2014 American Control Conference (ACC)

June 4-6, 2014. Portland, Oregon, USA

Exploiting Additional Actuators and Sensors

for Nano-Positioning Robust Motion Control

Robbert van Herpen, Tom Oomen, Edward Kikken, Marc van de Wal, Wouter Aangenent, Maarten Steinbuch

Wafer stage Metrology frame

Abstract— The ongoing need for miniaturization and an
increase of throughput in IC-manufacturing is obstructed by s
performance limitations in motion control of nano-positioning
wafer stages. These limitations are imposed by flexible dy-
namical behavior, associated with structural deformations of
the nano-positioning stages. The aim of this research is to
investigate limits on achievable performance in a conventional s
control configuration and to mitigate these limits through the
use of additional actuators and sensors. To this end, a systematic
framework for control design using additional actuators and
sensors in the generalized plant configuration is presented,
which leads to a well-posed H∞ -control optimization problem
that extends conventional design approaches in a natural way
and exploits physical insight to address structural deformations Fig. 1. A prototype lightweight wafer stage, enabling high accelerations.
in weighting filter design. Through an experimental confronta-
tion of the design framework with a prototype next-generation encompasses two degree-of-freedom control configurations,
nano-positioning motion system, successful performance en- as for example used in inferential control to account for
hancement beyond the conventional limits is demonstrated.
unmeasured performance variables [7].
I. I NTRODUCTION Although fundamental performance limitations are well-
understood, the standard formulation does not immediately
Increasing demands for high machine throughput of nano-
indicate that performance limitations in motion control result
positioning systems necessitates explicit control of flexible
from structural deformations. For example, flexible dynam-
dynamical behavior. An important application domain where
ical behavior does not necessarily induce right-half plane
positioning with nanometer accuracy plays a central role
poles and zeros, to which bandwidth limitations have been
is the lithography step in IC-manufacturing, see, e.g., [1].
associated, see [8, Sect. 5.7, 5.9]. Nevertheless, it is known
Herein, a desired IC-pattern is etched into a photosensitive
from practical experience that the attainable bandwidths in
layer on a substrate, using a device known as a wafer scanner.
motion control are dominantly restricted by flexible dynam-
For this purpose, the substrate, a silicon wafer, is placed
ical behavior, as is confirmed by the results in [9], [10],
on top of a wafer stage, see Fig. 1, to position it with
[11]. To formalize this limitation, it is essential to explicitly
respect to the light source. On the one hand, a nanometer
involve model uncertainty in a robust control setting, as is
positioning accuracy has to be achieved in view of the
also argued in [12] and [8, Sect. 5.3.2].
required feature size of current ICs. On the other hand, it is
vital for market viability of wafer scanners to achieve a high The main contribution of this paper is a systematic robust
machine throughput. Since a high throughput demands for control design approach to beat conventional limits through
large accelerations of the wafer stage, next-generation stages active control of flexible dynamical behavior using additional
are designed to be lightweight, as is further motivated in actuators and sensors. Based on physical insight, new weight-
[2]. However, lightweight stages tend to undergo structural ing filters for H∞ control design are proposed. The design
deformations upon large accelerations, which obstruct the framework is confronted with an industrial high-precision
intended positioning accuracy. Therefore, it is essential to motion system, demonstrating an enhancement of motion
explicitly control these structural deformations. performance beyond conventional bandwidth limitations.
Flexible dynamical behavior of lightweight positioning The paper is organized as follows. In Sect. II, a generalized
stages leads to limitations on the achievable motion perfor- plant framework is adopted. To address these limits in an
mance. Fundamental performance limitations for feedback extended control configuration, theoretical aspects of the
control are formulated in terms of sensitivity integrals, which generalized plant are addressed in Sect. III. Moreover, based
depend on right-half plane poles and zeros of the system, [3], on physical insight, an approach is proposed that exploits
[4]. Extensions of these classical results towards the general- the additional actuators and sensors to counteract undesired
ized plant framework for model-based control are provided deformations on the system. In Sect. V, this design approach
in, e.g., [5], [6]. In turn, the generalized plant framework is formalized in a loop-shaping framework for weighting
filter design in H∞ control. Subsequently, this framework
Van Herpen, Oomen, Kikken, and Steinbuch are with the Eind- is confronted with an industrial high-precision positioning
hoven University of Technology, Department of Mechanical Engineer- device in Sect. VII, where performance enhancement beyond
ing, Control Systems Technology Group, Eindhoven, The Netherlands
(email: [email protected]). Van de Wal and Aangenent conventional limitations is successfully shown. Conclusions
are with ASML Research, Veldhoven, The Netherlands. are drawn in Sect. VIII.

