Control Engineering Practice 15 (2007) 349361

On automating atomic force microscopes: An adaptive

control approach
Osamah M. El Rifai

, Kamal Youcef-Toumi
Department of Mechanical Engineering, Massachusetts Institute of Technology, 77 Massachusetts Ave. Rm. 3-348, Cambridge, MA 02139, USA
Received 19 March 2005; accepted 21 October 2005
Available online 19 December 2005
Abstract
In this paper, modeling and experimental results are given to reveal the structure of atomic force microscope (AFM) dynamics and
uncertainties which are strongly impacted by the users choice of scan and controller parameters. A robust adaptive controller is designed
to eliminate the need for the user to manually tune controller gains for different sample cantilever combinations and compensate for
uncertainties arising from the user choice of different scan parameters. The performance of the designed adaptive controller is studied in
simulation and veried through experiments. A substantial reduction in contact force can be achieved with the adaptive controller in
comparison with an integral controller.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Atomic force microscope (AFM); Adaptive control; Model
1. Introduction
Atomic force microscope (AFM) (Binning, Quate, &
Gerber, 1986) has been a key enabling tool in nano-sciences
and nanotechnology. AFM has been used in numerous
elds and applications including nano-manipulation (Sitti,
2003), imaging and dissection of DNA (Hansma, 1992),
measuring nano-scale friction and adhesion (Homola,
Israelachvili, Gee, & McGuiggan, 1989), studies of micro-
mechanics of single molecules (Fisher, Marszalek, Ober-
hauser, Carrion-Vazquez, & Fernandez, 1999), and in-
vestigation of nanotribology and nanomechanics of
MEMS devices (Bhushan, 1996).
The performance of AFM relies strongly on its dynamic
response and hence its control system. Several authors
(Schitter, Allgo wer, & Stemmer, 2004; Sebastian, Salapaka,
& Cleveland, 2003) have used linear robust control
strategies to control AFM dynamics. In Szuchi, Qingze,
and Devasia (2004), iterative control was used to compen-
sate for coupling dynamic effects in AFM.
The dynamics of AFM is strongly inuenced by the
users choice of cantilever, sample properties, environment,
and scan and controller parameters. As a result, the AFM
exhibits large level of uncertainties. Ultimately, the
objective is to be able to automatically select scan
parameters such as rate (or speed), force set-point, and
controller parameters in order to consistently achieve a
good image. Characteristics of achieving a good image
include probe and sample remain in contact during
scanning and the set-point error is maintained small at all
times. In addition, the control signal which is used to create
the sample image should be free from mode oscillations.
More so, high-frequency noise level in the image signal
should not be amplied by feedback. Consequently, system
uncertainties should be compensated for.
This paper presents initial work on automatically
selecting scan and controller parameters in order to
consistently achieve a good image. The approach is to
improve the performance of AFM in the presence of large
uncertainties by utilizing robust adaptive control to avoid
manual tuning of controller parameters. In addition,
guidelines for selecting scan parameters are provided. In
contrast to a xed robust control approach, adaptive
control can handle a larger range of parametric uncertainties
ARTICLE IN PRESS
www.elsevier.com/locate/conengprac
0967-0661/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2005.10.006

Corresponding author. Fax: +1 617 258 5802.

E-mail addresses: osamah@mit.edu (O. M. El Rifai), youcef@mit.edu
(K. Youcef-Toumi).
without leading to a conservative design or to the worst
case feedback instability.
The contributions of this paper are as follows: it reviews
earlier work (El-Rifai & Youcef-Toumi, 2000, 2001, 2002)
on modeling of AFM dynamics and builds on it to identify
an appropriate plant structure for adaptive control. A
robust adaptive control based on results in the literature
(Krstic, Kanellakopoulos, & Kokotovic, 1995; Xu & Yao,
1999; Yao & Tomizuka, 1997) is designed and tailored to
the AFM dynamics and practical implementation. In
addition, user specied scan parameters are tied with
control design and guidelines are given for scan parameter
selection. Moreover, a practical method for uncertainty
modeling and identication of range of uncertain para-
meters for use in the robust adaptive control design is
provided. Finally, both simulation and experimental results
are presented to demonstrate the performance of the
adaptive control.
The paper is organized as follows. Section 2 describes the
AFM and its operation. Probe sample interactions are
discussed in Section 2.1. In Section 2.2 a model for in-
contact dynamics of AFM is briey presented. In addition,
sources of uncertainties and their impact on dynamics are
discussed along with supporting experimental data. A
discussion on scan parameter selection and AFM perfor-
mance in presented in Section 3. Section 4 presents a robust
adaptive output control algorithm. The adaptive controller
is applied to AFM feedback system in Section 5. In
addition, discussion on estimating bounds on uncertain
parameters is given and both simulation and experimental
results are presented and discussed. Finally, summary and
concluding remarks are given in Section 6.
2. Atomic force microscope
An AFM, Fig. 1, has three main components, namely, a
scanner, a cantilever with a sharp probe, and a cantilever
deection sensor composed of a laser source and a position
sensitive diode (PSD). The scanner, typically a piezoelectric
tube, provides three-dimensional motion of the probe
relative to the sample. A piezo amplier is used to provide
the high-voltage needed for driving the piezoelectric
scanner. Information on sample topography or local
properties is obtained based on probesample interactions.
One of the main operating modes of AFM is contact mode.
In this mode, the probe presses against a sample exerting a
vertical force proportional to the cantilever deection. The
probe is then dragged against the sample along each scan
line in a raster fashion. The angle at the cantilevers free-
end is measured and fed back. During scanning, a control
system is used to maintain a constant angle by adjusting
the vertical deection of the piezoelectric scanner. Changes
in the deection of the scanner are taken as measure of
sample topography.
The performance of AFM is strongly inuenced by the
users choice of scan and controller parameters. Fig. 2(a)
shows some of the effects of scanning speed and controller
gains on the image of a calibration step. Higher gains result
in oscillations as the cantilever falls along the right edge of
the step, with peaks indicating momentary loss of contact
between the probe and the sample. However, the higher
gains improve tracking as the sharp left edge of the step is
resolved more accurately. On the other hand, choosing a
small contact force set-point reduces contact deformation
and friction; however, it reduces stability of the contact. As
seen in Fig. 2(b), the image generated with a small contact
force has erroneous height information due to loss of
contact between the probe and the sample.
In pursuit of consistently achieving an accurate and
artifact free image the AFM dynamics need to be under-
stood. Moreover, sources of uncertainties and their effects
on the dynamics need to be identied. This will be
discussed in the following two subsections.
2.1. Probesample interaction
The probesample interaction force is a nonlinear
function that depends on probesample separation, geo-
metry, operating environment, and probe and sample
material properties. A model for the contact interaction
force based on the work of Maugis was presented in El-
Rifai and Youcef-Toumi (2000). Out of contact van der
Waals forces are assumed. A nondimensional composite
forceseparation curve was generated using the model and
is shown in Fig. 3(a). The model can predict an instability
that has been observed in quasi-static experiments. This
quasi-static instability, as seen in the experimental results
of Fig. 3(b), occurs when an approaching/receding probe
jumps in/out of contact (pull-in/pull-off points), with the
sample surface corresponding to a sudden jump in
the contact area. The actual point of instability on the
forceseparation curve will depend on the stiffness of the
cantilever k
c
as shown in Fig. 3(a). The cantilever stiffness
is estimated from Fig. 3(b) as the slope of the line just after
the pull-off point. It can be seen from both gures that the
ARTICLE IN PRESS
X
Y
Z
Laser
PSD
Sample
Controller
Piezo Amplifier
Set-point
u
z
u
y
PSD
Scanner
Cantilever
Fig. 1. Schematic of the main components of an AFM.
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 350
model captures the main characteristics of the experimental
curve. However, the difference in the approach and retract
lines (i.e. hysteresis) is not captured by the model. This
behavior may be attributed to viscoelastic behavior of the
sample in addition to scanner hysteresis.
The pull-off force plays an important role in the choice
of contact force (feedback) set-point. The difference
between the set-point cantilever deection y
s
and pull-off
deection y
po
provides an estimate of the maximum output
error e
max
for the feedback system. If the output error
exceeds e
max
the probe would loose contact with the sample
and erroneous image data would result. However, the value
of e
max
is only an estimate as the pull-off deection is based
on quasi-steady conditions and the corresponding value in
dynamics situations may differ. In addition, the value of
y
po
may vary at different spots on the sample due to sample
variations and/or adhesion strength.
2.2. In-contact dynamics of atomic force microscopes
Previously, models were developed describing the
dynamics of AFM (El-Rifai & Youcef-Toumi, 2000,
2001, 2002). The main dynamics of interest are those
describing the vertical motion, i.e. from the scanner input
voltage u
z
to the output of the PSD y
PSD
. It has been
shown in El-Rifai and Youcef-Toumi (2002) that these
dynamics can be accurately described by the following
reduced order model:
y
PSD
%
d
z
p
s
2
2z
y
o
y
s d
z
p
o
2
y
k
y
1 d
y

