50 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 2, NO. 1, MARCH 2008
Defect-Aware High-Level Synthesis and Module
Placement for Microfluidic Biochips
Tao Xu, Krishnendu Chakrabarty, Fellow, IEEE, and Fei Su
Abstract—Recent advances in microfluidics technology have led
to the emergence of miniaturized biochip devices, also referred to
as lab-on-a-chip, for biochemical analysis. A promising category of
microfluidic biochips relies on the principle of electrowetting-on-
dielectric, whereby discrete droplets of nanoliter volumes can be
manipulated using an array of electrodes. As chemists adapt more
bioassays for concurrent execution on such “digital” droplet-based
microfluidic platforms, system integration, design complexity, and
the need for defect tolerance are expected to increase rapidly. Auto-
mated design tools for defect-tolerant and multifunctional biochips
are important for the emerging marketplace, especially for low-
cost, portable, and disposable devices for clinical diagnostics. We
present a unified synthesis method that combines defect-tolerant
architectural synthesis with defect-aware physical design. The pro-
posed approach allows architectural-level design choices and de-
fect-tolerant physical design decisions to be made simultaneously.
We use a large-scale protein assay and the polymerase chain reac-
tion procedure as case studies to evaluate the proposed synthesis
method. We also carry out simulations based on defect injection to
evaluate the robustness of the synthesized biochip designs.
Index Terms—Lab-on-a-chip, microfluidics.
I. INTRODUCTION
M
ICROFLUIDICS technology and bio-microelectrome-
chanical systems (MEMS) have seen tremendous
growth in the past few years [1]–[6]. A major application
driver for microfluidics is biochemical analysis and fluidic
processing in miniaturized devices, which are referred to in the
literature as biochips or lab-on-a-chip. Biochips can be used
for immunoassays, high-throughput sequencing, proteomic
analysis involving proteins and peptides, and environmental
toxicity monitoring. Another emerging application area for
microfluidics-based biochips is clinical diagnostics, especially
immediate point-of-care diagnosis of diseases [7].
A popular class of microfluidic biochips is based on con-
tinuous fluid flow in permanently etched microchannels. Fluid
flow in these devices is controlled either using micropumps and
microvalves, or by electrical methods based on electrokinetics
Manuscript received January 4, 2008; revised January 28, 2008. This work
was supported in part by the National Science Foundation under Grant IIS-
0312352 and Grant CCF-0541055. A preliminary version of this paper was pre-
sented at the 2007 IEEE International Conference on VLSI Design. This paper
was recommended by S. DeWeerth.
T. Xu and K. Chakrabarty are with the Department of Electrical and Computer
Engineering, Duke University, Durham, NC 22708 USA (e-mail: tx@ee.duke.
edu; krish@ee.duke.edu).
F. Su is with the Intel Corporation, Folsom, CA 95630 USA (e-mail: fei.
su@intel.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TBCAS.2008.918283
and electroosmosis [2]. An alternative category of microfluidic
biochips relies on the principle of electrowetting-on-dielectric.
Discrete droplets of nanoliter volume can be manipulated using
an array of electrodes. Following the analogy of digital elec-
tronics, this technology is referred to as “digital microfluidics”
[1]. Because each droplet can be controlled independently, these
systems also have dynamic reconfigurability, whereby groups of
unit cells in a microfluidic array can be reconfigured to change
their functionality during the concurrent execution of a set of
bioassays.
