
damping (f = 3, d = 0.2) our non-dimensional model shows the
presence of dynamic symmetry breaking bifurcation (SBB),
along with other secondary resonances (figure 13(b)). In that
case the dynamic symmetry breaking appears around ω = 1.
The chronological evolution of the symmetry breaking phe-
nomena is shown in the inset of figure 13(b) from ω = 0.9 to
1.05. There is a qualitative difference between the way in
which symmetry breaking appears in this numerical model
and the time histories we obtained experimentally in
figure 12; the relative heights of the maxima and minima in
the numerical and experimental time traces do not fully agree.
However, the symmetry broken state at ω = 1 is of a similar
nature to that found experimentally at 194 Hz. Our system
does not predict the existence of any additional secondary
peak from this one degree of freedom non-dimensional
model, and it is seen that the underlying trajectories of the
symmetry broken states differ from that of dynamical sym-
metry breaking. It is further noticed that the additional peak is
seen only during frequency up-sweep and disappears during
down-sweep. We also observe experimentally higher-order
linear modes in the output voltage at 384 Hz and 484 Hz
(figure 11(a)), which could be due to the different twisting
modes of the device in agreement with the eigen frequency
modes in COMSOL (figures 2(b) and (c)). These higher-order
resonances could also explain the peak observed at 196 Hz
due to the coupled interaction with the first nonlinear reso-
nance. However, we would then expect that this peak occurs
in both the up- and down-sweep directions.
Our assembled prototype has an in-built asymmetry in its
structure, as shown in figure 1(a). Magnets are placed on one
side of the resonator, and inductive transduction is carried out
from that side. So, damping is much more complex on this
side of the resonator compared to the opposite side. This
inherent asymmetry could have also led to the symmetry
broken states.
It is a unique advantage to use a nonlinear dynamical
system that can produce a large-amplitude oscillation at low
frequency components in addition to its inherent wideband
frequency response; normally, it is perceived that triggering
the symmetry breaking peaks for monostable systems is not
easy, even by moderately high acceleration levels [42].
However, in our system such peaks are observed in the
vicinity of the nonlinear peaks due to the first fundamental
mode; hence, sudden large amounts of power could be
obtained around those peaks. The height and position of those
secondary peaks in the frequency scale remain the same for
the increasing input acceleration levels, bringing it closer to
the nonlinear peaks for lower acceleration levels while
increasing the effective output power. It is seen that 15.8%,
24%, 40.6% and 110% of the peak power can be obtained at
the secondary peak for 2 g, 1.5 g, 1 g and 0.5 g accelerations,
respectively. So, for example, at 0.5 g acceleration, 0.1 μW is
generated at 214 Hz (nonlinear peak), whereas 0.11 μW is
12
D Mallick et al
obtained at 196 Hz (secondary peak), which are only 18 Hz
apart. Also, symmetry breaking states are useful in generating
significantly large power compared to symmetric states. It can
be observed that just before or after the secondary peak, when
time traces are symmetric, only a small power is obtained
compared to the symmetry broken peak point. For 1 g
acceleration, 0.14 μW is recorded at the secondary peak
(194 Hz), which would have been ∼3 times higher than the
extrapolated symmetric state at that point from the frequency
response. So, with further optimization, the symmetry
breaking phenomenon has the potential for exploitation to
achieve the enhanced performance of energy harvesting
devices by increasing the output power compared to sym-
metric states.
5.5. Comparative study
To establish the bandwidth enhancement ability of our spe-
cially designed energy harvester with fixed-fixed spring arms,
we have designed and fabricated a linear electromagnetic
energy harvester (figure 14(b)) exactly of the same volume
(0.78 cm3) as our nonlinear counterpart. The linear resonator
consists of four clamped-free configuration cantilever arms
(as shown in figure 14(a)). So, compared to our specially
designed clamped-guided cantilever, the restoring force
(F = kx ) in this case consists of only the linear term. Simi-
larly oriented magnets with exactly similar sizes are used as
the proof mass, whereas an identical coil is used to transduce
electrical power. A comparison between the frequency
responses of the open circuit voltage at 1 g acceleration of
both the linear and nonlinear prototype are plotted in
figure 14(c). By using a nonlinear stretching-based electro-
magnetic energy harvester of exactly the same size, a half-
power bandwidth that is 5 times wider is obtained.
A comparison between the output performances of
some of the reported nonlinear wideband works and our
results has been provided in table 3 in terms of half-power
bandwidth and power densities. The micro-scale devices
offer a large bandwidth compared to the macro-sized pro-
totypes due to the scaling effect and lower damping,
whereas the power densities of those are quite small. A
reasonably good bandwidth of operation is obtained with
our developed device among the macro-prototypes, whereas
the output power can be further enhanced with the optimi-
zation of the magnetic configuration.
6. Conclusions
In this paper, we have discussed the design, modelling and
experimental validation of a stretching strain-based nonlinear
EM energy harvester using FR4 as the spring material. An
analytical formulation of large deformation of the spring arms
shows that a cubic nonlinearity results from the contributions
of stretching, in addition to bending. This nonlinearity
increases the operational bandwidth of the VEH significantly,
up to 10 Hz at 1 g acceleration. Further, a numerical investi-
gation of a deterministic model based on a monostable
Smart Mater. Struct. 24 (2015) 015013 D Mallick et al

