Beruflich Dokumente
Kultur Dokumente
I. INTRODUCTION
N A large-volume multistage manufacturing system, such as
semiconductor manufacturing, a product has to go through
a number of processing steps before it can be transformed
to a finished product. In such a system, running individual
processes under optimal conditions may not necessarily result
in an optimal objective for the factory-wide metrics, such as
throughput and yield, cycle time and fabrication cost, and
electrical characteristics of the product [7]. To pursue an optimal objective for these factory-wide metrics, a factory-wide
control strategy is required. Some of the factors that influence
the factory-wide objectives are addressed in the product and
process design phase, such as design of the products, inventory
control, scheduling maintenance events, optimized product
Manuscript received April 1, 2007; revised July 12, 2007. This work was supported in part by the National Science Foundation under Grant EEC95-92125.
J. R. Moyne and D. M. Tilbury are with the Mechanical Engineering Department, University of Michigan, Ann Arbor, MI 48109-2125 USA (e-mail:
aftabak@umich.edu; tilbury@umich.edu).
A. A. Khan was with the Mechanical Engineering Department, University of
Michigan, Ann Arbor, MI 48109-2125 USA. He is now with Advanced Engineering Research Organization, Wah Cantt. 47040, Pakistan.
Digital Object Identifier 10.1109/TSM.2007.907609
scheduling, and efficient use of human resources. Other factors are addressed in the manufacturing phase. Among these,
advanced process control (APC)1 is rapidly emerging as the
predominant tool for continuous process improvement. Today,
APC strategies are applied in a process-centric fashion on an
R2R basis to increase accuracy, minimize equipment downtime,
ensure highly efficient processes, and reduce variability in the
processes [6], [21], [28]. However, these process-centric goals
of APC may not translate directly in the improvement of factory-wide objectives [7], [20]. Thus, researchers have proposed
a factory-wide supervisory control strategy that coordinates
individual R2R process controllers utilizing the end-of-run or in
situ metrology as well as final electrical testing data (ETest) of
the finished product [26], [20]. A schematic diagram of a typical factory-wide control solution is shown in Fig. 1, [26], [20].
Fig. 1 shows a general layout of the manufacturing steps that
a wafer might go through in a semiconductor manufacturing
factory. For a typical wafer, the series of processes shown in the
figure are repeated several times in order to produce multiple
interconnect layers on the wafer. A factory-wide controller
delivers control strategies to maximize an objective function
consisting of factors such as throughput, yield, cycle time, and
fabrication cost. These control strategies along with market
demands drive the electrical quality characteristics of the finished wafer. The desired electrical quality characteristics are
then translated into individual target values for each processing
step by a supervisory controller. The supervisory controller
recalculates optimized target values at each processing step
using the ETest (final metrology) data to achieve the desired
electrical characteristics of the finished wafer. Depending upon
the importance of a particular processing step, a metrology tool
might be employed to monitor the quality of sampled wafers.
This quality measurement (shown as the dashed lines in Fig. 1)
can be utilized as feed-forward data for the next process and as
feedback data for the current process [22]. However, the cost
and time associated with metrology greatly limits the number
of wafers that can be measured at a particular metrology station (usually one to three wafers per lot; the dashed lines in
Fig. 1 represent sampled metrology data). Integrated metrology
(IM) can provide wafer level metrology data to enable R2R
process control [14], [30]; however, issues such as increased
cycle time, lost throughput, and cost of metrology makes it
impractical in many process environments. Furthermore, all
types of metrology introduce delay into the feedback loop due
to wafer transport, measurement, and quality parameter calculation times. Thus, in most applications the process controller
1Advanced process control includes both run-to-run (R2R) control and fault
detection and classification (FDC) systems [21], [6].
365
Fig. 1. Supervisory L2L control of product going through a number of processing steps.
fault detection and classification (FDC) system collects equipment data (referred to as process variables in this paper) for
every process run. This enormous amount of process data (involving hundreds of variables) along with upstream metrology
data can be used for VM purposes (referred to as type-1 and
type-2 data [20] respectively; see Fig. 2). The VM data obtained
for every process can then be used in a feed-forward and feedback control scheme to provide R2R control for every wafer.