As chemists adapt more bioassays for concurrent execution
on a digital microfluidic platform, system integration and design
complexity are expected to increase steadily [8], [9]. A digital
microfluidic biochip platform has been developed for on-chip
gene sequencing through synthesis [10], which targets execution
of 10
fluidic operations. Other digital microfluidic biochips
are being designed for protein crystallization which requires
concurrent execution of hundreds of operations, with only mi-
croliter sample consumptions. The chip size is also increasing
sharply. A recently demonstrated droplet-based biochip embeds
more than 600 000 20
mby20 m electrodes [11]. Moreover,
next-generation biochips are expected to be multifunctional and
adaptive “biochemical processing” devices. For example, inex-
pensive biochips for clinical diagnostics integrate hematology,
pathology, molecular diagnostics, cytology, microbiology, and
serology onto the same platform. The emergence of such inte-
grated, multifunctional, and reconfigurable platforms provides
the electronic design automation community with a new appli-
cation driver and market for research into new algorithms and
design tools.
As in the case of integrated circuits (ICs), increase in density
and area of microfluidics-based biochips will reduce yield, es-
pecially for new technology nodes. Low yield will deter large-
scale and high-volume production, and it will increase produc-
tion cost. It will take time to ramp up yield learning based on
an understanding of defect types in such mixed-technology de-
vices. Therefore, defect-tolerant designs are important for the
emerging marketplace, especially for low-cost, portable, and
disposable devices for clinical diagnostics.
Another reason for the importance of defect tolerance lies in
the projected use of microfluidic biochips for safety-critical ap-
plications, e.g., patient health monitoring, neo-natal care, and
the monitoring of environmental toxins. Therefore, defect tol-
erance must be integrated into the automated design tools to
ensure high levels of system dependability. Despite the recent
emergence of automated synthesis methods for biochips [9],
[12], [13], defect tolerance has largely been overlooked in the
literature.
1932-4545/$25.00 © 2008 IEEE
XU et al.: DEFECT-AWARE HIGH-LEVEL SYNTHESIS AND MODULE PLACEMENT FOR MICROFLUIDIC BIOCHIPS 51
Fig. 1. Schematic of a digital microfluidic biochip [14].
As in traditional IC design, the automated design of a
digital microfluidic biochip can be divided into two major
phases. High-level synthesis is first used to generate a macro-
scopic structure of the biochip from the behavioral model
for a bioassay; this structure is analogous to a structural reg-
ister-transfer level model in electronics computer-aided design
(CAD). Physical design creates the final layout of the biochip,
consisting of the placement of microfluidic modules such as
mixers and storage units, as well as the routes that droplets take
between different modules, locations of on-chip reservoirs,
dispensing ports and integrated optical detectors, and other
geometrical details. These synthesis tools can relieve biochip
users from the burden of manually optimizing a set of assays
for increased throughout. Users can describe bioassays at a
sufficiently high level of abstraction; synthesis tools can then
map the behavioral description to the microfluidic array and
generate an optimized schedule of bioassay operations, the
binding of assay operations to resources, and a layout of the
microfluidic biochip. Thus, the biochip user can concentrate
on the development of the nano- and micro-scale bioassays,
leaving implementation details to the synthesis tools.
In the past few years, several automated design tools have
been proposed for microfluidic biochips. These design automa-
tion methods address operation scheduling and module place-
ment for digital microfluidics. However, most of these tools have
been focused on device-level physical modeling and simula-
tion of single components [3]. Few efforts have been reported
on developing system-level CAD tools for digital microfluidic
biochips design. Moreover, defect tolerance has largely been
overlooked in these tools.
In this paper, we propose a synthesis methodology that uni-
fies operation scheduling, resource binding, module placement,
and defect tolerance. The proposed automated design technique
is based on parallel recombinative simulated annealing (PRSA)
and its uses defect tolerance as a design criterion. All three
tasks, i.e., resource binding, scheduling, and placement, are car-
ried out at each step of the algorithm in a defect-aware fashion.
Thus, exact placement information, instead of a crude area es-
timate, is used to judge the quality of architectural-level syn-
thesis. This information is utilized by the annealing process to
select resources and schedule bioassay operations to produce
a high-quality design. The proposed method allows defect-tol-
erant architectural design choices and defect-aware physical de-
sign decisions to be made simultaneously. Defect tolerance is
incorporated during synthesis since it is integrated it into the
simulated annealing procedure. We use a representative protein
assay the polymerase chain reaction (PCR) procedure to eval-
uate the proposed synthesis methodology.