s
2
2z
y
o
y
s o
2
y
u
z
G
p
su
z
, 1
where z
y
is the damping ratio, o
y
is the scanner rst
bending mode natural frequency, d
y
and d
z
p
are constants
which depend on probesample contact properties and
surface forces between the sample and probe.
Results presented in El-Rifai and Youcef-Toumi (2002)
have shown that changes in contact force set-point or
ARTICLE IN PRESS
Fig. 2. (a) Experimental AFM images of calibration steps: (a) high feedback gains, (b) small contact force set-point.
Force
Probe-sample
Separation
Penetration
Region
Out-of-Contact
Region
-k
c
N/m
-k
c
N/m
Pull-in point
Pull-off point

0 50 100 150 200 250 300 350 400 450 500

5
0
5
10
15
Probe-sample Separation, [nm

]
F
o
r
c
e
,

[
n

N
]
Pull-in Point
Pull-off Point
(a)
(b)
Fig. 3. (a) Simulated normalized forceseparation curve, (b) experimental
forceseparation curve.
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 351
amplitude of disturbance cause changes in DC gain of the
frequency response and changes in the frequency of the
system zeros in addition to changes in the resonance peak
of system modes. The effect of these uncertainties can be
seen in the sample experimental results shown in Figs. 4
and 5. The change in the polezero structure (polezero
ipping) in Fig. 4 due to changes in the contact properties
(nominal contact force set-point) can be explained from
Eq. (1). Changes in the contact properties affect the free
term of the zeros in Eq. (1) and hence their frequency.
However, the rst system resonance is the scanner bending
mode and as seen in Fig. 4 is not affected. Physically, this is
true since the probe sample forces are orders of magnitude
smaller than the force the scanner can provide ($10 nN vs.
$1 N).
3. On scan parameters and AFM performance
The performance of AFM feedback system is strongly
coupled to user specied scan parameters, users choice of
cantilever and the sample being investigated. The user can
freely specify scan size, scan rate, image resolution (number
of data points per scan line), force set-point, and controller
gains. Scan size and resolution depend on the sample being
imaged, hence, their values depend on the application.
The choice of contact force is motivated by four main
factors, namely, contact stresses, sample deformation,
lateral friction force between the probe and the sample,
and most importantly maintaining probesample contact
during scanning. Using the Maugis model of Section 2.1,
the force required to theoretically achieve zero deformation
between the probe and sample was computed. The ratio
between the zero-deformation force and the pull-off force
is plotted versus l
c
in Fig. 6(a). l
c
may be viewed as the
ratio of elastic deformation to the effective range
over which surface forces act. Large values of l
c
indicates
ARTICLE IN PRESS
10
2
10
3
10
2
10
1
10
0
Frequency, [Hz]
M
a
g
n
i
t
u
d
e
10
2
10
3
200
100
0
100
Frequency, [Hz]
P
h
a
s
e
,

[
d
e
g
r
e
e
]
36 nN
113 nN
Fig. 4. In-contact frequency response with a PDMS sample: 17 nm
amplitude for different set-points.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
F

=
0
/

F
p
u
l
l
-
o
f
f

c
-2 -1 0 1 2 3 4 5 6
0
0.5
1
1.5
2
Force
C
o
n
t
a
c
t

R
a
d
i
u
s

c
= 3

c
= 0.01
(a)
(b)
Fig. 6. (a) Ratio of zero-deformation force to pull-off force vs. l
c
, (b)
contact radius vs. contact force.
10
2
10
3
Frequency, [Hz]
10
2
10
3
Frequency, [Hz]
u
z
= 6 nm u
z
= 10 nm
10
1
10
0
M
a
g
n
i
t
u
d
e
200
100
0
P
h
a
s
e
,