A. VM Module Development for Individual Process
The formulation of VM and R2R control at the wafer level for
an individual process is pictured schematically in Fig. 3 (refer
to [12] for further details). In Fig. 3, the process has outputs
inputs
, and process variables
. is the process
run index (which represents processing of a single wafer), is
the lot number (or the event when a wafer is measured at the
are the recipe (tool) settings
metrology station),
are the actual measurement of outputs
at the start of run
are the predicted values
at the metrology run ,
are the
for the outputs at the end of run , and
366
statistical or other characteristic values for process (FD) variables at the end of run . The FD system collects traces of
process variables during the process run. At the end of run the
trace signals for each of these process variables can be analyzed
in the time and/or frequency domain to identify characteristic
values that have the most dominant effect on the process outrefers to those characteristic values found
puts. The vector
at the end of process run [10]. Most of the FD systems calculate summary statistics (mean, standard deviation, etc.), which
are usually used as the characteristic values. The process is then
modelled by the following equation:
(2.1)
is the inputs gain matrix,
is the
where
is a vector of average
process variables gain matrix,
(constant) drift rates (per run ), and is a multivariate white
noise sequence with zero mean and variance .
An initial VM module is built from a data set of process
variables, inputs, and outputs, obtained with the help of design of experiments (DOE) [19]. If the upstream metrology is
not available, then it would not be used in building the initial
VM module. R2R control at wafer level is made possible by
as well as the actual
feeding back the predicted outputs
outputs
. The VM module is recursively updated and the
is
controller parameters retuned as new metrology data
available, as shown by the dashed line in Fig. 3 [12].
To obtain the initial VM module, define
and arrange the DOE data for
runs in
and
two matrices,
. Let
and
, where
and
are the mean values of the
columns of
and , respectively. Then, a linear regression
model of the given process can be written as
(2.2)
is a matrix of regression coefficients
and is an
matrix of errors whose elements are independently and identically distributed with mean
zero and variance .
Different regression methods, such as multiple linear regression (MLR), principal component regression (PCR), and partial
least squares (PLS) can be applied to the matrices V and Y to
estimate the coefficient matrix in (2.2). For a typical semiconductor manufacturing process the number of FD variables is
generally large. Thus, for most processes the FD variables will
be correlated and this fact coupled with measurement noise can
lead to imprecise estimation of the matrix using MLR technique [9]. In the PCR approach, the regressor variables in the
rank deficient V are replaced by a reduced set of principal components (PC) of V [8], [25]. But the PCs are derived without
any reference to the output matrix Y. Thus, the derived PCs that
explain much of the variation in V may not be related to the variation in Y, giving rise to imprecise estimation of . In case of
noisy and highly correlated data, projection to latent structures
where
(PLS) has been shown to be a robust multivariate linear regression technique [13]. The PLS technique has been successfully
applied in the modelling, prediction, and statistical control of the
behavior of a wide variety of processes [33], [3], [14]. In contrast to PCR, PLS regression performs a simultaneous decomposition of V and Y with the constraint that these components
explain as much as possible of the covariance between V and Y
[32], [13], [17]. The reduced set of PLS components can then
of the regression coefficient
be used to find an estimate
matrix in (2.2), [12]. Thus, the given process is modelled as
(2.3)
is the PLS estimate of the process output Y. In the
where
online implementation, as new FD data is collected, characteristic values for the run are found, and a new vector is formed.
The process output is then predicted using the PLS regression
coefficient matrix
(2.4)
The regression coefficient matrix
can be recursively updated using a moving window approach when actual measurement of output is available. The size of this moving window is
found by reaching a tradeoff between the speed of adaptation of
the PLS model (requiring smaller size) and avoiding response
to noisy data (requiring larger size). In one approach for the
moving window, the number of data points in Y and V remains
fixed while old data is replaced by new data. Alternatively, some
of the DOE data can be retained while new values of Y and V
replaces older ones; this second approach seeks to preserve the
richness of the DOE data in the reformulation. The procedure
for building PLS model and its update for VM purposes is summarized as follows.
1) Identify process inputs, outputs, and FD variables that are
important for VM purposes.
2) Using the allowable range of input values, conduct DOE
for the process and collect process data and measured outputs for every process run [1], [19].
3) Construct Y and V from the DOE data and apply PLS tech.
nique to find the PLS regression coefficient matrix
4) In the online implementation, use (2.4) to predict the output
from one metrology run to the next.
construct the new data matrices
5) At the metrology run
and
and find the PLS regression coefficient
.
matrix
367
at the metrology
prediction error
, where
is a preruns for process-1;
defined maximum allowable prediction error for process-1.