The rest of the paper is organized as follows. Section II pro-
vides an overview of digital microfluidic biochips. In Section III
we discuss related prior work on biochip design automation.
Section IV presents an overview of a defect-oblivious unified
synthesis method. In Section V, we introduce a new criterion of
evaluating defect tolerance for a digital microfluidic biochip and
incorporate it in an enhanced version of the synthesis technique.
In Section VI, we use a protein assay and PCR as case studies to
evaluate the proposed synthesis method. We also carry out sim-
ulations based on defect injection to evaluate the robustness of
the synthesized biochip designs. Finally, conclusions are drawn
in Section VII.
II. D
IGITAL MICROFLUIDIC BIOCHIPS
The microfluidic biochips discussed in this paper are based
on the manipulation of nanoliter droplets using the principle
of electrowetting, i.e., modulation of the interfacial tension
between a conductive fluid and a solid electrode through an
electric field. The basic cell of a digital microfluidic biochip
consists of two parallel glass plates, as shown in Fig. 1. The
bottom plate contains a patterned array of individually control-
lable electrodes, and the top plate is coated with a continuous
ground electrode. The droplets containing biochemical samples
and the filler medium, such as the silicone oil, are sandwiched
between the plates. The droplets travel inside the filler medium.
By varying the electrical potential along a linear array of
electrodes, droplets can be moved along this line of electrodes.
The velocity of the droplet can be controlled by adjusting the
control voltage (0–90 V), and droplets can be moved at speeds
of up to 20 cm/s [1]. Based on this principle, microfluidic
droplets can be moved freely to any location of a 2-D array
without the need for micropumps and microvalves.
Using a 2-D array, many basic microfluidic operations for
different bioassays can be performed, such as sample intro-
duction (dispense), sample movement (transport), temporarily
sample preservation (store), sample dilution with buffer (dilute)
and the mixing of different samples (mix). For instance, the
mix operation is used to route two droplets to the same location
and then turn them around some pivot points. Note that these
operations can be performed anywhere on the array, whereas
in continuous-flow biochips they must operate in a specific mi-
cromixer or microchamber fixed on a substrate. This property
52 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 2, NO. 1, MARCH 2008
is referred to as the reconfigurability of a digital microfluidic
biochip. The configurations of the microfluidic array, i.e., the
routes that droplets transport and the rendezvous points of
droplets, are programmed into a microcontroller that controls
the voltages of electrodes in the array. In this sense, these
mixers and storage units used during the operations can be
viewed as reconfigurable virtual devices. In addition, a digital
microfluidic biochip also contains on-chip reservoirs/dis-
pensing ports that are used to generate and dispense sample or
reagent droplets, as well as integrated optical detectors such as
LEDs and photodiodes. In contrast to mixers or storage units,
these resources are nonreconfigurable.
III. R
ELATED
PRIOR WORK
Biochips belong to the class of MEMS, which is a relatively
young field compared to IC design. Nevertheless, a number of
MEMS CAD tools are available today for modeling [15] and
synthesis [16]. Attempts have also been made to make MEMS
defect-tolerant [17]. However, because of the differences in ac-
tuation methods between MEMS and digital microfluidics, these
techniques cannot be directly used for the design of microfluidic
biochips.
Although research on microfluidic biochips has gained
momentum in recent years, CAD tools for biochips are still in
their infancy [18]. Recent years have seen growing interest in
the this area [9], [12], [13], [19], [20]. One of the first published
methods for biochip synthesis decouples high-level synthesis
from physical design [13]. It is based on rough estimates for
placement costs such as the areas of the microfluidic modules.
