Proceedings of the

44th IEEE Conference on Decision and Control, and

WeIA18.3
the European Control Conference 2005
Seville, Spain, December 12-15, 2005

Modeling of an inkjet printhead for Iterative Learning Control

using bilaterally coupled multiports
M.B. Groot Wassink* and O.H. Bosgra* and D.J. Rixen** and S.H. Koekebakker***

Abstract— Modeling an inkjet printhead usually comprises a things, fluid-structure interaction, acoustics, fluid-mechanics,
trade-off between high accuracy and small computational load. free surface flow, and drop-formation. To deal with these
In face of the control purpose in mind, this paper presents a phenomena, almost all modelers resort to the use of finite
modeling strategy that breaks this boundary. For that purpose,
an inkjet printhead is modeled as series connection of bilaterally element packages and computational fluid dynamics (e.g.
coupled multiports using first principles only. By experimental [4]). The use of these packages induce large computational
validation, it is shown that the adopted model structure and time. Additionally, since various physical domains have to
parameters result in an accurate fit of the actual system. Next, be coupled, usually it also comprises the use of staggered
the resulting model and the accompanying physical insight are schemes (e.g. [7]). This further increases the computational
used to improve the actuation of an inkjet printhead. The
successful implementation of ILC based on the obtained model load. The resulting models, though accurate, are not suitable
not only shows the model’s applicability to control, but also for the application of control, e.g. ILC.
provides new input pulses that enable higher jetting frequencies.
In this paper, a modeling strategy is presented that
I. INTRODUCTION
does not suffer from the aforementioned computational
Inkjet technology is an important key-technology from burden yet offers the required accuracy. Using first principle
an industrial point of view. Its ability to deposit various modeling only, an inkjet channel is modeled as a system of
types of material on a substrate in certain patterns makes it bilaterally coupled multiports. The basic concepts employed
a very versatile technology. Not surprisingly, applications in this paper originate from the analysis of fluid transmission
of this technology cover a wide range from the traditional lines ([15], [16]). For the inkjet channel, these concepts
document printing to the manufacturing of electronics such are further elaborated and formulated within a multiport
as Flat Panel Displays (e.g. [1]), the production of organic framework ([9], [14]). The coupling of these subsystems
electronics (e.g. [2]), and the use for rapid prototyping and their physical domains is achieved by the application
(e.g. [3]). Each specific field of application imposes its of Redheffer’s star product ([8]). Furthermore, by using the
own performance requirements on the inkjet printhead. resulting model and the physical insight it provides, the
First, specifications in terms of timing, positioning, and actuation of an inkjet printhead is improved such that higher
volume have to be met. Often, these criteria are quite tight. jetting frequencies can be attained. This is accomplished by
Typically, one can think of an accuracy to be met in terms the implementation of lifted ILC ([18], [19], [20]).
of fractions of microseconds, micrometers and picoliters.
Second, requirements play a role concerning reproducibility The paper is organized as follows. First, a system de-
in face of aging, material and ink variations, and the like. scription is provided in Section II. Then, the theoretical
In the future, these performance criteria will become even modeling is discussed in Section III. In Section IV, the
tighter. The requirements for future applications motivate obtained model is validated using experimental results. Its
ongoing research into inkjet technology. applicability to ILC is demonstrated in Section V. Finally,
Section VI presents the conclusions and gives an outlook on
Modeling plays an essential role in the development of future work.
inkjet technology. For example, models can provide insight
into a specific printhead design that is required for its further II. SYSTEM DESCRIPTION
optimization. Also, they can be used for the synthesis of
x=0 x=L
input wave forms or controllers, such as for Iterative [reservoir]
piezo actuator/sensor
[nozzle]
Learning Control (ILC) ([11]). Though for the applications ink
Vpuls
mentioned above a relatively simple yet accurate model 1

suffices, modeling an inkjet printhead still is a challenging

2
task given its complex nature. It comprises, among other
3
* M.B. Groot Wassink and O.H. Bosgra are with Delft Center for
Systems and Control, Delft University of Technology, 2628 CD Delft, The 4
Nederlands [email protected]
** D.J. Rixen is with Delft University of Technology, 2628 CD Delft, t
5