Once the VM results are available for every process run (or
every wafer), the next step is to use them in the APC paradigm
to enable R2R control at the wafer level. Fig. 5 shows a general
layout of an R2R control solution using VM outputs. The predicted outputs from individual VM modules can be provided to
the process R2R controllers, which in turn get their target values
from the supervisory controller. The VM results can also be provided to the R2R controller of the next process as a feed-forward
term. This will provide the incoming wafer quality variation to
the R2R controller of the next process for better control of the
process [29].
A. Example: VM and R2R Control for TWO Consecutive
Processes
A simulation study is performed to illustrate the use of VM
in providing the predicted outputs to the R2R controllers and
the coordination between the VM modules of two processes
in tandem. Each process is assumed to act on one wafer at a
time. The two processes are such that the outputs of the second
process depend upon the outcome of the first process. The processes are modelled as
Process-1:
(3.2)
Process-2:
(3.3)
for
for
368
TABLE I
TYPICAL INPUT, OUTPUT, AND PROCESS VARIABLES FOR TWO PROCESSES
where
is a matrix of parameters that represents
on
and is the wafer index. As shown
the effects of
in Fig. 4, process-1 could be the development process in photolithography carried out for patterning photoresist on the wafer,
while process-2 could be a plasma etch process that etches the
underlying oxide layer according to the lithography patterns.
Both of the processes have a number of input, output, and FD
variables, as listed in Table I.
Note that in practice some of the FD variables are highly correlated while others are not, and some variables impact one or
more outputs while others may not affect the outputs at all. The
process control simulation is set up to explore the capability of
PLS method to handle these types of input/output (I/O) relationships, as described in the following.
For both processes, let the number of input, output, and
process variables and target values be given as
Tgt
, and
. Thus,
Tgt
, and
. To simulate the outputs of the
two processes, the following values are used for the different
parameters in their respective models:
TABLE II
PROCESS VARIABLE VALUES
(3.4)
and
(3.6)
found
where is an estimate of the process gain matrix
from DOE and
is an estimate of
in
(2.1) at run . These estimates are updated at every run using
two multivariate EWMA equations
(3.7)
(3.8)
(3.5)
where is a
identity matrix, and
and
are
diagonal matrices [4] having EWMA weights on the main
diagonals.
369
TABLE III
CORRESPONDENCE BETWEEN ACTUAL PROCESS BEHAVIOR AND SIMULATION STUDY
of process1 is significantly improved with no additional physical metrology when VM is used on R2R basis.
for every process run, which is then fed back to the dEWMA
response when a
R2R controller. The plot also shows the
dEWMA control scheme uses only the actual metrology (from
metrology measurement of every tenth wafer). In the latter case,
the control value remains unchanged between metrology runs,
which results in the deviation of output from the target value due
. In comparison, the R2R controller using
to the drift term in
VM data provides an output closer to the target.
Fig. 7 shows the response plots of the second output for
. The controller uses VM data on an R2R basis
process-2
but the VM module itself is built with two different approaches:
1) when the process-2 VM module uses process-1 outputs
370
Fig. 7. Simulation results show that when process-1 metrology is used in VM for process-2, output y
of process-2 is significantly improved with no additional physical metrology when VM is used on R2R basis.
wafer to compensate for a systematic drift. However, such a control scheme can result in significantly degraded system behavior
when the drift is not accurately predicted or the process undergoes a sudden shift in drift due to maintenance events or other
external disturbances [4], [10]. While the VM is also using a
prediction of the drift, the control inputs are updated after every
wafer based on actual measurements of the process variables.
Thus, the VM does not present the same danger for divergence.
In fact, if the VM indicates that drastic changes have occurred in
the process, steps can be taken to identify and correct any problems at once, instead of continuing to produce wafers until the
next metrology event.
IV. OPPORTUNITIES, ISSUES, AND SOLUTIONS IN REALIZING
FACTORY-WIDE VM AND R2R CONTROL
The methods and concepts developed for VM provide a
number of opportunities that can be leveraged in realizing factory-wide R2R control and achieving factory level objectives.
However, these opportunities also bring a number of issues
to the forefront that must be overcome. Some of these issues
relate to the VM module building while other relate to their
use in the R2R control solution. Understanding the current
state-of-the-art as to how VM solutions can be applied as well
371
TABLE IV
OPPORTUNITIES, ISSUES, AND APPROACHES IN EMPLOYING VM ON FACTORY
LEVEL (NTS: NEAR TERM SOLUTION, LTS: LONG-TERM SOLUTION)
372
..