978-1-4799-3271-9/$31.00 ©2014 AACC 984

 w z w z w z
wext zext
G Gext G̃ext
u y u y
C u y
Cext Cext
uext yext uext yext
(a) Standard configuration. (b) Configuration with additional control inputs and (c) Setup corresponding to the proposed design pro-
outputs, violating assumptions for synthesis. cedure in this paper.
Fig. 2. Generalized plant configuration.

II. P ROBLEM FORMULATION AND BASIC CONCEPT Definition 2 (Extended configuration) Consider Fig. 2(b),
A. Experimental setup where for Gext it holds that  
  I 0
In this paper, the prototype lightweight wafer stage in I 0 0
Gext 0 I  = G.
Fig. 1 is considered. This device is used to position a wafer, 0 I 0
0 0
on top of which ICs are to be produced, with respect to a
light source. To increase productivity, the IC manufacturing The performance Jext that is achieved by a controller Cext
industry is currently moving towards the use of wafers with a in the extended configuration is given by
diameter of 450 mm. Therefore, the wafer stage in Fig. 1 has Jext (Cext ) := k F` (Gext , Cext )k.
dimensions 600 × 600 × 60 mm. Yet, it weighs 13.5 kg only
to enable high accelerations. Due to the lightweight design, It is straightforward to show that increasing the number of
active control of structural deformations is necessary. control inputs and outputs enables performance optimization.
B. Problem formulation Lemma 3 (Performance enhancement) Consider the gen-
As is argued in Sect. I, flexible dynamical behavior eralized plant configurations in Def. 1 and Def. 2. Then,
introduces performance limitations in motion systems. In tra- min Jext (Cext ) ≤ J (C ? ).
ditional control configurations, flexible dynamics introduce Cext
bandwidth limitations in controlling the rigid-body motion
Although conceptually straightforward, the design of a
degrees-of-freedom (DOFs) of the wafer stage.
norm-based controller that achieves performance beyond
The goal of this paper is to go beyond the limits of a
conventional limits is non-trivial. It requires the formulation
traditional control configuration by explicitly accounting for
of a well-posed H∞ synthesis problem, in which the freedom
flexible dynamical behavior in control design. In particular,
provided by additional inputs and outputs is exploited in
to counteract undesired deformations of the wafer stage,
judicious weighting filter design. In Sect. III and IV, these
control using a large number of actuators and sensors is
aspects are further addressed through the development of a
investigated. Therefore, the system has been designed to pro-
framework for control synthesis, based on Fig. 2(c), in which
vide abundant opportunities for actuator placement; Lorentz-
additional actuators and sensors are exploited.
actuators can be easily mounted at 81 distinct locations
underneath the stage in Fig. 1. With respect to sensors, linear III. T HEORETICAL ASPECTS FOR CONTROL SYNTHESIS
encoders with nanometer resolution are available for position IN THE EXTENDED GENERALIZED PLANT
measurements at all four corners of the stage. Indeed, the use Although it is clear that additional actuators and sen-
of additional control inputs and outputs provides freedom for sors enable performance enhancement, cf. Sect. II-C, the
performance enhancement, as is formalized next. configuration in Fig. 2(b) is unsuitable for the synthesis
of an optimal controller that exploits new control inputs
C. Extra inputs-outputs enable performance improvement and outputs in a meaningful way. To further explain this,
In this section, the merits of adding extra inputs and standard assumptions for H2 and H∞ optimal control, see
outputs for control are analyzed systematically. To this end, [14, Sect. III.A] and [13, Sect. 16.2.3], are reviewed.
the generalized plant configuration in Fig. 2(a) is considered,
see, e.g., [13], [8]. Herein, w are exogenous inputs and z ex- Assumption 4 Consider the standard generalized plant con-
ogenous outputs of the generalized plant G. The controller C figuration in Fig. 2(a). Let G have a state-space realization
is designed by minimizing a certain norm between w and z.
 
A B1 B2
G =  C1 0 D12 .
Definition 1 (Standard configuration) Consider Fig. 2(a). C2 D21 0
The performance J achieved by a controller C is given by The following assumptions are standard in optimal control.
J (C) := k F` (G, C)k.