[
d
e
g
r
e
e
]
Fig. 5. In-contact frequency response with a Glass sample: 14 nN set-
point, different disturbance amplitude.
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 352
a soft contact while small values corresponds to hard
contacts. It is seen that except for very hard contacts
(l
c
o0:1), the force ratio is about 0:89. This implies that in
order to achieve zero sample deformation, the force set-
point would be such that the cantilever is pulling the probe
away from the sample. This condition demands that during
scanning the set-point error be maintained very small at all
times, consequently requiring a high-bandwidth typically
beyond the systems mechanical resonances, and hence is
impractical. On the other hand, the set-point should be
chosen to minimize the friction force between the probe
and the sample. In Homola et al. (1989) experiments on
mica have shown that in the absence of wear, the average
friction force is directly proportional to contact area. The
contact area is non-zero as long as the probe and sample
remain in contact. Hence, there is no set-point which will
make the friction force zero. However, the larger the
contact force, the greater the contact area, Fig. 6(b), and
the higher the friction force will be. Accordingly, a small
contact force is desired. The minimum contact area for a
stable contact will also depend on the cantilever stiffness.
Based on the above arguments, it is seen that the contact
force should be selected as small as possible. The maximum
achievable feedback bandwidth should then be identied
for that particular set-point. The scan speed should be
chosen to be smaller or equal to the maximum value for
which contact is maintained at all times during scanning.
The maximum feedback bandwidth will depend on the
structure of the controller.
For a given scan size, the scan rate will dictate the lateral
speed at which the probe moves across the sample.
Assuming negligible probesample deformations, the
vertical speed of the probe v
z
is related to the constant
lateral scan speed v
scan
as shown in Fig. 7 by the local slope
at the point of probesample contact given by tana
Dz=Dx or tana Dz=Dy (depending on scan direction);
hence, v
z
is given by
v
z

v
scan
tana
. (2)
A lower bound on the slope is given using the included
angle of the probe a
p
. The assumption of negligible
contact deformations may be reasonable if the controller
manages to keep the set-point error small at all
times. Alternatively, if the ratio between contact to
cantilever stiffnesses is high, the assumption would be
reasonable.
The objective is to be able to automatically select scan
rate (or speed), force set-point, and controller parameters
in order to consistently achieve a good image. Character-
istics of achieving a good image include probe and sample
remain in contact during scanning, the set-point error is
maintained small at all times, and the control signal used to
create the image is free from mode oscillations. In addition,
system uncertainties due to cantilever and sample proper-
ties, variations in contact stiffness and dissipation, and
polezero ipping, should all be compensated for. More so,
high-frequency noise level in the image signal should not be
amplied by feedback.
4. Robust adaptive output control
Modeling and analysis of AFM dynamics presented in
El-Rifai and Youcef-Toumi (2000, 2001, 2002) have led
to identifying model structure and various sources of
uncertainties. This valuable information will be used in the
controller design to maximize possible gains in perfor-
mance. The controller to be designed is based only on
output measurement y
PSD
and have both adaptive and
robust parts to compensate for parametric and time-
varying bounded uncertainties and disturbances.
The controller which will be presented is based on
several main results in the literature (Krstic et al., 1995; Xu
& Yao, 1999; Yao & Tomizuka, 1997). However, several
simplications were made to ensure a practically viable
controller while achieving improved performance over a
xed controller.
The reduced-order plant model of Eq. (1) is used as the
design model and presented more generally as
y
b
n
s
n
b
n1
s
n1
b
0
s
n
a
n1
s
n1
a
0
u
z
, 3
where y is the plant output (y
PSD
in Eq. (1)). A rst
order low-pass lter, Eq. (4), is augmented with the plant
model.
u
z

p
s p
u. (4)
The resulting design model will have a relative degree of
one. The lter bandwidth can be selected to reduce the
effect of unmodeled dynamics and sensor noise by
ltering high-frequency components. Disturbance at the
plant input may arise from actuator nonlinearities
and hysteresis. Both input and output are included
in the model. Therefore, the design model considered for
ARTICLE IN PRESS
Sample
Probe
z
Scan direction
v
z
v
scan

p
Fig. 7. Relation between scan speed and vertical probe speed.
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 353
controller design is given as
y
b
0
n
s
n
b
0
n1
s
n1
b
0
0
s
n1
a
0
n1
s
n
a
0
0
u d
o

j
r
0
q
j
s
q1
r
0
q1
j
s
q
r
0
q0
j
s
n1
a
0
n1
s
n
a
0
0
d
i
j
. 5
Eq. (5) can be written in state space form as
_ x Ax a
0
y b
0
u a
0
d
o

j
r
0
j
d
j
, (6)
y x
1
d
o
; A
0 . . .
.
.
.
I
n1
_ _
, (7)
where d
o
is an output disturbance, d
j
is the jth input
disturbance, r
0
j
is vector of known constant parameters, and
a
0
and b
0
are vectors of unknown constant parameters. The
main assumptions involved are:
(A.1) The plant is minimum phase, i.e. B
0
s b
0
n
s
n

b
0
n1
s
n1
b
0
0
is Hurwitz.
(A.2) The sign of the high-frequency gain, sgnb
0
n
, is
known and taken as positive.
(A.3) Plant order n is known and plant relative degree
is 1.
(A.4) Bounds on the uncertain parameters and dis-
turbances are known.
It is important to realize the practical applicability of
assumptions (A.1)(A.4) to AFM. Experimental frequency
response data such as Figs. 4 and 5 demonstrate that in the
frequency range of interest (up to the rst two resonances)
the AFM dynamics are minimum phase. Non-minimum
phase behavior would be expected at higher frequencies. In
addition, with an augmented rst order lter as suggested
above, the dynamics are of relative degree 1. With the fact
that AFM dynamics are open loop stable in addition to
satisfying assumption (A.1) the sign of the high frequency
gain can be known from the sign of the DC gain. The DC
gain can be estimated as will be described in Section 5.1.
Guidelines on estimating bounds on uncertain parameters
and disturbances (assumption (A.4)) are described in
Section 5.1.
The details of the adaptive controller design and stability
results are given in Appendix A. The adaptive control law
has the following form:
u ^ Ra
1
a
2
, (8)
where a
1
is the adaptive part of the controller and is given
by Eq. (A.23) and a
2
is the robust part and is given by Eq.
(A.32). In addition, a parameter projection is used for the
update law for the unknown parameters
^
H and ^ R (see
Appendix A for denitions):
_
^
H Proj
^
H; Cz
1
K, (9)
_
^ R Proj^ R; sgnb
0
n
g
R
z
1
a
1
, (10)
Proj
^
Y
i
; g
ii
z
1
K
i