.
..
..
.
(4.1)
can be found
In an ideal environment these drift rates
from the DOE data. In reality, however, it is impractical since it
will require an enormous number of experiments to accurately
estimate the drift rates. A reasonable approach would be to initially assume zero values for all the entries in matrix . Then, in
the online implementation of VM, estimate the individual drift
terms as the maintenance events are performed [34] (this approach is a subject of future work).
Multirecipe and Multistep Processes: Many processes in
semiconductor manufacturing consist of multiple process steps
conducted in a sequence before a physical measurement is
made on the quality attributes (or metrology variables). For
example, the contact etch process has two similar steps: an
oxide etch and a subsequent nitride etch. Both processes are
performed in similar processing tools one after the other, but
wafer CD is measured only after the second process. This situation is schematically shown in Fig. 9. In such cases it becomes
difficult to correlate two sets of process variables with one set
of metrology variables. A multiprocess multirecipe approach
is therefore needed for the formulation of VM problem. In the
example case of contact etch, the second process, nitride etch,
determines the final CD. It seems logical to use process data
from this process only in building the VM module. However,
in general, data from all process steps should be used in the
VM module and each step components should be weighted
according to the corresponding level of their influence on the
final quality variables, i.e.,
(4.2)
where is the weight factor for the th process step. In addition,
the R2R controller must be capable of generating two sets of
control values, one for each step of the process. Note that this is
a future research topic.
Use of Upstream Metrology in Process Control: In the example study of Section III, it was demonstrated that the use of
preprocess metrology in VM module enhances the control performance. However, there might be cases when the immediate
pre- and post-process metrology variables are not the same. For
example, quality variable of interest after plasma etch process
is the CD and CD uniformity across the wafer, while the following process is usually metal implant or chemicalmechanical polishing (CMP) process where the output variable is film
thickness and thickness uniformity. In such cases, either a relational model should be developed between the two metrology
variables (e.g., performing etching DOE or simple regression
techniques) [15] or FD data of the current process could be fed
forward to the VM module of the following process to augment
its wafer level process data.
Furthermore, it may not be feasible to employ VM and subsequent R2R control on wafer level for every processing tool in
the manufacturing line [23]. Both the nature of the processes and
the tools can be responsible for the infeasibility of VM. Some
processes, such as CVD, are lot-based processes, i.e., they inherently take a lot for processing as opposed to a single wafer.
On the other hand, the processing tool itself may either not allow
automatic recipe adjustment for every wafer in the lot, may not
have purposeful process variables to be used in the VM module,
or may not have data collection and analysis system (such as
FDC) available to support VM. For example, with respect to
process variables, in the photolithography overlay process, the
stepper tool FD variables do not have a reported relationship
with the overlay measurements. Here, additional sensors might
be needed to derive a VM module for the overlay process. An
additional downside in a lot-based process is that the wafer
level metrology will not be available for the VM modules of
the following processes. To circumvent this problem either the
lot-level metrology could be used for all wafers in the lot or
wafer level metrology be derived from the lot level metrology
using historical data and the physics-based process model.
Design of Factory-Wide Controller: Improvement in process
control at a particular processing step does not necessarily
imply improvement in factory-wide objectives or final product
quality. For example, in [7] the authors note that designing
a controller for thickness uniformity at CMP process had a
negative effect on the uniformity after the etch process. Thus,
the CMP process controller was adjusted from an objective
of optimizing the CMP process to providing adequate control
so that uniformity after etch process could be improved. In
such situations, it becomes crucial to design a factory-wide
controller that can allow tradeoffs between process centric
goals and the final product or factory-wide objectives. This will
require extending the process centric data consolidation and
synchronization strategies to include all processes in the factory. Furthermore, appropriate models are needed to project the
factory-wide objectives onto individual process goals. With the
help of historical data, the approach and methods adopted for
process level VM can be extended to determine such models.
VM Application to Factory-Wide Metrics: In this paper, the
VM methodology is shown to be applied to estimate process or
multiprocess metrology. However, the methodology and concepts can be extended to estimate other metrics that play important roles in achieving factory-wide objectives. Factory-wide
metrics such as process yield, overall factory yield, final ETest
values, rework jobs, and R2R control quality (in terms of the observed deviation from target) can be estimated by tracking necessary parameters throughout the factory that impact these metrics. The estimation models for these metrics would provide dynamic information to the factory-wide controller, which in turn
could alter process-centric goals to achieve the desired values
for these metrics.