T C
  
• Assumption 4.1: D12 1 D12 = 0 I .
In addition, the optimal nominal controller C ? is defined as 
B1
  
0
T
• Assumption 4.2: D21 = .
C ? = arg min J (C). D21 I
C

Next, the generalized plant is extended with additional Assumption 4 ensures a bounded control effort in optimal
inputs uext and outputs yext for control, see Fig. 2(b). control design. In particular, it follows from Ass. 4.1 that D12

985
 −100 −100
kc −125 −125

m1 m2 −150 −150

Magnitude (dB)
F1 Fdiff k −175 −175
1 2 3 1 2 3
10 10 10 10 10 10
x1 b x2
−100 −100
xdiff −125 −125

Fig. 3. Benchmark system with additional actuator and sensor, which −150 −150
enables a change of the system’s stiffness through high-gain control kc . −175 −175
1 2 3 1 2 3
10 10 10 10 10 10
is a tall matrix with full column rank. Hence, the perfor- Frequency (Hz) Pflex
mance outputs z = C1 x + D12 u include a nonsingular Fig. 4. Extended non-collocated plant Pncol,ext (solid) and equivalent
penalty on the control signals u. Similarly, from Ass. 4.2 plant Pncol,eq (dash-dotted) obtained after closing feedback loop for Pflex .
it follows that D21 is a wide matrix with full row rank.
F1 x2
Thus, the sensors y = C2 x + D21 w are all affected by the
performance inputs w.
Pncol,ext
Fdiff xdiff
Typically, Ass. 4 is violated for the extended generalized Fd
plant Gext in Fig. 2(b). In particular, z does not include + −
Cflex Pncol,eq
a penalty on the additional control signals uext , while w
Fig. 5. Equivalent plant Pncol,eq obtained by control of Pflex .
does in general not affect the additional sensors yext . As
a consequence, Fig. 2(b) generally leads to an ill-posed is that the transfer function F1 7→ x2 is non-collated. This
optimal control design problem, in which unbounded signals leads to performance limitations as robustness has to be
uext , yext may result in an attempt to minimize the norm J accounted for. To anticipate on the design framework in
of the transfer matrix from w to z. For well-posed optimal Sect. V, two steps are taken to address conventional per-
control synthesis, the setup in Fig. 2(c) is considered, where formance limitations using the extra plant input and output.
• zext includes a nonsingular penalty on uext , and i) A controller Cflex is designed for the flexible mode Pflex .
• wext affects yext in a nonsingular way. The aim is to enhance the stiffness of the system, such
Under these assumptions, Fig. 2(c) is effectively in the that flexible dynamics prevail at higher frequencies.
standard format of Fig. 2(a) and enables the exploitation of ii) The equivalent non-collocated plant Pncol,eq is deter-
additional actuators and sensors in optimal control. mined, see Fig. 5. This equivalent plant provides new
The setup in Fig. 2(c) extends Fig. 2(b) with auxiliary freedom for enhancement of motion performance, e.g.,
performance variables wext —zext that in essence serve to by enabling higher bandwidths using PID-type control.
bound the control effort of uext —yext in H∞ synthesis. In view of i), the following result is motivated by an
Herein it is not yet clear how to select wext and zext , and investigation of the physics of the benchmark motion system.
how to dictate suitable performance goals for these auxiliary Result 6 Consider Fig. 5, with static controller Cflex = kc .
channels such that performance beyond conventional limits At low frequencies, undesired structural deformations xdiff ,
is achieved. Next, a suitable design philosophy is developed which result from disturbance forces Fd that excite flexible
based on physical insight. dynamics of the system, are reduced with a factor (1 + kkc ).
IV. E XPLOITING PHYSICAL INSIGHT TOWARDS
To support Result 6, observe that
CONTROL DESIGN BEYOND THE CONVENTIONAL LIMITS
2
In this section, physical insight is exploited to develop a Pflex,cont : Fd 7→ xdiff = (2)
design strategy that exploits additional actuators and sensors + 2bs + 2(k + kc ) ms2
to go beyond conventional performance limits for flexible As a consequence, low-frequent disturbances of magnitude
motion systems. To develop this insight, the simple model Fd result in a deformation xdiff ≈ Fd /(k + kc ), compared
in Fig. 3 is considered. This model is chosen such that it to xdiff ≈ Fd /k without the controller Cflex , see (1) and (2).
captures all relevant aspects of the wafer stage system in From a physical point of view, an additional stiffness kc is
Fig. 1 in view of the development of weighting functions. added to the system through control, see Fig. 3.
Here, the position xdiff := (x2 − x1 ) can be measured. In In view of ii), in Pncol,eq
q , the resonance frequency mani-
addition, a force Fdiff := F2 − F1 can be applied. This leads fests itself at fres,eq ≈ 2(k+k c)
/(2π) Hz. From a physical
m
to the following definition of the extended plant. point of view, by actively counteracting internal deformations
Definition 5 The extended non-collocated plant is defined as of the system, the controlled system behaves as a rigid-body
        to a much higher frequency. Indeed, this provides freedom to
u y F1 x2
Pncol,ext : 7→ = 7→ . enhance the control bandwidth. This concept is used in the
uext yext Fdiff xdiff
next section to design weighting filters for H∞ synthesis.
A Bode diagram of Pncol,ext is shown in Fig. 4. Herein, V. T RANSLATION TO WEIGHTING FILTERS FOR
2 MULTIVARIABLE H∞ SYNTHESIS
Pflex : uext 7→ yext = (1)
ms2 + 2bs + 2k In this section, a novel weighting filter design framework
describes the system’s flexible mode. The key point in Fig. 4 for multivariable H∞ control synthesis is proposed, in which

986
w1 w2 z2 z1 s1 a1
+ + z
+ Rx
Cs W1 Pext W2 s2 s4
- s3
Ry
a3
a4 a2

Fig. 6. Feedback configuration with Ps indicated by dashed box.