0;
^
Y
i
pY
i
min
and g
ii
z
1
K
i
o0;
0;
^
Y
i
XY
i
max
and g
ii
z
1
K
i
40;
g
ii
z
1
K
i
otherwise;
_

_
where g
ii
and g
R
are constants and ^ R is an estimate of
R 1=b
0
n
. The designed robust adaptive output controller
guarantees global boundedness and asymptotic regulation
in the presence and absence of bounded disturbances,
respectively. The performance of this controller is de-
scribed in the next section. It is shown that it compensates
for uncertainty, disturbances and eliminates the need for
manual tuning of controller gains.
5. Robust adaptive output control implementation
The dynamic order of the adaptive controller will depend
on the number of uncertain parameters and the number of
states of the model. The latter governs the order of
observer dynamics needed to estimate the unmeasured
states since only output measurement is utilized. Therefore,
in order to minimize the dynamic order of the controller a
simple low-order model is desired. The model chosen
consists of the scanner bending polezero pair as in Eq. (1).
In addition, a rst order lter is augmented with the model
with a pole at s p. To include integration action in the
controller a constant input disturbance

d
i
is included. The
projection law of Eq. (A.29) will provide a mechanism for
integrator anti-wind. Therefore, the system dynamic
equations become
_ x Ax a
0
y b
0
u r,
a
0
a
1
p; a
o
pa
1
; pa
o

T
; b
0
pb
2
; b
1
; b
o

T
,
r
0

d
i
; 0; 0
T
. (11)
The results of Section 4, can be applied directly to obtain
the adaptive controller. Note that the update law for

d
i
is
_
^

d
i
g
^

d
i
z
1
; )
^

d
i
g
^

d
i
_
z
1
dt, provides the desired integra-
tion action of the output error as part of the control signal.
The robust adaptive controller was implemented on a
commercial AFM from Quesant Instrument Corporation
model Q-250. The controller was coded in Simulink and
implemented on a dSPACE DSP board with 16-bit
Analog-to-Digital and 14-bit Digital-to-Analog converters.
In the following subsections guidelines on estimating
bound on uncertain parameters will be presented. In
addition, simulation and experimental results will be given
and discussed.
5.1. Selection of controller parameters
Estimates of bounds on the unknown parameters may be
obtained as follows. The scanner is excited in bending and
the PSD signal is collected. The rst resonance frequency
ARTICLE IN PRESS
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 354
o
r
1
may be identied automatically by software. Alter-
natively, the collected data may be graphically presented to
the user. The user then identies the rst resonance by
clicking on the rst peak in the amplitude data. The value
of o
r
1
can be stored for future usage. This procedure could
be done during calibration of the AFM and need not be
repeated before each scan experiment. Therefore, a
o
o
2
r
1
and a
1
2zo
r
1
. Damping ratio z of the scanner bending is
expected to be low as in typical mechanical structures say
0.1%. However, an accurate estimate is not needed. The
probe is then brought into contact with the sample and
then retracted until the contact is broken. Using this data
an estimate of the cantilever deection at the pull-off point
can be obtained. The estimate may be reduced by a factor
of safety to account for variations in the sample properties,
hence different pull-off forces at different points on the
sample surface. In addition, an estimate of the sensitivity of
u
z
to PSD signal can be obtained from probe retraction
data. The sensitivity is the DC gain of the transfer function
G
p
0. Due to possible variations in the DC gain based on
set-point and sample properties the range for the DC gain
may be increased to vary for example within an order of
magnitude. Therefore, b
o
G
p
0a
o
. The frequency of the
modal zeros (o
z
) has been shown to vary around the
resonance frequency. Hence, b
2
b
o
=o
z
. Finally,
b
1
2z
z
o
z
b
2
, where again z
z
is made to vary as for z.
Bounds on sample topography disturbance or equiva-
lently d
o
may be obtained by considering that while in-
contact, the cantilever deection would take the minimum
value of pull-off force and would vary during the short
response time of the feedback system ($4=o
b
) with a
maximum rate of j
_
d
o
j V
scan
= tana
p
. Hence, the bounds
are d
o
2 y
po
; y
s
V
scan
= tana
p
4=o
b
. Additional mar-
gins can be included to account for variations in the pull-
off point. The constant input disturbance

d
i
was intro-
duced to realize an integral action in the controller. For
practical implementation, a mechanism for integrator anti-
wind up need to be included. The mechanism could be
devised by appropriately choosing the range of the
unknown disturbance

d
i
. The integral part of the control
signal is u
I
^ R
^

d
i
^ Rg
d
i
_
z
1
dt, where, ^ R 2 1=b
2
max
; 1=b
2
min
.
Hence, u
min
=b
2
max
p
^