373
For example, the final metrology or ETest parameter of a particular wafer is often used as a quality metric to assess against
factory-wide objectives and is thus used to adjust process goals
for the wafers-in-process (WIP); however, there is an inherent
delay of multiple lots in the factory-wide control. To circumvent this problem, a VM module can be built that can project
the final ETest parameters onto individual process steps in the
manufacturing line
(4.3)
where represents the unique number assigned to a wafer,
is the contribution of the th process output in determining the
final ETest parameters
for that wafer, is the VM data
is the total number of processes for
of the th process, and
the wafer to become a finished product. Using historical data,
in
PLS methods can be applied to calculate the parameters
the above equation. As a wafer progresses through the manufacturing line, (4.3) can be used to find the anticipated ETest
values which can then be provided to the factory-wide controller
at each process step. At a particular process step, using these
predicted ETest values and the process VM data of the wafer,
the factory-wide controller can dynamically optimize process
goals (or Tgt values) for the remaining processes to achieve the
desired Etest values.
V. CONCLUSION AND FUTURE WORK
In this paper, we have provided a factory-wide control approach for semiconductor manufacturing that uses VM generated from FD data and upstream metrology for R2R control at
the wafer level. An overview of the VM module building was
presented for an individual process and the coordination of two
VM modules in tandem was described with a simulation example; inference can be used to extend this two-module example
to a factory-wide solution. Issues and challenges that might arise
in the practical implementation of this approach on a factory
level were discussed along with possible solutions and suggestions for future work.
VM alone can increase metrology data availability, reduce
send-ahead wafers, improve quality guarantee levels, and reduce
cycle time, but its use in process control can bring additional
benefits of reduced product variability, increased yield, and improved product Cpk, which contribute significantly in achieving
factory-wide goals. The main barrier to using VM for process
control is its expected lower prediction quality and the danger
of aggressive control in the face of outliers in VM data. As suggested in this paper, this issue can be addressed when the VM
data is accompanied by a quality metric, which the process controller can use in its control scheme. The VM data provides indication of the sudden shifts and changes in drift direction, which
would have been missed in the traditional metrology strategies
due to consistent metrology delays, severely reducing quality
of control. Thus, VM-based process control can also result in
reduced scrap by alerting the process engineer in time to shut
down the controller or adjust the controller parameters in case
of sudden changes in the process conditions.
374
[15] J.-S. Lin, P.-H. Chen, S. Wu, F. Ko, M.-S. Zhou, M.-S. Liang, J. Zou,
T. Mullins, J. Moyne, and K. Edwards, Model based wafer-to-wafer
control for Cu CMP, in Proc. VLSI Multilevel Interconnect Conf., Sept.
2003.
[16] T.-H. Lin, M.-H. Hung, R.-C. Lin, and F.-T. Cheng, A virtual
metrology scheme for predicting CVD thickness in semiconductor
manufacturing, in Proc. IEEE Int. Conf. Robotics Automation, May
2006, pp. 10541059.
[17] A. Lorber, L. E. Wangent, and B. R. Kowalski, A theoretical foundation for the PLS algorithm, J. Chemometrics, vol. 1, pp. 1931, 1987.
[18] J. Mao, Run-to-run control with fault detection and rejection, in Proc.
AEC/APC Symp. XV, Colorado Springs, CO, Sep. 2003.
[19] D. C. Montgomery, Introduction to Statistical Quality Control. New
York: Wiley, 1996.
[20] J. Moyne, Making the move to fab-wide APC, Solid State Technol.,
vol. 47, no. 9, p. 47, Sept. 2004.
[21] J. Moyne, E. del Castillo, and A. M. Hurwitz, Run-to-Run Control in
Semiconductor Manufacturing. Boca Raton, FL: CRC, 2001.
[22] J. Moyne, V. Solakhian, A. Yershov, M. Anderson, and D.
Mockler-Hebert, Development and deployment of a multi-component advanced process control system for an epitaxy tool, in Proc.
Advanced Semiconductor Manufacturing 2002 IEEE/SEMI Conf.
Workshop, Apr./May l2, 2002, pp. 125130.
[23] N. Patel and M. Anderson, Considerations in deploying wafer-level
control across the fab. In, in Proc. AEC/APC Symp. XV, Colorado
Springs, CO, Sep. 2003.