Figure 14. (a) Spring configuration of the linear prototype on FR4. (b) Assembled prototype for testing. (c) Comparison of the open circuit
voltage response for the linear and nonlinear prototypes under 1 g input acceleration.

Table 3. Comparison of output performances of some of the reported nonlinear wideband energy harvesters with our work.

Reference Prototype Scale Type of Nonlinearity Acceleration Bandwidth (Hz) Power Density (μW cm−3)
[43] Macro End-Stop (Hardening) 0.1 g 5 —
[32] Macro Magnetic Force (Softening) 0.3 g 3 —
[22] Macro Magnetic Force (Bistable) 0.35 g 1.5 —
[12] MEMS Spring Hardening 3g 25 0.056
[44] MEMS Stretching (Hardening) 0.8 g 28 0.045
[45] Macro Magnetic Force (Bistable) 0.3 g 7 —
[46] Macro Spring Hardening 0.038 g 0.6 0.57
This Work Macro Stretching (Hardening) 1g 10 0.45

Duffing oscillator support this increase in bandwidth. A
qualitative discussion based on the numerical model shows
that the presence of dynamical noise in the experimental
system limits this bandwidth. Additionally, a non-character-
istic secondary peak (at 196 Hz) was observed experimentally
in which 110% of the peak power at the nonlinear jump point
can be achieved at 0.5 g acceleration. Almost three times
more power is generated at this symmetry broken peak
compared to the nearby symmetric states, showing a high
potential of dynamical symmetry breaking in generating more
power compared to the symmetric states; this can be utilized
by further optimization. A detailed examination of this
secondary peak shows the SBB at that particular position.
However, an elementary dimensional and non-dimensional
one degree of freedom model could not justify the presence of
any such additional peak. It is inferred that the symmetry
broken secondary resonance may appear due to the dynamic
symmetry breaking of the oscillator or to the inherent asym-
metry of the built prototype.
This work shows the potential of a FR4-based, small
footprint, wide bandwidth, nonlinear EM harvester generating
a useful amount of power, which can be increased sig-
nificantly by further optimization of the magnet coil
assembly.