(a) Actuator-sensor configuration. (b) Torsion bending mode.
the philosophy presented in Sect. IV is applied to beat Fig. 7. Inputs and outputs that can be used to compensate for torsion of
conventional performance limits using additional actuators the wafer stage.
and sensors. This control design philosophy has a direct Weighting filter design for flexible motion systems
interpretation in terms of a desired loop-shape P C. A natural
framework to incorporate loop-shape requirements in H∞ D1) Consider the flexible behavior Pflex : uext 7→ yext , see
synthesis is the shaped-plant framework, [15, Chap. 6]. Ass. 8. Shape the desired loop-gain of the flexible plant
A. Specifying loop-shape requirements in H∞ control design Pflex,s = W2,flex Pflex W1,flex (3)
The shaped-plant framework is based on the feedback by designing weighting filters W1,flex and W2,flex .
configuration in Fig. 6. Herein, Pext is obtained by extending Essentially, W1,flex and W2,flex should incorporate a
a rigid-body configuration with additional inputs and outputs. high gain, hereby dictating a reduction of internal
system deformations through active feedback control.
Definition 7 The extended plant is defined as This concept is further explained by Result 6 in Sect. II.
   
u y D2) First, construct the equivalent plant
Pext : ũ 7→ ỹ , ũ = , ỹ = .
uext yext
Peq = F` (Pext , W2,flex W1,flex ), (4)
Assumption 8 Without loss of generality, inputs and outputs which represents the behavior in the motion DOFs
u, y ∈ Rn are used to control n motion DOFs of the system, under control of the system’s flexibilities, see Ass. 8.
whereas uext , yext ∈ Rm are m additional inputs and outputs Subsequently, shape the loop-gain
for control of flexible dynamical behavior, see Sect. IV.
Peq,s = W2,eq Peq W1,eq (5)
Next, Pext is appended with weighting functions W1 , W2 .
by designing the weighting filters W1,eq , W2,eq such that
Definition 9 The shaped plant is Ps = W2 Pext W1 . new freedom for bandwidth enhancement is exploited.
The shaped plant reflects a desired loop-shape that is After completing these design steps, the weights
dictated through W1 and W2 . Subsequently, a H∞ -controller 
W1,eq
 
W2,eq

Cs is designed that attains this loop-shape as close as W1 = , W2 = .
W1,flex W2,flex
possible while stabilizing the system Ps in Fig. 6. This design
problem can be cast in the configuration shown in Fig. 2(c). form the shaped extended plant in Def. 9. On the basis of
this shaped plant, a H∞ -norm based controller is synthesized
Definition 10 The performance criterion J, ˜ used to design along the lines explained in Sect. V-A.
the feedback loop in Fig. 6, is given by In Sect. VII, the proposed design framework is further
J˜(Ps , Cs ) := k T (Ps , Cs ) k∞ , discussed and illustrated on the industrial lightweight posi-
where   tioning device in Fig. 1. First, a suitable control configuration
Ps for this device is presented in Sect. VI.
(I + Cs Ps )−1 Cs I .
 
T (Ps , Cs ) : w 7→ z =
I VI. C ONTROL CONFIGURATION AND DYNAMICS
? OF THE EXPERIMENTAL SETUP
Using Def. 10, the H∞ -optimal controller Cext for the
extended plant Pext is determined by In this section, a control configuration for the lightweight
?
wafer stage in Fig. 1 is presented, by selecting suitable
Cext = W1 Cs? W2 , actuator and sensor locations. The input-output configuration
is chosen in such a way, that a clear illustration of the new
where concepts presented in this paper is facilitated.
Cs? = arg min J˜(Ps , Cs ),
Cs A. Actuation principle and actuator-sensor placement
see [15, Chap. 6] for details. In the next section, a framework State-of-the-art wafer stages are designed to operate in a
for the design of weighting filters W1 and W2 for high- vacuum. To facilitate this, the system in Fig. 1 is actuated
performance control of flexible motion systems is presented. on the basis of magnetic levitation [16]. Hereby, the wafer
stage can be positioned contactless. As a result, there are
B. Weighting filter design framework that exploits freedom 6 motion degrees-of-freedom (DOFs), i.e., 3 translations
provided by additional actuators and sensors [x, y, z] and 3 rotations [Rx , Ry , Rz ]. By design, the stroke
In this section, the physical insight presented in Sect. IV of the system in the x − y plane is limited, as the primary
is integrated into a framework for multivariable control. In purpose of the prototype is to investigate the effect of
essence, the same two steps taken in Sect. IV are followed structural deformations. Since structural deformations are
for weighting filter design, succeeded by a full multivariable most manifest in the DOFs [z, Rx , Ry ], see Fig. 7(a), this
H∞ control synthesis. coordinate frame is considered for control design.