d
i
pu
max
=b
2
min
,where g

d
i
acts as an
integral gain. Therefore, the output of the integrator is
saturated at the minimum and maximum values of the
control signal u
min
and u
max
, respectively.
The observer gain vector k is chosen to place the
eigenvalues of the observer lters at 1 kHz. In general, the
observer eigenvalues should be faster than the rst bending
resonance. However, excessive observer bandwidth would
result in amplifying measurement noise and may increase
the effect of unmodeled dynamics on the feedback system.
The rst order lter cut off frequency p should also be
chosen similarly. It is interesting to note that the low-pass
lter can help reduce exciting unmodeled dynamics of the
AFM and reduce the effect of high-frequency measurement
noise. Therefore, the lter is practically valuable, in
addition it is simple and intuitive to tune. The gain k
z
1
dictates the decay rate of transient response with a time
constant of 1=2k
z
1
. So it should be chosen about the order
of the rst resonance response time.
The values of controller gains E
1
and E
2
(see Appendix A)
may be chosen small to reduce the nal set-point error, at
least theoretically. Choosing small values would increase
the feedback bandwidth, hence, may violate the validity of
the assumed design model. As a result, unmodeled
dynamics may cause feedback instability. Further work is
needed to understand how values of E
1
and E
2
may be
chosen in the presence of unmodeled dynamics. In the
simulation study which will be presented shortly, E
1
3
and E
2
k
z
1
were used.
The scan speed may be used to control the frequency
content of sample topography disturbance as seen by the
AFM feedback system. The adaptive controller contains an
inverse of the plant dynamics which is tuned in real time to
compensate for parametric uncertainties. Thus, if the
disturbance contains frequencies close to the plant complex
zeros, the response of the control signal is expected to be
oscillatory, as it strives to maintain the output error small.
Consequently, in contrast to I-control the scan speed
should be selected rst and then the contact force set-point
may be chosen accordingly. The minimum sample features
that the probe can detect laterally depends on the probe
radius of curvature. There maybe different denitions for
the size of the minimum detectable feature. However, from
a dynamics point of view it is desired that the probe track
the nest sample features it can detect without introducing
oscillations in the control signal. Hence, the scan speed
may be estimated by setting the scan speed such that the
probe (of radius R
p
) moves laterally a distance of 4R
p
within the response time of the scanner 1=o
r
1
Hz, hence,
the maximum v
scan
max
is given by
v
scan
max
4R
p
o
r
1
Hz. (12)
As has been discussed previously, within the response time
1=o
r
1
Hz an estimate of the maximum distance the
cantilever can deect from the set-point Dy
PSD
under the
sharpest feature is given by
jDy
PSD
j jy
s
y
po
j
v
scan
DC gain
o
r
1
tana
p
k
z
p
=u
z
, (13)
where DC gain is the estimate of sensitivity between u
z
and
y
PSD
, and k
z
p
=u
z
is the small amplitude linear calibration
factor between u
z
and the scanner vertical displacement z
p
.
Hence, the minimum force set-point can be estimated. The
actual set-point may be chosen slightly larger to allow for
variations in the pull-off force over the sample surface and
approximations made in the above equations.
5.2. Simulation results
In order to test the performance of the adaptive
controller simulations were performed. The model used
for simulations was far more realistic than the low-order
ARTICLE IN PRESS
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 355
design model. The model (El-Rifai & Youcef-Toumi,
2002), is tenth order with a relative degree of 6 in contrast
to the second order design model of Eq. (1). The simulation
model retained the rst two bending modes at 380 Hz and
3.4 kHz and the associated zeros at 430 Hz and 3.8 kHz. In
addition, a single mode for the cantilever was also included
at 12 kHz. The phase lag introduced by the dynamics of the
drive amplier was accounted for using two second order
lters with resonance frequency of 8 kHz and damping
ratio of 0.707. This matches the measured frequency
response of the amplier within the frequency range of
interest. Finally, a bandwidth-limited white noise source
was used as a input to the cantilever. This simulates the
effect of thermal and measurement noise on the cantilever
output. The resulting cantilever noise oscillations were 2

A
peak-to-peak at a bandwidth of 10 kHz. These values were
chosen to reect typical specications of commercial
AFMs. The sample shape used in the simulation is that
of the calibration steps. The sample contains both high-
aspect ratio features in addition to ne features, which
would allow testing the controller performance for
different types of samples.
The scanning simulations were performed using both the
adaptive and an integral controller (pure integrator). As
has been shown in El-Rifai and Youcef-Toumi (2004) an I-
controller provides higher bandwidth and improved
performance over PI and PID controllers commonly used
in commercial AFMs. The gain of the I-controller was
maximized to just before oscillations were observed in the
image. This allows a fair comparison between both
controllers. Fig. 8 shows the retrace image at scan speed
of 30 mm=s for both controllers. The results show that both
images portray the general shape correctly; however, some
of the details are lost with the I-controller. It is seen that
the sharpness of the edges is not tracked well. Hence, for
quantitative measurements they may result in large errors.
In Fig. 8, the slope of the right side of the step-like sample
is measured with a 10.4% error using the I-controller. The
set-point error is shown in Fig. 9, where it is seen that a
substantial reduction in output error is achieved with the
adaptive controller (4:7

A) compared to I-controller
(297

A). The probe scans the sample in a raster
fashion, hence the same scan line is scanned twice. The
rst pass is called the trace line and the second pass is
referred to as the retrace line. Fig. 9 shows the set-point
error for both the trace and retrace lines. It can be seen that
in the retrace (second pass) the tracking error is substan-
tially improved compared to the trace image as a result of
adaptation; 4:7

A versus 24

A. Hence, with the adaptive
controller it is advantageous to create the image from the
retrace lines. In addition, the so-called convolution errors
that result in the image due to the nite size of the probe
are important for sample features with characteristic
dimension similar to that of the probe or having a high-
aspect ratio. There are several procedures for deconvolving
the image (Villarubia, 1996) in order to recover an estimate
of the true sample shape. The quantitative errors produced
above by the I-controller would translate to more error and
uncertainties in the deconvolved sample shape estimate.
The above results also demonstrate that the noise level in
the image is low. This is a result of choosing a reasonable
bandwidth of 1 kHz for the observer and the rst order
lter dynamics.
The adaptive controller eliminates the need for the
user to manually tune the controller gains in a trial
and error fashion. This manual tuning is usually required
when different sample and/or cantilevers are used and in
cases when dynamics are changed due to changes in
the set-point shown in Figs. 4 and 5. To demonstrate
the performance of the adaptive controller under
such conditions, the cantilever stiffness was changed from
its previous value of k
c
0:04 to k
c
1:25 N=m. Fig. 10
shows the image and output error for both the adaptive
ARTICLE IN PRESS
0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
0
20
40
60
80
100
Time, [s]
H
e
i
g
h
t
,

[
n
m
]
I-Control
RAOC
Sample
Fig. 8. Image of a step scanned at 30 mm=s with cantilever stiffness
k
c
0:04 N=m.
0 50 100 150 200 250 300 350 400
400
300
200
100
0
100
200
300
Time, [ms]
S
e
t
-
p
o
i
n
t

E
r
r
o
r
,

[
A
n
g
s
t
r
o
m
]
0 200 400
10
0
10
20
30
Time, [ms]
I-Controller
RAOC
Fig. 9. Set-point error in

A for image of a step scanned at 30 mm=s with
cantilever stiffness k
c
0:04 N=m: trace and retrace images.
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 356
control and the I-controller. The I-controller gain was kept
at its previous value.
5.3. Experimental results
The developed robust adaptive control was implemented
on a commercial AFM and its performance is tested by
scanning calibration steps. The resulting 3-D image is
shown in Fig. 11. The steps were also scanned using a xed
integral controller with gain adjusted to just before
ARTICLE IN PRESS
Fig. 11. Experimental 3-D image of a step scanned at 28 mm=s.
0 500 1000 1500 2000 2500 3000 3500
0
20
40
60
80
100
120
Displacement, [nm]
H
e
i
g
h
t
,