13
Smart Mater. Struct. 24 (2015) 015013
Acknowledgment
This work is financially supported by the Science Foundation
Ireland (SFI) Principal Investigator (PI) project on ‘Vibration
Energy Harvesting,’ grant no. SFI-11/PI/1201.
References
[1] Glynne-Jones P, Tudor M J, Beeby S P and White N M 2004
Sensors Actuators A 110 344–9
[2] Roundy S, Wright P K and Rabaey J M 2003 Energy
Scavenging for Wireless Sensor Networks (Berlin: Springer)
[3] Erturk A and Inman D J 2009 Smart Mater. Struct. 18 025009
[4] Elfrink R, Kamel T M, Goedbloed M, Matova S, Hohlfeld D,
Andel Y V and Schaijk R V 2009 J. Micromech. Microeng.
19 094005
[5] Mitcheson P D, Miao P, Stark B H, Yeatman E M, Holmes A S and
Green T C 2004 Sensors Actuators A 115 523–9
[6] Nguyen S D and Halvorsen E 2011 J. Microelectromech. Syst.
20 6
[7] Kulkarni S, Roy S, O’Donnell T, Beeby S and Tudor J 2006
J. Appl. Phys. 99 08P511
[8] Kulkarni S, Koukharenko E, Torah R, Tudor J, Beeby S,
O'Donnell T and Roy S 2008 Sensors Actuators A 145–146
336–42
[9] Beeby S P, Torah R N, Tudor M J, Glynne-Jones P,
O'Donnell T, Saha C R and Roy S 2007 J. Micromech.
Microeng. 17 1257–65
[10] Koukharenko E, Beeby S P, Tudor M J, White N M,
O’Donnell T, Saha C, Kulkarni S and Roy S Microsyst.
Technol. 12 1071–7
[11] Kazmierski T J and Beeby S 2011 Energy Harvesting Systems
Principles, Modeling and Applications (Berlin: Springer)
pp 9–11
[12] Liu H, Qian Y, Wang N and Lee C 2014 J. Microelectromech.
Syst. 23 740–9
[13] Liu H, Koh K H and Lee C 2014 Appl. Phys. Lett. 104 053901
[14] Liu H, Qian Y and Lee C 2013 Sensors Actuators A 204 37–43
[15] Zhu D, Beeby S, Tudor J and Harris N 2012 Smart Mater.
Struct. 21 075020
[16] Cheng S and Arnold D P 2010 J. Micromech. Microeng. 20 025015
[17] Roundy S and Zhang Y 2005 Proc. SPIE 5649, pp 373–84
[18] Leland S E and Wright P K Smart Mater. Struct. 15 1413–20
[19] Challa V R, Prasad M G, Shi Y and Fisher F T 2008 Smart
Mater. Struct. 17 015035
14
D Mallick et al
[20] Sari I, Balkan T and Kulah H 2008 Sensors Actuators A 145-
146 405–13
[21] Yang B, Lee C, Xiang W, Xie J, He J H, Kotlanka R K,
Low S P and Feng H 2009 J. Micromech. Microeng. 19
035001
[22] Erturk A, Hoffmann J and Inman D J 2009 Appl. Phys. Lett. 94
254102
[23] Mann B P and Sims N D 2009 J. Sound Vib. 319 515–30
[24] Marinkovic B and Koser H 2009 Appl. Phys. Lett. 94 103505
[25] Stanton S C, McGehee C C and Mann B P 2010 Physica D 239
640–53
[26] Ramlan R, Brennan M J, Mace B R and Kovacic I 2010
Nonlinear Dyn. 59 545–58
[27] Gammaitoni L, Neri I and Vocca H 2009 Appl. Phys. Lett. 94
164102
[28] Cottone F, Vocca H and Gammaitoni L 2009 Phys. Rev. Lett.
102 080601
[29] Cottone F, Gammaitoni L, Vocca H, Ferrari M and Ferrari V
2012 Smart Mater. Struct. 21 035021
[30] Marzencki M, Defosseux M and Basrour S 2009
J. Microelectromech. Syst. 18 1444
[31] Tang L, Yang Y and Soh C K 2010 J. Intell. Mater. Syst. and
Struc. 21 1867
[32] Stanton S C, McGehee C C and Mann B P 2009 Appl. Phys.
Lett. 95 174103
[33] Liu H, Lee C, Kobayashi T, Tay C J and Quan C 2012 Sensors
Actuators: A 186 242–8
[34] Hatipoglu G and Urey H 2010 Smart Mater. Struct. 19 015022
[35] www.magdev.co.uk/sintered-ndfeb-magnets
[36] Harne R L and Wang K W 2013 Smart Mater. Struct. 22
023001
[37] Sebald G, Kuwano H, Guyomar D and Ducharne B 2011
Smart Mater. Struct. 20 102001
[38] Barton D A W, Burrow S G and Clare L R 2010 J. Vib. Acoust.
132 021009
[39] Nguyen D S, Halvorsen E, Jensen G U and Vogl A 2010
J. Micromech. Microeng. 20 125009
[40] Parlitz U and Lauterborn W 1985 Phys. Lett. 107A 8
[41] Olson C L and Olson M G 1991 Am. J. Phys. 59 907
[42] Masana R and Daqaq M F 2012 J. Appl. Phys. 111 044501
[43] Soliman M S M, Abdel-Rahman E M, El-Saadany E F and
Mansour R R 2008 J. Micromech. Microeng. 18 115021
[44] Dai X, Miao X, Sui L, Zhou H, Zhao X and Ding G 2012 Appl.
Phys. Lett. 100 031902
[45] Ferrari M, Ferrari V, Guizzetti M, Ando B, Baglio S and
Trigona C Sensors Actuators A 162 425
[46] Saha C R, O’Donnell T, Wang N and McCloskey P 2008 Sens.
Actuators: A 147 248–53