987
 −100 −100 −100 −100 1) FRF identification Po
−150 −150 −150 −150

Identification
−200
1 2 3
−200
1 2 3
−200
1 2 3
−200
1 2 3 2) Weighting filter design W1 , W2
10 10 10 10 10 10 10 10 10 10 10 10
data
−100 −100 −100 −100

3) Nominal identification P̂
Magnitude (dB)

−150 −150 −150 −150

−200 −200 −200 −200

10
1 2
10 10
3
10
1
10
2 3
10 10
1
10
2
10
3
10
1
10
2 3
10 Validation
4) Embedding Po in model set P
−100 −100 −100 −100
data

−150 −150 −150 −150

5) Robust control synthesis C RP
−200 −200 −200 −200
1 2 3 1 2 3 1 2 3 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Fig. 9. Identification and robust control design procedure, where enhanced
−100 −100 −100 −100 weighting filter design plays an important role.
−150 −150 −150 −150
are measured indeed. On the other hand, the fourth input of
−200
10
1 2
10 10
3
−200
10
1
10
2 3
10
−200
10
1
10
2
10
3
−200
10
1
10
2 3
10 Pflex Pext does not excite rigid-body behavior, since upon applying
Frequency (Hz) a force F = [1, 2, 1, −4]T N to the actuators a,
Fig. 8. Measured frequency response function Po,ext (dotted) and control-
Fz = F1 + F2 + F3 + F4 = 0,
relevant 8th order parametric model P̂ext (solid).
MRx = −F1 + F2 + F3 + 21 F4 = 0,
Next, an input-output configuration is selected. In this
paper, 4 actuators are placed as indicated in Fig. 7(a). MRy = −F1 − F2 + F3 − 21 F4 = 0,
Specifically, a1 , a2 , and a3 are placed below corners of the i.e., the net force and rotational moments on the stage are
stage, whereas a4 is positioned at the middle of the line zero in the rigid-body coordinates, cf. Fig. 7(a).
between the center and a2 . Moreover, corner sensors s1 , Before presenting the merits of using the additional control
s2 , and s3 are used as indicated in Fig. 7(a). Finally, a input-output, the essential steps towards synthesis of a robust
piezo sensor s4 , which measures strain of the wafer stage, is controller are reviewed.
available at the middle of the line between the center and s2 . C. Steps towards high-performance robust control design
Since 4 actuators and sensors are used to control 3 rigid-body
The framework in Fig. 9 describes the essential steps
DOFs, there is freedom to actively control flexible dynamical
towards synthesis of a high-performance robust controller.
behavior of the wafer stage.
First, a frequency response function (FRF) of the true system
B. Coordinate frame for extended control design Po is determined (Step 1). On the basis of this FRF, the
In this section, the input-output behavior of the system is proposed weighting filter design approach in Sect. V-B can
transformed to a coordinate frame that satisfies Ass. 8. be applied (Step 2). Subsequently, a model set for robust
control is constructed, which requires nominal modeling
Result 11 Let Pa,s : a 7→ s be the mapping from the (Step 3) and uncertainty modeling (Step 4). Finally, the
system’s actuators to sensors, where a = [a1 , a2 , a3 , a4 ]T optimal robust controller is computed (Step 5).
and s = [s1 , s2 , s3 , s4 ]T . The extended plant, see Def. 7, is The system identification for robust control design proce-
obtained through the decoupling dure in [2] is employed. Herein, the performance criterion J˜
Pext = Sy Ty Pa,s Tu Su , in Def. 10 is jointly optimized during the modeling step and
where the robust control synthesis step. In particular, robust-control-
 1 0 1 0
 1 −1 0 1 relevant identification is pursued, see [2] for further details.
 