The Netherlands
*** S.H. Koekebakker is with Océ-Technologies B.V., 5900 MA Venlo,
The Netherlands Fig. 1. A schematic view of an inkjet channel and its working principle

0-7803-9568-9/05/$20.00 ©2005 IEEE 4766

 A schematic view of a channel of an inkjet printhead divided into several subsystems that make up the complete
is depicted in Fig. 1. It consists of a channel of several model. Next, each subsystem is discussed shortly. Finally, the
millimeters, a nozzle, and a piezo-actuator. Typically, around coupling of these subsystems is discussed and the resulting
75 nozzles per inch are integrated in an array that forms a overall model is analyzed. Note that to obtain a model of a
printhead. To fire a droplet, a trapezoidal pulse is provided complete printhead, as many as appropriate channel models
to the piezo actuator. Then, ideally, the following occurs. can be merged. This will be addressed in this section as well.
To start with, a negative pressure wave is generated in
the channel by enlarging the volume in the channel (step A. The model structure
1). This pressure wave splits up and propagates in both
directions (step 2). These pressure waves are reflected at Piezo
Piezo
actuator
actuator
the reservoir that acts as an open end and at the nozzle
that acts as a closed end (step 3). Note that the pressure Drop
Reservoir Channel Connection Nozzle
wave reflecting at the nozzle is not large enough to result formation

in a droplet yet. Next, by decreasing the channel’s volume

to its original value a positive pressure wave is superposed Fig. 2. A schematic overview of the twoport model of an inkjet channel
on the reflected waves exactly when they are located in
the middle of the channel (step 4). Consequently, the wave In Fig. 2, a schematic overview is given of the twoport
travelling towards the reservoir is cancelled whereas the model of the inkjet channel depicted in Fig. 1. This partition
wave travelling towards the nozzle is amplified such that into subsystems is based on the specific design of the inkjet
it is large enough to result in a droplet (step 5). Typically, channel under consideration. To start with, the segment of
a printhead is operated at 10 kHz. It takes about 20 µ s to the channel that is actuated by the piezo-actuator is called
fire a droplet and around 200 µ s for the pressure waves to the channel block. It differs only from the connection block
be completely damped. Most of the damping occurs in the by the fact that the latter is not actuated. The reservoir forms
nozzle of an ink channel. the physical boundary of an inkjet channel and also forms
the boundary of the model. As all these three blocks can
According to [5], the piezo-actuator is concurrently be modeled using acoustics, they are referred to as acoustic
used as sensor. Physically, it senses the force that results path. The following two blocks, those of the nozzle and
from the pressure distribution in the channel acting on the droplet formation, are modeled using the basic equations in
piezo’s surface that borders the channel. This force creates fluid-mechanics. Together, they make up the fluidic path.
a charge on the piezo unit. Since only changes in charge are The last block is that of the piezo-actuator. Note that a
measured, in fact the time derivative of the instantaneous block representing the electrical part of the inkjet channel
present force is sensed. Furthermore, since the resulting is omitted since it does not contribute significantly to the
voltage drop of this current over a resistance is measured, dynamics of the printhead. If required though, it can be
we have that a voltage is the eventual sensor signal. For the incorporated straightforwardly.
trapezoidal pulse used for actuation, a typical sensor signal
is depicted in Fig. 9. The following remarks concerning this B. The acoustic path
sensor signal are noteworthy. First, due to the integrating The acoustic path consists of the channel, connection,
character of the sensor the resulting signal is an average and reservoir. To start with, it is assumed that for these
of the pressure that is present in a channel. Second, since blocks a one dimensional approach can be used. This
all the piezo’s are connected to the same substrate, the implies that only plane waves occur during operation of
actuation as well as sensing is influenced by structural an inkjet channel. Flow3D simulations confirm the validity
cross-talk. Despite all these facts, the current sensor signal of this approach. Second, it is assumed that there is no
can be regarded as representative for the jetting process. mean flow and that only small perturbations occur. This is a
Consequently, in this paper a model is derived that has the valid assumption, since the volume that is jetted is so small
pulse sent to the piezo as input and has the described sensor that this is hardly noticeable as mean flow in the channel.
signal as output. The accompanying measured transfer Finally, the dissipation also is assumed to be negligible.
function will serve as validation for the twoport model.
The modeling of the channel is treated first. To that pur-
During the derivation, it is assumed that a printhead pose, we start with the conservation of mass and momentum
behaves linearly. In [11], it was demonstrated that the non- for a channel with variable cross-section A(x,t):
linear effect of jetting a droplet is so small that from a
control perspective the operation of a printhead indeed can
∂ A(x,t)ρ (x,t) ∂ A(x,t)ρ (x,t)v(x,t)
be regarded as linear. + =0 (1)
∂t ∂x
III. THEORETICAL MODELING ∂ A(x,t)ρ (x,t)v(x,t) ∂ A(x,t)v2 (x,t)ρ (x,t) ∂ A(x,t)p(x,t)
+ + =0
∂t ∂x ∂x
In this section, the theoretical modeling of an inkjet (2)
channel is discussed. To start with, the inkjet channel is