These estimates provide lower bounds on the exact biochip
area, since the overheads due to spare cells and cells used
for droplet transportation are not known
a priori. However,
it cannot be accurately predicted if the biochip design meets
system specifications, e.g., maximum allowable array area and
upper limits on assay completion times, until both high-level
synthesis and physical design are carried out. When design
specifications are not met, time-consuming iterations between
high-level synthesis and physical design are required. A link
between these steps is especially necessary if defect tolerance is
to be considered during synthesis. Moreover, defect-tolerance
must be incorporated to guarantee the reliability of the synthesis
result.
IV. U
NIFIED
SYNTHESIS METHOD
A. Problem Formulation
Fig. 2 illustrates the design flow for the proposed synthesis
method. A sequencing graph is first obtained from the protocol
for a bioassay [4]. This acyclic graph
has vertex set
in one-to-one correspondence
with the set of assay operations, and edge set
representing dependencies between assay
operations. The weight for each node,
, denotes the time
taken for operation
; note however that this value is not as-
signed until resource binding has been performed during syn-
thesis. Since droplet movement is very fast in contrast to assay
operations [1], [21], we can ignore the droplet transportation
time between different assay operations. In addition, a microflu-
idic module library is also provided as an input of the synthesis
procedure. This module library, analogous to a standard/custom
cell library used in cell-based VLSI design, includes different
microfluidic functional modules, such as mixers and storage
units. Each module is characterized by its function (mixing,
storing, detection, etc.) and parameters such as width, length
and operation duration. Moreover, some design specifications
are also given a priori, e.g., an upper limit on the bioassay com-
pletion time
, an upper limit on the size of microfluidic
array
, and the set of nonreconfigurable resources such as
on-chip reservoirs/dispensing ports and integrated optical detec-
tors.
The proposed synthesis tool performs scheduling, resource
binding, and placement in a unified manner. As in the case
of high-level synthesis for ICs, resource binding refers to
the mapping from bioassay operations to available functional
resources. Note that there may be several types of resources for
any given bioassay operation. For example, a 2
2-array mixer,
a2
3-array mixer and a 2 4-array mixer can be used for a
droplet mixing operation [1]. These mixers differ in their areas
as well as mixing times. In such cases, a resource selection
procedure must be used. On the other hand, due to the resource
constraints, a resource binding may associate one functional
resource with several assay operations; this necessitates re-
source sharing. Once resource binding is carried out, the time
duration for each bioassay operation can be easily determined.
Scheduling determines the start times and stop times of all
assay operations, subject to the precedence constraints imposed
by the sequencing graph. In a valid schedule, assay operations
that share a microfluidic module cannot execute concurrently.
Scheduling and resource binding also need to be tied to the
placement problem for biochips; placement determines the
various configurations of a microfluidic array as well as the lo-
cations of integrated optical detectors and reservoirs/dispensing
ports. The property of virtual devices makes the placement of
reconfigurable microfluidic modules, such as mixers or storage
units, on a 2-D microfluidic array quite different from the
traditional placement problem in VLSI design.
The output of the synthesis tool includes the mapping of assay
operation to resources, a schedule for the assay operations, and
the placement of the modules. The synthesis procedure attempts
to find a desirable design point that satisfies the input specifica-
tions. If such a solution does not exist, the synthesis tool out-
puts the best solution that can be achieved. In order to measure
the quality of a synthesis flow, we need to consider the min-
imization of the array area
and the completion time for
the bioassay. For this multi-objective optimization problem, a
weighting approach is used. Here weights
and , where
, are assigned to the criteria of normalized area (de-
noted by
) and normalized bioassay time (denoted by
), respectively. The solution with the lowest value of the
metric
is considered to be
an acceptable solution.