[
n
m
] Adaptive Control
I-Control
0 0.05 0.1 0.15
30
20
10
0
10
20
30
Time, [s]
O
u
t
p
u
t

E
r
r
o
r
,

[
n
m
]
I-Control
Adaptive Control
(a)
(b)
Fig. 12. Experimental image of a step scanned at 28 mm=s: (a) height, (b)
set-point error.
0 0.05 0.1 0.15
5
0
5
10
15
20
x 10
3
Time, [s]
E
s
t
i
m
a
t
e

a

1
Fig. 13. Sample of parameter estimates: a
0
1
.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
20
0
20
40
60
80
100
Time, [s]
H
e
i
g
h
t
,

[
n
m
]
RAOC
I-Controller
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
40
30
20
10
0
10
20
30
Time, [s]
O
u
t
p
u
t

E
r
r
o
r

[
A
n
g
s
t
r
o
m
]
0 0.05 0.1 0.15 0.2
1
0.5
0
0.5
1
O
u
t
p
u
t

E
r
r
o
r

[
A
n
g
s
t
r
o
m
]
(a)
(b)
Fig. 10. Image of a step scanned at 30 mm=s with cantilever stiffness
k
c
1:25 N=m: (a) height, (b) set-point error.
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 357
oscillation starts appearing in the image. Cross section of
the images for both controllers are compared in Fig. 12(a).
The adaptive controller provides more accuracy of details
of the steps and slopes. This is evident in Fig. 12(b) where
the set-point error is substantially smaller for the adaptive
controller a factor of 7. The improvement is less than
predicted by simulations. This in part is due to the use of
14-bit (instead of 16-bit) Analog-to-Digital converter in
controller implementation. This has substantially increased
the noise level in the system (from about 2

Arms to more
than about 8

Arms). As a result, lter bandwidth and gains
used in the adaptive controller were modest hence reducing
the achievable performance. A sample parameter estimate
is shown in Fig. 13 for a
0
1
. Improvements to the
experimental setup will be addressed in future work.
6. Conclusions
The paper addressed automating the process of selecting
AFM scan and control parameters. Modeling and experi-
mental results are shown revealing the structure of AFM
dynamics and how it is affected by different choices of scan
and controller parameters. The resulting system uncertain-
ties are also discussed. Changes in contact force set-point
or amplitude of disturbance cause changes in DC gain of
the frequency response and changes in the frequency of the
system zeros in addition to changes in the resonance peak
of system modes. A robust adaptive output controller was
designed to guarantee global boundedness and asymptotic
regulation in the presence and absence of bounded
disturbances, respectively. The performance of the de-
signed adaptive controller is veried by simulating scan-
ning experiments. The adaptive controller eliminates the
need for the user to manually tune controller gains for
different sample cantilever combinations and compensates
for uncertainties arising from the user choice of different
scan parameters. In addition, a substantial reduction in
contact force and the retrace line set-point error can be
achieved with the adaptive controller in comparison with a
well-tuned integral controller. Experimental implementa-
tion conrmed improved performance. Future work will
focus on detailed experimental performance validation of
the algorithm under various sample-cantilever combination
and operating conditions.
Appendix A. Derivation of the adaptive controller
The plant model is given by Eq. (6). Since only output
measurements are available, an observer needs to be
designed to estimate the unmeasured states. The so-called
K-lters (Krstic et al., 1995) will be used for this task. Eq.
(6) can be rewritten as
_ x Ax Fy; u
T
Ha
0
d
o

j
r
0
j
d
j
, (A.1)
Fy; u
T
I
m1
u; I
n
y; H b
0
; a
0

T
. (A.2)
A.1. Observer design
Consider an observer of the form
_
^ x A^ x Fy; u
T
Hky e
T
1
^ x, (A.3)
where e
1
is the standard basis vector in R
n1
, and k
k
1
; . . . ; k
n

T
such that A
o
A ke
T
1
is Hurwitz, which
implies that 9 PA
o
A
T
o
P I, P P
T
40. However,
Eq. (A.3) involves the unknown parameter vector H, and
therefore cannot be directly used. Furthermore, consider
the lters n and X given by
_
n A
o
n ky, A:4
_
X
T
A
o
X
T
Fy; u
T
. A:5
The state estimation is given by ^ x n X
T
H. This may be
seen by considering
^ x n X
T
H, (A.6)
_
^ x
_
n
_
X
T
H, (A.7)
_
^ x A
o
n ky A
o
X
T
HFy; u
T
H A:8
A ke
T
1
n X
T
H ky Fy; u
T
H A:9
A^ x Fy; u
T
Hky e
T
1
^ x. A:10
The dynamics of estimation error e x ^ x are governed
by
_ e A
o
e a
0
kd
o

j
r
0
j
d
j
. (A.11)
Since A
o
is Hurwitz, hence, e is exponentially stable in the
absence of disturbances. With assumption of bounded
disturbances the disturbed dynamics of e are bounded but
the bound is unknown. The response of e to disturbances e
d
can be bounded by a known function. However, response
due to initial conditions e
o
is not known a priori. Consider
the columns of X
T
u
n
; . . . ; u
o
; N. The rst n 1 columns
are given by
_ u
j
A
o
u
j
e
nj
; j 0; . . . ; n. (A.12)
Also note that A
j
o
e
n
e
nj
; j 0; . . . ; n 1. The following
lter can be used in addition to a single algebraic equation
to evaluate u
j
:
_
l A
o
l e
n
u, A:13
u
j
A
j
o
l; j 0; . . . ; n. A:14
Similarly, N can be obtained as follows:
_ Z A
o
Z e
n
y, A:15
N A
n1
o
Z; . . . ; A
o
Z; Z. A:16
Using the identity A
n
o
e
n
k leads to n A
n
o
Z.
A.2. Controller design
Now that the observer lters are available, the controller
design may be carried out. First dene the output error
ARTICLE IN PRESS
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 358
z
1
y y
s
. The output dynamics are given by
_ z
1
x
2
a
0
n1
y b
0
n
u