4767
 Here, v(x,t), p(x,t), A(x,t), and ρ (x,t) are the velocity,




pressure, channel cross-section, and density, respectively. (1) ∂ z1 (x,t) v + cw 0 ∂ z1 (x,t)
+ = (10)
can be written as: ∂t z2 (x,t) 0 v − cw ∂ x z2 (x,t)
⎡ ⎤

v−cw  ∂ ρ v2 −p
∂ A(x,t)ρ (x,t) ∂ v(x,t) ∂ A(x,t)ρ (x,t) 2 A(x,t) + ⎣ 2ρ ⎦ ∂ A(x,t)
+ A(x,t)ρ (x,t) + v(x,t) =0 v+cw
∂ t ρ v2 −p ∂x
∂t ∂x ∂x 2 2ρ
(3)
Note that z1 (x,t) and z2 (x,t) have the physical dimension
and (2) as: of flow. After application of the Laplace transform while
assuming zero initial conditions and some reshuffling we
∂ A(x,t)ρ (x,t) ∂ v(x,t) ∂ v(x,t) obtain:
v(x,t) + A(x,t)ρ (x,t) + 2A(x,t)ρ (x,t)v(x,t)
∂t ∂t ∂x
(4)



∂ A(x,t)ρ (x,t) ∂ A(x,t)p(x,t) ∂ z1 (x, s) − cws+v 0 z1 (x, s)
+ v (x,t)
2
+ =0 = (11)
∂x ∂x ∂x z2 (x, s) 0 s
cw −v z2 (x, s)
⎡ ⎤ ⎡ ⎤
−s(cw −v) ρ v −p
2

Using the mass balance (3), (4) can be written as: +⎣ 2(cw +v) ⎦ A(x, s) + ⎣ 2ρ (cw +v) ⎦ ∂ A(x, s)
−s(cw +v) p−ρ v2 ∂x
2(cw −v) 2ρ (cw −v)

∂ v(x,t) ∂ v(x,t) ∂ A(x,t)p(x,t)

A(x,t)ρ (x,t) +A(x,t)ρ (x,t)v(x,t) + = 0 This renders the partial differential equation to an ordinary
∂t ∂x ∂x
(5) one that can be solved straightforwardly. Prior to that, the
forcing function A(x, s) is defined to be the product of A(x)
Now, A(x,t)v(x,t) is replaced by φ (x,t). Furthermore, it and A(s). A(x) represents the shape of the piezo-actuator
is assumed that the variations in density and pressure under when actuated. It is assumed that the piezo creates a uniform
adiabatic conditions are related through: cross-sectional variation K over its complete length. The
amplitude of this mode as well as the trajectory in time is
dp dρ 1 dp dρ 1 dp determined by A(s), though being Laplace transformed. The
| = c2w → = 2 and = 2 (6)
d ρ adiabatic dt cw dt dx cw dx solution to (11) can be computed straightforwardly. If A(x, s)
is replaced by KA(s), the first ordinary differential equation
where cw is the wave propagation velocity. After the (ODE) of (11) reads as:
appropriate substitutions, both equations (1) and (5) are lin- ∂ s −s(cw − v)
c2 ρ (x,t) z1 (x, s) + z1 (x, s) = KA(s) (12)
earized. Consequently, the coefficients v(x,t), wA(x,t) , ρA(x,t)
(x,t) , ∂x cw + v 2(cw + v)
p(x,t)
and ρ (x,t) are considered constant having the nominal values
c2w ρ A p
Using the solution at x = 0, z1 (0, s), as boundary condition,
v, A , ρ , and ρ , respectively. If the variables are from now the solution to (12) can be written as:
assumed to represent deviations from their nominal value,
the set of conservation laws can be written in vector form −sL KA(s)(cw − v)  −sL