B. PRSA-Based Algorithm
The resource-constrained scheduling problem and the
module placement problem have been shown in the literature
XU et al.: DEFECT-AWARE HIGH-LEVEL SYNTHESIS AND MODULE PLACEMENT FOR MICROFLUIDIC BIOCHIPS 53
Fig. 2. Example illustrating system-level synthesis [13].
to be NP-complete [22]. Therefore, heuristics are needed to
solve the optimization problem in a computationally efficient
manner. In this paper, we propose a synthesis algorithm based
on simulated annealing, which is widely used in traditional
electronic design automation [23]. Since we are dealing with
multi-objective optimization (chip area, assay time, defect
tolerance), we combine simulated annealing algorithm with
a genetic algorithm to better represent candidate designs, as
has been done earlier in electronic design automation [24].
We therefore focus on a parallel recombinative simulated
annealing (PRSA) based algorithm. PRSA is a well-studied
combinational optimization method that has some of the best
attributes of both genetic algorithms and simulated annealing
algorithms [25]. This class of algorithms is best viewed as
genetic algorithms that use Boltzmann trials between modified
and existing solutions to select the solutions that exist in the
next generation.
We present a PRSA-based algorithm to solve the optimization
problem for biochip synthesis. The pseudocode for this heuristic
approach is shown in Fig. 3. Some details of the procedure are
listed as follows, and they also will be illustrated in Section V.
1) Representation of a Chromosome: A robust repre-
sentation technique called random keys is used in this
algorithm [26]. A random key is a random number sam-
pled from
. Each chromosome in the population
can be encoded as a vector of random keys, named
, where is
the number of assay operations. Here the first set of
genes
are used to determine resource binding, i.e.,
,
to . The second set of genes are to set the delay time
of the operations, which is calculated as follows: delay value
of operation
, to , where
is a constant that can be fine-tuned through experiments. The
Fig. 3. Pseudo-code for the PRSA-based heuristic algorithm.
last
genes are used to determine the placement priorities, i.e.,
priority value of operation
, to .
2) Construction Procedure: The goal of this procedure is to
carry out resource binding, scheduling and placement under de-
pendency and resource constraints, by using a vector of random
numbers (i.e., genes from a chromosome). It consists of the fol-
lowing three phases.
1) Phase I—Resource binding: To simplify the synthesis pro-
cedure, in this phase we temporarily do not consider an
upper limit on the number of available reconfigurable re-
sources. A reconfigurable resource type for a bioassay is
selected based on its associated gene value, i.e.,
.For
example, for a mixing operation
,a2 2-array mixer is
selected if
;a2 3-array mixer is chosen
54 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 2, NO. 1, MARCH 2008
if ;a2 4-array mixer is selected if
; a 4-electrode linear array mixer is
selected if
.
Reservoirs/dispensing ports and optical detectors are non-
reconfigurable resources. The number of such resources is
fixed, and it is determined by the system design specifica-
tions. The gene values for the corresponding operations de-
termine the selection of resource instance. For example, if
there are two optical detectors available, namely
and
, a optical detection operation is bound to if
, and to if .
After Phase I, a weight
, i.e., the duration time
for the corresponding operation, has been assigned to
each node
of the sequencing graph. Thus, an original
sequencing graph without node weights is modified to a
weighted sequencing graph.
2) Phase II—
Scheduling: In this phase, a feasible bioassay
schedule, satisfying temporal precedence constraints
as well as nonreconfigurable resource constraints, is
constructed by using the delay values
from a chro-
mosome. Due to its low computational complexity of
, where is the number of operations to schedule,
a list scheduling algorithm is used in this step [27]. As in
Phase I, only constraints for nonreconfigurable resources
are taken into account here.
To schedule the operation
, we set its start time to be
either is the predecessor of ,
or it used the same resource as
and stop time as
. After this phase, a scheduled
sequencing graph with resource binding is obtained.
3) Phase III—Placement: Based on the results from resource
binding and scheduling, we attempt to place the microflu-
idic modules on a 2-D array to satisfy the design specifi-
cations. A greedy algorithm referred to as Kamer-BF al-
gorithm is used in this phase [28]. Microfluidic modules
are first sorted in the descending order of their priority
values
. In each step, the module with the highest
priority among the unplaced ones is selected and placed.