0 −1 1 0  0 1 −1 2 An important implication of this approach is that the novel
Ty =  −1 1 0 0 , Tu =

1 0 1 1 ,
−4
weighting filter design framework influences all subsequent
0 0 0 1 0 0 0
design steps towards the synthesis of a robust controller.
and Su , Sy are diagonal scaling matrices.
VII. B EYOND CONVENTIONAL BANDWIDTH LIMITATIONS
Using the geometry of the wafer stage, see Fig. 7(a), it USING THE EXTRA ACTUATOR - SENSOR PAIR
is immediate that the first three coordinates of Pext contain In this section, the framework for weighting filter design
decoupled rigid-body behavior in the n = 3 motion DOFs in Sect. V-B is confronted with the prototype wafer stage.
[z, Rx , Ry ]. In Fig. 8, collocated behavior is observed in the The aim is to beat conventional bandwidth limits.
z-translation, whereas non-collocated behavior is observed
in the Rx - and Ry - rotations. Step 1. Frequency response function identification
After applying rigid-body decoupling, there is m = 1 The identified frequency response function Po,ext of the
additional input-output directionality available for control of system in Fig. 1 is depicted in Fig. 8. The torsion bending
flexible dynamics. The first flexible mode is a torsion bending mode of the stage manifests itself as a lightly damped res-
of the stage, depicted in Fig. 7(b). This mode can effectively onance phenomenon at 143 Hz, whereas remaining flexible
be controlled using the fourth input-output direction of Pext , phenomena occur above 500 Hz.
indicated by Pflex in Fig. 8. On the one hand, the fourth Due to the actuator-sensor configuration in Fig. 7(a), non-
output of Pext is the piezo sensor s4 , which measures collocated behavior is observed beyond the torsion frequency
strain. As a consequence, no rigid-body displacements are in the rotational coordinates Rx and Ry . Hence, the torsion
observed at this output, while the structural deformations mode imposes restrictions on the achievable control band-
associated with the flexible dynamical behavior of the system widths using conventional rigid-body control.

988
 160 −100 −100 −100

Magnitude (dB)
140
−150 −150 −150
120

100 −200 −200 −200

1 2 3 1 2 3 1 2 3
10 10 10 10 10 10 10 10 10
80

Magnitude (dB)
1 2 3
10 10 10 −100 −100 −100
Frequency (Hz)
Fig. 10. Magnitude of W2,flex W1,flex = k? (dotted) and Cflex (solid). −150 −150 −150

50 −200 −200 −200

1 2 3 1 2 3 1 2 3
10 10 10 10 10 10 10 10 10
Magnitude (dB)

−100 −100 −100

0
-9 −150 −150 −150

−200 −200 −200

1 2 3 1 2 3 1 2 3
−50 10 10 10 10 10 10 10 10 10
1 2 3
10 10 10 Frequency (Hz)
Fig. 12. Equivalent plant Po,eq (dotted) and parametric fit P̂eq (solid)
Phase (deg)

0
−180 under active control of the torsion loop. The initial model of the motion
−360 DOFs [z, Rx , Ry ] from Fig. 8 is shown for comparison (dashed).
−540
−720
10
1
10
2
10
3 sufficient phase-lead around 200 Hz while at the same time
Frequency (Hz) achieving sufficient suppression of higher order modes using
Fig. 11. Bode diagram of the loop-gain Pflex Cflex . notches and roll-off, which is needed to enable a stable high-
Step 2. Application of the new design framework gain torsion loop.
In this section, weighting filters are designed for robust ◦ Design step D2
control, following the two steps in Sect. V. First, the stiffness In design step D2, Sect. V-B, the FRF of the equivalent
of the wafer stage in the direction of the torsion bending plant under control of the torsion loop is determined. Instead
mode is enlarged. Second, the equivalent plant in the motion of (4),
DOFs [z, Rx , Ry ] is determined, for which the obtained Peq = F` (Po,ext , Cflex )
freedom for bandwidth enhancement is exploited. is determined by means of a system identification experi-
◦ Design step D1 ment, where the manually designed controller Cflex is im-
In design step D1, Sect. V-B, the additional control input plemented on the true system. The FRF is shown in Fig. 12.
and output available on top of the rigid-body configuration Result 12 By active control of the torsion mode, the me-
are exploited to control the flexible dynamical behavior Pflex . chanical stiffness of the stage is successfully modified. As
As observed in Fig. 8, below 200 Hz, a result, the frequency at which the torsion mode manifests
1 itself shifts from 143 to 193 Hz.
Pflex ≈
m s + d? s + k ?
? 2
Theoretically, under proportional control of the torsion
where the compliance 1/k ? is approximately -145 dB, i.e., mode with a gain k ? = 1.78 · 107√ , see (6), the resonance
k ? ≈ 1.78 · 107 N/m. By controlling Pflex with a high loop- frequency would manifest itself at 2 · 143 = 202 Hz. The
gain, the apparent stiffness of the system can be enhanced, manually designed torsion controller Cflex achieves a shift of
which reduces internal deformations, cf. Result 6 in Sect. II. the resonance frequency that is slightly lower only. Also, the
Here, it is aimed to double the stiffness of the torsion mode. controller adds damping to the torsion mode, as is clearly
Thus, for Pflex,s in (3), the weights observed in Fig. 12.
√ Now that new freedom has been generated for bandwidth
W1,flex = W2,flex = k ? = 4.22 · 103 , (6) enhancement in the motion DOFs [z, Rx , Ry ], this freedom
dictate the desired loop-gain for the torsion mode. should be exploited in the design of weighting filters W1,eq
For the considered control configuration, it is challenging and W2,eq , see (5). These weighting filters are designed
to achieve the specified gain in the torsion loop. If Pflex is along the lines of [2, Sect. III.A] and [15], and reflect
multiplied with a gain k ? of 145 dB, then the torsion mode common loop-shaping goals for PID-type of motion control
at 143 Hz, as well as all parasitic higher order flexible modes as also encountered in, e.g., [11], [17], [10]. To obtain a
beyond 500 Hz, will have a loop-gain that is lifted above the meaningful robust performance optimization problem, it
0 dB line, see Fig. 8. Hence, all these phenomena become is needed to dictate attainable control goals. On the other
relevant for stability. hand, it is desired to achieve control bandwidths that are as
To achieve a stable high-gain torsion loop using robust high as possible. By using a procedure similar to [18], target
control synthesis, a non-conservative model set is required bandwidths of [64, 55, 56] Hz for [z, Rx , Ry ] are selected.
that encompasses all flexible phenomena with little uncer-
tainty. This demands for an accurate nominal model that ∗ Remark on notation for 3 × 3 robust control design
captures all these phenomena. Due to the involved modeling As motivated in design step D1, instead of performing a
complexity, this is beyond the scope of this paper. Instead, formal 4 × 4 robust control design for Pext in Fig. 8, the
the torsion feedback loop is designed on the basis of manual controller for Pflex is designed using manual loop-shaping.
loop-shaping. The resulting controller is depicted in Fig. 10. It then remains to design a 3 × 3 robust controller for Peq
Even using SISO manual design, it is non-trivial to generate in Fig. 12.