as: z1 (L, s) = z1 (0, s)e cw +v − 1 − e cw +v (13)
2

 
 
c2w ρ
∂ p(x,t) v ∂ p(x,t) A similar computation reveals the solution for the second
+ A = (7)
∂t φ (x,t) A
v ∂ x φ (x,t) ODE of (11):
ρ



cw ρ
2
∂ 0 ∂
− A A(x,t) + v2 − p A(x,t) −sL KA(s)(cw + v)  −sL

v ∂t ρ ∂x z2 (0, s) = z2 (L, s)e cw −v + 1 − e cw −v (14)
2
 
c2w ρ
v
The eigenvalues of matrix A
A have the values The solution to (11) can be written in vector form as:
ρ v
λ1,2 = v ± cw . Its corresponding eigenvectors are:  


 −sL 

cw ρ 
 z1 (L, s) e cw +v 0 z1 (0, s)
− cwAρ z2 (0, s)
= −sL
z2 (L, s)
(15)
m1 = A m2 = (8) 0 cw −v
e
1 1 ⎡   ⎤
(cw −v) −sL
− 2  1−e
cw +v
+⎣ (cw +v) −sL
 ⎦ KA(s)
If we now define: 1 − e cw −v
2

  

z1 (x,t)  −1 A
2cw ρ
1
2 p(x,t) Now, (15) represents a twoport system as depicted as
= m1 m2 =
z2 (x,t) − 2cAw ρ 1
2
φ (x,t) in Fig. 3 (block 1). Here, Lch , Lco , Aco , A p represent the
(9) length of the channel, the length of the connection, its
cross-section, and the surface of the piezo bordering the
then (7) can be brought to the form: channel, respectively. Note that v = 0 since we assumed

4768
that there is no mean flow. As can be seen, the solution
admits a nice interpretation as travelling waves. To obtain p(s) ρ Ln A2n s + 8π µ Ln
Z(s) = = (17)
the original physical states p(x, s) and φ (x, s), the inverse v(s) A2n
transformation of (9) can be applied to the states z1 (x, s)
and z2 (x, s) (block 3). To compute output φ (s) given the input p(s), we get:
p(s) A3n
For the connection, a similar approach can be used, except φ (s) = An v(s) = An = (18)
Z(s) ρ Ln An s + 8π µ Ln
2
that the cross-section remains constant and can be left out of
the mass and momentum equations. The solution is depicted In Fig. 4, the fluidic path is depicted (block 5).
in Fig. 3 (block 2). The last subsystem of the acoustic path
is the reservoir. For the waves that come from the channel p
it acts as open end as the reservoir contains a large amount
of ink compared to that within an ink channel. In Fig. 3, the An
reservoir is taken into account (block 4). Z(s)

φ 5
F A(s)
−sLch Fig. 4. Graphical interpretation of the fluidic path
Ap Kcw (1 − e cw )
2

−sLch -
−sLco 2cw ρ p D. The actuation
z1 (0, s) e cw +v z1 (L, s) e cw +v Aco
Piezoelectric material acts as a twoport system quite
1 −cw ρ naturally. It becomes electrically polarized when subjected to
1 2 −1 3
Aco
4 mechanical strain and strained when subjected to an electric
−sLch −sLco field (e.g. [13]). For the piezoelectric actuator of an inkjet
z2 (0, s) e cw −v z2 (L, s) e cw −v 1 φ channel, the voltage V of the input pulse and the force F that
acts on the piezo surface bordering the inkjet channel are
Fig. 3. Graphical interpretation of the acoustic path selected as inputs. The electrical charge q measured and the
displacement u of the piezo actuator are regarded as outputs.
The behavior of the actuator then can be described by the
C. The fluidic path following set of equations:




The fluidic path consists of the nozzle and droplet q
=
C d V
(19)
formation. However, droplet formation is accounted for u d 1/k F
slightly differently compared to the other subsystems, since
it does not actually make part of an inkjet channel. Droplet with d, k, and C the piezoelectric charge coefficient, the
formation is computed as postprocessing step according stiffness of the piezo, and the piezo capacity, respectively.
to [6]. On the basis of the meniscus velocity trajectory The values of these parameters depend on the specific piezo
computed by an inkjet channel model, the droplet properties material used. However, their values are highly influenced
can be computed. The remainder of the fluidic path, the by the specific structure that surrounds the piezo actuator,
nozzle, is modeled as an impedance. Note that there are such as for example the substrate to which the actuator
many other options (e.g. [12]). is attached. To account for these effects, packages like
Ansys or Femlab can be invoked to compute the so called
During the derivation of the nozzle impedance it is as- effective parameters of the piezo actuator. Irrespective of
sumed that the nozzle is completely filled with ink at all the complexity of the actuator, as long as the effective
times. Though this assumption is violated in the actual parameters can be computed, (19) can be used.
system, the amount of ink is so small that this assumption
still holds. Newton’s second law for the nozzle, stated in The following remarks are noteworthy. First, note that
terms of p(s) and v(s), reads as: the fluid-structure interaction is taken into account via the
stiffness of the piezo. A displacement of the piezo results
8π µ Ln via the ink in a force sensed by the piezo. This force on
p(s)An = ρ An Ln sv(s) + v(s) (16)
An its turn causes a displacement of the piezo via the piezo’s
stiffness. This way, the fluid-structure accounted for. Second,
with An , Ln , and µ being the nozzle’s cross-section, length, in this paper it is assumed that the piezo actuator deforms
and viscosity, respectively. The friction due to the pressure according to its zeroth order mode. However, the approach
gradient across the nozzle is accounted for in the second presented in this paper admits using other modes as well.
term assuming a Poiseuille flow profile. According to the This can be achieved by using another forcing function A(x)
definition, the nozzle impedance can be written as: that represents the desired actuation shape. Third, cross-talk

4769
can also be accounted for. Basically, cross-talk results in v5 a1 v3 v3 a2 v1

a change of forcing function as well. This effect can be

b1 c1 b2 c2
quantified using for example a FEM package. Accounting for
cross-talk is important for the design of the printhead itself v6 d1 v4 v4 d2 v2

as well as for the design of input pulses or controllers. Also,

accounting for cross-talk is the key to generate a complete Fig. 7. The coupling of two subsystems using Redheffer’s star product
printhead model. Once the cross-talk is known, an array of
channels can be used instead of a single one. Finally, as can Finally, the resulting overall model is provided with addi-
be seen in Fig. 5, the piezo capacity is left out of the twoport tional viscous damping by shifting its poles away from the
model. In the measurement setup, this piezo is compensated imaginary axis without altering their natural frequency. Each
for. resonance frequency can be tuned individually. To analyze
q V the resulting system, the transfer function from a pulse sent
to the piezo to the sensor signal measured is taken as measure
d
6
d to compare it to the actual system. In Fig. 8, the theoretically
derived inkjet channel model is depicted in gray. In the
1/k next section, a further analysis and comparison to the actual
system is performed.
F u
IV. EXPERIMENTAL VALIDATION
Fig. 5. Graphical interpretation of the actuation path
In this section, the theoretically derived model is compared
to the actual system. To that purpose, the frequency response
of an inkjet channel is measured using the experimental
E. The complete twoport system setup. To identify the frequency response function (FRF)
In the previous subsections, the subsystems that make up from the piezo used as actuator to the piezo used as sensor,
the inkjet channel model have been discussed. To couple the a sine-sweep was used. The amplitude of the sinusoids
various subsystems, normally one uses a staggered scheme was chosen such that the inkjet channel was not jetting.
of some kind ([10]). For example, a sequential staggered The measured FRF includes among other things the piezo
scheme is depicted in Fig. 6 and comprises the following amplifier and a low-pass filter with a cut-off frequency of 500
steps. First, the response of system 1 to a certain input is kHz. The latter is used to eliminate the high-frequent piezo
computed. Second, this response is used as input for system behavior. The amplifier and low-pass filter cause a significant
2. Next, after having computed the response of system 2 phase drop at high frequencies.
to this input, system 1 can be provided with a new input a
timestep ∆t later. As long as the timestep is kept sufficiently
−15