To minimize the chip area, the selected module can only
be placed adjacent to modules which have already been lo-
cated on the chip layout to avoid waste of space. Note that
there might be multiple locations for the selected module.
In this case, the greedy algorithm evaluates each place-
ment and selects the one which result in the smallest chip
area. Resource constraints must be satisfied, e.g., there
should be no spatial overlap between the module with pre-
viously placed ones if their usage overlaps in the schedule.
The placement problem can also be modeled by a 3-D
packing problem, which will be illustrated by an example
in Section V. In addition, we add a segregation region be-
tween two active modules. This additional area not only
isolates the functional module from its neighbors, thereby
avoiding unexpected cross-contamination, but it also pro-
vides a transportation path for droplet movement between
different modules.
The above greedy algorithm not only deals with the place-
ment of reconfigurable resources, but it can also adapt to the
location of nonreconfigurable resources such as optical detec-
tors. For a fabricated chip, the locations of the optical detectors
are fixed. The placement algorithm views these optical detectors
as preplaced modules and ensures that the location of optical
detectors will not be used by other modules during their sched-
uled operation time. On the other hand, the locations of reser-
voirs/dispensing ports can be determined manually after syn-
thesis, since they do not affect the area of microfluidic array or
the processing time for the bioassay.
Therefore, based on the information provided by a chromo-
some, the synthesis procedure can be carried out based on the
above three phases. The fitness value of this chromosome is de-
termined by the synthesis results. Through a series of genera-
tions of evolution controlled by a simulated annealing process,
we can find a best chromosome, i.e., with the smallest fitness
value, from the final population. The synthesis results obtained
from this chromosome represent the solution to our optimiza-
tion problem.
V. D
EFECT-TOLERANT
SYNTHESIS
Digital microfluidic biochips are fabricated using standard
microfabrication techniques [1]. Due to the underlying mixed
technology and multiple energy domains, they exhibit unique
failure mechanisms and defects. A manufactured microfluidic
array may contain several defective cells. We have observed
defects such as dielectric breakdown, shorts between adjacent
electrodes, and electrode degradation, particle contamination
and residue, etch variations, etc. [29], [30].
Reconfiguration techniques can be used to bypass faulty cells
or faulty optical detectors to tolerate manufacturing defects.
Bioassay operations bound to these faulty resources in the
original design need to be remapped to other fault-free re-
sources. Due to the strict resource constraints in the fabricated
biochip, alterations in the resource binding operation, schedule
and placement must be carried out carefully. Our proposed
system-level synthesis tool can be modified to deal with this
issue by introducing defect tolerance schemes.
To reconfigure a defective biochip, a PRSA-based algorithm
along the lines of the one described in Section IV-B was used
in [31]. The following additional considerations must be taken
into account.
1) The objective during reconfiguration is to minimize the
bioassay completion time while accommodating all mi-
crofluidic modules and optical detectors in the fabricated
microfluidic array.
2) As resource constraints, the defect-free parts of the mi-
crofluidic array and the number of fabricated fault-free
nonreconfigurable resources replace the original design
specifications.
3) In the placement phase, the locations of the defective cells
are no longer available. Note that the locations of nonre-
configurable resources such as integrated optical detectors
and reservoirs/ dispensing ports are fixed in the fabricated
biochip.
Using this enhanced synthesis tool, a set of bioassays can be
easily mapped to a biochip with a few defective cells; thus we
do not need to discard the defective biochip.
The enhanced defect tolerance schemes in this paper are com-
posed of two attributes: defect-aware synthesis, i.e., anticipate
![Fig. 1. Schematic of a digital microfluidic biochip [14].](/figures/fig-1-schematic-of-a-digital-microfluidic-biochip-14-79byqpo9.webp)