j
r
0
n
j
d
j

_
d
o
a
0
n1
d
o
, (A.17)
x n X
T
He, (A.18)
x
2
x
2
X
T
2
He
x
2
, (A.19)
_ z
1
x
2
X
T
2
He
x
2
a
0
n1
y b
0
n
u

j
r
0
n
j
d
j

_
d
o
a
0
n1
d
o
, A:20
_ z
1
x
2
H
T
K e
x
2
b
0
n
u D, (A.21)
where K
T
X
T
2
e
T
n2
y, and D

j
r
0
n
j
d
j

_
d
o
a
0
n1
d
o
.
Dene
~
H H
^
H, R 1=b
n
, and ~ R R ^ R. Now consider
the choice of control law u (Yao & Tomizuka, 1997) given
by
u ^ Ra
1
a
2
(A.22)
and
b
0
n
^ Ra
1
b
0
n
^ Ra
1
a
1
a
1
a
1
~ Rb
0
n
a
1
.
Substituting the above control law into Eq. (A.21) results
in
_ z
1
x
2
H
T
K e
x
2
b
0
n
^ Ra
1
a
2
D,
_ z
1
x
2
H
T
K e
x
2
a
1
b
0
n
~ Ra
1
b
0
n
a
2
D.
Choosing a
1
as
a
1
x
2

^
H
T
K, A:23
_ z
1

~
H
T
K e
x
2
b
0
n
~ Ra
1
b
0
n
a
2
D. A:24
In order to prove closed loop stability with the above
control law, consider the Lyapunov function candidate V
1
given by
V
1

1
2
z
2
1

jb
n
j
2g
R
~ R
2

1
2
~
H
T
C
1
~
He
T
Pe, (A.25)
_
V
1
z
1

~
H
T
K e
x
2
b
0
n
~ Ra
1
b
0
n
a
2
D

jb
0
n
j
g
R
~ R
_
^ R
~
H
T
C
1
_
^
He
T
e. A:26
The adaptation laws have to ensure the following two
properties, namely:
~
H
T
z
1
K C
1
_
^
Hp0, (A.27)
^
Ht 2 P, (A.28)
where C C
T
is a constant gain matrix. Eq. (A.27) ensures
that the adaptation law will render the Lyapunov function
non-increasing. A parameter projection (Khalil, 1997) is
used for the update law as
_
^
H Proj
^
H; Cz
1
K, (A.29)
_
^ R sgnb
0
n
g
R
z
1
a
1
, (A.30)
Proj
^
Y
i
; g
ii
z
1
K
i

0;
^
Y
i
pY
i
min
and g
ii
z
1
K
i
o0;
0;
^
Y
i
XY
i
max
and g
ii
z
1
K
i
40;
g
ii
z
1
K
i
otherwise;
_

_
where g
ii
are constants.
_
V
1
pz
1
e
x
2o
b
n
a
2
D e
T
e. (A.31)
Now a
2
could be chosen to ensure that V
1
p0. a
2
is chosen
to compensate for disturbances D, uncertain parameter
estimation error, and state estimation error e
x
2o
. The
following form is chosen for a
2
:
a
2
k
z
1
=b
n
min
z
1
a
21
a
22
; k
z
1
40. (A.32)
Since bounds on uncertain parameters and disturbances
are known a
21
can be chosen to satisfy the following two
conditions:
z
1
b
0
n
a
21

~
H
T
K b
0
n
~ Ra
1
DpE
1
, (A.33)
z
1
a
21
o0, (A.34)
where E
1
40 is a design parameter. Eq. (A.33) is used to
bound the disturbances and uncertainties by a known
function to enable controlling transient performance at all
times. Eq. (A.34) is used to ensure that when adaptation is
on a
21
is dissipative in nature, so V
1
will remain negative
semidenite. To handle e
x
2
, a
22
is used and is chosen to
satisfy the following condition:
z
1
b
0
n
a
22
e
x
2
pE
2
e
2
x
2
, (A.35)
z
1
a
22
o0, (A.36)
where E
2
40 is a design parameter, and e
x
2
is bounded
(Eq. (A.11) and assumption (A.4)) but unknown. The
stability analysis can now be completed. With the proper
choice of a
21
and a
22
, Eq. (A.26) reduces to
_
V
1
pk
z
1
z
2
1
E
1
E
2
e
2
x
2
, (A.37)
_
V
1
p2k
z
1
V
1
E
1
E
2
e
2
x
2
, (A.38)
V
1
tpV
1
0 e
2k
z
1
t

_
t
0
e
2k
z
1
tt
E
1
E
2
e
2
x
2
t dt, A:39
V
1
tpV
1
0 e
2k
z
1
t

1
2k
z
1
1 e
2k
z
1
t

E
1
E
2
ke
x
2
k
2
max
e
T
e. A:40
Hence, z
1
y y
s
is bounded. From Eq. (A.11) e is
bounded and also H is bounded as a result of the
adaptation laws of Eqs. (A.29) and (A.30). The boundedness
ARTICLE IN PRESS
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 359
of z
1
leads to boundedness of Z, Eq. (A.15). To prove the
boundedness of l consider the transfer function represen-
tation of Eq. (A.13) given by
l
i
s
s
i1
k
1
s
i2
k
i1
s
n1
k
1
s
n
k
n1
u, (A.41)
y
b
0
s
a
0
s
u, (A.42)
l
i
s
s
i1
k
1
s
i2
k
i1
a
0
s
s
n1
k
1
s
n
k
n1
b
0
s
y.
Based on assumption (A.1), l
i
, i 2; . . . ; n 1, is also
bounded. By referring to Eq. (A.18) it is seen that the states
x are also bounded. Finally, u is also bounded. Conse-
quently, the closed loop system is stable and all signals are
bounded. Now assume that the disturbances D are zero,
and consider the following Lyapunov function candidate V
and its time derivative both given as
V
1
2
z
2
1

jb
0
n
j
2g
R
~ R
2

1
2
~
H
T
C
1
~
H
E
2
2
e
T
Pe, (A.43)
_
V z
1

~
H
T
K e
x
2
b
0
n
~ Ra
1
b
n
a
2
,

jb
n
j
g
R
~ R
_
^ R
~
H
T
C
1
_
^
HE
2
e
T
e. A:44
Substituting Eq. (A.32) into Eq. (A.44) leads to
_
Vpk
z
1
z
2
1