−20
small, this is a valid approach. However, the computational −25
Magnitude [dB]

load is very high, especially if more than two systems are to −30

be coupled. −35

−40

−45
system 1
−50
5 6
10 10
Frequency [Hz]

200
system 2
timestep time 0
Phase [Deg.]

−200

Fig. 6. Sequential staggered solution procedure −400

−600
For most inkjet printhead models, a computational fluid
−800
dynamics package is used to model the behavior of the 5
10
6
10

acoustic and fluidic path and a finite element package for

Fig. 8. FRF from the piezo actuator to the piezo sensor; measured (black)
the actuation path. In this paper, only first principle modeling and theoretically derived (gray)
has been used such that analytical expressions are available
for the formulation of twoports. One major advantage of the To compare the measured and theoretically derived FRF,
presented approach is that the use of staggered scheme can the low-pass filter is incorporated in the frequency response
be avoided. Instead, the Redheffer star product can be used. of the theoretical model as well. Also, since the derivative of
Given two twoports as depicted in Fig. 7, the coupled system the electrical charge is measured instead of the charge itself,
can then be computed according to: a differentiator is added to the theoretical model. In Fig. 8,
   the resulting theoretical and measured FRF are depicted.
v1 a2 (1 − c1 b2 )−1 a1 c2 + a2 c1 (1 − b2 c1 )−1 d2 v5
= Based on Fig. 8, it is concluded that the actual system is
v6 b1 + d1 b2 (1 − c1 b2 )−1 a1 d1 (1 − b2 c1 )−1 d2 v2
(20) modeled accurately. Except for some small differences,

4770
 1
seems a logical choice as control strategy to accomplish
0.8 active damping of the fluid-mechanics.

0.6
In this paper, use is made of ILC in the lifted setting, see
[18], [19]. The accompanying control structure is depicted in
0.4
Fig. 10. The mapping H is the impulse response matrix of
Sensor signal [V]

0.2 the process sensitivity and S is the sensitivity. In case of the

printhead, only open-loop dynamics are present such that H
0
just equals the plant and S the identity. The learning matrix
is represented by L and may be non-causal and time-varying.
z−1 is one trial delay operator and can be seen as memory
−0.2

−0.4 block. The trial length N equals 1000 corresponding with

the sample rate of 10 MHz and the DOD frequency of 10
−0.6
0 50 100 150 kHz. Signal uk is a vector containing the system’s inputs or
Time [µs]
states of the ILC system. Signal yk is the system output, ŷre f
Fig. 9. Sensor signal resulting from a standard trapezoidal pulse (black the reference trajectory, and d the disturbance. ek is the error
dotted, scaled); measured (black) and simulated response (gray) output. The update of the system’s input is ∆uk and uk+1 is
the input for the next trial k + 1. At the k-th trial, signal uk is
provided to the system, resulting in the integrated output yk .
the location and magnitude of the resonance frequencies The output yk is then subtracted from the reference yre f to
match. In Fig. 9, the measured and simulated response to a obtain the error ek . Based on this error, the learning controller
standard trapezoidal actuation pulse are depicted. Despite computes the adjustments to the input ∆uk that, added to the
the fact that this pulse results in a droplet being jetted by previous input, forms the input for the next trail uk+1 .
the inkjet channel, the measured response is predicted by
the theoretical model quite accurately. Altogether, for the yref-d

control purpose in mind, the theoretically derived model S

offers an excellent starting point. The remaining model uk+1 uk yk
yref
+ -1 +
inaccuracies can be handled by ILC quite well. z I H -
+