[24] N. S. Patel and S. T. Jenkins, Adaptive optimization of run-to-run
controllers: The EWMA example, IEEE Trans. Semiconduct. Manuf.,
vol. 13, no. 1, pp. 97107, Feb. 2000.
[25] A. Phatak and S. D. Jong, The geometry of partial least squares, J.
Chemometrics, vol. 11, pp. 311338, 1997.
[26] S. J. Qin, G. Cherry, R. Good, J. Wang, and C. A. Harrison, Semiconductor manufacturing process control and monitoring: A fab-wide
framework, J. Process Contr., vol. 16, no. 3, pp. 179191, 2006.
[27] S. Ruegsegger, A. Wagner, J. S. Freudenberg, and D. S. Grimard,
Feedforward control for reduced run-to-run variation in microelectronics manufacturing, IEEE Trans. Semiconduct. Manuf., vol. 12,
no. 4, pp. 493502, Nov. 1999.
[28] M. Sarfaty, A. Shanmugasundram, A. Schwarm, J. Z. J. Paik, R. Pan,
M. Seamons, H. Li, R. Hung, and S. Parikh, Advance process control
solutions for semiconductor manufacturing, in Proc. IEEE Int. Symp.
Semiconductor Manufacturing Conf., Apr. 2002, pp. 101106.
[29] H. Sasano, W. Liu, D. S. L. Mui, K. Yoo, and J. Yamartino, Advanced gate process critical dimension control in semiconductor manufacturing, in Proc. IEEE Int. Symp. Semiconductor Manufacturing
(ISSM), Sep. 2003, pp. 382385.
[30] C. Schneider, L. Pfitzner, and H. Ryssel, Integrated metrology: An
enabler for advanced process control (APC), in Proc. SPIEInt. Soc.
Opt. Eng., 2001, vol. 4406, pp. 118130.
[31] H. Stark and J. Woods, Probability, Random Processes, and Estimation
Theory for Engineers, 2nd ed. Englewood Cliffs, NJ: Prentice Hall,
1994.
[32] X. Wang, U. Kruger, and B. Lennox, Recursive partial least squares
algorithms for monitoring complex industrial processes, Contr. Eng.
Practice, vol. 11, no. 6, pp. 613632, 2003.
[33] L. L. Zheng, T. J. McAvoy, Y. Huang, and G. Chen, Application
of multivariate statistical analysis in batch processes, Industr. Eng.
Chemistry Res., vol. 40, no. 7, pp. 16411649, 2001.
[34] J. Zou and J. A. Mullins, Modeling process operation in model based
R2R control, in Proc. AEC/APC Symp. XVII, Palm Springs, CA, Sep.
2005.
Aftab A. Khan received the B.S. degree in mechanical engineering from NWFP University of
Engineering and Technology, Peshawar, Pakistan, in
1994, and the M.S. and Ph.D. degrees in mechanical
engineering from the University of Michigan, Ann
Arbor, in 2003 and 2007, respectively. He conducted
his Ph.D. thesis research on virtual metrology and
control for manufacturing processes.
He is currently working at Advanced Engineering
Research Organization, Pakistan. His research interests include virtual metrology, predictive control,
system modeling, and run-to-run control and their application to semiconductor
process manufacturing.
375
Dawn M. Tilbury received the B.S. degree in electrical engineering, summa cum laude, from the University of Minnesota, Minneapolis, in 1989, and the
M.S. and Ph.D. degrees in electrical engineering and
computer sciences from the University of California,
Berkeley, in 1992 and 1994, respectively.
In 1995, she joined the faculty of the Mechanical Engineering Department at the University of
Michigan, Ann Arbor, where she is currently holds
the rank of Professor. Her research interests include
distributed control of mechanical systems with
network communication, logic control of manufacturing systems, performance
management and control of computing systems, and uncertainty modeling
in cooperative control. She is coauthor of the textbook Feedback Control of
Computing Systems.
Dr. Tilbury won the EDUCOM Medal (jointly with Professor W. Messner
of Carnegie Mellon University) in 1997 for her work on the web-based Control
Tutorials for Matlab (published by Addison-Wesley, 1999). She received an NSF
CAREER award in 1999, and is the 2001 recipient of the Donald P. Eckman
Award of the American Automatic Control Council. She was a member of the
20042005 class of the Defense Science Study Group (DSSG) and is a current
member of DARPAs Information Science and Technology Study Group (ISAT).
She belongs to ASME, IEEE, and SWE, and is an elected member of the IEEE
Control Systems Society Board of Governors.