~
H
T
z
1
K C
1
_
^
H
~ Rb
0
n
z
1
a
1

1
g
R
_
^ R
_ _
b
0
n
z
1
a
21
z
1
e
x
2
b
0
n
a
22
E
2
e
T
e. A:45
The second and third terms on the right hand side of
inequality (A.45)p0 using Eqs. (A.27) and (A.30),
respectively. The fourth term o0 from Eq. (A.34) while
the fth term is pE
2
e
2
x
2
from Eq. (A.35). Hence, Eq. (A.45)
reduces to
_
Vpk
z
1
z
2
1
E
2
e
2
x
2
E
2
e
2
x
2
E
2

i1;3;4;...;n
e
2
i
, A:46
_
Vpk
z
1
z
2
1
. A:47
Therefore, z
1
is bounded. By inspecting Eq. (A.24), _ z
1
is
also bounded. Hence, by Barbalats lemma, z
1
! 0 as
t ! 1, achieving asymptotic regulation. In the absence of
disturbances the closed loop system achieves asymptotic
regulation. It is worth nothing that assumption (A.2) can
be relaxed using results of Nussbaum (1983). One possible
choice for the robust function a
21
is
gXj
~
H
T
K b
0
n
~ Ra
1
Dj,
a
21

g
b
0
n
min
tanh
0:2785gz
1
E
1
_ _
z
1
b
0
n
a
21

~
H
T
K b
0
n
~ Ra
1
D A:48
pz
1
g tanh
0:2785gz
1
E
1
_ _ _

~
H
T
K b
0
n
~ Ra
1
D
_
A:49
pjz
1
j j
~
H
T
K b
0
n
~ Ra
1
Dj gjz
1
j tanh
0:2785gz
1
E
1
_ _
pE
1
, A:50
tanh0 0; tanh1 1; tanh1 1, (A.51)
where for E40, x 2 R,
0pjxj x tanh
x
E
_ _
p0:2785E. (A.52)
Using the fact that when the argument of tanhx ! 1
as x ! 1 and tanhx % x for x % 0, alternatively, a
12
may be chosen as
gXj
~
H
T
K b
0
n
~ Ra
1
Dj
2
, (A.53)
a
21

gz
1
4b
n
0
min
E
1
. (A.54)
Similarly, a
22
can be chosen to satisfy Eqs. (A.35) and
(A.36), using the following form:
a
22

z
1
4b
n
min
E
2
, (A.55)
z
1
b
n
a
22
e
x
2o
pz
1

z
1
4E
2
e
x
2o
_ _
p
jz
1
j
2
4E
2
jz
1
j je
x
2o
j
max
. A:56
References
Bhushan, B. (1996). Nanotribology and nanomechanics of MEMS
devices. Proceedings of the ninth annual international workshop on
micro-electro-mechanical systems (pp. 9198).
Binning, G., Quate, C. F., & Gerber, C. (1986). Atomic force microscope.
Physical Review Letters, 56(9), 930933.
El-Rifai, O. M., & Youcef-Toumi, K. (2000). Dynamics of contact-mode
atomic force microscopes. Proceedings of the American control
conference (pp. 21182122). Chicago, Illinois, USA, June 2830.
El-Rifai, O. M., & Youcef-Toumi, K. (2001). Coupling in piezoelectric
tube scanners used in scanning probe microscopes. Proceedings of the
American control conference (pp. 32513255). Arlington, Virginia,
USA, June 2527.
El-Rifai, O. M., & Youcef-Toumi, K. (2002). Dynamics of atomic force
microscopes: Experiments and simulations. Proceedings of the IEEE
conference on control applications (pp. 11261131). Glasgow, Scotland,
September 1820.
El-Rifai, O. M., & Youcef-Toumi, K. (2004). Trade-offs and performance
limitations in mechatronic systems: A case study. Annual Reviews in
Control, 28(2), 181192.
ARTICLE IN PRESS
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 360
Fisher, T., Marszalek, P., Oberhauser, A., Carrion-Vazquez, M., &
Fernandez, J. (1999). The micro-mechanics of single molecules
studied with atomic force microscopy. Journal of Physiology, 520(1),
514.
Hansma, H. (1992). Reproducible imaging and dissection of plasmid
DNA under liquid using atomic force microscope. Science, 256,
11801184.
Homola, A. M., Israelachvili, J. N., Gee, M. L., & McGuiggan, P. M.
(1989). Measurements of and relation between the adhesion and
friction of 2 surfaces separated by molecularly thin liquid-lms.
Journal of Tribology, 111(4), 675682.
Khalil, H. (1997). Adaptive output feedback control of nonlinear systems
represented by input-output models. IEEE Transactions on Automatic
Control, 41(2), 177188.
Krstic, M., Kanellakopoulos, I., & Kokotovic, P. V. (1995). Nonlinear and
adaptive control design. New York: Wiley.
Nussbaum, R. D. (1983). Some remarks on a conjecture in parameter
adaptive control. Systems and Control Letters, 3, 243246.
Schitter, G., Allgo wer, F., & Stemmer, A. (2004). A new control strategy
for high-speed atomic force microscopy. Nanotechnology, 15, 108114.
Sebastian, A., Salapaka, M., & Cleveland, J. (2003). Robust control
approach to atomic force microscopy. Proceedings of the 42nd IEEE
conference on decision and control (pp. 34433444). Maui, Hawaii,
USA, December 912.
Sitti, M. (2003). Teleoperated and automatic nanomanipulation systems
using atomic force microscope probes. IEEE conference on decision and
control (pp. 21182123). Maui, Hawaii, December 912.
Szuchi, T., Qingze, Z., & Devasia, S. (2004). Iterative control of dynamics-
coupling effects in piezo-based nano-positioners for high-speed AFM.
Proceedings of the IEEE international conference on control applications
(pp. 711717). Taipei, Taiwan, September 24.
Villarubia, J. S. (1996). Scanned probe microscope tip characterization
without calibrated tip characterizers. Journal of Vacuum Science and
Technology B, 14(2), 15181521.
Xu, L., & Yao, B. (1999). Output feedback adaptive robust control of
uncertain linear systems with large disturbances. Proceedings of the
American control conference (pp. 556560). San Diego, California, USA.
Yao, B., & Tomizuka, M. (1997). Adaptive robust control of nonlinear
systems: Effective use of information. Proceedings of the 11th IFAC
symposium on system identication (pp. 913918).
ARTICLE IN PRESS
O. M. El Rifai, K. Youcef-Toumi / Control Engineering Practice 15 (2007) 349361 361