To improve the current theoretical model, the following uk ek

refinements can be applied to it. First, the deformation of L
the actuator A(x) can be modeled more accurately such that
it matches the actual deformation of the piezo more closely. Fig. 10. Lifted ILC control structure in the trial domain
Second, the quite coarse model used for the nozzle block can
be improved. Note that these refinements are not required The control goal is to achieve higher jetting frequencies
for a successful application of ILC, as is demonstrated in while preserving the droplet’s properties. By choosing the
section V. current sensor signal as controlled variable, the control goal
has to be translated in a suitable reference trajectory for
V. ILC CONTROL FOR AN INKJET CHANNEL the pressure in the channel. This trajectory now consists of
In Section II, the functioning of an inkjet channel was two parts. During the first part, the trajectory is constructed
explained using the concept of travelling waves. Based on such that a droplet of certain predefined properties results.
this interpretation, a theoretical model has been derived in During the second part of the reference trajectory, the
Section III that matches the measured FRF of an actual fluid-mechanics is brought to a rest as soon as possible after
system accurately. Apparently, a relatively simple physical the firing of a droplet. By doing so, the conditions for higher
model suffices to that purpose. This model and the physical jetting frequencies are created. To actually construct the
insight it provides can be now be used for control. reference trajectory the following procedure is adopted. As
starting point for the construction of the reference trajectory,
An important issue in inkjet technology is the strive for the one depicted in Fig. 9 is used. The first part up to the
higher jetting frequencies. As discussed in Section II, inkjet firing of a droplet is left unchanged (up to 20 µ s), whereas
printheads function thanks to resonating fluid-mechanics. during the remainder of the time the pressure is forced to a
At the same time, however, these limit the attainable jetting rest by speeding up the damping. The pressure is not forced
frequency since the residual vibrations must be damped first to a rest immediately but somehow gradually. This is done
prior to jetting a next droplet. Typically, it takes around to ensure the refill of the nozzle with ink and to avoid too
200 µ s for the pressure waves to be damped such that high actuation voltages. The resulting reference trajectory is
a next droplet can be jetted: this to guarantee the same depicted in Fig. 11.
droplet properties each time a channel is actuated. Given
the highly repetitive character of the jetting process, ILC For a detailed description of ILC controller synthesis, one

4771
is referred to [11], [18], [19]. The resulting ILC controller been employed within an ILC framework to design input
is implemented on the experimental setup. Next to the ref- wave forms that leave the droplet formation undisturbed
erence trajectory, the sensor signal resulting from a standard while bringing the channel to a rest quickly after droplet
trapezoidal and the learned ILC pulse are shown in Fig. 11. ejection. Consequently, a higher jetting frequency can be
The accompanying actuation pulses are shown in Fig. 12. attained. The successful implementation of ILC based on
Based on Fig. 11, the conclusion is drawn that the reference the obtained model demonstrates the model’s applicability
trajectory is attained. As discussed, the first part of reference to control. Further improvement of the model as well as its
trajectory up to the firing of a droplet is the same as is use for control is subject to ongoing research.
realized by the standard trapezoidal pulse. Consequently,
VII. ACKNOWLEDGMENTS
it is not surprisingly that the learned ILC pulse resembles
the standard trapezoidal pulse for the first part. After that, The authors are grateful to Océ-Technologies B.V. for their
the ILC controller adjusts the actuation pulse such that the support of the research reported here.
fluid-mechanics follow the desired trajectory. Apparently, the R EFERENCES
theoretical model is accurate enough to ensure convergence
[1] R. Bennett, ”Precision Industrial Ink-jet Printing Technology for Full
and thus can be used to design input wave forms that enable Color PLED display and TFT-LCD manufacturing”, in Proc. of the
higher jetting frequencies. 3rd International Display Manufacturing Conf., Taipei, Japan, 2003.
[2] P.W. Cooley and D.B. Wallace and B.V. Antohe, ”Application of Ink-
1.2 jet Printing Technology to BioMEMS and Microfluidic Systems”, in
1
Proc. SPIE Microfluidics and BioMEMS Conference, San Francisco,
CA, October 2001.
0.8 [3] W. Voit et al., Application of inkjet technology for the deposition of
0.6
magnetic nanoparticles to form micron-scale structures, IEEE Proc.
Science Measurement and Technology, vol. 150, 2003, pp. 252-256.
0.4
[4] H. Wijshoff, Free surface flow and acousto-elastic interaction in piezo
Sensor signal [V]