989
 150 150 150
motion control design. The framework exploits additional
125 125 125 inputs and outputs. A systematic weighting function design
100 100 100 procedure for H∞ control design is presented and embedded
75
0 1 2 3
75
0 1 2 3
75
0 1 2 3
in an identification and robust control design framework.
10 10 10 10 10 10 10 10 10 10 10 10
The approach is demonstrated on an industrial wafer stage
Magnitude (dB)

150 150 150

system, confirming a significant performance enhancement.
125 125 125
ACKNOWLEDGEMENTS
100 100 100 The authors gratefully acknowledge the contributions by Okko Bosgra in
an early stage of this research. This research is supported by ASML research
75
0 1 2 3
75
0 1 2 3
75
0 1 2 3
and by the Innovational Research Incentives Scheme under the VENI grant
10 10 10 10 10 10 10 10 10 10 10 10 Precision Motion: Beyond the Nanometer (no. 13073) awarded by NWO
150 150 150 (The Netherlands Organisation for Scientific Research) and STW (Dutch
Science Foundation).
125 125 125 R EFERENCES
100 100 100
[1] V. M. Martinez and T. F. Edgar, “Control of lithography in semi-
75 75 75 conductor manufacturing,” IEEE Control Systems Magazine, vol. 26,
0 1 2 3 0 1 2 3 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 no. 6, pp. 46–55, 2006.
Frequency (Hz) [2] T. Oomen, R. van Herpen, S. Quist, M. van de Wal, O. Bosgra, and
Fig. 13. Comparison of the robust controllers in motion DOFs [z, Rx , Ry ] M. Steinbuch, “Connecting system identification and robust control
for conventional plant Po (dashed) and equivalent plant Peq,o (solid). for next-generation motion control of a wafer stage,” IEEE Trans. on
x 10
−6 Control Systems Technology, vol. 22, no. 1, pp. 102–118, 2014.
1.5 [3] J. S. Freudenberg and D. P. Looze, “Right half plane poles and zeros
and design tradeoffs in feedback systems,” IEEE Trans. on Automatic
1 Control, vol. 30, no. 6, pp. 555–565, 1985.
[4] R. Middleton, “Trade-offs in linear control system design,” Automat-
Amplitude (m)