0.2 inkjet, Proc. NSTI Nanotechnology Conf. and Trade Show, vol. 2, pp.
215-218, Boston, MA, 2004.
0
[5] European Patent 1 378 360 A1
−0.2 [6] J.F. Dijksman, Hydrodynamics of small tubular pumps, Journal of
Fluid Mechanics, vol 139, pp. 173-191, 1984.
−0.4
[7] H. Seitz and J. Heinzl, Modelling of a microfluidic device with piezo-
−0.6 electric actuators, Journal of Micromechanics and Microengineering,
vol. 14, pp. 1140-1147, 2004.
−0.8
0 10 20 30 40 50
Time [µs]
60 70 80 90 100 [8] R.M. Redheffer, On a certain linear fractional transformation, Journal
of Math. and Phys., vol. 39, pp. 269-286, 1960.
[9] H.M. Paynter, Analysis and Design of Engineering Systems, The MIT
Fig. 11. Sensor signal; without ILC (black), with ILC (gray), and the
Press, Cambridge, MA, 1961.
reference trajectory (black dotted)
[10] C.A. Felippa and K.C. Park and C. Farhat, Partitioned Analysis
of Coupled Systems, Computer Methods in Applied Mechanics and
Engineering, vol. 190, issues 24-25, pp. 3247-3270, 2001.
35 [11] M.B. Groot Wassink and N.J.M. Bosch and O.H. Bosgra and S.H.
Koekebakker, Enabling higher jetting frequencies for an inkjet print-
30
head using Iterative Learning Control, accepted for Conf. Contr.
25 Applications, Toronto, Canada, 2005.
[12] J.E. Fromm, Numerical Calculations of the Fluid Dynamics of Drop-
20
on-Demand Jets, IBM Journal of Research and Development, vol 28,
15
no. 3, pp. 322-333, 1984.
Input [V]

[13] J.W. Waanders, Piezoelectric Ceramics. Properties and applications,

10 Philips Components, Eindhoven, 1991.
[14] D. Karnopp and R.C. Rosenberg, Analysis and Simulation of Multiport
5
Systems. The Graph Approach To Physical System Dynamics, The
0 MIT Press, Cambridge, MA, 1968.
[15] F.T. Brown, The Transient Response of Fluid Lines, Journal of Basic
−5 Engineering, December 1962, pp. 547-553.
−10
[16] F.T. Brown, Step Responses of Liquid Lines with Frequency-
0 10 20 30 40 50
Time [µs]
60 70 80 90 100
Dependent Effects of Viscosity, Transactions of the ASME, June 1965,
pp. 504-510.
[17] C. van Oosten, O.H. Bosgra, B.G. Dijkstra, Reducing Residual Vibra-
Fig. 12. Actuation signal; standard trapezoidal (black) and ILC pulse (gray)
tions through Iterative Learning Control with Application to a Wafer
Stage, in Proc. American Control Conference 2004, Boston, Ma, 2004.
[18] R. Tousain, E. van der Meché, O.H. Bosgra, Design strategies for
VI. CONCLUSIONS AND RECOMMENDATIONS iterative learning control based on optimal control, in Proc. 40th IEEE
Conf. on Decision and Control, pp. 4463-4468, Orlando, December
In this paper, an inkjet printhead model has been derived 2001.
using a series connection of bilaterally coupled multiports. [19] B.G. Dijkstra, Iterative Learning Control with applications to a wafer-
It has been shown that by restricting to first principle stage, PhD thesis, Delft University of Technoloy, 2004.
[20] K.L. Moore, Itertive Learning Control for Deterministic Systems,
modeling only and the use of the Redheffer star product Advances in Industrial Control. Springer-Verlag, 1993.
for the coupling of the subsystems, the computational load
could be kept low while achieving good accuracy. The
resulting model and the physical insight it provides have

4772