0.5
ica, vol. 27, no. 2, pp. 281–292, 1991.
0 [5] J. Hong and D. S. Bernstein, “Bode integral constraints, collocation,
and spillover in active noise and vibration control,” IEEE Trans. on
−0.5 Control Systems Technology, vol. 6, no. 1, pp. 111–120, 1998.
[6] J. S. Freudenberg, C. V. Hollot, R. H. Middleton, and V. Toochinda,
−1
“Fundamental design limitations of the general control configuration,”
−1.5
IEEE Trans. on Automatic Control, vol. 48, no. 8, pp. 1355–1370, 2003.
0 0.5 1 1.5 2 2.5 3 3.5 4 [7] T. Oomen, E. Grassens, F. Hendriks, R. van Herpen, and O. Bosgra,
Time (s) “Inferential motion control: Identification and robust control with un-
Fig. 14. Standstill error on application of a disturbance signal: conventional measured performance variables,” Proc. IEEE Conference on Decision
(blue) and extended (green) control configuration. and Control, Orlando, FL, USA, pp. 964–969, 2011.
[8] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control –
Step 3 - 5. Modeling and robust control design Analysis and Design, 2nd ed. John Wiley & Sons, 2005.
Next, step 3-5 in Fig. 9 are performed using the framework [9] R. A. de Callafon and P. M. J. Van den Hof, “Suboptimal feedback
in [2]. The robust controller is synthesized for the equivalent control by a scheme of iterative identification and control design,”
Mathematical Modelling of Systems, vol. 3, no. 1, pp. 77–101, 1997.
plant under control of the torsion mode, with target band- [10] M. van de Wal, G. van Baars, F. Sperling, and O. Bosgra, “Multivari-
widths of [64, 55, 56] Hz, see Step 2. D-K iterations is used able H∞ /µ feedback control design for high-precision wafer stage
to compute the optimal robust controller, see Fig. 13. motion,” Control Engineering Practice, vol. 10, no. 7, pp. 739–755,
To evaluate the merits of the proposed design procedure, 2002.
[11] U. Schönhoff and R. Nordmann, “A H∞ -weighting scheme for PID-
a comparison is made with conventional rigid-body control. like motion control,” Proc. IEEE International Conference on Control
Therefore, a 3 × 3 robust controller is designed for the first 3 Applications, Glasgow, Scotland, UK, pp. 192–197, 2002.
input-output directions of Po,ext in Fig. 8. It is emphasized [12] G. C. Goodwin, M. E. Salgado, and J. I. Yuz, “Performance limitations
for linear feedback systems in the presence of plant uncertainty,” IEEE
that no control of the torsion mode is applied. The same Trans. on Automatic Control, vol. 48, no. 8, pp. 1312–1319, 2003.
steps as outlined in Fig. 9 are followed. In the weighting [13] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control.
filter design, again a procedure similar to [18], [19] is used Prentice Hall, Upper Saddle River, NJ, USA, 1996.
to establish meaningful target bandwidths for [z, Rx , Ry ] of [14] J. C. Doyle, K. Glover, P. P. Khargonekar, and B. A. Francis, “State-
space solutions to standard H2 and H∞ control problems,” IEEE
[54, 44, 44] Hz. The rigid-body controller is shown in Fig. 13. Trans. on Automatic Control, vol. 34, no. 8, pp. 831–847, 1989.
[15] D. McFarlane and K. Glover, “Robust controller design using normal-
Result 13 The conventional robust rigid-body controller ized coprime factor plant descriptions,” Lecture Notes in Control and
achieves bandwidths of [50, 30, 30] Hz, which is in line with Information Sciences, vol. 138, 1990.
the results in [20] on the same setup. By enhancing the stiff- [16] J. C. Compter, “Electro-dynamic planar motor,” Precision Engineer-
ness of the wafer stage through active control of the torsion ing, vol. 28, no. 2, pp. 171–180, 2004.
[17] M. Steinbuch and M. Norg, “Advanced motion control: An industrial
mode, a bandwidth enhancement towards [60, 39, 39] Hz is perspective,” European Journal of Control, vol. 4, pp. 278–293, 1998.
successfully obtained. [18] P. Date and A. Lanzon, “A combined iterative scheme for identification
and control redesigns,” International Journal of Adaptive Control and
Time domain results with the controllers implemented, Signal Processing, vol. 18, no. 8, pp. 629–644, 2004.
see Fig. 14, corroborate the observations, i.e., the extended [19] M. Graham and R. de Callafon, “Performance weight adjustment for
control configuration significantly enhances the performance. iterative cautious control design,” Proc. European Control Conference,
Kos, Greece, pp. 217–222, 2007.
VIII. C ONCLUSIONS [20] F. Boeren, R. van Herpen, T. Oomen, M. van de Wal, and O. Bosgra,
“Enhancing performance through multivariable weighting function
Control performance in traditional motion systems is lim- design in H∞ loop-shaping: with application to a motion system,”
ited by flexible dynamical behavior. In this paper, a new Proc. American Control Conference, Washington, D.C., USA, pp.
framework is proposed that enables high performance robust 6051–6056, 2013.

990